id
string | image
string | input
string | system
string | output
string | boxes
sequence | box_texts
sequence | detail
dict |
|---|---|---|---|---|---|---|---|
1610.08904-en-4 | 1610.08904-en-4.jpg | optimality of Fisher-like criteria relies on the assumption that the data of each class is of a Gaussian $\hat{\text{distribution, which is obviously not satisfied in our case. Also, a high cost }O(m^{2})}$ is entailed to compute the Fisher loss in a mini-batch with $m$ samples by computing all the pairwise distances.
Consequently, we propose a faster-to-compute loss function that approximately maximizes the margin between the positive and negative similarity distributions without making any assumption about the distribution's shape or pattern. Specifically, we retrieve one hard quadruplet from a random batch during each forward pass. Please see the illustration in Figure 1(c). The quadruplet consists of the most dissimilar positive sample pair in the batch $(\hat{i},\hat{j})=\operatorname*{argmin}_{(i,j)\in\hat{P}}\hat{S}_{i,j}$,which means their similarity score is most likely to cross the“safe margin”towards the negative similarity distribution in this local range. Next, we build a similarity neighborhood graph that links the chosen positive pair with their respective negative neighbors in the batch, and choose the hard negatives as the other two quadruplet members $\hat{k}=\operatorname{argmax}_{(\hat{i},k)\in\hat{N}}S_{\hat{i},k}$, and $\hat{l}=\operatorname{argmax}_{(\hat{j},l)\in\hat{N}}S_{\hat{j},l}.$ Using this hard quadruplet $(\hat{i},\hat{j},\hat{k},\hat{l})$, we can now locally approximate the inter-distribution margin as $\operatorname*{min}(S_{i,\hat{j}}-S_{\hat{i},\hat{k}},\hat{S}_{\hat{i},\hat{j}}-S_{\hat{j},\hat{l}})$ in a robust manner. This makes us immediately come to a double-header hinge loss $E_m$ to discriminate the target similarity distributions under the large margin criterion
(3)
$$\begin{aligned}&min\quad E_{m}=\sum_{i,\hat{j}}(\varepsilon_{\hat{i},\hat{j}}+\tau_{\hat{i},\hat{j}}),\\&s.t.:\:\forall(\hat{i},\hat{j}),\quad\max\left(0,\alpha+S_{\hat{i},\hat{j}}-S_{\hat{i},\hat{j}}\right)\leq\varepsilon_{\hat{i},\hat{j}},\:\max\left(0,\alpha+S_{\hat{j},\hat{l}}-S_{\hat{i},\hat{j}}\right)\leq\tau_{\hat{i},\hat{j}},\\&(\hat{i},\hat{j})=\operatorname*{argmin}_{(i,j)\in\hat{P}}S_{i,j},\:\hat{k}=\operatorname*{argmax}_{(i,k)\in\hat{N}}S_{(j,l)\in\hat{N}},\:\hat{l}=\operatorname*{argmax}_{(j,l)\in\hat{N}}S_{j,l},\:\varepsilon_{\hat{i},\hat{j}}\geq0,\:\tau_{\hat{i},\hat{j}}\geq0\end{aligned}$$
where $\varepsilon_{\hat{i},\hat{j}},\tau_{\hat{i},\hat{j}}$ are the slack variables, and $\alpha$ is the enforced margin.
The discriminative loss has four main benefits: 1) The discrimination of similarity distributions is assumption-free. 2) Hard samples are simultaneously found during the loss minimization. 3) The loss function incurs a low computational cost and encourages faster convergence. Specifically, the searching cost of the hard positive pair $(\hat{i},\hat{j})$ is very small since the positive pair set $\hat{P}$ is usually much smaller than the negative pair set $\hat{N}$ in an $m$-sized mini-batch. While the hard negative mining only incurs an $O(m)$ complexity. 4) Eqs. (2, 3) can be easily optimized through the standard stochastic gradient descent to adjust the CNN parameters.
3.2 Joint metric and embedding optimization
Given the learned PDDM and mined hard samples in a mini-batch, we can use them to solve for a better, local similarity-aware feature embedding at the same time. For computational efficiency, we reuse the hard quadruplet's features for metric optimization (Eq. (3)) in the same forward pass. What follows is to use the double-header hinge loss again, but to constrain the deep features this time, see Figure 2. The objective is to ensure the Éuclidean distance between hard negative features $(D_{\hat{i},\hat{k}})$ or $D_{\hat{j},\hat{l}})$ is larger than that between hard positive features $D_{\hat{i},\hat{j}}$ by a large margin. Combining the embedding loss $E_{e}$ and metric loss $E_{m}$ (Eq. (3)) gives our final joint loss function:
(4)
$$\begin{aligned}&\mathrm{min}&&E_{m}+\lambda E_{e}+\gamma\|\widetilde{W}\|_{2},\quad\mathrm{where}\:E_{e}=\sum_{\hat{i},\hat{j}}(o_{\hat{i},\hat{j}}+\rho_{\hat{i},\hat{j}}),\\s.t.&&\forall(\hat{i},\hat{j}),\quad\max\left(0,\beta+D_{\hat{i},\hat{j}}-D_{\hat{i},\hat{k}}\right)\leq o_{\hat{i},\hat{j}},\:\max\left(0,\beta+D_{\hat{i},\hat{j}}-D_{\hat{j},\hat{l}}\right)\leq\rho_{\hat{i},\hat{j}},\\&&&D_{\hat{i},\hat{j}}=\|f(x_{\hat{i}})-f(x_{\hat{j}})\|_{2},\:o_{\hat{i},\hat{j}}\geq0,\:\rho_{\hat{i},\hat{j}}\geq0,\end{aligned}$$
where $\widetilde{W}$ are the CNN parameters for both the PDDM and feature embedding, and $o_{\hat{i},\hat{j}},\rho_{\hat{i},\hat{j}}$ and $\beta$ are the slack variables and enforced margin for $E_e$, and $\lambda,\gamma$ are the regularization parameters. Since all features are $\ell _{2}$- normalized ( see Figure 2) , we have $\beta + D_{\hat{i} , \hat{j} }- D_{\hat{i} , \hat{k} }= \beta - 2f( x_{\hat{i} }) \hat{f} ( x_{\hat{j} }) + 2f( x_{\hat{i} }) f( x_{\hat{k} }) $, and can conveniently derive the gradients as those in triplet-based methods [27,33].
This joint objective provides effective supervision in two domains, respectively at the score level and feature level that are mutually informed. Although the score level supervision by $E_m$ alone is already capable of optimizing both our metric and feature embedding, we will show the benefits of adding the feature level supervision by $E_e$ in experiments. Note we can still enforce the large margin relations of quadruple features in $E_e$ using the simple Euclidean metric. This is because the quadruple features are selected by our PDDM that is learned in the local Euclidean space as well.
5 | [
[
"0.1765",
"0.09216",
"0.8228",
"0.09216",
"0.8228",
"0.1061",
"0.1765",
"0.1061"
],
[
"0.1765",
"0.108",
"0.8228",
"0.108",
"0.8228",
"0.1187",
"0.1765",
"0.1187"
],
[
"0.1748",
"0.1206",
"0.8057",
"0.1206",
"0.8057",
"0.1338",
"0.1748",
"0.1338"
],
[
"0.1732",
"0.1395",
"0.822",
"0.1395",
"0.822",
"0.1528",
"0.1732",
"0.1528"
],
[
"0.1748",
"0.1547",
"0.8228",
"0.1547",
"0.8228",
"0.1686",
"0.1748",
"0.1686"
],
[
"0.1748",
"0.1686",
"0.8228",
"0.1686",
"0.8228",
"0.1818",
"0.1748",
"0.1818"
],
[
"0.1748",
"0.1831",
"0.8228",
"0.1831",
"0.8228",
"0.1963",
"0.1748",
"0.1963"
],
[
"0.1724",
"0.1957",
"0.8267",
"0.1957",
"0.8267",
"0.2166",
"0.1724",
"0.2166"
],
[
"0.1732",
"0.2122",
"0.8228",
"0.2134",
"0.8228",
"0.2279",
"0.1732",
"0.2267"
],
[
"0.1765",
"0.2291",
"0.822",
"0.2291",
"0.822",
"0.2393",
"0.1765",
"0.2393"
],
[
"0.1748",
"0.2412",
"0.824",
"0.2412",
"0.824",
"0.2551",
"0.1748",
"0.2551"
],
[
"0.1708",
"0.254",
"0.826",
"0.2551",
"0.826",
"0.276",
"0.1708",
"0.2747"
],
[
"0.1732",
"0.2747",
"0.822",
"0.2747",
"0.822",
"0.289",
"0.1732",
"0.289"
],
[
"0.1732",
"0.2903",
"0.8228",
"0.2903",
"0.8228",
"0.305",
"0.1732",
"0.305"
],
[
"0.1732",
"0.3062",
"0.8037",
"0.3062",
"0.8037",
"0.3193",
"0.1732",
"0.3193"
],
[
"0.2467",
"0.324",
"0.4656",
"0.3276",
"0.4656",
"0.3499",
"0.2467",
"0.3452"
],
[
"0.8037",
"0.3289",
"0.826",
"0.3289",
"0.826",
"0.3435",
"0.8037",
"0.3435"
],
[
"0.1887",
"0.3542",
"0.687",
"0.3542",
"0.687",
"0.375",
"0.1887",
"0.375"
],
[
"0.6846",
"0.3591",
"0.739",
"0.3591",
"0.739",
"0.3738",
"0.6846",
"0.3738"
],
[
"0.286",
"0.3801",
"0.564",
"0.3806",
"0.564",
"0.3984",
"0.286",
"0.3977"
],
[
"0.5513",
"0.379",
"0.8",
"0.3801",
"0.8",
"0.401",
"0.5513",
"0.3997"
],
[
"0.3423",
"0.3965",
"0.3945",
"0.3965",
"0.3945",
"0.411",
"0.3423",
"0.411"
],
[
"0.4631",
"0.3977",
"0.5186",
"0.3977",
"0.5186",
"0.4092",
"0.4631",
"0.4092"
],
[
"0.5825",
"0.3965",
"0.6357",
"0.3965",
"0.6357",
"0.411",
"0.5825",
"0.411"
],
[
"0.1732",
"0.4136",
"0.6216",
"0.4136",
"0.6216",
"0.4312",
"0.1732",
"0.4312"
],
[
"0.1748",
"0.435",
"0.8228",
"0.435",
"0.8228",
"0.449",
"0.1748",
"0.449"
],
[
"0.1748",
"0.4507",
"0.822",
"0.4507",
"0.822",
"0.464",
"0.1748",
"0.464"
],
[
"0.1748",
"0.464",
"0.8228",
"0.464",
"0.8228",
"0.4778",
"0.1748",
"0.4778"
],
[
"0.1732",
"0.4797",
"0.8228",
"0.4797",
"0.8228",
"0.4937",
"0.1732",
"0.4937"
],
[
"0.1748",
"0.4956",
"0.824",
"0.4956",
"0.824",
"0.5103",
"0.1748",
"0.5103"
],
[
"0.1748",
"0.5103",
"0.826",
"0.5103",
"0.826",
"0.5234",
"0.1748",
"0.5234"
],
[
"0.1724",
"0.5234",
"0.4878",
"0.523",
"0.4878",
"0.537",
"0.1724",
"0.5386"
],
[
"0.1724",
"0.552",
"0.5024",
"0.553",
"0.5024",
"0.5674",
"0.1724",
"0.5664"
],
[
"0.1748",
"0.5776",
"0.8228",
"0.5776",
"0.8228",
"0.5913",
"0.1748",
"0.5913"
],
[
"0.1732",
"0.5913",
"0.826",
"0.5923",
"0.826",
"0.607",
"0.1732",
"0.606"
],
[
"0.1748",
"0.607",
"0.8267",
"0.607",
"0.8267",
"0.6206",
"0.1748",
"0.6206"
],
[
"0.1748",
"0.6206",
"0.826",
"0.6206",
"0.826",
"0.634",
"0.1748",
"0.634"
],
[
"0.1732",
"0.6333",
"0.822",
"0.6353",
"0.822",
"0.6494",
"0.1732",
"0.648"
],
[
"0.1765",
"0.6494",
"0.8228",
"0.6494",
"0.8228",
"0.664",
"0.1765",
"0.664"
],
[
"0.1748",
"0.6655",
"0.716",
"0.6655",
"0.716",
"0.6797",
"0.1748",
"0.6797"
],
[
"0.267",
"0.6816",
"0.6953",
"0.686",
"0.6953",
"0.707",
"0.267",
"0.7026"
],
[
"0.7983",
"0.6875",
"0.824",
"0.683",
"0.8286",
"0.7007",
"0.8022",
"0.705"
],
[
"0.21",
"0.7134",
"0.7837",
"0.7134",
"0.7837",
"0.7344",
"0.21",
"0.7344"
],
[
"0.3228",
"0.738",
"0.6313",
"0.738",
"0.6313",
"0.756",
"0.3228",
"0.756"
],
[
"0.1724",
"0.7603",
"0.824",
"0.761",
"0.824",
"0.78",
"0.1724",
"0.779"
],
[
"0.1748",
"0.78",
"0.824",
"0.78",
"0.824",
"0.7935",
"0.1748",
"0.7935"
],
[
"0.1732",
"0.7915",
"0.826",
"0.7935",
"0.826",
"0.808",
"0.1732",
"0.806"
],
[
"0.1732",
"0.808",
"0.722",
"0.808",
"0.722",
"0.8228",
"0.1732",
"0.8228"
],
[
"0.1748",
"0.829",
"0.8228",
"0.829",
"0.8228",
"0.8423",
"0.1748",
"0.8423"
],
[
"0.1748",
"0.8433",
"0.8228",
"0.8433",
"0.8228",
"0.857",
"0.1748",
"0.857"
],
[
"0.1748",
"0.857",
"0.8228",
"0.857",
"0.8228",
"0.8706",
"0.1748",
"0.8706"
],
[
"0.1748",
"0.8706",
"0.824",
"0.8706",
"0.824",
"0.884",
"0.1748",
"0.884"
],
[
"0.1732",
"0.8833",
"0.824",
"0.884",
"0.824",
"0.8984",
"0.1732",
"0.898"
],
[
"0.1765",
"0.8994",
"0.683",
"0.8994",
"0.683",
"0.91",
"0.1765",
"0.91"
],
[
"0.4944",
"0.936",
"0.5063",
"0.936",
"0.5063",
"0.9478",
"0.4944",
"0.9478"
]
] | [
"optimality of Fisher-like criteria relies on the assumption that the data of each class is of a Gaussian",
"distribution,whichis obviouslynot satisfiedin our case.Also,ahighcost O(m)isentailedto",
"compute the Fisher loss in a mini-batch with m samples by computing all the pairwise distances.",
"Consequently, we propose a faster-to-compute loss function that approximately maximizes the",
"margin between the positive and negative similarity distributions without making any assumption",
"about the distribution's shape or pattern.Specifically, we retrieve one hard quadruplet from a",
"random batch during each forward pass. Please see the illustration in Figure 1(c). The quadruplet",
"consistsofthemost dissimilarpositivesamplepairinthebatch ()=armin(i,)e Si.jwhich",
"means their similarity score is most likely to cross the\"safe margin\"towards the negative similarity",
"distributionin thislocal range.Next.webuild asimilarity neighborhood graphthat links the chosen",
"positive pair with their respective negative neighbors in the batch, and choose the hard negatives",
"as the other two quadruplet members h = argmax(z,k)en Si,k, and i = argmax(3,)en Ss,. Using",
"this hard quadruplet (. i, k, ), we can now locally approximate the inter-distribution margin as",
"min(Ss - S, k, Sz - S;i) in a robust manner. This makes us immediately come to a double-header",
"hinge loss Em to discriminate the target similarity distributions under the large margin criterion:",
"min Em=,(ei,3 + Ta),",
"(3)",
"s.t. : v(2,3), max (o,a+ St, - St3) ≤e.3, max (o,a + Ss.i - St.3)",
" ≤T3,3",
"(3,3) = argmin Si,j, k = argmax Sik, i",
"i = argmax Sin, es,3 ≥ 0, Ta, ≥0,",
"(i,j)eP",
"(i,k)EN",
"(G,I)EN",
"where es.j, Ti,j are the slack variables, and a is the enforced margin.",
"The discriminative loss has four main benefits: ) The discrimination of similarity distributions is",
"assumption-free. 2) Hard samples are simultaneously found during the loss minimization. 3) The",
"loss function incurs a low computational cost and encourages faster convergence. Specifically, the",
"searching cost of the hard positive pair (i, ) is very small since the positive pair set P is usually much",
"smaller than the negative pair set N in an m-sized mini-batch. While the hard negative mining only",
"incurs an O(m) complexity. 4) Eqs. (2, 3) can be easily optimized through the standard stochastic",
"gradient descent to adjust the CNN parameters.",
"3.2Jointmetric and embedding optimization",
"Given the learned PDDM and mined hard samples in a mini-batch, we can use them to solve for",
"a better, local similarity-aware feature embedding at the same time. For computational efficiency,",
"we reuse the hard quadruplet's features for metric optimization (Eq. (3)) in the same forward pass.",
"What follows is to use the double-header hinge loss again, but to constrain the deep features this time,",
"see Figure 2. The objective is to ensure the Euclidean distance between hard negative features (Di.",
"or D;) is larger than that between hard positive features Di; by a large margin. Combining the",
"embedding loss Ee and metric loss Em (Eq. (3) gives our final joint loss function:",
"min Em+Ee+llWll2,where Ee=,(oi3+pa,3),",
"(4)",
"8.t. : V(,3), max (o, β+ D,3 -De) ≤o%,3, max (o, β+ De3 -D)≤p.3,",
"D3 = lf(s) -f(ag)l2, 0元3 ≥0, Pr3 ≥0,",
"where W are the CNN parameters for both the PDDM and feature embedding, and os.3, Pi., and β are",
"the slack variables and enforced margin for Ee, and A, are the regularization parameters. Since all",
"features are l2-normalized (seeFigure2),we haveβ+D3-D=β-2f(α)f(α)+2f(r)f(),",
"and can conveniently derive the gradients as those in triplet-based methods [27, 33].",
"This joint objective provides effective supervision in two domains, respectively at the score level and",
"feature level that are mutually informed. Although the score level supervision by Em alone is already",
"capable of optimizing both our metric and feature embedding, we will show the benefits of adding the",
"feature level supervision by Ee in experiments. Note we can still enforce the large margin relations",
"of quadruple features in Ee using the simple Euclidean metric. This is because the quadruple features",
"are selectedbyourPDDMthatislearnedinthelocalEuclideanspace aswell.",
"5"
] | {
"arxiv_id": "1610.08904",
"language": "en",
"page_idx": 4
} |
||
2306.04695-en-0 | 2306.04695-en-0.jpg | $\textbf{CoNCEPTBED: Evaluating Concept Learning Abilities of Text- to- Image Diffusion}$
$\mathbf{Models}$
Maitreya Patel$^{1*}$,Tejas Gokhale$^2$,Chitta Baral$^1$,Yezhou Yang$^{1}$
1 Arizona State University
$^{2}$ University of Maryland Baltimore County
### $\mathbf{Abstract}$
The ability to understand visual concepts and replicate and compose these concepts from images is a central goal for computer vision. Recent advances in text-to-image (T2I) models have lead to high definition and realistic image quality generation by learning from large databases of images and their descriptions. However, the evaluation of T2I models has focused on photorealism and limited qualitative measures of visual understanding. To quantify the ability of T2I models in learning and synthesizing novel visual concepts $(a.k.a.$ personalized T2I), we introduce CONCEPTBED, a large-scale dataset that consists of 284 unique visual concepts, and 33K composite text prompts. Along with the dataset, we propose an evaluation metric, Concept Confidence Deviation (CCD), that uses the confidence of oracle concept classifiers to measure the alignment between concepts generated by T2I generators and concepts contained in target images. We evaluate visual concepts that are either objects, attributes, or styles, and also evaluate four dimensions of compositionality: counting, attributes, relations, and actions. Our human study shows that CCD is highly correlated with human understanding of concepts. Our results point to a trade-off between learning the concepts and preserving the compositionality which existing approaches struggle to overcome. The data, code, and interactive demo is available at: https://conceptbed.github.io/
# 1 Introduction
Humans reason about the visual world by aggregating entities that they see into“visual concepts”: both cats and elephants are animals, and both palms and pines are trees. We use natural language to describe images and things that we see. Although this type of visual concept learning is well-defined in human psychology (Murphy 2004), it remains elusive in the context of data-driven techniques capable of learning and reasoning from images and their natural language descriptions.
Text-to-Image (T2I) generative models are trained to translate natural language phrases into images that correspond to that input. High-quality T2I models, therefore, serve as a link between human-level concepts (expressed in natural language) and their visual representations and are one way to reproduce visual concepts. On the other hand,
$^*$Corresponding Author: maitreya.patel@asu.edu
Copyright $\dot{\odot}$ 2024, Association for the Advancement of Artificial
Intelligence (www.aaai.org). All rights reserved.
<FigureHere>
Figure 1: Visual concept learners such as textual inversion models learn to“invert”a set of images (about a concept $c)$ into a text embedding V*, and then use this learned textual concept in new text prompts to generate images of concept $c$ under different contexts and by performing novel compositions with other concepts. The proposed CONCEPTBED dataset along with the evaluation metric ccD allows us to comprehensively and quantifiably evaluate concept learning abilities of text-to-image diffusion models.
this has also sparked interest in visual concept learning $(a.k.a.$ personalized T2I) through the procedure of“image $inversion"-$to translate one or many images corresponding to a visual concept into a latent representation of that visual concept. While earlier methods primarily explored image inversion using generative adversarial networks (Xia et al. 2022), methods such as Textual Inversion (Gal et al. 2022) and Dreambooth (Ruiz et al. 2022) combine image inversion with T2I- this has led to an effective way to quickly learn concepts from a few images and reproduce them in novel combinations and compositions with other concepts, attributes, styles, etc. These methods aim to learn concepts with minimal reference images by fine-tuning pre-trained text-conditioned diffusion models (Figure 1). Therefore this paradigm of T2I and image inversion is a powerful new way of learning and reproducing concepts.
Within this paradigm of novel visual concept learning via image inversion, two primary evaluation criteria have emerged:(1) concept alignment, which assesses the corre- | [
[
"0.0907",
"0.1218",
"0.9087",
"0.1225",
"0.9087",
"0.1401",
"0.0907",
"0.1395"
],
[
"0.4631",
"0.1445",
"0.5376",
"0.1445",
"0.5376",
"0.159",
"0.4631",
"0.159"
],
[
"0.2418",
"0.1812",
"0.7607",
"0.1812",
"0.7607",
"0.1957",
"0.2418",
"0.1957"
],
[
"0.4182",
"0.202",
"0.5864",
"0.2026",
"0.5864",
"0.2172",
"0.4182",
"0.2166"
],
[
"0.353",
"0.2134",
"0.6436",
"0.2153",
"0.6436",
"0.2329",
"0.353",
"0.2311"
],
[
"0.02287",
"0.2615",
"0.06046",
"0.2615",
"0.06046",
"0.3726",
"0.02287",
"0.3726"
],
[
"0.25",
"0.2734",
"0.3145",
"0.2734",
"0.3145",
"0.2886",
"0.25",
"0.2886"
],
[
"0.5317",
"0.2727",
"0.626",
"0.2727",
"0.626",
"0.2842",
"0.5317",
"0.2842"
],
[
"0.10297",
"0.2942",
"0.4617",
"0.2942",
"0.4617",
"0.308",
"0.10297",
"0.308"
],
[
"0.10297",
"0.3093",
"0.4631",
"0.3093",
"0.4631",
"0.3225",
"0.10297",
"0.3225"
],
[
"0.614",
"0.3145",
"0.649",
"0.3145",
"0.649",
"0.318",
"0.614",
"0.318"
],
[
"0.732",
"0.315",
"0.7876",
"0.315",
"0.7876",
"0.3257",
"0.732",
"0.3257"
],
[
"0.10297",
"0.3213",
"0.4631",
"0.3213",
"0.4631",
"0.3352",
"0.10297",
"0.3352"
],
[
"0.10297",
"0.334",
"0.4631",
"0.334",
"0.4631",
"0.348",
"0.10297",
"0.348"
],
[
"0.10297",
"0.3467",
"0.4631",
"0.3467",
"0.4631",
"0.3599",
"0.10297",
"0.3599"
],
[
"0.5254",
"0.3484",
"0.601",
"0.3484",
"0.601",
"0.3591",
"0.5254",
"0.3591"
],
[
"0.6807",
"0.3435",
"0.7524",
"0.3435",
"0.7524",
"0.3547",
"0.6807",
"0.3547"
],
[
"0.7036",
"0.3542",
"0.7256",
"0.3542",
"0.7256",
"0.3623",
"0.7036",
"0.3623"
],
[
"0.826",
"0.3484",
"0.899",
"0.3484",
"0.899",
"0.3591",
"0.826",
"0.3591"
],
[
"0.10297",
"0.3591",
"0.4631",
"0.3591",
"0.4631",
"0.3726",
"0.10297",
"0.3726"
],
[
"0.5483",
"0.358",
"0.585",
"0.358",
"0.585",
"0.3687",
"0.5483",
"0.3687"
],
[
"0.8423",
"0.3591",
"0.8813",
"0.3591",
"0.8813",
"0.3662",
"0.8423",
"0.3662"
],
[
"0.1013",
"0.3706",
"0.4631",
"0.3726",
"0.4631",
"0.3865",
"0.1013",
"0.3838"
],
[
"0.02533",
"0.3623",
"0.058",
"0.3623",
"0.058",
"0.3965",
"0.02533",
"0.3965"
],
[
"0.528",
"0.3718",
"0.652",
"0.3718",
"0.652",
"0.385",
"0.528",
"0.385"
],
[
"0.10297",
"0.3838",
"0.4631",
"0.3838",
"0.4631",
"0.3977",
"0.10297",
"0.3977"
],
[
"0.10297",
"0.3965",
"0.4617",
"0.3965",
"0.4617",
"0.4104",
"0.10297",
"0.4104"
],
[
"0.63",
"0.3933",
"0.706",
"0.3933",
"0.706",
"0.4048",
"0.63",
"0.4048"
],
[
"0.01878",
"0.3997",
"0.058",
"0.3984",
"0.0621",
"0.4849",
"0.02287",
"0.486"
],
[
"0.10297",
"0.4092",
"0.4631",
"0.4092",
"0.4631",
"0.4224",
"0.10297",
"0.4224"
],
[
"0.5933",
"0.411",
"0.709",
"0.411",
"0.709",
"0.4248",
"0.5933",
"0.4248"
],
[
"0.738",
"0.4136",
"0.7915",
"0.4136",
"0.7915",
"0.4236",
"0.738",
"0.4236"
],
[
"0.10297",
"0.4216",
"0.4631",
"0.4216",
"0.4631",
"0.435",
"0.10297",
"0.435"
],
[
"0.10297",
"0.435",
"0.4631",
"0.435",
"0.4631",
"0.449",
"0.10297",
"0.449"
],
[
"0.5254",
"0.4312",
"0.7075",
"0.4312",
"0.7075",
"0.4443",
"0.5254",
"0.4443"
],
[
"0.1013",
"0.4463",
"0.4617",
"0.4475",
"0.4617",
"0.461",
"0.1013",
"0.4602"
],
[
"0.1013",
"0.459",
"0.4617",
"0.4602",
"0.4617",
"0.4734",
"0.1013",
"0.473"
],
[
"0.10297",
"0.473",
"0.4617",
"0.473",
"0.4617",
"0.486",
"0.10297",
"0.486"
],
[
"0.5195",
"0.4697",
"0.911",
"0.4702",
"0.911",
"0.4849",
"0.5195",
"0.4841"
],
[
"0.1013",
"0.4849",
"0.4617",
"0.4849",
"0.4617",
"0.4988",
"0.1013",
"0.4988"
],
[
"0.02287",
"0.4912",
"0.06046",
"0.4912",
"0.06046",
"0.712",
"0.02287",
"0.712"
],
[
"0.5215",
"0.4873",
"0.911",
"0.4873",
"0.911",
"0.4976",
"0.5215",
"0.4976"
],
[
"0.10297",
"0.4976",
"0.4617",
"0.4976",
"0.4617",
"0.511",
"0.10297",
"0.511"
],
[
"0.5195",
"0.4988",
"0.9126",
"0.4988",
"0.9126",
"0.512",
"0.5195",
"0.512"
],
[
"0.10297",
"0.5103",
"0.4617",
"0.5103",
"0.4617",
"0.5234",
"0.10297",
"0.5234"
],
[
"0.5195",
"0.513",
"0.914",
"0.513",
"0.914",
"0.5264",
"0.5195",
"0.5264"
],
[
"0.10297",
"0.523",
"0.4631",
"0.523",
"0.4631",
"0.536",
"0.10297",
"0.536"
],
[
"0.5195",
"0.5264",
"0.91",
"0.5264",
"0.91",
"0.5405",
"0.5195",
"0.5405"
],
[
"0.1013",
"0.535",
"0.46",
"0.535",
"0.46",
"0.549",
"0.1013",
"0.549"
],
[
"0.5215",
"0.5405",
"0.9126",
"0.5405",
"0.9126",
"0.5537",
"0.5215",
"0.5537"
],
[
"0.1013",
"0.5474",
"0.4631",
"0.5474",
"0.4631",
"0.561",
"0.1013",
"0.561"
],
[
"0.1013",
"0.56",
"0.4631",
"0.56",
"0.4631",
"0.5737",
"0.1013",
"0.5737"
],
[
"0.5215",
"0.5547",
"0.914",
"0.5547",
"0.914",
"0.5684",
"0.5215",
"0.5684"
],
[
"0.10046",
"0.5728",
"0.4617",
"0.5713",
"0.4617",
"0.586",
"0.10046",
"0.587"
],
[
"0.5195",
"0.5684",
"0.9126",
"0.5684",
"0.9126",
"0.582",
"0.5195",
"0.582"
],
[
"0.1013",
"0.5854",
"0.4355",
"0.5854",
"0.4355",
"0.5986",
"0.1013",
"0.5986"
],
[
"0.5195",
"0.582",
"0.8",
"0.582",
"0.8",
"0.595",
"0.5195",
"0.595"
],
[
"0.2157",
"0.6187",
"0.2279",
"0.6187",
"0.2279",
"0.629",
"0.2157",
"0.629"
],
[
"0.2385",
"0.616",
"0.3506",
"0.617",
"0.3506",
"0.632",
"0.2385",
"0.631"
],
[
"0.5215",
"0.6255",
"0.9126",
"0.6255",
"0.9126",
"0.6396",
"0.5215",
"0.6396"
],
[
"0.08496",
"0.6353",
"0.4763",
"0.636",
"0.4763",
"0.651",
"0.08496",
"0.6494"
],
[
"0.5215",
"0.6396",
"0.9126",
"0.6396",
"0.9126",
"0.654",
"0.5215",
"0.654"
],
[
"0.08496",
"0.651",
"0.4795",
"0.651",
"0.4795",
"0.664",
"0.08496",
"0.664"
],
[
"0.5195",
"0.654",
"0.9126",
"0.654",
"0.9126",
"0.6675",
"0.5195",
"0.6675"
],
[
"0.08826",
"0.664",
"0.478",
"0.664",
"0.478",
"0.678",
"0.08826",
"0.678"
],
[
"0.5195",
"0.6675",
"0.9126",
"0.6675",
"0.9126",
"0.681",
"0.5195",
"0.681"
],
[
"0.0866",
"0.678",
"0.478",
"0.678",
"0.478",
"0.6914",
"0.0866",
"0.6914"
],
[
"0.5186",
"0.681",
"0.9126",
"0.6816",
"0.9126",
"0.696",
"0.5186",
"0.6943"
],
[
"0.08496",
"0.6914",
"0.478",
"0.6914",
"0.478",
"0.705",
"0.08496",
"0.705"
],
[
"0.5195",
"0.6943",
"0.9126",
"0.6943",
"0.9126",
"0.7085",
"0.5195",
"0.7085"
],
[
"0.0866",
"0.705",
"0.478",
"0.705",
"0.478",
"0.7197",
"0.0866",
"0.7197"
],
[
"0.5195",
"0.709",
"0.9126",
"0.709",
"0.9126",
"0.7227",
"0.5195",
"0.7227"
],
[
"0.0866",
"0.7197",
"0.4763",
"0.7197",
"0.4763",
"0.733",
"0.0866",
"0.733"
],
[
"0.5186",
"0.7217",
"0.911",
"0.7227",
"0.911",
"0.7373",
"0.5186",
"0.7363"
],
[
"0.08826",
"0.735",
"0.478",
"0.735",
"0.478",
"0.7456",
"0.08826",
"0.7456"
],
[
"0.5195",
"0.7363",
"0.911",
"0.7363",
"0.911",
"0.75",
"0.5195",
"0.75"
],
[
"0.08417",
"0.7456",
"0.3179",
"0.747",
"0.3179",
"0.761",
"0.08417",
"0.7603"
],
[
"0.5215",
"0.7505",
"0.9126",
"0.7505",
"0.9126",
"0.7646",
"0.5215",
"0.7646"
],
[
"0.10297",
"0.761",
"0.478",
"0.761",
"0.478",
"0.7744",
"0.10297",
"0.7744"
],
[
"0.5186",
"0.763",
"0.9126",
"0.7646",
"0.9126",
"0.779",
"0.5186",
"0.778"
],
[
"0.0866",
"0.7744",
"0.4763",
"0.7744",
"0.4763",
"0.7886",
"0.0866",
"0.7886"
],
[
"0.5195",
"0.778",
"0.9126",
"0.778",
"0.9126",
"0.7915",
"0.5195",
"0.7915"
],
[
"0.08496",
"0.7886",
"0.4763",
"0.7886",
"0.4763",
"0.803",
"0.08496",
"0.803"
],
[
"0.5195",
"0.7925",
"0.914",
"0.7925",
"0.914",
"0.806",
"0.5195",
"0.806"
],
[
"0.0866",
"0.803",
"0.4795",
"0.803",
"0.4795",
"0.8164",
"0.0866",
"0.8164"
],
[
"0.5195",
"0.806",
"0.9126",
"0.806",
"0.9126",
"0.8193",
"0.5195",
"0.8193"
],
[
"0.08496",
"0.8174",
"0.4795",
"0.8174",
"0.4795",
"0.8306",
"0.08496",
"0.8306"
],
[
"0.5195",
"0.821",
"0.911",
"0.821",
"0.911",
"0.834",
"0.5195",
"0.834"
],
[
"0.08496",
"0.83",
"0.478",
"0.83",
"0.478",
"0.8433",
"0.08496",
"0.8433"
],
[
"0.5195",
"0.834",
"0.7686",
"0.834",
"0.7686",
"0.8477",
"0.5195",
"0.8477"
],
[
"0.10706",
"0.851",
"0.4495",
"0.8516",
"0.4495",
"0.866",
"0.10706",
"0.8657"
],
[
"0.5376",
"0.8477",
"0.9126",
"0.8477",
"0.9126",
"0.8623",
"0.5376",
"0.8623"
],
[
"0.0866",
"0.863",
"0.4795",
"0.863",
"0.4795",
"0.877",
"0.0866",
"0.877"
],
[
"0.5195",
"0.8623",
"0.9126",
"0.8623",
"0.9126",
"0.8755",
"0.5195",
"0.8755"
],
[
"0.0866",
"0.877",
"0.3718",
"0.877",
"0.3718",
"0.89",
"0.0866",
"0.89"
],
[
"0.5195",
"0.8755",
"0.9126",
"0.8755",
"0.9126",
"0.8896",
"0.5195",
"0.8896"
]
] | [
"CoNCEPTBED: Evaluating Concept Learning Abilities of Text-to-Image Diffusion",
"Models",
"Maitreya Patell*, Tejas Gokhale?, Chitta Baral', Yezhou Yang",
"Arizona State University",
"2 University of Maryland Baltimore County",
"",
"Abstract",
"CONCEPT LEARNING",
"The ability to understand visual concepts and replicate and",
"compose these concepts from images is a central goal for",
"A",
"GENERATOR",
"computer vision. Recent advances in text-to-image (T21)",
"models have lead to high definition and realistic image qual-",
"ity generation by learning from large databases of images",
"Concept Image",
"Reconstruction",
"Loss",
"Reconstructed",
"and their descriptions. However, the evaluation of T2I models",
"(Input)",
"(Output)",
"has focused on photorealism and limited qualitative measures",
"2",
"CONCEPTCOMPOSITION",
"of visual understanding. To quantify the ability of T2I mod-",
"els in learning and synthesizing novel visual concepts (a.k.a.",
"Aphotoof V",
"[']",
"personalized T2I), we introduce CONCEPTBED, a large-scale",
"\" fuying in Hhe yky!",
"GENERATOR",
"dataset that consists of 284 unique visual concepts, and 33K",
"composite text prompts. Along with the dataset, we propose",
"Two V sitting on a tree branch.",
"an evaluation metric, Concept Confidence Deviation (CcD),",
"that uses the confidence of oracle concept classifiers to mea-",
"sure the alignment between concepts generated by T2I gener-",
"Figure 1: Visual concept learners such as textual inversion",
"ators and concepts contained in target images. We evaluate vi-",
"690900",
"modelslearnto\"invertasetofimages(aboutaconceptc)",
"sual concepts that are either objects, attributes, or styles, and",
"into a text embedding V*, and then use this learned textual",
"also evaluate four dimensions of compositionality: counting,",
"concept innewtextprompts togenerate images of conceptc",
"attributes, relations, and actions. Our human study shows that",
"under different contexts and by performing novel compo-",
"CCD is highly correlated with human understanding of con-",
"sitions with other concepts. The proposed CoNCEPTBED",
"cepts. Our results point to a trade-off between learning the",
"concepts and preserving the compositionality which existing",
"dataset along with the evaluation metric cCD allows us to",
"approaches struggle to overcome. The data, code, and inter-",
"comprehensively and quantifiably evaluate concept learning",
"active demo is available at: https://conceptbed.github.io/",
"abilities of text-to-image diffusion models.",
"1.",
"Introduction",
"this has also sparked interest in visual concept learning",
"Humans reason about the visual world by aggregating en-",
"(a.k.a. personalized T2I) through the procedure of “image",
"tities that they see into “visual concepts\": both cat s and",
"inversion\"-to translate one or many images corresponding",
"elephants are animals, and both palms and pines",
"to a visual concept into a latent representation of that visual",
"are trees. We use natural language to describe images",
"concept. While earlier methods primarily explored image",
"and things that we see. Although this type of visual con-",
"inversion using generative adversarial networks (Xia et al.",
"cept learning is well-defined in human psychology (Murphy",
"2022), methods such as Textual Inversion (Gal et al. 2022)",
"2004), it remains elusive in the context of data-driven tech-",
"and Dreambooth (Ruiz et al. 2022) combine image inver-",
"niquescapableoflearningandreasoningfromimagesand",
"sion with T21 - this has led to an effective way to quickly",
"their natural language descriptions.",
"learn concepts from a few images and reproduce them in",
"Text-to-Image (T2I) generative models are trained to",
"novel combinations and compositions with other concepts,",
"translate natural language phrases into images that corre-",
"attributes, styles, etc. These methods aim to learn concepts",
"spond to that input. High-quality T2I models, therefore,",
"with minimal reference images by fine-tuning pre-trained",
"serve as a link between human-level concepts (expressed in",
"text-conditioned diffusion models (Figure 1). Therefore this",
"natural language) and their visual representations and are",
"paradigm of T2I and image inversion is a powerful new way",
"one way to reproduce visual concepts. On the other hand,",
"of learning and reproducing concepts.",
"*Corresponding Author:maitreya.patel@asu.edu",
"Within this paradigm of novel visual concept learning",
"Copyright 2024, Association for the Advancement of Artificial",
"via image inversion, two primary evaluation criteria have",
"Intelligence (www.aaai.org).All rights reserved.",
"emerged: (1) concept alignment, which assesses the corre-"
] | {
"arxiv_id": "2306.04695",
"language": "en",
"page_idx": 0
} |
||
2306.04695-en-1 | 2306.04695-en-1.jpg | <table>
<tbody>
<tr>
<th rowspan="5">Domain Art Painting Cartoon Photo Sketch</th>
<th colspan="4">ConceptBed Dataset</th>
<th rowspan="5">Attributes Yellow Wing Orange Beak 1 Brown Eyes</th>
</tr>
<tr>
<th> </th>
<th>Objec</th>
<th>$S$</th>
<th> </th>
</tr>
<tr>
<th>Dogs</th>
<th>Aircraft</th>
<th>Car</th>
<th>Fruit</th>
</tr>
<tr>
<th>Cats</th>
<th>Boat</th>
<th>Clock</th>
<th>Bird</th>
</tr>
<tr>
<th>Monkeys</th>
<th>Bottle</th>
<th>Fish</th>
<th>888</th>
</tr>
<tr>
<th>nte</th>
<th> </th>
<th>80 conc</th>
<th>pts</th>
<th> </th>
<th> </th>
</tr>
<tr>
<td>- T</td>
<td>Total</td>
<td>284 unic</td>
<td>ue c conce</td>
<td>pts</td>
<td>L TE IINW</td>
</tr>
<tr>
<td> </td>
<td>ConceF</td>
<td>$\mathbf{otBedCO}$</td>
<td>mpositi</td>
<td>ions</td>
<td> </td>
</tr>
<tr>
<td>Attribute</td>
<td>Relation</td>
<td> </td>
<td>1</td>
<td>n</td>
<td>Countine</td>
</tr>
<tr>
<td>TIUULO with red beak ears are up AredV* eating 18987 Promots</td>
<td>Meldu $V^{*}$ with fish $V^{*}$ sitting in be Ahorse besid 16902 Prom</td>
<td>$V$ ucket $V$ $A$ $=V^*$</td>
<td>ACTIO licking is is running $redV^{*}$ eat 8014</td>
<td>herself g ing</td>
<td>ILILILE Two V* s sitting A photo of three $eV^{*}$ 1 withacat $TwoV^*v$ 1083 Prompts</td>
</tr>
<tr>
<td>18987 Prompts</td>
<td>16902 Prom</td>
<td>pts</td>
<td>8014 Pron</td>
<td>npts</td>
<td>1083 F 1 Prompts</td>
</tr>
</tbody>
</table>
Figure 2: A summary of the CONCEPTBED dataset for largescale grounded evaluations of concept learners. The collection of concepts is categorized into three classes:(1) Domain, ( 2) Objects, and ( 3) Attributes. CoNCEPTBED has 284 unique concepts and four compositional categories. Here, $\mathbf{V}^{*}\hat{\mathrm{~is~a~learned~concept.}}$
spondence between the generated images and the target concept images, and (2) composition alignment, which evaluates whether the generated images maintain compositionality. Previous studies have been small scale, evaluating only a small number of hand-picked concepts and compositions; as such making generic claims via such findings is diffcult. Furthermore, the established evaluation metrics such as DINO-based cosine similarity (Ruiz et al. 2022) (for measuring concept alignment), KID (Kumari et al. 2022) (for measuring the amount of concept overfitting), and CLIPScore (Hessel et al. 2021) (for evaluating compositionality), have encountered challenges in accurately capturing human preferences. Consequently, there is a growing need for better automated evaluations.
Therefore, we introduce CoNCEPTBED, comprehensive dataset and evaluation framework that is aligned with human preferences. The CoNCEPTBED dataset comprises 284 distinct concepts and approximately 33,000 composite text prompts, which can be further extended using the provided automatic realistic dataset creation pipeline. The dataset focuses on four diverse concept learning evaluation tasks: learning styles, learning objects, learning attributes, and compositional reasoning. To gain a deeper understanding of previous methodologies, we incorporate four composition categories-action, attribution, counting, and relations.
We use our large-scale dataset to evaluate concept learners, by developing a novel evaluation metric called Concept Confidence Deviation (CCD). We conduct a human study and find that relative evaluations of models in terms of cci are well aligned with human preferences. Therefore, CCD combined with the CONCEPTBED dataset, offers an alternative to existing evaluation strategies, facilitating more effective large-scale evaluations. For each evaluation criteria, we train supervised classifiers (oracles) to detect whether generated concept images are accurate. Subsequently, the confidence scores from these oracles are utilized to calculate the instance-level concept deviations of the generated concept images in relation to the reference target ground truth images using the proposed CCD metric. This approach enables us to assess concept and composition alignment more effectively. We further show that CCD calculated using a pretrained few-shot classifier also maintains a high correlation with human preferences. This allows CCD to measure concept alignment on unseen concepts.
We conduct extensive experiments on four recently proposed concept learning methodologies. In total, we fine-tune approximately 1100 models (one model per concept) and generate over 500,000 concept-specific images. Our results reveal a surprising trade-off between concept alignment and composition alignment, wherein methods excelling at concept alignment tend to fall short in preserving compositions and vice versa. This suggests that previous concept learning approaches are either highly overfitted or severely underfitted. Furthermore, our experiments demonstrate that utilizing a pre-trained CLIP (Radford et al. 2021) textual encoder aids in maintaining compositionality, but it lacks the flexibility required to learn complex concepts, such as sket ch.
In summary, we make the following key contributions: $\cdot$ We introduce CONCEPTBED, a comprehensive bench-
mark for grounded quantitative evaluations of text-
conditioned concept learners.
$\bullet$ The Concept Confidence Deviation (CCD) evaluation metric, measures the learners' ability to preserve concepts and compositions. We demonstrate a strong correlation between CCD and human preferences.
$\bullet$ Through extensive experiments with 1,100+ models, we identify shortcomings in prior works and suggest future research directions. CONCEPTBED sets a standard for evaluating personalized text-to-image generative models.
## 2 $\textbf{Preliminaries}$
Prior studies on concept learning have focused on textconditioned diffusion models, such as Textual Inversion (Gal et al. 2022), DreamBooth (Ruiz et al. 2022), and Custom Diffusion (Kumari et al. 2022). These models operate within the T2I paradigm, where a text prompt $(y)$ serves as input to generate the corresponding image $(x_{gen})$ representing thegiven presenting then presenting the given prompt $y.$ A popular approach within T2I is the Latent Diffusion Model (LDM) (Rombach et al. 2022), which incorporates two key modules:
1. Textual Encoder $(C_\theta){:}$ This module generates embed-
dings corresponding to the input text prompt;
2. Generator $(\tilde\epsilon_\phi):$The generator estimates the noise itera-
tively from the input randomly sampled matrix at times-
tamp $t\left(z_{t}\right)$,conditioned on the text.
Since T2I models solely consider text input, the target concept $(c)$ is represented in terms of text tokens. These tokens can subsequently be employed to generate images associated with concept $c.$ Therefore, in Textual Inversion, the concept learning task is approached as an image inversion problem, aiming to map the target concept back to the text- ${\mathrm{embedding~space.}}$
Let V$^*$ denote the text tokens corresponding to the learned concept $c.$ Once the optimal mapping from V' to the target concept is determined, we can generate concept-specific images using the LDM by providing V$^*$ in the text prompt. | [
[
"0.0866",
"0.06946",
"0.478",
"0.06946",
"0.478",
"0.0833",
"0.0866",
"0.0833"
],
[
"0.652",
"0.07135",
"0.773",
"0.07135",
"0.773",
"0.0852",
"0.652",
"0.0852"
],
[
"0.5474",
"0.07764",
"0.5947",
"0.07764",
"0.5947",
"0.0884",
"0.5474",
"0.0884"
],
[
"0.8364",
"0.077",
"0.9004",
"0.077",
"0.9004",
"0.09155",
"0.8364",
"0.09155"
],
[
"0.0866",
"0.084",
"0.478",
"0.084",
"0.478",
"0.09784",
"0.0866",
"0.09784"
],
[
"0.5415",
"0.089",
"0.6006",
"0.089",
"0.6006",
"0.101",
"0.5415",
"0.101"
],
[
"0.694",
"0.0871",
"0.738",
"0.0871",
"0.738",
"0.0985",
"0.694",
"0.0985"
],
[
"0.7793",
"0.09344",
"0.808",
"0.0985",
"0.8037",
"0.1111",
"0.7754",
"0.1061"
],
[
"0.839",
"0.0966",
"0.898",
"0.0966",
"0.898",
"0.108",
"0.839",
"0.108"
],
[
"0.6323",
"0.09784",
"0.661",
"0.09784",
"0.661",
"0.1092",
"0.6323",
"0.1092"
],
[
"0.0866",
"0.09784",
"0.478",
"0.09784",
"0.478",
"0.1111",
"0.0866",
"0.1111"
],
[
"0.689",
"0.0985",
"0.7344",
"0.0985",
"0.7344",
"0.1092",
"0.689",
"0.1092"
],
[
"0.7305",
"0.09973",
"0.7563",
"0.09973",
"0.7563",
"0.108",
"0.7305",
"0.108"
],
[
"0.687",
"0.1061",
"0.7183",
"0.108",
"0.7173",
"0.1206",
"0.6855",
"0.1187"
],
[
"0.7344",
"0.1061",
"0.7686",
"0.1061",
"0.7686",
"0.1218",
"0.7344",
"0.1218"
],
[
"0.8364",
"0.1067",
"0.9004",
"0.1048",
"0.9004",
"0.1193",
"0.8364",
"0.1218"
],
[
"0.6323",
"0.108",
"0.66",
"0.108",
"0.66",
"0.1206",
"0.6323",
"0.1206"
],
[
"0.7754",
"0.108",
"0.8037",
"0.108",
"0.8037",
"0.1193",
"0.7754",
"0.1193"
],
[
"0.5537",
"0.1111",
"0.591",
"0.1111",
"0.591",
"0.1218",
"0.5537",
"0.1218"
],
[
"0.5483",
"0.101",
"0.5947",
"0.101",
"0.5947",
"0.11237",
"0.5483",
"0.11237"
],
[
"0.08826",
"0.113",
"0.4763",
"0.113",
"0.4763",
"0.1237",
"0.08826",
"0.1237"
],
[
"0.734",
"0.11743",
"0.7646",
"0.1193",
"0.763",
"0.1307",
"0.732",
"0.1288"
],
[
"0.839",
"0.1187",
"0.898",
"0.1187",
"0.898",
"0.13",
"0.839",
"0.13"
],
[
"0.6807",
"0.1193",
"0.722",
"0.1193",
"0.722",
"0.13",
"0.6807",
"0.13"
],
[
"0.5537",
"0.1218",
"0.589",
"0.1218",
"0.589",
"0.1332",
"0.5537",
"0.1332"
],
[
"0.6323",
"0.1193",
"0.6855",
"0.1193",
"0.6855",
"0.13",
"0.6323",
"0.13"
],
[
"0.0866",
"0.127",
"0.4763",
"0.127",
"0.4763",
"0.137",
"0.0866",
"0.137"
],
[
"0.544",
"0.1332",
"0.5947",
"0.1332",
"0.5947",
"0.1445",
"0.544",
"0.1445"
],
[
"0.693",
"0.1307",
"0.732",
"0.1307",
"0.732",
"0.1383",
"0.693",
"0.1383"
],
[
"0.08496",
"0.1395",
"0.478",
"0.1395",
"0.478",
"0.1528",
"0.08496",
"0.1528"
],
[
"0.835",
"0.1351",
"0.899",
"0.1351",
"0.899",
"0.1459",
"0.835",
"0.1459"
],
[
"0.661",
"0.1427",
"0.786",
"0.1427",
"0.786",
"0.1528",
"0.661",
"0.1528"
],
[
"0.0866",
"0.154",
"0.4763",
"0.154",
"0.4763",
"0.1642",
"0.0866",
"0.1642"
],
[
"0.642",
"0.1559",
"0.7974",
"0.1559",
"0.7974",
"0.1692",
"0.642",
"0.1692"
],
[
"0.08826",
"0.1654",
"0.4763",
"0.1654",
"0.4763",
"0.1787",
"0.08826",
"0.1787"
],
[
"0.5474",
"0.1704",
"0.596",
"0.1704",
"0.596",
"0.1818",
"0.5474",
"0.1818"
],
[
"0.6445",
"0.1686",
"0.6855",
"0.1704",
"0.6846",
"0.1831",
"0.6436",
"0.1812"
],
[
"0.7466",
"0.1718",
"0.7837",
"0.1718",
"0.7837",
"0.1831",
"0.7466",
"0.1831"
],
[
"0.841",
"0.1704",
"0.886",
"0.1704",
"0.886",
"0.1818",
"0.841",
"0.1818"
],
[
"0.625",
"0.1787",
"0.6807",
"0.1812",
"0.6787",
"0.1957",
"0.6235",
"0.1932"
],
[
"0.7197",
"0.1799",
"0.8037",
"0.1818",
"0.8037",
"0.1963",
"0.7197",
"0.1945"
],
[
"0.8164",
"0.1787",
"0.884",
"0.1812",
"0.884",
"0.1957",
"0.8164",
"0.1932"
],
[
"0.536",
"0.1843",
"0.609",
"0.1843",
"0.609",
"0.1957",
"0.536",
"0.1957"
],
[
"0.08496",
"0.1812",
"0.4795",
"0.1812",
"0.4795",
"0.1945",
"0.08496",
"0.1945"
],
[
"0.535",
"0.1925",
"0.6055",
"0.1945",
"0.6055",
"0.2058",
"0.535",
"0.2039"
],
[
"0.6235",
"0.1913",
"0.705",
"0.1925",
"0.705",
"0.2039",
"0.6235",
"0.2026"
],
[
"0.721",
"0.1925",
"0.7793",
"0.1925",
"0.7793",
"0.2026",
"0.721",
"0.2026"
],
[
"0.819",
"0.1932",
"0.8945",
"0.1932",
"0.8945",
"0.2039",
"0.819",
"0.2039"
],
[
"0.535",
"0.202",
"0.603",
"0.2026",
"0.603",
"0.214",
"0.535",
"0.2134"
],
[
"0.08826",
"0.1945",
"0.4763",
"0.1945",
"0.4763",
"0.2084",
"0.08826",
"0.2084"
],
[
"0.721",
"0.2026",
"0.789",
"0.2026",
"0.789",
"0.2134",
"0.721",
"0.2134"
],
[
"0.819",
"0.2026",
"0.8945",
"0.2026",
"0.8945",
"0.2134",
"0.819",
"0.2134"
],
[
"0.626",
"0.2039",
"0.701",
"0.2039",
"0.701",
"0.2108",
"0.626",
"0.2108"
],
[
"0.08826",
"0.2084",
"0.478",
"0.2084",
"0.478",
"0.2216",
"0.08826",
"0.2216"
],
[
"0.539",
"0.214",
"0.6006",
"0.2153",
"0.6006",
"0.226",
"0.539",
"0.2247"
],
[
"0.6313",
"0.2153",
"0.6978",
"0.2153",
"0.6978",
"0.226",
"0.6313",
"0.226"
],
[
"0.728",
"0.2153",
"0.7915",
"0.2153",
"0.7915",
"0.226",
"0.728",
"0.226"
],
[
"0.8325",
"0.214",
"0.8936",
"0.214",
"0.8936",
"0.2247",
"0.8325",
"0.2247"
],
[
"0.0866",
"0.2229",
"0.478",
"0.2229",
"0.478",
"0.2361",
"0.0866",
"0.2361"
],
[
"0.08826",
"0.2361",
"0.478",
"0.2361",
"0.478",
"0.25",
"0.08826",
"0.25"
],
[
"0.5215",
"0.2437",
"0.91",
"0.2437",
"0.91",
"0.2568",
"0.5215",
"0.2568"
],
[
"0.08826",
"0.252",
"0.2361",
"0.252",
"0.2361",
"0.262",
"0.08826",
"0.262"
],
[
"0.10297",
"0.2632",
"0.4795",
"0.2632",
"0.4795",
"0.2778",
"0.10297",
"0.2778"
],
[
"0.523",
"0.2588",
"0.9087",
"0.2588",
"0.9087",
"0.2695",
"0.523",
"0.2695"
],
[
"0.5215",
"0.2727",
"0.91",
"0.2727",
"0.91",
"0.2827",
"0.5215",
"0.2827"
],
[
"0.08826",
"0.2798",
"0.4756",
"0.2798",
"0.4756",
"0.2903",
"0.08826",
"0.2903"
],
[
"0.5195",
"0.2854",
"0.9126",
"0.2854",
"0.9126",
"0.2986",
"0.5195",
"0.2986"
],
[
"0.0866",
"0.2922",
"0.478",
"0.2922",
"0.478",
"0.3062",
"0.0866",
"0.3062"
],
[
"0.5215",
"0.2998",
"0.9126",
"0.2998",
"0.9126",
"0.3132",
"0.5215",
"0.3132"
],
[
"0.08496",
"0.3062",
"0.4795",
"0.3062",
"0.4795",
"0.3193",
"0.08496",
"0.3193"
],
[
"0.5215",
"0.3113",
"0.6787",
"0.3125",
"0.6787",
"0.327",
"0.5215",
"0.3257"
],
[
"0.08417",
"0.3193",
"0.478",
"0.318",
"0.478",
"0.332",
"0.08417",
"0.3333"
],
[
"0.08826",
"0.3352",
"0.4756",
"0.3352",
"0.4756",
"0.3452",
"0.08826",
"0.3452"
],
[
"0.0866",
"0.348",
"0.478",
"0.348",
"0.478",
"0.361",
"0.0866",
"0.361"
],
[
"0.5317",
"0.3542",
"0.9126",
"0.3547",
"0.9126",
"0.3694",
"0.5317",
"0.3687"
],
[
"0.0866",
"0.361",
"0.478",
"0.361",
"0.478",
"0.375",
"0.0866",
"0.375"
],
[
"0.0866",
"0.3757",
"0.4795",
"0.3757",
"0.4795",
"0.3894",
"0.0866",
"0.3894"
],
[
"0.5337",
"0.3718",
"0.7305",
"0.3718",
"0.7305",
"0.3818",
"0.5337",
"0.3818"
],
[
"0.08826",
"0.3894",
"0.4795",
"0.3894",
"0.4795",
"0.4028",
"0.08826",
"0.4028"
],
[
"0.5215",
"0.3838",
"0.911",
"0.3838",
"0.911",
"0.3977",
"0.5215",
"0.3977"
],
[
"0.0866",
"0.4028",
"0.4495",
"0.4028",
"0.4495",
"0.4167",
"0.0866",
"0.4167"
],
[
"0.5337",
"0.3977",
"0.9126",
"0.3977",
"0.9126",
"0.411",
"0.5337",
"0.411"
],
[
"0.5337",
"0.411",
"0.91",
"0.411",
"0.91",
"0.4248",
"0.5337",
"0.4248"
],
[
"0.10297",
"0.4172",
"0.478",
"0.4172",
"0.478",
"0.4312",
"0.10297",
"0.4312"
],
[
"0.5337",
"0.4236",
"0.7646",
"0.4248",
"0.7646",
"0.4382",
"0.5337",
"0.4375"
],
[
"0.0866",
"0.4312",
"0.4795",
"0.4312",
"0.4795",
"0.4443",
"0.0866",
"0.4443"
],
[
"0.5215",
"0.4382",
"0.914",
"0.4382",
"0.914",
"0.4526",
"0.5215",
"0.4526"
],
[
"0.08826",
"0.4463",
"0.4763",
"0.4463",
"0.4763",
"0.457",
"0.08826",
"0.457"
],
[
"0.536",
"0.4539",
"0.911",
"0.4539",
"0.911",
"0.464",
"0.536",
"0.464"
],
[
"0.0866",
"0.4602",
"0.4756",
"0.4602",
"0.4756",
"0.4702",
"0.0866",
"0.4702"
],
[
"0.5337",
"0.4666",
"0.911",
"0.4666",
"0.911",
"0.4797",
"0.5337",
"0.4797"
],
[
"0.0866",
"0.4734",
"0.4763",
"0.4734",
"0.4763",
"0.4841",
"0.0866",
"0.4841"
],
[
"0.08496",
"0.486",
"0.4739",
"0.4873",
"0.4739",
"0.4988",
"0.08496",
"0.4976"
],
[
"0.5337",
"0.481",
"0.9087",
"0.481",
"0.9087",
"0.4944",
"0.5337",
"0.4944"
],
[
"0.0866",
"0.502",
"0.4739",
"0.502",
"0.4739",
"0.512",
"0.0866",
"0.512"
],
[
"0.08417",
"0.513",
"0.4795",
"0.5146",
"0.4795",
"0.528",
"0.08417",
"0.5264"
],
[
"0.6436",
"0.509",
"0.789",
"0.509",
"0.789",
"0.5234",
"0.6436",
"0.5234"
],
[
"0.08496",
"0.528",
"0.4756",
"0.528",
"0.4756",
"0.5415",
"0.08496",
"0.5415"
],
[
"0.5215",
"0.53",
"0.911",
"0.53",
"0.911",
"0.5435",
"0.5215",
"0.5435"
],
[
"0.08826",
"0.5435",
"0.4763",
"0.5435",
"0.4763",
"0.5537",
"0.08826",
"0.5537"
],
[
"0.523",
"0.5454",
"0.911",
"0.5454",
"0.911",
"0.556",
"0.523",
"0.556"
],
[
"0.08826",
"0.5566",
"0.4763",
"0.5566",
"0.4763",
"0.5674",
"0.08826",
"0.5674"
],
[
"0.5195",
"0.556",
"0.9126",
"0.5566",
"0.9126",
"0.5713",
"0.5195",
"0.571"
],
[
"0.0866",
"0.571",
"0.4739",
"0.571",
"0.4739",
"0.5806",
"0.0866",
"0.5806"
],
[
"0.5215",
"0.571",
"0.9126",
"0.571",
"0.9126",
"0.584",
"0.5215",
"0.584"
],
[
"0.08826",
"0.5854",
"0.4763",
"0.5854",
"0.4763",
"0.595",
"0.08826",
"0.595"
],
[
"0.5215",
"0.587",
"0.911",
"0.587",
"0.911",
"0.5977",
"0.5215",
"0.5977"
],
[
"0.0866",
"0.5977",
"0.4763",
"0.5977",
"0.4763",
"0.6113",
"0.0866",
"0.6113"
],
[
"0.5195",
"0.601",
"0.9126",
"0.601",
"0.9126",
"0.6143",
"0.5195",
"0.6143"
],
[
"0.0866",
"0.6113",
"0.478",
"0.6123",
"0.478",
"0.6235",
"0.0866",
"0.6226"
],
[
"0.5195",
"0.613",
"0.911",
"0.6123",
"0.911",
"0.627",
"0.5195",
"0.6274"
],
[
"0.08826",
"0.625",
"0.4756",
"0.6255",
"0.4756",
"0.637",
"0.08826",
"0.636"
],
[
"0.5195",
"0.627",
"0.9126",
"0.627",
"0.9126",
"0.64",
"0.5195",
"0.64"
],
[
"0.08496",
"0.638",
"0.4795",
"0.638",
"0.4795",
"0.6514",
"0.08496",
"0.6514"
],
[
"0.5195",
"0.6416",
"0.7256",
"0.6416",
"0.7256",
"0.655",
"0.5195",
"0.655"
],
[
"0.08826",
"0.6514",
"0.478",
"0.6514",
"0.478",
"0.6655",
"0.08826",
"0.6655"
],
[
"0.0866",
"0.6665",
"0.3186",
"0.6675",
"0.3186",
"0.6787",
"0.0866",
"0.678"
],
[
"0.523",
"0.661",
"0.91",
"0.661",
"0.91",
"0.675",
"0.523",
"0.675"
],
[
"0.10297",
"0.6797",
"0.4763",
"0.6797",
"0.4763",
"0.6943",
"0.10297",
"0.6943"
],
[
"0.5415",
"0.675",
"0.839",
"0.675",
"0.839",
"0.688",
"0.5415",
"0.688"
],
[
"0.5215",
"0.6895",
"0.546",
"0.6895",
"0.546",
"0.7007",
"0.5215",
"0.7007"
],
[
"0.539",
"0.6875",
"0.911",
"0.688",
"0.911",
"0.7026",
"0.539",
"0.702"
],
[
"0.0866",
"0.6943",
"0.478",
"0.6943",
"0.478",
"0.7085",
"0.0866",
"0.7085"
],
[
"0.0866",
"0.7085",
"0.4795",
"0.7085",
"0.4795",
"0.7217",
"0.0866",
"0.7217"
],
[
"0.5415",
"0.7026",
"0.91",
"0.7026",
"0.91",
"0.7163",
"0.5415",
"0.7163"
],
[
"0.0866",
"0.7217",
"0.4763",
"0.7217",
"0.4763",
"0.735",
"0.0866",
"0.735"
],
[
"0.5435",
"0.7183",
"0.7754",
"0.7183",
"0.7754",
"0.729",
"0.5435",
"0.729"
],
[
"0.0866",
"0.7373",
"0.4763",
"0.7373",
"0.4763",
"0.7476",
"0.0866",
"0.7476"
],
[
"0.5376",
"0.7363",
"0.914",
"0.7363",
"0.914",
"0.75",
"0.5376",
"0.75"
],
[
"0.0866",
"0.752",
"0.4763",
"0.752",
"0.4763",
"0.762",
"0.0866",
"0.762"
],
[
"0.5215",
"0.75",
"0.9126",
"0.75",
"0.9126",
"0.763",
"0.5215",
"0.763"
],
[
"0.08496",
"0.763",
"0.4763",
"0.763",
"0.4763",
"0.7764",
"0.08496",
"0.7764"
],
[
"0.5215",
"0.7666",
"0.91",
"0.7666",
"0.91",
"0.7764",
"0.5215",
"0.7764"
],
[
"0.0866",
"0.778",
"0.4763",
"0.779",
"0.4763",
"0.7905",
"0.0866",
"0.789"
],
[
"0.523",
"0.78",
"0.911",
"0.78",
"0.911",
"0.7905",
"0.523",
"0.7905"
],
[
"0.0866",
"0.7915",
"0.4763",
"0.7915",
"0.4763",
"0.8047",
"0.0866",
"0.8047"
],
[
"0.523",
"0.7935",
"0.911",
"0.7935",
"0.911",
"0.8037",
"0.523",
"0.8037"
],
[
"0.08496",
"0.8047",
"0.478",
"0.8047",
"0.478",
"0.8184",
"0.08496",
"0.8184"
],
[
"0.5215",
"0.806",
"0.9126",
"0.806",
"0.9126",
"0.8193",
"0.5215",
"0.8193"
],
[
"0.0866",
"0.8184",
"0.478",
"0.8184",
"0.478",
"0.832",
"0.0866",
"0.832"
],
[
"0.5195",
"0.8193",
"0.6396",
"0.821",
"0.6396",
"0.832",
"0.5195",
"0.8306"
],
[
"0.08496",
"0.8335",
"0.4763",
"0.8335",
"0.4763",
"0.8467",
"0.08496",
"0.8467"
],
[
"0.536",
"0.834",
"0.9126",
"0.834",
"0.9126",
"0.8477",
"0.536",
"0.8477"
],
[
"0.08496",
"0.8467",
"0.4395",
"0.8467",
"0.4395",
"0.86",
"0.08496",
"0.86"
],
[
"0.5186",
"0.8477",
"0.9126",
"0.8467",
"0.9126",
"0.8613",
"0.5186",
"0.8623"
],
[
"0.1013",
"0.8613",
"0.468",
"0.86",
"0.468",
"0.8735",
"0.1013",
"0.875"
],
[
"0.5195",
"0.8623",
"0.911",
"0.8623",
"0.911",
"0.8755",
"0.5195",
"0.8755"
],
[
"0.08826",
"0.8755",
"0.478",
"0.8755",
"0.478",
"0.8896",
"0.08826",
"0.8896"
],
[
"0.5195",
"0.8755",
"0.9126",
"0.8755",
"0.9126",
"0.8896",
"0.5195",
"0.8896"
]
] | [
"spondence between the generated images and the target con-",
"ConceptBed Dataset",
"Domain",
"Attributes",
"cept images, and (2) composition alignment, which evalu-",
"Art Painting",
"Obiects",
"Fruit",
"YellowWing",
"Dogs",
"ates whether the generated images maintain compositional-",
"Aircraft",
"Car",
"Boat",
"Clock",
"Orange Beak",
"Cats",
"Bird",
"Photo",
"Cartoon",
"ity.Previous studieshavebeen small scale,evaluating only",
"Fish",
"Brown Eyes",
"Bottle",
"Sketch",
"Monkeys",
"asmallnumberofhand-pickedconceptsandcompositions:",
"4concepts",
"80.conce",
"as such making generic claims via such findings is diff-",
"112attributes",
"Total:284uniqueconcepts",
"cult.Furthermore,theestablishedevaluationmetrics suchas",
"ConceptBed Compositions",
"DINO-based cosine similarity (Ruiz et al. 2022) (for mea-",
"Attribute",
"Relation",
"Action",
"Counting",
"with fish",
"vis licking herself",
"TwoVsiting",
"V* with red beak",
"suring concept alignment), KID (Kumari et al. 2022) (for",
"Va's ears are up",
"V*sitting in bucket",
"V*is running",
"Aphotoofthree",
"Ared v*eating",
"measuring the amount of concept overfitting), and CLIP-",
"A red V*eating",
"Two V*withacat",
"Ahorsebesidey*",
"Score (Hessel et al. 2021) (for evaluating compositionality),",
"18987Prompt",
"16902Prompts",
"8014Prompts",
"1083Prompts",
"have encountered challenges in accurately capturing human",
"preferences. Consequently, there is a growing need for better",
"Figure 2: A summary of the CONCEPTBED dataset for large-",
"automatedevaluations.",
"Therefore, we introduce CoNCEPTBED, comprehensive",
"scalegroundedevaluationsofconceptlearners.Thecollec",
"tion of concepts is categorized into three classes:(1)Do-",
"dataset and evaluation framework that is aligned with hu-",
"main, (2)Objects, and (3)Attributes. CoNCEPTBED has 284",
"man preferences. The CoNCEPTBED dataset comprises 284",
"unique concepts and four compositional categories. Here,",
"distinct concepts and approximately 33,000 composite text",
"V* is a learned concept.",
"prompts, which can be further extended using the provided",
"automaticrealisticdatasetcreationpipeline.The datasetfo-",
"cuses on four diverse concept learning evaluation tasks:",
"mark for grounded quantitative evaluations of text-",
"learning styles, learning objects, learning attributes, and",
"compositional reasoning. To gain a deeper understanding of",
"conditionedconceptlearners.",
"previous methodologies,we incorporate four composition",
": The Concept Confidence Deviation (CCD) evaluation met-",
"categories - action, attribution, counting, and relations.",
"ric, measures the learners' ability to preserve concepts and",
"compositions. We demonstrate a strong correlation be-",
"We use our large-scale dataset to evaluate concept learn-",
"tween ccD andhuman preferences",
"ers, by developing a novel evaluation metric called Concept",
": Through extensive experiments with 1,100+ models, we",
"Confidence Deviation(CCD).We conduct a human study",
"identify shortcomings in prior works and suggestfuture",
"andfindthatrelativeevaluationsofmodels intermsof ccL",
"research directions. CoNCEPTBED sets a standard for",
"are well aligned with human preferences.Therefore,CcD",
"combinedwiththeCoNCEPTBEDdataset,offersanalterna",
"evaluating personalized text-to-image generative models.",
"tivetoexisting evaluation strategies,facilitatingmoreeffec",
"tive large-scale evaluations. For each evaluation criteria, we",
"2Preliminaries",
"train supervised classifiers (oracles) to detect whether gen",
"Prior studies on concept learning have focused on text-",
"erated conceptimages areaccurate.Subsequently,thecon-",
"conditioneddiffusionmodels,suchasTextual Inversion(Gal",
"fidencescoresfromtheseoracles areutilizedtocalculate",
"et al. 2022), DreamBooth (Ruiz et al. 2022), and Custom",
"theinstance-level concept deviations of the generated con",
"Diffusion (Kumari et al. 2022). These models operate within",
"cept images in relation to the reference target ground truth",
"theT2Iparadigm,whereatextprompt(y)servesasinput",
"images using the proposed ccD metric. This approach en-",
"to generate the corresponding image ( gen)representing the",
"ables us toassess conceptand composition alignmentmore",
"given prompt y. A popular approach within T2I is the La-",
"effectively.Wefurther show that ccD calculated using apre-",
"tent Diffusion Model (LDM) (Rombach et al. 2022), which",
"trained few-shot classifier also maintains a high correlation",
"incorporates two key modules:",
"with human preferences. This allows ccD to measure con-",
"ceptalignmenton unseen concepts.",
"1. Textual Encoder (Ce): This module generates embed-",
"We conduct extensive experiments on four recently pro-",
"dings corresponding to the input text prompt;",
"2.",
"Generator (e): The generator estimates the noise itera-",
"posed concept learning methodologies. In total, we fine-tune",
"approximately 1100 models (one model per concept) and",
"tively from the input randomly sampled matrix at times-",
"generate over 500,000 concept-specific images.Our results",
"tampt (zt),conditioned onthetext.",
"reveal asurprisingtrade-offbetweenconceptalignmentand",
"Since T2I models solely consider text input, the target",
"composition alignment,wherein methods excelling at con-",
"concept (c) is represented in terms of text tokens. These to-",
"cept alignment tend to fall short in preserving compositions",
"kens cansubsequentlybe employedto generateimages as-",
"and viceversa.This suggests that previous concept learning",
"sociated withconcept c.Therefore,in TextualInversion,the",
"approaches are either highly overfitted or severely underfit-",
"conceptlearningtaskis approached as animageinversion",
"ted. Furthermore, our experiments demonstrate that utilizing",
"problem, aiming to map the target concept back to the text-",
"a pre-trained CLIP (Radford et al. 2021) textual encoder aids",
"embedding space.",
"in maintaining compositionality, but it lacks the flexibility",
"Let V* denote the text tokens corresponding to the learned",
"required to learn complex concepts, such as sketch.",
"concept c. Once the optimal mapping from V* to the target",
"In summary, we make the following key contributions:",
"concept is determined, we can generate concept-specific im-",
".We introduce CoNCEPTBED, a comprehensive bench-",
"ages using the LDM by providing V* in the text prompt."
] | {
"arxiv_id": "2306.04695",
"language": "en",
"page_idx": 1
} |
||
2306.04695-en-2 | 2306.04695-en-2.jpg | <table>
<tbody>
<tr>
<td>$\operatorname{Algorithn}$</td>
<td> </td>
</tr>
<tr>
<td>6: for</td>
<td> </td>
</tr>
<tr>
<td>7: $T$</td>
<td> </td>
</tr>
<tr>
<td>8: 'e $\mathcal{p}$ T</td>
<td> </td>
</tr>
<tr>
<td>9: for</td>
<td> </td>
</tr>
<tr>
<td>10:</td>
<td> </td>
</tr>
<tr>
<td>11:</td>
<td>$\{ / / /$ Eq. 3$\}$</td>
</tr>
<tr>
<td>12: $CCD$</td>
<td> </td>
</tr>
</tbody>
</table>
Suppose we are provided with $m$ images $(X_1:m)$ of the target concept $c.$ Now, in order to learn the text tokens V* corresponding to the concept $c$ from the set of images $X_1:m$,the Textual Inversion methodology aims to optimize V$^*$ by reconstructing $X_{1:m}$ using the objective function of the LDM with frozen parameters $\theta$ and $\phi:$
(1)
$$V^{*}=\mathop{\mathrm{argmin}}_{v}\mathop{\mathbb{E}}_{\epsilon\sim\mathcal{N}(0,1),\stackrel{x\in X_{1:m},\:t,}{z\sim\mathcal{E}(x)}}||\epsilon-\epsilon_{\phi}(z_{t},t,x,C_{\theta}(y))||_{2}^{2}$$
In the case of DreamBooth and Custom Diffusion, instead of finding the optimal V$^*$, it optimizes the model parameter $\phi$ associated with the noise estimator $(\epsilon_\phi).$ This optimization process enables the model to learn the mapping between randomly initialized V$^*$ and the target concept $c.$
$\begin{array}{cc}{\vdots}&{(2)}\end{array}$
$$\phi^{*}=\underset{\phi}{\operatorname*{argmin}}\underset{x\in X_{1:m},\:t,}{\operatorname*{\mathbb{E}}}||\epsilon-\epsilon_{\theta}(z_{t},t,x,C_{\phi}(y))||_{2}^{2}$$
Once $\phi^*$ is obtained, it can be used to generate images re-
lated to the target concept.$^{1}$
Once the images are generated, in order to evaluate these generated images, it is essential to verify whether they align with the learned concepts while maintaining compositional- $ity.$
## 3 CoNCEPTBED
In this section, we introduce CONCEPTBED, a comprehensive collection of concepts, designed to accurately estimate concept and composition alignment by quantifying deviations in the generated images. Later, we introduce the novel evaluation framework associated with CONCEPTBED. Please refer to the Appendix for additional insights on the proposed dataset and evaluation framework.
l DreamBooth and Custom-Diffusion use additional regularizer to improve compositionally by using same objective function on a $\bar{\text{diverse set of image-caption pairs.}}$
<FigureHere>
Figure 3: Qualitative examples showcasing the effectiveness of concept learners on the CONCEPTBED dataset. The leftmost column displays four instances of ground truth target concept images ( $\dot{\mathrm{V}}^*).$ Subsequent columns exhibit target concept-specific images generated by all baseline methods.
3.1 CoNCEPTBED: Dataset Construction
CONCEPTBED incorporates existing datasets such as ImageNet (Deng et al. 2009), PACS (Li et al. 2017), CUB (Wah et al. 2011), and Visual Genome (Krishna et al. 2017), enabling the creation of a labeled dataset. Figure 2 provides an overview of the CoNCEPTBED dataset.
Learning Styles. We use styles from the PACS dataset: Art Painting, Cartoon, Photo, and Sketch. Each style contains images corresponding to seven categories. The concept learner aims to use examples from one style as a reference and generate style-specific images for all seven entities. Learning Objects. Extracting object-level concepts is accomplished through the utilization of the ImageNet dataset. It comprises 1000 low-level concepts from the Word- $\hat{\text{Net (Fellbaum 2010) hierarchy. However, due to the pres-}}$ ence of noise in ImageNet images and the lack of relevance to daily life for many concepts, we employ an automated filtering pipeline to ensure the usefulness and quality of the reference concept images. The pipeline involves extracting a list of low-level concepts and their parent concepts from | [
[
"0.08496",
"0.0682",
"0.3774",
"0.0682",
"0.3774",
"0.0821",
"0.08496",
"0.0821"
],
[
"0.539",
"0.0707",
"0.596",
"0.0707",
"0.596",
"0.07764",
"0.539",
"0.07764"
],
[
"0.6274",
"0.0707",
"0.687",
"0.0707",
"0.687",
"0.07764",
"0.6274",
"0.07764"
],
[
"0.7236",
"0.0707",
"0.761",
"0.0707",
"0.761",
"0.07764",
"0.7236",
"0.07764"
],
[
"0.7793",
"0.0745",
"0.8267",
"0.0745",
"0.8267",
"0.0821",
"0.7793",
"0.0821"
],
[
"0.841",
"0.0726",
"0.9067",
"0.0726",
"0.9067",
"0.0833",
"0.841",
"0.0833"
],
[
"0.5415",
"0.0789",
"0.5933",
"0.0789",
"0.5933",
"0.0859",
"0.5415",
"0.0859"
],
[
"0.6436",
"0.0789",
"0.669",
"0.0789",
"0.669",
"0.0871",
"0.6436",
"0.0871"
],
[
"0.721",
"0.0789",
"0.74",
"0.0789",
"0.74",
"0.0871",
"0.721",
"0.0871"
],
[
"0.10046",
"0.0884",
"0.4756",
"0.0884",
"0.4756",
"0.1029",
"0.10046",
"0.1029"
],
[
"0.11273",
"0.09973",
"0.4272",
"0.1016",
"0.4272",
"0.11615",
"0.11273",
"0.11426"
],
[
"0.6357",
"0.0985",
"0.6445",
"0.0985",
"0.6445",
"0.1016",
"0.6357",
"0.1016"
],
[
"0.11273",
"0.1111",
"0.4004",
"0.113",
"0.4004",
"0.1307",
"0.11273",
"0.1288"
],
[
"0.1152",
"0.1288",
"0.3105",
"0.1288",
"0.3105",
"0.1427",
"0.1152",
"0.1427"
],
[
"0.8813",
"0.1256",
"0.892",
"0.1256",
"0.892",
"0.1288",
"0.8813",
"0.1288"
],
[
"0.10046",
"0.1401",
"0.2484",
"0.1401",
"0.2484",
"0.154",
"0.10046",
"0.154"
],
[
"0.694",
"0.1445",
"0.749",
"0.1445",
"0.749",
"0.1483",
"0.694",
"0.1483"
],
[
"0.751",
"0.1433",
"0.831",
"0.1433",
"0.831",
"0.1509",
"0.751",
"0.1509"
],
[
"0.0907",
"0.1515",
"0.3105",
"0.1515",
"0.3105",
"0.1692",
"0.0907",
"0.1692"
],
[
"0.08905",
"0.1654",
"0.2524",
"0.1674",
"0.2524",
"0.1843",
"0.08905",
"0.1818"
],
[
"0.09235",
"0.1818",
"0.1054",
"0.1818",
"0.1054",
"0.1925",
"0.09235",
"0.1925"
],
[
"0.125",
"0.1818",
"0.1584",
"0.1755",
"0.1626",
"0.1875",
"0.1283",
"0.1932"
],
[
"0.09235",
"0.1945",
"0.1054",
"0.1945",
"0.1054",
"0.2052",
"0.09235",
"0.2052"
],
[
"0.125",
"0.1913",
"0.2378",
"0.1925",
"0.2378",
"0.207",
"0.125",
"0.2058"
],
[
"0.0907",
"0.2052",
"0.10864",
"0.2052",
"0.10864",
"0.2184",
"0.0907",
"0.2184"
],
[
"0.1422",
"0.207",
"0.1814",
"0.207",
"0.1814",
"0.2172",
"0.1422",
"0.2172"
],
[
"0.7256",
"0.209",
"0.8057",
"0.209",
"0.8057",
"0.2166",
"0.7256",
"0.2166"
],
[
"0.09235",
"0.2197",
"0.10706",
"0.2197",
"0.10706",
"0.2299",
"0.09235",
"0.2299"
],
[
"0.1259",
"0.2172",
"0.2484",
"0.2172",
"0.2484",
"0.2317",
"0.1259",
"0.2317"
],
[
"0.0907",
"0.2299",
"0.10864",
"0.2299",
"0.10864",
"0.2443",
"0.0907",
"0.2443"
],
[
"0.1396",
"0.226",
"0.255",
"0.2279",
"0.255",
"0.2468",
"0.1396",
"0.2443"
],
[
"0.0907",
"0.2456",
"0.10864",
"0.2456",
"0.10864",
"0.2588",
"0.0907",
"0.2588"
],
[
"0.1259",
"0.2468",
"0.156",
"0.2393",
"0.1602",
"0.2507",
"0.1299",
"0.2583"
],
[
"0.1854",
"0.2443",
"0.2255",
"0.2443",
"0.2255",
"0.2551",
"0.1854",
"0.2551"
],
[
"0.1814",
"0.252",
"0.2051",
"0.252",
"0.2051",
"0.2588",
"0.1814",
"0.2588"
],
[
"0.2279",
"0.2474",
"0.2566",
"0.2456",
"0.2573",
"0.2583",
"0.2296",
"0.26"
],
[
"0.09235",
"0.2588",
"0.1054",
"0.2588",
"0.1054",
"0.2695",
"0.09235",
"0.2695"
],
[
"0.1274",
"0.2588",
"0.1748",
"0.2588",
"0.1748",
"0.2695",
"0.1274",
"0.2695"
],
[
"0.1896",
"0.2568",
"0.2443",
"0.2588",
"0.2443",
"0.2708",
"0.1896",
"0.2683"
],
[
"0.0866",
"0.2708",
"0.1095",
"0.2708",
"0.1095",
"0.2842",
"0.0866",
"0.2842"
],
[
"0.1438",
"0.2727",
"0.2345",
"0.2708",
"0.2345",
"0.2842",
"0.1438",
"0.286"
],
[
"0.694",
"0.2727",
"0.8286",
"0.2734",
"0.8286",
"0.2842",
"0.694",
"0.2827"
],
[
"0.2059",
"0.2842",
"0.3652",
"0.281",
"0.3652",
"0.2986",
"0.2059",
"0.3018"
],
[
"0.4182",
"0.2842",
"0.4795",
"0.2842",
"0.4795",
"0.2986",
"0.4182",
"0.2986"
],
[
"0.1447",
"0.2874",
"0.1978",
"0.2854",
"0.1978",
"0.2966",
"0.1447",
"0.2986"
],
[
"0.0866",
"0.2854",
"0.10864",
"0.2854",
"0.10864",
"0.2974",
"0.0866",
"0.2974"
],
[
"0.0825",
"0.2935",
"0.2695",
"0.2935",
"0.2695",
"0.318",
"0.0825",
"0.318"
],
[
"0.7134",
"0.3403",
"0.8145",
"0.3403",
"0.8145",
"0.351",
"0.7134",
"0.351"
],
[
"0.0866",
"0.3467",
"0.478",
"0.3467",
"0.478",
"0.361",
"0.0866",
"0.361"
],
[
"0.08496",
"0.3599",
"0.4763",
"0.3591",
"0.4763",
"0.3738",
"0.08496",
"0.375"
],
[
"0.0866",
"0.375",
"0.4795",
"0.375",
"0.4795",
"0.3882",
"0.0866",
"0.3882"
],
[
"0.603",
"0.377",
"0.6113",
"0.377",
"0.6113",
"0.3801",
"0.603",
"0.3801"
],
[
"0.08496",
"0.3882",
"0.4763",
"0.3882",
"0.4763",
"0.4028",
"0.08496",
"0.4028"
],
[
"0.5312",
"0.3933",
"0.5483",
"0.3933",
"0.5483",
"0.3977",
"0.5312",
"0.3977"
],
[
"0.08496",
"0.4028",
"0.4795",
"0.4028",
"0.4795",
"0.4167",
"0.08496",
"0.4167"
],
[
"0.746",
"0.406",
"0.79",
"0.4077",
"0.79",
"0.4153",
"0.746",
"0.4136"
],
[
"0.08496",
"0.4167",
"0.2974",
"0.4167",
"0.2974",
"0.43",
"0.08496",
"0.43"
],
[
"0.6436",
"0.4192",
"0.6685",
"0.4192",
"0.6685",
"0.4224",
"0.6436",
"0.4224"
],
[
"0.2303",
"0.4526",
"0.246",
"0.4526",
"0.246",
"0.4634",
"0.2303",
"0.4634"
],
[
"0.2852",
"0.4526",
"0.4468",
"0.4526",
"0.4468",
"0.4666",
"0.2852",
"0.4666"
],
[
"0.4558",
"0.455",
"0.4756",
"0.455",
"0.4756",
"0.4634",
"0.4558",
"0.4634"
],
[
"0.2115",
"0.4622",
"0.2632",
"0.464",
"0.2632",
"0.4749",
"0.2115",
"0.473"
],
[
"0.10455",
"0.4539",
"0.1912",
"0.4539",
"0.1912",
"0.4653",
"0.10455",
"0.4653"
],
[
"0.1912",
"0.4702",
"0.2876",
"0.4702",
"0.2876",
"0.481",
"0.1912",
"0.481"
],
[
"0.7305",
"0.4749",
"0.794",
"0.4749",
"0.794",
"0.4817",
"0.7305",
"0.4817"
],
[
"0.10297",
"0.4912",
"0.478",
"0.4912",
"0.478",
"0.505",
"0.10297",
"0.505"
],
[
"0.5215",
"0.4988",
"0.914",
"0.4988",
"0.914",
"0.512",
"0.5215",
"0.512"
],
[
"0.08496",
"0.505",
"0.478",
"0.505",
"0.478",
"0.5195",
"0.08496",
"0.5195"
],
[
"0.5195",
"0.513",
"0.9126",
"0.513",
"0.9126",
"0.5264",
"0.5195",
"0.5264"
],
[
"0.0866",
"0.5195",
"0.4795",
"0.5195",
"0.4795",
"0.5327",
"0.0866",
"0.5327"
],
[
"0.5195",
"0.5264",
"0.911",
"0.5264",
"0.911",
"0.5405",
"0.5195",
"0.5405"
],
[
"0.0866",
"0.5327",
"0.478",
"0.5327",
"0.478",
"0.547",
"0.0866",
"0.547"
],
[
"0.08496",
"0.547",
"0.3857",
"0.547",
"0.3857",
"0.56",
"0.08496",
"0.56"
],
[
"0.5195",
"0.5415",
"0.914",
"0.5415",
"0.914",
"0.5547",
"0.5195",
"0.5547"
],
[
"0.5186",
"0.5547",
"0.9067",
"0.5547",
"0.9067",
"0.5684",
"0.5186",
"0.5684"
],
[
"0.2296",
"0.5835",
"0.2443",
"0.5835",
"0.2443",
"0.5933",
"0.2296",
"0.5933"
],
[
"0.282",
"0.5835",
"0.4453",
"0.5835",
"0.4453",
"0.5977",
"0.282",
"0.5977"
],
[
"0.4436",
"0.584",
"0.4763",
"0.584",
"0.4763",
"0.595",
"0.4436",
"0.595"
],
[
"0.2075",
"0.5947",
"0.2615",
"0.5967",
"0.2615",
"0.604",
"0.2075",
"0.6016"
],
[
"0.10455",
"0.5854",
"0.1896",
"0.5854",
"0.1896",
"0.5967",
"0.10455",
"0.5967"
],
[
"0.1602",
"0.5986",
"0.1683",
"0.5986",
"0.1683",
"0.604",
"0.1602",
"0.604"
],
[
"0.1847",
"0.601",
"0.2834",
"0.601",
"0.2834",
"0.6113",
"0.1847",
"0.6113"
],
[
"0.5195",
"0.595",
"0.8486",
"0.5967",
"0.8486",
"0.6113",
"0.5195",
"0.61"
],
[
"0.08496",
"0.6206",
"0.478",
"0.622",
"0.478",
"0.636",
"0.08496",
"0.6353"
],
[
"0.5195",
"0.6157",
"0.91",
"0.616",
"0.91",
"0.63",
"0.5195",
"0.629"
],
[
"0.08417",
"0.6353",
"0.2632",
"0.6353",
"0.2632",
"0.6484",
"0.08417",
"0.6484"
],
[
"0.5195",
"0.63",
"0.914",
"0.629",
"0.914",
"0.643",
"0.5195",
"0.6445"
],
[
"0.5186",
"0.6426",
"0.911",
"0.643",
"0.911",
"0.6577",
"0.5186",
"0.657"
],
[
"0.10297",
"0.651",
"0.4795",
"0.651",
"0.4795",
"0.664",
"0.10297",
"0.664"
],
[
"0.5195",
"0.6577",
"0.9126",
"0.6577",
"0.9126",
"0.672",
"0.5195",
"0.672"
],
[
"0.08496",
"0.664",
"0.4795",
"0.664",
"0.4795",
"0.678",
"0.08496",
"0.678"
],
[
"0.08417",
"0.677",
"0.478",
"0.678",
"0.478",
"0.6914",
"0.08417",
"0.69"
],
[
"0.5195",
"0.674",
"0.778",
"0.674",
"0.778",
"0.6846",
"0.5195",
"0.6846"
],
[
"0.08496",
"0.6934",
"0.1095",
"0.6934",
"0.1095",
"0.7056",
"0.08496",
"0.7056"
],
[
"0.5195",
"0.69",
"0.9126",
"0.69",
"0.9126",
"0.704",
"0.5195",
"0.704"
],
[
"0.5215",
"0.704",
"0.914",
"0.704",
"0.914",
"0.7183",
"0.5215",
"0.7183"
],
[
"0.2018",
"0.7197",
"0.3594",
"0.72",
"0.3594",
"0.735",
"0.2018",
"0.7344"
],
[
"0.5195",
"0.7183",
"0.911",
"0.7183",
"0.911",
"0.732",
"0.5195",
"0.732"
],
[
"0.5215",
"0.7344",
"0.911",
"0.7344",
"0.911",
"0.744",
"0.5215",
"0.744"
],
[
"0.08496",
"0.7407",
"0.478",
"0.741",
"0.478",
"0.756",
"0.08496",
"0.755"
],
[
"0.5195",
"0.7456",
"0.9126",
"0.7456",
"0.9126",
"0.759",
"0.5195",
"0.759"
],
[
"0.08496",
"0.756",
"0.478",
"0.756",
"0.478",
"0.7695",
"0.08496",
"0.7695"
],
[
"0.5186",
"0.763",
"0.91",
"0.7646",
"0.91",
"0.779",
"0.5186",
"0.778"
],
[
"0.08496",
"0.7695",
"0.478",
"0.7695",
"0.478",
"0.7827",
"0.08496",
"0.7827"
],
[
"0.08496",
"0.7827",
"0.4795",
"0.7827",
"0.4795",
"0.797",
"0.08496",
"0.797"
],
[
"0.5195",
"0.779",
"0.911",
"0.779",
"0.911",
"0.7925",
"0.5195",
"0.7925"
],
[
"0.08496",
"0.797",
"0.4763",
"0.797",
"0.4763",
"0.81",
"0.08496",
"0.81"
],
[
"0.5195",
"0.7925",
"0.9126",
"0.7925",
"0.9126",
"0.806",
"0.5195",
"0.806"
],
[
"0.08496",
"0.8096",
"0.478",
"0.81",
"0.478",
"0.8237",
"0.08496",
"0.8228"
],
[
"0.5195",
"0.806",
"0.9126",
"0.806",
"0.9126",
"0.8193",
"0.5195",
"0.8193"
],
[
"0.08417",
"0.8247",
"0.3743",
"0.8237",
"0.3743",
"0.837",
"0.08417",
"0.8384"
],
[
"0.5195",
"0.821",
"0.914",
"0.821",
"0.914",
"0.834",
"0.5195",
"0.834"
],
[
"0.5195",
"0.834",
"0.914",
"0.834",
"0.914",
"0.8477",
"0.5195",
"0.8477"
],
[
"0.10706",
"0.8486",
"0.4795",
"0.8496",
"0.4795",
"0.8643",
"0.10706",
"0.863"
],
[
"0.5215",
"0.8477",
"0.9126",
"0.8477",
"0.9126",
"0.8623",
"0.5215",
"0.8623"
],
[
"0.08496",
"0.863",
"0.4805",
"0.863",
"0.4805",
"0.877",
"0.08496",
"0.877"
],
[
"0.5186",
"0.8623",
"0.9126",
"0.8623",
"0.9126",
"0.8755",
"0.5186",
"0.8755"
],
[
"0.08417",
"0.875",
"0.29",
"0.8755",
"0.29",
"0.89",
"0.08417",
"0.8896"
],
[
"0.5186",
"0.8755",
"0.914",
"0.8755",
"0.914",
"0.8896",
"0.5186",
"0.8896"
]
] | [
"Algorithm 1:Concept ConfidenceDeviation",
"InputReference",
"Textualinversior",
"Hinversion",
"DreamBooth",
"Custom Diffusion",
"Concepts:V*",
"(LDM)",
"(SD)",
"Input: Concept fine-tuned models G E (ge), c E CconcEPTBeD;",
"Oracles F, {FpAC, Frmagenet, FcuBs,FvQA];",
"",
"Reference set of concept images Xref e (re]:",
"Target set of prompts Y e (ye]:",
"",
"Output: Estimated CCD",
"Anhotoofar",
"loginthestyleofV",
"I: Initialize: score = ; preal - ",
"2:forcE CcONCEPTBED do",
"3:",
"preal",
"4:",
"ift = VQA then",
"5:",
"c=3",
"V* flying inthesky",
"6:",
"for r = 1... M do",
"7:",
"preal F(ai,c)",
"8:",
"preal",
"1M",
"M",
"1pre",
"9:",
"forn=",
"1Ndo",
"10:",
"Egen = ge(ye)",
"Two Vsitting ona tree branch.",
"-1 * (F(rgen,c) -preal)",
"(// Eq. 3]",
"score←",
"11:",
"12: CCD= NN score.",
" is sitting in thegrass.",
"Suppose we are provided with m images (Xi:m) of the tar-",
"get concept c. Now, in order to learn the text tokens * cor-",
"responding to the concept c from the set of images Xi:m, the",
"",
"Textual Inversion methodology aims to optimize v* by re-",
"",
"constructing Xi:m using the objective function of the LDM",
"drivingV",
"with frozen parameters and :",
"",
"E",
"le - E(zt, t, c, Ce(y))l2",
"(1)",
"cEXi:m,t",
"V*= argmin",
"~N(0,1),Z~E()",
"CatisintheV*",
"In the case of DreamBooth and Custom Diffusion, instead",
"Figure 3: Qualitative examples showcasing the effectiveness",
"of finding the optimal v*, it optimizes the model parameter",
"of concept learners on the CONCEPTBED dataset. The left-",
"β associated with the noise estimator (e). This optimization",
"most column displays four instances of ground truth tar-",
"process enables the model to learn the mapping between ran-",
"domly initialized V* and the target concept c.",
"get concept images (V*). Subsequent columns exhibit target",
"concept-specific images generated by all baseline methods.",
"E",
"Ile - Ee(zt, t, a, Cs(y)l2",
"(2)",
"zEXi:m,",
"@* = argmin",
"c",
"~N(0,1),z~E()",
"3.1 CoNCEPTBED: Dataset Construction",
"Once @* is obtained, it can be used to generate images re-",
"CoNCEPTBED incorporates existing datasets such as Ima-",
"lated to the target concept.",
"geNet (Deng et al. 2009), PACS (Li et al. 2017), CUB (Wah",
"et al. 2011), and Visual Genome (Krishna et al. 2017), en-",
"Once the images are generated, in order to evaluate these",
"abling the creation of a labeled dataset. Figure 2 provides an",
"generated images, it is essential to verify whether they align",
"with thelearned concepts whilemaintaining compositional-",
"overvieWoftheCoONCEPTBEDdataset.",
"ity.",
"Learning Styles. We use styles from the PACS dataset: Art",
"Painting, Cartoon, Photo, and Sketch, Each style",
"3CONCEPTBED",
"contains images corresponding to seven categories. The con-",
"ceptlearneraimstouseexamplesfromonestyleasarefer-",
"In this section, we introduce CONCEPTBED, a compre-",
"ence and generate style-specific images for all seven entities.",
"hensive collection of concepts, designed to accurately es-",
"Learning Objects. Extracting object-level concepts is ac-",
"timate concept and composition alignment by quantifying",
"deviations in the generated images. Later, we introduce the",
"complished through the utilization of the ImageNet dataset.",
"novel evaluation framework associated with CoNCEPTBED",
"It comprises 1000 low-level concepts from the Word-",
"Please refer tothe Appendixfor additional insights on the",
"Net (Fellbaum 2010) hierarchy. However, due to the pres-",
"proposed dataset and evaluation framework.",
"ence of noise in ImageNet images and the lack of relevance",
"to daily life for many concepts, we employ an automated",
"'DreamBooth and Custom-Diffusion use additional regularizer",
"filtering pipeline to ensure the usefulness and quality of the",
"to improve compositionally by using same objective function on a",
"reference concept images. The pipeline involves extracting",
"diverse set of image-caption pairs.",
"a list of low-level concepts and their parent concepts from"
] | {
"arxiv_id": "2306.04695",
"language": "en",
"page_idx": 2
} |
||
2306.04695-en-4 | 2306.04695-en-4.jpg | <table>
<tbody>
<tr>
<th rowspan="2">$\mathbf{Model}$</th>
<th colspan="2">$\mathbf{Concepts}$</th>
<th colspan="2">$\mathbf{Fine-grained}_{\mathrm{CUB}}$</th>
<th rowspan="2">Composition</th>
</tr>
<tr>
<th>$\mathbf{Domain}_{\mathrm{PACS}}$</th>
<th>$\mathbf{Objects}_{\mathrm{ImageNet}}$</th>
<th>$\mathbf{Object-level}$</th>
<th>$\mathbf{Attribute-level}$</th>
</tr>
<tr>
<td>$\operatorname{TI}\left(\mathrm{LDM}\right)$</td>
<td>0.0478</td>
<td>0.0955</td>
<td>0.2289</td>
<td>0.1174</td>
<td>0.1906</td>
</tr>
<tr>
<td>$\operatorname{TI}(\operatorname{SD})$</td>
<td>0.2456</td>
<td>0.0472</td>
<td>0.0859</td>
<td>0.0332</td>
<td>0.1090</td>
</tr>
<tr>
<td>$DB$</td>
<td>0.6825</td>
<td>0.0678</td>
<td>0.0963</td>
<td>0.0469</td>
<td>0.3527</td>
</tr>
<tr>
<td>CD</td>
<td>0.6206</td>
<td>0.2085</td>
<td>0.3934</td>
<td>0.1743</td>
<td>0.4916</td>
</tr>
<tr>
<td>Original</td>
<td>0.0000</td>
<td>0.0000</td>
<td>0.0000</td>
<td>0.000</td>
<td>0.0000</td>
</tr>
</tbody>
</table>
Table 1: Results of Concept Alignment Evaluation. The table shows the performance of concept learners evaluated using $\text{the CCD () metric for Concepts (Domain}_{\mathrm{PACS}},\mathrm{~and~Object}_{\mathrm{ImageNel}}),\hat{\text{Fine-grained}}_{\mathrm{CUB~(Object-level,~and~Attribute-level),~and}}$ Comositian Thbot bort and wofraris dorforis andlos indis indis bold and adunda dordin dordin d mordis d Composition. The best and worst performing models are indicated by bold and underlined numbers, respectively.
<table>
<tbody>
<tr>
<th>$\mathbf{Models}$</th>
<th>$CLIP$</th>
<th>$\mathbf{Relation}$ $VQA.$</th>
<th>$CCD$</th>
<th>$CLIP$</th>
<th>$\mathbf{Action}$ $VQA$</th>
<th>$CCD$</th>
<th>$CLIP$</th>
<th>$\mathbf{Attribute}$ $VQA.$</th>
<th>$CCD$</th>
<th>$CLIP$</th>
<th>Counting $VQA.$</th>
<th>$CCD$</th>
</tr>
<tr>
<td>TI $(\mathbf{L}$DM)</td>
<td>0.6589</td>
<td>$66.60\%$</td>
<td>0.2074</td>
<td>0.6523</td>
<td>$68.69\%$</td>
<td>0.2098</td>
<td>0.6599</td>
<td>$72.22\%$</td>
<td>0.1331</td>
<td>0.6515</td>
<td>$65.78\%$</td>
<td>0.1231</td>
</tr>
<tr>
<td>TI (SD)</td>
<td>0.6294</td>
<td>$70.09\%$</td>
<td>0.1735</td>
<td>0.6274</td>
<td>$70.81\%$</td>
<td>0.1884</td>
<td>0.6360</td>
<td>$74.75\%$</td>
<td>0.1091</td>
<td>0.6301</td>
<td>$68.38\%$</td>
<td>0.1020</td>
</tr>
<tr>
<td>$DB$</td>
<td>0.7051</td>
<td>$82.20\%$</td>
<td>0.0542</td>
<td>0.6995</td>
<td>$84.61\%$</td>
<td>0.0496</td>
<td>0.6862</td>
<td>$82.24\%$</td>
<td>0.0355</td>
<td>0.6924</td>
<td>$78.90\%$</td>
<td>-0.0016</td>
</tr>
<tr>
<td>CD</td>
<td>$\overline{0.7065}$</td>
<td>$82.94\%$</td>
<td>0.0471</td>
<td>$\underline{0.7053}$</td>
<td>$86.35\%$</td>
<td>0.0347</td>
<td>0.6940</td>
<td>$84.20\%$</td>
<td>0.0163</td>
<td>0.6921</td>
<td>$\underline{79.36\%}$</td>
<td>-0.0054</td>
</tr>
<tr>
<td>SD</td>
<td>0.7222</td>
<td>$83.42\%$</td>
<td>0.0403</td>
<td>0.7178</td>
<td>$87.39\%$</td>
<td>0.0256</td>
<td>0.7053</td>
<td>$83.85\%$</td>
<td>0.0184</td>
<td>0.7085</td>
<td>$81.07\%$</td>
<td>-0.0206</td>
</tr>
<tr>
<td>Original</td>
<td>0.6626</td>
<td>$87.45\%$</td>
<td>0.0000</td>
<td>0.6831</td>
<td>$89.78\%$</td>
<td>0.0000</td>
<td>0.6306</td>
<td>$85.79\%$</td>
<td>0.0000</td>
<td>0.6553</td>
<td>$78.32\%$</td>
<td>0.0000</td>
</tr>
</tbody>
</table>
Table 2: Compositional Reasoning Evaluation Results. The table shows the performance of the prior works for Composition Alignment. $\bar{\mathrm{CLIP}}\left(\uparrow\right)$ is the traditional image-text alignment metric. VQA ($\bar{\uparrow})$ is the accuracy of the ViLT VQA classifier on generated boolean questions. And CCD (J) is the composition deviations reported from the ViLT model with respect to its performance on original images. The best-performing model is indicated by bold numbers, while the performance that is higher than the original data is reported with underline.
# 3.4 Task Specific Evaluation Settings
To efficiently leverage the CONCEPTBED evaluation pipeline, we trained separate oracles on the corresponding CONCEPTBED datasets. Two different types of evaluations are conducted, each with its respective set of oracles: 1) concept alignment, measured by concept classifiers, and 2) compositional reasoning, measured by a VQA model.
ous works, CLIP was not used due to its inability to capture compositions (Thrush et al. 2022). Instead, taking after (Cho, Zala, and Bansal 2022), we propose to use a pretrained ViLT (Kim, Son, and Kim 2021) as a VQA model for composition evaluations. Specifically, from each composite prompt, the boolean questions with positive answers are generated (Banerjee et al. 2021). As ViLT is essentially a classifier, the CCD can be calculated with respect to the confidence of the model associated with a“yes”answer.
# 4 Experiments & Results
In this section, we benchmark four state-of-the-art concept learning methodologies. We frst explain the experimental setup and report the evaluation results using the CONCEPTBED framework along with human preferences. Additional details about the experimental setup, results, and human evaluations are in the appendix.
Concept Alignment: Concept alignment evaluation was performed on all tasks, including the generated concept images with different composite text prompts. To evaluate the style, a ResNet18 (He et al. 2015) model is trained to distinguish the images between four style concepts. To evaluate the objects, a ConvNeXt (Liu et al. 2022) model is fine-tuned on 80 classes from the CONCEPTBED using the ImageNet training subset. The Concept Embedding Model (CEM) (Zarlenga et al. 2022) was trained on CUB to detect the concepts and attributes. Images corresponding to the concepts were generated for each task by following the prompts:“A photo of V”” for objects and“A photo of a $<entity-name>$ in the style of V$^*,,$,for styles. Here,$<entity-$ name> belongs to the seven classes from PACS. The remaining task, composition, utilizes the same pre-trained ConvNeXt model for concept alignment, as CONCEPTBED compositions are specifically for 80 ImageNet concepts.
## 4.1 Experimental Setup
In our experiments, we study four text-conditioned diffusion modeling-based concept learning strategies: Textual Inversion (TI) on LDM and SD, DreamBooth (DB) (Ruiz et al. 2022), and Custom Diffusion (CD) (Kumari et al. 2022). We generate $N=100$ images for all concepts to measure the concept alignment and $N=3$ images for 33K composite text prompts. For a total of 284 concepts, we train all four baselines. This leads to 1100+ concept-specific fine-tuned models and we generate a total of 500,000 images for evaluations. To show the stability of CCD, we report the mean performance across the three seeds of oracle training.
Compositional Reasoning: To measure the image-text alignment with respect to the input prompts, the conceptspecific token (V$^*)$ was removed and replaced with the corresponding ground truth label (i.e., dogs, cats, etc.). The image-text similarity was then measured. Unlike previ- | [
[
"0.335",
"0.0682",
"0.4004",
"0.0707",
"0.3987",
"0.0859",
"0.3342",
"0.084"
],
[
"0.1929",
"0.0745",
"0.2385",
"0.0745",
"0.2385",
"0.089",
"0.1929",
"0.089"
],
[
"0.5376",
"0.06946",
"0.6445",
"0.0707",
"0.6445",
"0.0852",
"0.5376",
"0.084"
],
[
"0.7236",
"0.077",
"0.8057",
"0.077",
"0.8057",
"0.0884",
"0.7236",
"0.0884"
],
[
"0.2712",
"0.0852",
"0.353",
"0.0871",
"0.353",
"0.1016",
"0.2712",
"0.09973"
],
[
"0.367",
"0.0859",
"0.4617",
"0.0884",
"0.4617",
"0.1029",
"0.367",
"0.101"
],
[
"0.4968",
"0.0871",
"0.576",
"0.0871",
"0.576",
"0.1016",
"0.4968",
"0.1016"
],
[
"0.5923",
"0.0871",
"0.6855",
"0.0871",
"0.6855",
"0.0985",
"0.5923",
"0.0985"
],
[
"0.1929",
"0.1016",
"0.2573",
"0.1042",
"0.2566",
"0.1193",
"0.1912",
"0.11743"
],
[
"0.29",
"0.1042",
"0.3367",
"0.1042",
"0.3367",
"0.1193",
"0.29",
"0.1193"
],
[
"0.3906",
"0.1042",
"0.437",
"0.1042",
"0.437",
"0.1187",
"0.3906",
"0.1187"
],
[
"0.511",
"0.1042",
"0.558",
"0.1016",
"0.5596",
"0.11743",
"0.513",
"0.1193"
],
[
"0.6167",
"0.1042",
"0.6626",
"0.1042",
"0.6626",
"0.1187",
"0.6167",
"0.1187"
],
[
"0.739",
"0.1042",
"0.786",
"0.1042",
"0.786",
"0.1187",
"0.739",
"0.1187"
],
[
"0.1929",
"0.113",
"0.2443",
"0.11554",
"0.2427",
"0.1307",
"0.1912",
"0.1288"
],
[
"0.2893",
"0.11554",
"0.3367",
"0.11554",
"0.3367",
"0.13",
"0.2893",
"0.13"
],
[
"0.3896",
"0.11554",
"0.437",
"0.11554",
"0.437",
"0.13",
"0.3896",
"0.13"
],
[
"0.511",
"0.11554",
"0.5596",
"0.11554",
"0.5596",
"0.13",
"0.511",
"0.13"
],
[
"0.615",
"0.11554",
"0.6626",
"0.11554",
"0.6626",
"0.13",
"0.615",
"0.13"
],
[
"0.739",
"0.11554",
"0.786",
"0.11554",
"0.786",
"0.13",
"0.739",
"0.13"
],
[
"0.2893",
"0.1256",
"0.335",
"0.1276",
"0.3342",
"0.1427",
"0.2876",
"0.1401"
],
[
"0.511",
"0.1256",
"0.5596",
"0.1276",
"0.558",
"0.1427",
"0.5107",
"0.1401"
],
[
"0.738",
"0.1256",
"0.786",
"0.1276",
"0.785",
"0.1427",
"0.7363",
"0.1401"
],
[
"0.615",
"0.127",
"0.6626",
"0.127",
"0.6626",
"0.1414",
"0.615",
"0.1414"
],
[
"0.1936",
"0.1276",
"0.2197",
"0.1276",
"0.2197",
"0.1401",
"0.1936",
"0.1401"
],
[
"0.3896",
"0.127",
"0.4387",
"0.127",
"0.4387",
"0.1414",
"0.3896",
"0.1414"
],
[
"0.1912",
"0.1383",
"0.2197",
"0.1383",
"0.2197",
"0.1547",
"0.1912",
"0.1547"
],
[
"0.2893",
"0.1383",
"0.335",
"0.1383",
"0.335",
"0.154",
"0.2893",
"0.154"
],
[
"0.511",
"0.1383",
"0.5596",
"0.1383",
"0.5596",
"0.154",
"0.511",
"0.154"
],
[
"0.615",
"0.1383",
"0.6626",
"0.1383",
"0.6626",
"0.154",
"0.615",
"0.154"
],
[
"0.739",
"0.137",
"0.7876",
"0.1395",
"0.786",
"0.1547",
"0.738",
"0.1528"
],
[
"0.1929",
"0.1559",
"0.2484",
"0.1559",
"0.2484",
"0.1704",
"0.1929",
"0.1704"
],
[
"0.3896",
"0.1395",
"0.437",
"0.1395",
"0.437",
"0.154",
"0.3896",
"0.154"
],
[
"0.2893",
"0.1559",
"0.3367",
"0.1559",
"0.3367",
"0.1704",
"0.2893",
"0.1704"
],
[
"0.3896",
"0.1559",
"0.437",
"0.1559",
"0.437",
"0.1704",
"0.3896",
"0.1704"
],
[
"0.511",
"0.1559",
"0.5596",
"0.1559",
"0.5596",
"0.1704",
"0.511",
"0.1704"
],
[
"0.6177",
"0.1559",
"0.66",
"0.1559",
"0.66",
"0.1704",
"0.6177",
"0.1704"
],
[
"0.739",
"0.1559",
"0.7876",
"0.1559",
"0.7876",
"0.1704",
"0.739",
"0.1704"
],
[
"0.08417",
"0.1843",
"0.9126",
"0.1849",
"0.9126",
"0.1995",
"0.08417",
"0.1989"
],
[
"0.08496",
"0.1989",
"0.915",
"0.1989",
"0.915",
"0.2166",
"0.08496",
"0.2166"
],
[
"0.0866",
"0.2134",
"0.824",
"0.2134",
"0.824",
"0.2267",
"0.0866",
"0.2267"
],
[
"0.4182",
"0.2424",
"0.4656",
"0.2443",
"0.464",
"0.26",
"0.4167",
"0.2583"
],
[
"0.7793",
"0.2437",
"0.846",
"0.2456",
"0.846",
"0.26",
"0.7793",
"0.2583"
],
[
"0.597",
"0.2443",
"0.66",
"0.2443",
"0.66",
"0.2588",
"0.597",
"0.2588"
],
[
"0.0972",
"0.2488",
"0.1503",
"0.2507",
"0.1487",
"0.2651",
"0.09644",
"0.2632"
],
[
"0.2255",
"0.2443",
"0.2834",
"0.2443",
"0.2834",
"0.2588",
"0.2255",
"0.2588"
],
[
"0.1732",
"0.2551",
"0.2141",
"0.2551",
"0.2141",
"0.2708",
"0.1732",
"0.2708"
],
[
"0.232",
"0.2551",
"0.2737",
"0.2568",
"0.273",
"0.2727",
"0.2303",
"0.2708"
],
[
"0.361",
"0.2551",
"0.4004",
"0.2551",
"0.4004",
"0.2708",
"0.361",
"0.2708"
],
[
"0.4207",
"0.2551",
"0.46",
"0.2568",
"0.4592",
"0.2727",
"0.4192",
"0.2708"
],
[
"0.5474",
"0.2551",
"0.5884",
"0.2551",
"0.5884",
"0.2708",
"0.5474",
"0.2708"
],
[
"0.7344",
"0.2551",
"0.774",
"0.2551",
"0.774",
"0.2708",
"0.7344",
"0.2708"
],
[
"0.4805",
"0.2556",
"0.5146",
"0.2556",
"0.5146",
"0.2715",
"0.4805",
"0.2715"
],
[
"0.2942",
"0.2556",
"0.327",
"0.2556",
"0.327",
"0.2715",
"0.2942",
"0.2715"
],
[
"0.6685",
"0.2556",
"0.701",
"0.2556",
"0.701",
"0.2715",
"0.6685",
"0.2715"
],
[
"0.859",
"0.2556",
"0.8906",
"0.2556",
"0.8906",
"0.2715",
"0.859",
"0.2715"
],
[
"0.7935",
"0.2568",
"0.8384",
"0.2568",
"0.8384",
"0.2715",
"0.7935",
"0.2715"
],
[
"0.6045",
"0.2568",
"0.648",
"0.2568",
"0.648",
"0.2715",
"0.6045",
"0.2715"
],
[
"0.09644",
"0.2715",
"0.1642",
"0.2734",
"0.1642",
"0.2886",
"0.09644",
"0.286"
],
[
"0.1724",
"0.2747",
"0.2181",
"0.2747",
"0.2181",
"0.286",
"0.1724",
"0.286"
],
[
"0.2296",
"0.2747",
"0.2793",
"0.2747",
"0.2793",
"0.286",
"0.2296",
"0.286"
],
[
"0.2893",
"0.2747",
"0.335",
"0.2747",
"0.335",
"0.286",
"0.2893",
"0.286"
],
[
"0.3586",
"0.2747",
"0.4043",
"0.2747",
"0.4043",
"0.286",
"0.3586",
"0.286"
],
[
"0.415",
"0.2747",
"0.464",
"0.2747",
"0.464",
"0.286",
"0.415",
"0.286"
],
[
"0.4756",
"0.2734",
"0.524",
"0.2734",
"0.524",
"0.2886",
"0.4756",
"0.2886"
],
[
"0.546",
"0.2747",
"0.5923",
"0.2747",
"0.5923",
"0.286",
"0.546",
"0.286"
],
[
"0.603",
"0.2747",
"0.652",
"0.2747",
"0.652",
"0.286",
"0.603",
"0.286"
],
[
"0.661",
"0.2747",
"0.7075",
"0.2747",
"0.7075",
"0.286",
"0.661",
"0.286"
],
[
"0.734",
"0.2747",
"0.778",
"0.2747",
"0.778",
"0.286",
"0.734",
"0.286"
],
[
"0.789",
"0.2747",
"0.839",
"0.2747",
"0.839",
"0.286",
"0.789",
"0.286"
],
[
"0.853",
"0.2747",
"0.8965",
"0.2747",
"0.8965",
"0.286",
"0.853",
"0.286"
],
[
"0.09644",
"0.2827",
"0.1503",
"0.2854",
"0.1487",
"0.2998",
"0.0948",
"0.2974"
],
[
"0.2279",
"0.2827",
"0.2793",
"0.2854",
"0.2778",
"0.2998",
"0.2263",
"0.2974"
],
[
"0.2893",
"0.2827",
"0.335",
"0.2854",
"0.3342",
"0.2998",
"0.2876",
"0.2974"
],
[
"0.601",
"0.2827",
"0.653",
"0.2854",
"0.652",
"0.2998",
"0.6006",
"0.2974"
],
[
"0.3586",
"0.286",
"0.4043",
"0.286",
"0.4043",
"0.2974",
"0.3586",
"0.2974"
],
[
"0.1724",
"0.286",
"0.2181",
"0.286",
"0.2181",
"0.2974",
"0.1724",
"0.2974"
],
[
"0.4167",
"0.286",
"0.4656",
"0.286",
"0.4656",
"0.2974",
"0.4167",
"0.2974"
],
[
"0.4756",
"0.286",
"0.523",
"0.286",
"0.523",
"0.2974",
"0.4756",
"0.2974"
],
[
"0.546",
"0.286",
"0.5923",
"0.286",
"0.5923",
"0.2974",
"0.546",
"0.2974"
],
[
"0.6626",
"0.286",
"0.709",
"0.286",
"0.709",
"0.2974",
"0.6626",
"0.2974"
],
[
"0.734",
"0.286",
"0.778",
"0.286",
"0.778",
"0.2974",
"0.734",
"0.2974"
],
[
"0.789",
"0.286",
"0.839",
"0.286",
"0.839",
"0.2974",
"0.789",
"0.2974"
],
[
"0.853",
"0.286",
"0.899",
"0.286",
"0.899",
"0.2974",
"0.853",
"0.2974"
],
[
"0.2279",
"0.2935",
"0.2793",
"0.2954",
"0.2778",
"0.31",
"0.2263",
"0.308"
],
[
"0.4143",
"0.2942",
"0.4656",
"0.2966",
"0.464",
"0.3113",
"0.4126",
"0.3093"
],
[
"0.4739",
"0.2942",
"0.524",
"0.2966",
"0.523",
"0.3113",
"0.4722",
"0.3093"
],
[
"0.601",
"0.2935",
"0.653",
"0.2954",
"0.652",
"0.31",
"0.6006",
"0.308"
],
[
"0.661",
"0.2942",
"0.71",
"0.2966",
"0.709",
"0.3113",
"0.66",
"0.3093"
],
[
"0.2893",
"0.2954",
"0.3367",
"0.2954",
"0.3367",
"0.31",
"0.2893",
"0.31"
],
[
"0.1724",
"0.2954",
"0.2181",
"0.2954",
"0.2181",
"0.31",
"0.1724",
"0.31"
],
[
"0.3586",
"0.2954",
"0.406",
"0.2954",
"0.406",
"0.31",
"0.3586",
"0.31"
],
[
"0.544",
"0.2954",
"0.5933",
"0.2954",
"0.5933",
"0.31",
"0.544",
"0.31"
],
[
"0.732",
"0.2954",
"0.7793",
"0.2954",
"0.7793",
"0.31",
"0.732",
"0.31"
],
[
"0.8506",
"0.2954",
"0.902",
"0.2954",
"0.902",
"0.31",
"0.8506",
"0.31"
],
[
"0.79",
"0.2966",
"0.839",
"0.2966",
"0.839",
"0.308",
"0.79",
"0.308"
],
[
"0.0989",
"0.2966",
"0.125",
"0.2966",
"0.125",
"0.31",
"0.0989",
"0.31"
],
[
"0.1724",
"0.3062",
"0.2181",
"0.308",
"0.2173",
"0.3225",
"0.1708",
"0.3208"
],
[
"0.2893",
"0.3062",
"0.335",
"0.308",
"0.3342",
"0.324",
"0.2876",
"0.3213"
],
[
"0.4756",
"0.3062",
"0.523",
"0.308",
"0.5215",
"0.3225",
"0.4739",
"0.3208"
],
[
"0.601",
"0.3062",
"0.652",
"0.308",
"0.6504",
"0.3225",
"0.6006",
"0.3208"
],
[
"0.789",
"0.3062",
"0.841",
"0.308",
"0.839",
"0.3225",
"0.7876",
"0.3208"
],
[
"0.2279",
"0.307",
"0.2793",
"0.307",
"0.2793",
"0.3213",
"0.2279",
"0.3213"
],
[
"0.09644",
"0.307",
"0.1259",
"0.307",
"0.1259",
"0.324",
"0.09644",
"0.324"
],
[
"0.3586",
"0.308",
"0.406",
"0.308",
"0.406",
"0.3225",
"0.3586",
"0.3225"
],
[
"0.4143",
"0.308",
"0.4656",
"0.308",
"0.4656",
"0.3225",
"0.4143",
"0.3225"
],
[
"0.546",
"0.308",
"0.5933",
"0.308",
"0.5933",
"0.3225",
"0.546",
"0.3225"
],
[
"0.734",
"0.307",
"0.778",
"0.307",
"0.778",
"0.3225",
"0.734",
"0.3225"
],
[
"0.8506",
"0.308",
"0.902",
"0.308",
"0.902",
"0.3225",
"0.8506",
"0.3225"
],
[
"0.0989",
"0.3245",
"0.12335",
"0.3245",
"0.12335",
"0.3384",
"0.0989",
"0.3384"
],
[
"0.664",
"0.3093",
"0.709",
"0.3093",
"0.709",
"0.3208",
"0.664",
"0.3208"
],
[
"0.29",
"0.3245",
"0.3367",
"0.3245",
"0.3367",
"0.339",
"0.29",
"0.339"
],
[
"0.3594",
"0.3245",
"0.406",
"0.3245",
"0.406",
"0.339",
"0.3594",
"0.339"
],
[
"0.415",
"0.3245",
"0.4656",
"0.3245",
"0.4656",
"0.339",
"0.415",
"0.339"
],
[
"0.4756",
"0.3245",
"0.523",
"0.3245",
"0.523",
"0.339",
"0.4756",
"0.339"
],
[
"0.601",
"0.3245",
"0.652",
"0.3245",
"0.652",
"0.339",
"0.601",
"0.339"
],
[
"0.6626",
"0.3245",
"0.71",
"0.3245",
"0.71",
"0.339",
"0.6626",
"0.339"
],
[
"0.79",
"0.3245",
"0.839",
"0.3245",
"0.839",
"0.339",
"0.79",
"0.339"
],
[
"0.2296",
"0.3257",
"0.2778",
"0.3257",
"0.2778",
"0.3372",
"0.2296",
"0.3372"
],
[
"0.1732",
"0.3257",
"0.2173",
"0.3257",
"0.2173",
"0.3372",
"0.1732",
"0.3372"
],
[
"0.5474",
"0.3257",
"0.591",
"0.3257",
"0.591",
"0.3372",
"0.5474",
"0.3372"
],
[
"0.7344",
"0.3257",
"0.778",
"0.3257",
"0.778",
"0.3372",
"0.7344",
"0.3372"
],
[
"0.851",
"0.3257",
"0.899",
"0.3257",
"0.899",
"0.3372",
"0.851",
"0.3372"
],
[
"0.29",
"0.3352",
"0.335",
"0.3372",
"0.3342",
"0.3516",
"0.2893",
"0.3499"
],
[
"0.603",
"0.3352",
"0.652",
"0.3372",
"0.6504",
"0.3516",
"0.601",
"0.3499"
],
[
"0.1724",
"0.336",
"0.2181",
"0.336",
"0.2181",
"0.351",
"0.1724",
"0.351"
],
[
"0.0989",
"0.3384",
"0.152",
"0.3384",
"0.152",
"0.3499",
"0.0989",
"0.3499"
],
[
"0.2296",
"0.3384",
"0.2778",
"0.3384",
"0.2778",
"0.3499",
"0.2296",
"0.3499"
],
[
"0.3594",
"0.3384",
"0.4028",
"0.3384",
"0.4028",
"0.3499",
"0.3594",
"0.3499"
],
[
"0.4167",
"0.3384",
"0.464",
"0.3384",
"0.464",
"0.3499",
"0.4167",
"0.3499"
],
[
"0.4763",
"0.3384",
"0.523",
"0.3384",
"0.523",
"0.3499",
"0.4763",
"0.3499"
],
[
"0.5474",
"0.3384",
"0.5923",
"0.3384",
"0.5923",
"0.3499",
"0.5474",
"0.3499"
],
[
"0.6626",
"0.3384",
"0.709",
"0.3384",
"0.709",
"0.3499",
"0.6626",
"0.3499"
],
[
"0.7344",
"0.3384",
"0.778",
"0.3384",
"0.778",
"0.3499",
"0.7344",
"0.3499"
],
[
"0.79",
"0.3384",
"0.839",
"0.3384",
"0.839",
"0.3499",
"0.79",
"0.3499"
],
[
"0.853",
"0.3384",
"0.899",
"0.3384",
"0.899",
"0.3499",
"0.853",
"0.3499"
],
[
"0.0866",
"0.3655",
"0.914",
"0.3655",
"0.914",
"0.379",
"0.0866",
"0.379"
],
[
"0.0866",
"0.379",
"0.9126",
"0.379",
"0.9126",
"0.3933",
"0.0866",
"0.3933"
],
[
"0.0866",
"0.3933",
"0.9126",
"0.3933",
"0.9126",
"0.4072",
"0.0866",
"0.4072"
],
[
"0.08496",
"0.4072",
"0.9126",
"0.4072",
"0.9126",
"0.4204",
"0.08496",
"0.4204"
],
[
"0.08496",
"0.4216",
"0.4028",
"0.4216",
"0.4028",
"0.435",
"0.08496",
"0.435"
],
[
"0.08496",
"0.4602",
"0.38",
"0.4622",
"0.38",
"0.4766",
"0.08496",
"0.4749"
],
[
"0.5186",
"0.461",
"0.9126",
"0.4622",
"0.9126",
"0.4766",
"0.5186",
"0.476"
],
[
"0.0866",
"0.4817",
"0.4795",
"0.4817",
"0.4795",
"0.4956",
"0.0866",
"0.4956"
],
[
"0.5195",
"0.4766",
"0.911",
"0.4766",
"0.911",
"0.4905",
"0.5195",
"0.4905"
],
[
"0.5195",
"0.4905",
"0.9126",
"0.4905",
"0.9126",
"0.504",
"0.5195",
"0.504"
],
[
"0.0866",
"0.4968",
"0.4795",
"0.4968",
"0.4795",
"0.5103",
"0.0866",
"0.5103"
],
[
"0.5195",
"0.504",
"0.9126",
"0.504",
"0.9126",
"0.5176",
"0.5195",
"0.5176"
],
[
"0.08496",
"0.5103",
"0.478",
"0.5103",
"0.478",
"0.5234",
"0.08496",
"0.5234"
],
[
"0.5195",
"0.518",
"0.91",
"0.518",
"0.91",
"0.532",
"0.5195",
"0.532"
],
[
"0.0866",
"0.526",
"0.4763",
"0.526",
"0.4763",
"0.536",
"0.0866",
"0.536"
],
[
"0.5195",
"0.532",
"0.914",
"0.532",
"0.914",
"0.5454",
"0.5195",
"0.5454"
],
[
"0.08496",
"0.5386",
"0.478",
"0.5386",
"0.478",
"0.552",
"0.08496",
"0.552"
],
[
"0.5195",
"0.5454",
"0.911",
"0.5454",
"0.911",
"0.5596",
"0.5195",
"0.5596"
],
[
"0.0866",
"0.552",
"0.4102",
"0.552",
"0.4102",
"0.565",
"0.0866",
"0.565"
],
[
"0.5195",
"0.5596",
"0.9126",
"0.5596",
"0.9126",
"0.5728",
"0.5195",
"0.5728"
],
[
"0.08496",
"0.5737",
"0.4795",
"0.5747",
"0.4795",
"0.589",
"0.08496",
"0.5884"
],
[
"0.5195",
"0.5737",
"0.888",
"0.5737",
"0.888",
"0.587",
"0.5195",
"0.587"
],
[
"0.0866",
"0.589",
"0.478",
"0.589",
"0.478",
"0.603",
"0.0866",
"0.603"
],
[
"0.08496",
"0.603",
"0.4795",
"0.603",
"0.4795",
"0.616",
"0.08496",
"0.616"
],
[
"0.603",
"0.6016",
"0.8286",
"0.6016",
"0.8286",
"0.616",
"0.603",
"0.616"
],
[
"0.08496",
"0.616",
"0.478",
"0.616",
"0.478",
"0.63",
"0.08496",
"0.63"
],
[
"0.5186",
"0.6206",
"0.9126",
"0.622",
"0.9126",
"0.636",
"0.5186",
"0.6353"
],
[
"0.08496",
"0.631",
"0.478",
"0.631",
"0.478",
"0.6445",
"0.08496",
"0.6445"
],
[
"0.5195",
"0.636",
"0.91",
"0.636",
"0.91",
"0.6494",
"0.5195",
"0.6494"
],
[
"0.08417",
"0.643",
"0.478",
"0.6426",
"0.478",
"0.657",
"0.08417",
"0.6577"
],
[
"0.5195",
"0.6494",
"0.9126",
"0.6494",
"0.9126",
"0.6636",
"0.5195",
"0.6636"
],
[
"0.08496",
"0.6577",
"0.4795",
"0.6577",
"0.4795",
"0.672",
"0.08496",
"0.672"
],
[
"0.5195",
"0.6636",
"0.9126",
"0.6636",
"0.9126",
"0.678",
"0.5195",
"0.678"
],
[
"0.08496",
"0.6724",
"0.4795",
"0.6724",
"0.4795",
"0.686",
"0.08496",
"0.686"
],
[
"0.5195",
"0.678",
"0.911",
"0.678",
"0.911",
"0.6914",
"0.5195",
"0.6914"
],
[
"0.08496",
"0.686",
"0.4756",
"0.686",
"0.4756",
"0.6997",
"0.08496",
"0.6997"
],
[
"0.5195",
"0.69",
"0.763",
"0.69",
"0.763",
"0.704",
"0.5195",
"0.704"
],
[
"0.08496",
"0.7007",
"0.4795",
"0.7007",
"0.4795",
"0.714",
"0.08496",
"0.714"
],
[
"0.08417",
"0.712",
"0.4795",
"0.7134",
"0.4795",
"0.728",
"0.08417",
"0.7266"
],
[
"0.523",
"0.7183",
"0.556",
"0.7183",
"0.556",
"0.73",
"0.523",
"0.73"
],
[
"0.556",
"0.7173",
"0.714",
"0.7173",
"0.714",
"0.732",
"0.556",
"0.732"
],
[
"0.08496",
"0.7266",
"0.4795",
"0.7266",
"0.4795",
"0.741",
"0.08496",
"0.741"
],
[
"0.0866",
"0.741",
"0.478",
"0.741",
"0.478",
"0.755",
"0.0866",
"0.755"
],
[
"0.5215",
"0.7373",
"0.9126",
"0.7373",
"0.9126",
"0.7505",
"0.5215",
"0.7505"
],
[
"0.08496",
"0.755",
"0.478",
"0.755",
"0.478",
"0.768",
"0.08496",
"0.768"
],
[
"0.5195",
"0.7505",
"0.9126",
"0.7505",
"0.9126",
"0.7646",
"0.5195",
"0.7646"
],
[
"0.08496",
"0.7695",
"0.4795",
"0.7695",
"0.4795",
"0.7827",
"0.08496",
"0.7827"
],
[
"0.5195",
"0.765",
"0.9126",
"0.765",
"0.9126",
"0.779",
"0.5195",
"0.779"
],
[
"0.0866",
"0.782",
"0.4763",
"0.782",
"0.4763",
"0.7954",
"0.0866",
"0.7954"
],
[
"0.5195",
"0.778",
"0.914",
"0.778",
"0.914",
"0.7915",
"0.5195",
"0.7915"
],
[
"0.08417",
"0.797",
"0.4534",
"0.797",
"0.4534",
"0.81",
"0.08417",
"0.81"
],
[
"0.5195",
"0.7925",
"0.9126",
"0.7925",
"0.9126",
"0.806",
"0.5195",
"0.806"
],
[
"0.5195",
"0.806",
"0.9126",
"0.806",
"0.9126",
"0.8193",
"0.5195",
"0.8193"
],
[
"0.08496",
"0.8184",
"0.4795",
"0.8193",
"0.4795",
"0.834",
"0.08496",
"0.8335"
],
[
"0.5215",
"0.821",
"0.914",
"0.821",
"0.914",
"0.834",
"0.5215",
"0.834"
],
[
"0.08496",
"0.834",
"0.478",
"0.834",
"0.478",
"0.8477",
"0.08496",
"0.8477"
],
[
"0.5195",
"0.834",
"0.914",
"0.834",
"0.914",
"0.8477",
"0.5195",
"0.8477"
],
[
"0.08496",
"0.8477",
"0.4795",
"0.8477",
"0.4795",
"0.8613",
"0.08496",
"0.8613"
],
[
"0.5195",
"0.8477",
"0.9126",
"0.8477",
"0.9126",
"0.8613",
"0.5195",
"0.8613"
],
[
"0.08496",
"0.8623",
"0.4795",
"0.8623",
"0.4795",
"0.8755",
"0.08496",
"0.8755"
],
[
"0.5195",
"0.8623",
"0.914",
"0.8623",
"0.914",
"0.8755",
"0.5195",
"0.8755"
],
[
"0.08496",
"0.8755",
"0.478",
"0.8755",
"0.478",
"0.8896",
"0.08496",
"0.8896"
],
[
"0.517",
"0.875",
"0.8677",
"0.8755",
"0.8677",
"0.89",
"0.517",
"0.8896"
]
] | [
"Concepts",
"Model",
"Fine-grainedcUB",
"Composition",
"DomainpAcs",
"ObjectSImageNet",
"Object-level",
"Attribute-level",
"TI (LDM)",
"0.0478",
"0.0955",
"0.2289",
"0.1174",
"0.1906",
"TI (SD)",
"0.2456",
"0.0472",
"0.0859",
"0.0332",
"0.1090",
"0.6825",
"0.0963",
"0.3527",
"0.0469",
"DB",
"0.0678",
"CD",
"0.6206",
"0.3934",
"0.1743",
"0.4916",
"Original",
"0.2085",
"0.0000",
"0.0000",
"0.0000",
"0.000",
"0.0000",
"Table 1: Results of Concept Alignment Evaluation. The table shows the performance of concept learners evaluated using",
"the CCD (l) metric for Concepts (DomainpAcs, and ObjectimageNet), Fine-grainedcuB (Object-level, and Attribute-level), and",
"Composition. The best and worst performing models are indicated by bold and underlined numbers, respectively.",
"Action",
"Counting",
"Attribute",
"Models",
"Relation",
"CLIP",
"VQA.",
"CLIP",
"VQA",
"CLIP",
"CLIP",
"CCD",
"CCD",
"CCD",
"CCD",
"VQA.",
"VQA.",
"TI (LDM)",
"0.6589",
"66.60%",
"0.2074",
"0.6523",
"68.69%",
"0.2098",
"0.6599",
"72.22%",
"0.1331",
"0.6515",
"65.78%",
"0.1231",
"TI (SD)",
"70.09%",
"0.1735",
"74.75%",
"0.6274",
"0.6294",
"70.81%",
"0.1884",
"0.6360",
"0.1091",
"0.6301",
"68.38%",
"0.1020",
"82.20%",
"84.61%",
"0.0496",
"82.24%",
"0.0355",
"0.0542",
"0.7051",
"0.6995",
"0.6862",
"0.6924",
"-0.0016",
"78.90%",
"DB",
"0.7065",
"0.0471",
"0.0347",
"84.20%",
"79.36%",
"82.94%",
"CD",
"0.7053",
"86.35%",
"0.6940",
"0.6921",
"-0.0054",
"SD",
"0.0163",
"0.0403",
"0.7178",
"87.39%",
"0.0256",
"83.85%",
"0.0184",
"81.07%",
"83.42%",
"0.7222",
"0.7053",
"0.7085",
"-0.0206",
"0.0000",
"85.79%",
"0.6626",
"Original",
"87.45%",
"0.6831",
"89.78%",
"0.0000",
"0.6306",
"0.0000",
"0.6553",
"78.32%",
"0.0000",
"Table 2: Compositional Reasoning Evaluation Results. The table shows the performance of the prior works for Composition",
"Alignment. CLIP () is the traditional image-text alignment metric. VQA () is the accuracy of the ViLT VQA classifier on",
"generated boolean questions. And ccD () is the composition deviations reported from the ViLT model with respect to its",
"performance on original images. The best-performing model is indicated by bold numbers, while the performance that is higher",
"than the original data is reported with underline.",
"3.4 Task Specific Evaluation Settings",
"ous works, CLIP was not used due to its inability to cap-",
"To efficiently leverage the CONCEPTBED evaluation",
"ture compositions (Thrush et al. 2022). Instead, taking af-",
"ter (Cho, Zala, and Bansal 2022), we propose to use a pre-",
"pipeline, we trained separate oracles on the corresponding",
"trained ViLT (Kim, Son, and Kim 2021) as a VQA model",
"CoNCEPTBED datasets. Two different types of evaluations",
"for composition evaluations. Specifically, from each com-",
"areconducted,each withitsrespective setoforacles:1)con-",
"posite prompt, the boolean questions with positive answers",
"cept alignment, measured by concept classifiers, and 2) com-",
"are generated (Banerjee et al. 2021). As ViLT is essentially",
"positional reasoning, measured by a VQA model.",
"a classifier, the cCD can be calculated with respect to the",
"Concept Alignment:Concept alignment evaluation was",
"confidence of the model associated with a \"yes\" answer.",
"performed on all tasks, including the generated concept im-",
"ages with different composite text prompts. To evaluate the",
"4Experiments&Results",
"style, a ResNet18 (He et al. 2015) model is trained to dis-",
"In this section, we benchmark four state-of-the-art concept",
"tinguish the images between four style concepts. To eval-",
"learning methodologies. We first explain the experimen-",
"uate the objects, a ConvNeXt (Liu et al. 2022) model is",
"tal setup and report the evaluation results using the CoN-",
"fine-tuned on 8O classes from the CoNCEPTBED using the",
"CEPTBED framework along with human preferences. Addi-",
"ImageNet training subset. The Concept Embedding Model",
"tional details about the experimental setup, results, and hu-",
"(CEM) (Zarlenga et al. 2022) was trained on CUB to de-",
"man evaluations are in the appendix.",
"tect the concepts and attributes. Images corresponding to",
"the concepts were generated for each task by foilowing the",
"4.1",
"Experimental Setup",
"prompts: \"A photo of *\"for objects and \"A photo of a",
"<entity-name> in the style of V*\" for styles. Here, <entity-",
"In our experiments, we study four text-conditioned diffusion",
"name> belongs to the seven classes from PACS. The re-",
"modeling-based concept learning strategies: Textual Inver-",
"maining task, composition, utilizes the same pre-trained",
"sion (TI) on LDM and SD, DreamBooth (DB) (Ruiz et al.",
"ConvNeXt model for concept alignment, as CoNcEPTBED",
"2022), and Custom Diffusion (CD) (Kumari et al. 2022). We",
" compositions are specifically for 80 ImageNet concepts.",
"generate N = 100 images for all concepts to measure the",
"concept alignment and N = 3 images for 33K composite",
"Compositional Reasoning:To measure the image-text",
"text prompts. For a total of 284 concepts, we train all four",
"alignment with respect to the input prompts, the concept-",
"baselines. This leads to 1100+ concept-specific fine-tuned",
"specific token (V*) was removed and replaced with the",
"models and we generate a total of 500, 000 images for eval-",
"corresponding ground truth label (i.e., dogs, cat s, etc.).",
"uations. To show the stability of CCD, we report the mean",
"The image-text similarity was then measured. Unlike previ-",
"performance across the three seeds of oracle training"
] | {
"arxiv_id": "2306.04695",
"language": "en",
"page_idx": 4
} |
||
2209.09760-en-0 | 2209.09760-en-0.jpg | # Dynamic Graph Message Passing Networks for Visual Recognition
Li Zhang, Mohan Chen, Anurag Arnab, Xiangyang Xue, and Philip H.S. Torn
Abstract—Modelling long-range dependencies is critical for scene understanding tasks in computer vision. Although convolution neural networks (CNNs) have excelled in many vision tasks, they are still limited in capturing long-range structured relationships as they typically consist of layers of local kernels. A fully-connected graph, such as the self-attention operation in Transformers, is beneficial for such
modelling, however, its computational overhead is prohibitive. In this paper, we propose a dynamic graph message passing network, that significantly reduces the computational complexity compared to related works modelling a fully-connected graph. This is achieved by adaptively sampling nodes in the graph, conditioned on the input, for message passing. Based on the sampled nodes, we dynamically predict node-dependent filter weights and the affinity matrix for propagating information between them. This formulation allows us to design a self-attention module, and more importantiy a new Transformer-based backbone network, that we use for both image
classification pretraining, and for addressing various downstream tasks (e.$g$. object detection, instance and semantic segmentation).
Using this model, we show significant improvements with respect to strong, state-of-the-art baselines on four different tasks. Our
approach also outperforms fully-connected graphs while using substantially fewer floating-point operations and parameters. Code and
models will be made publicly available at https://github.com/fudan-zvg/DGMN2.
Index Terms—Dynamic Message Passing, Transformer
#
Although convolutional neural networks (CNNs) have excelled in graphs means that the long-range contextual information needed a wide range of scene understanding tasks [25],[33],[64],they are for complex vision tasks such as segmentation and detection[3], still limited by their ability to capture these long-range interactions. [54],[57] will only be partially captured. Along this direction, To improve the capability of CNNs in this regard, a recent, popular
model Non-local networks [73] proposes a generalisation of the based on graph sampling. However, the proposed sampling method
attention model of [69] and achieves significant advance in several
computer vision tasks.
relationships among all feature elements in a feature map to produce graph where the neighbours for each node were fixed and filter
Non-10cal networke sssentially model pariwise structured the modelling capacity wais restricted Telationships amongall fature elementsin alfeature mapto toroduce graph wherer then nejghbours foreach\\\text{The attention weigints which are used for feature agre
local networks effectively model a fully-connected feature graph novel dynamic graph message passing network (DGMN) model,
dense prediction tasks on high-resolution imagery, as commonly encountered in semantic segmentation [12]. Moreover, in dense prediction tasks, capturing relations between all pairs of pixels is on the node features. Intuitively, this learned sampling allows the
usually unnecessary due to the redundant information contained
within the image (Fig. 1). Simply subsampling the feature map to a subset of the most relevant nodes in the graph; (ii) Based on reduce the memory requirements is also suboptimal, as such naïve the nodes that have been sampled, we further dynamically predict subsampling would result in smaller objects in the image not being node-dependant, and thus position specific, filter weights and also represented adequately.
Graph convolution networks (GCNs)[21],[31]- which the feature nodes via message passing. The dynamic weights propagate information along graph-structured input data- can and affinities are especially beneficial to specifically model each
Fudan University.
$E{-mail:lizhangfd@fudan.edu.cn}$
Anurag Arnab is with Google Research. Mohan Chen and Xiangyang Xue are with School of Computer Science, Fudan University. Philip H.S. Torr is with the Department of Engineering Science, University of Oxford.
GraphSAGE[23] introduced an efficient graph learning model considered a uniform sampling strategy along the spatial dimension of the input, and was independent of the actual input. Consequently, the modelling capacity was restricted as it assumed a static input
To address the aforementioned shortcomings, we propose a targeting effective and efficient deep representational learning with joint modeling of two key dynamic properties as illustrated in Fig. 1. Our contribution is twofold: (i) We dynamically sample the neighbourhood of a node from the feature graph, conditioned network to efficiently gather long-range context by only selecting the affinity matrix, which are used to propagate information among sampled feature context, leading to more effective message passing. Both of these dynamic properties are jointly optimised in a single
$\bullet$ $Li$ Zhang $is$ the corresponding author with the School $of$ Data $Science$, model, and we modularise the DGMN as a network layer for simple
deployment into existing networks.
We demonstrate the proposed model on the tasks of image classification, semantic segmentation, object detection and instance segmentation on the challenging ImageNet [60], Cityscapes [12] | [
[
"0.12335",
"0.06946",
"0.88",
"0.06946",
"0.88",
"0.09784",
"0.12335",
"0.09784"
],
[
"0.3064",
"0.1048",
"0.6895",
"0.1061",
"0.6895",
"0.1307",
"0.3064",
"0.13"
],
[
"0.2009",
"0.1528",
"0.8",
"0.1528",
"0.8",
"0.1674",
"0.2009",
"0.1674"
],
[
"0.1111",
"0.1913",
"0.8896",
"0.1913",
"0.8896",
"0.2058",
"0.1111",
"0.2058"
],
[
"0.10864",
"0.2039",
"0.8896",
"0.2058",
"0.8896",
"0.2197",
"0.10864",
"0.2172"
],
[
"0.1095",
"0.2184",
"0.8677",
"0.2184",
"0.8677",
"0.2317",
"0.1095",
"0.2317"
],
[
"0.1111",
"0.2329",
"0.8896",
"0.2329",
"0.8896",
"0.2437",
"0.1111",
"0.2437"
],
[
"0.1111",
"0.2468",
"0.875",
"0.2468",
"0.875",
"0.2568",
"0.1111",
"0.2568"
],
[
"0.02124",
"0.26",
"0.0637",
"0.26",
"0.0637",
"0.4016",
"0.02124",
"0.4016"
],
[
"0.1095",
"0.2583",
"0.8823",
"0.2583",
"0.8823",
"0.2715",
"0.1095",
"0.2715"
],
[
"0.11273",
"0.2734",
"0.871",
"0.2734",
"0.871",
"0.281",
"0.11273",
"0.281"
],
[
"0.1095",
"0.2842",
"0.8384",
"0.2842",
"0.8384",
"0.2974",
"0.1095",
"0.2974"
],
[
"0.1111",
"0.2986",
"0.871",
"0.2986",
"0.871",
"0.3093",
"0.1111",
"0.3093"
],
[
"0.1111",
"0.3113",
"0.8506",
"0.3113",
"0.8506",
"0.3213",
"0.1111",
"0.3213"
],
[
"0.10864",
"0.324",
"0.884",
"0.3225",
"0.884",
"0.336",
"0.10864",
"0.3372"
],
[
"0.1111",
"0.3384",
"0.568",
"0.3384",
"0.568",
"0.3484",
"0.1111",
"0.3484"
],
[
"0.1095",
"0.3623",
"0.437",
"0.3623",
"0.437",
"0.377",
"0.1095",
"0.377"
],
[
"0.02124",
"0.4028",
"0.05637",
"0.4016",
"0.05963",
"0.483",
"0.0237",
"0.4841"
],
[
"0.10706",
"0.4216",
"0.2296",
"0.4216",
"0.2296",
"0.435",
"0.10706",
"0.435"
],
[
"0.10455",
"0.4475",
"0.5",
"0.4475",
"0.5",
"0.461",
"0.10455",
"0.461"
],
[
"0.5024",
"0.4475",
"0.9224",
"0.4475",
"0.9224",
"0.461",
"0.5024",
"0.461"
],
[
"0.1095",
"0.464",
"0.4968",
"0.464",
"0.4968",
"0.4749",
"0.1095",
"0.4749"
],
[
"0.5034",
"0.464",
"0.919",
"0.464",
"0.919",
"0.4749",
"0.5034",
"0.4749"
],
[
"0.0801",
"0.4792",
"0.4958",
"0.4792",
"0.4958",
"0.4893",
"0.0801",
"0.4893"
],
[
"0.5073",
"0.4792",
"0.921",
"0.4792",
"0.921",
"0.4893",
"0.5073",
"0.4893"
],
[
"0.0801",
"0.4937",
"0.4927",
"0.4937",
"0.4927",
"0.504",
"0.0801",
"0.504"
],
[
"0.505",
"0.4924",
"0.921",
"0.4924",
"0.921",
"0.506",
"0.505",
"0.506"
],
[
"0.02696",
"0.502",
"0.05554",
"0.502",
"0.05554",
"0.5884",
"0.02696",
"0.5884"
],
[
"0.0784",
"0.5083",
"0.4927",
"0.5083",
"0.4927",
"0.518",
"0.0784",
"0.518"
],
[
"0.5063",
"0.507",
"0.9233",
"0.507",
"0.9233",
"0.521",
"0.5063",
"0.521"
],
[
"0.03513",
"0.523",
"0.04004",
"0.523",
"0.04004",
"0.53",
"0.03513",
"0.53"
],
[
"0.0768",
"0.521",
"0.4983",
"0.521",
"0.4983",
"0.534",
"0.0768",
"0.534"
],
[
"0.5107",
"0.523",
"0.9224",
"0.523",
"0.9224",
"0.5327",
"0.5107",
"0.5327"
],
[
"0.0768",
"0.535",
"0.4927",
"0.535",
"0.4927",
"0.549",
"0.0768",
"0.549"
],
[
"0.5063",
"0.535",
"0.9224",
"0.535",
"0.9224",
"0.549",
"0.5063",
"0.549"
],
[
"0.03513",
"0.5474",
"0.04004",
"0.5474",
"0.04004",
"0.5547",
"0.03513",
"0.5547"
],
[
"0.0768",
"0.55",
"0.4944",
"0.55",
"0.4944",
"0.563",
"0.0768",
"0.563"
],
[
"0.5034",
"0.549",
"0.921",
"0.549",
"0.921",
"0.5625",
"0.5034",
"0.5625"
],
[
"0.0768",
"0.5645",
"0.4944",
"0.5645",
"0.4944",
"0.5776",
"0.0768",
"0.5776"
],
[
"0.505",
"0.5645",
"0.9224",
"0.5645",
"0.9224",
"0.5776",
"0.505",
"0.5776"
],
[
"0.0768",
"0.579",
"0.2214",
"0.579",
"0.2214",
"0.5923",
"0.0768",
"0.5923"
],
[
"0.5073",
"0.579",
"0.9233",
"0.579",
"0.9233",
"0.5923",
"0.5073",
"0.5923"
],
[
"0.0237",
"0.586",
"0.058",
"0.586",
"0.058",
"0.6494",
"0.0237",
"0.6494"
],
[
"0.10046",
"0.5967",
"0.4958",
"0.5967",
"0.4958",
"0.607",
"0.10046",
"0.607"
],
[
"0.5073",
"0.595",
"0.921",
"0.595",
"0.921",
"0.606",
"0.5073",
"0.606"
],
[
"0.076",
"0.61",
"0.4944",
"0.61",
"0.4944",
"0.6235",
"0.076",
"0.6235"
],
[
"0.5063",
"0.6094",
"0.9224",
"0.6094",
"0.9224",
"0.6226",
"0.5063",
"0.6226"
],
[
"0.0768",
"0.625",
"0.4958",
"0.625",
"0.4958",
"0.638",
"0.0768",
"0.638"
],
[
"0.5034",
"0.6235",
"0.7485",
"0.6235",
"0.7485",
"0.637",
"0.5034",
"0.637"
],
[
"0.0768",
"0.6396",
"0.4927",
"0.6396",
"0.4927",
"0.653",
"0.0768",
"0.653"
],
[
"0.5293",
"0.636",
"0.921",
"0.638",
"0.921",
"0.6514",
"0.5293",
"0.6494"
],
[
"0.03513",
"0.6426",
"0.05148",
"0.6426",
"0.05148",
"0.6665",
"0.03513",
"0.6665"
],
[
"0.076",
"0.653",
"0.4927",
"0.654",
"0.4927",
"0.6675",
"0.076",
"0.6665"
],
[
"0.5063",
"0.6514",
"0.9233",
"0.6514",
"0.9233",
"0.6655",
"0.5063",
"0.6655"
],
[
"0.03513",
"0.6636",
"0.05148",
"0.6636",
"0.05148",
"0.6797",
"0.03513",
"0.6797"
],
[
"0.0768",
"0.6685",
"0.4927",
"0.6685",
"0.4927",
"0.6816",
"0.0768",
"0.6816"
],
[
"0.5063",
"0.6655",
"0.9224",
"0.6655",
"0.9224",
"0.6797",
"0.5063",
"0.6797"
],
[
"0.02696",
"0.6753",
"0.05148",
"0.6724",
"0.058",
"0.7056",
"0.0335",
"0.709"
],
[
"0.0784",
"0.6846",
"0.4927",
"0.6846",
"0.4927",
"0.6943",
"0.0784",
"0.6943"
],
[
"0.505",
"0.681",
"0.921",
"0.681",
"0.921",
"0.6943",
"0.505",
"0.6943"
],
[
"0.076",
"0.696",
"0.4944",
"0.6963",
"0.4944",
"0.711",
"0.076",
"0.7104"
],
[
"0.5063",
"0.696",
"0.9224",
"0.696",
"0.9224",
"0.709",
"0.5063",
"0.709"
],
[
"0.0784",
"0.712",
"0.492",
"0.712",
"0.492",
"0.7227",
"0.0784",
"0.7227"
],
[
"0.5063",
"0.7104",
"0.9233",
"0.7104",
"0.9233",
"0.7236",
"0.5063",
"0.7236"
],
[
"0.0768",
"0.726",
"0.4944",
"0.726",
"0.4944",
"0.7393",
"0.0768",
"0.7393"
],
[
"0.505",
"0.7246",
"0.9224",
"0.7246",
"0.9224",
"0.738",
"0.505",
"0.738"
],
[
"0.0784",
"0.742",
"0.4927",
"0.742",
"0.4927",
"0.7524",
"0.0784",
"0.7524"
],
[
"0.5034",
"0.7393",
"0.9224",
"0.7393",
"0.9224",
"0.7524",
"0.5034",
"0.7524"
],
[
"0.0768",
"0.756",
"0.4927",
"0.756",
"0.4927",
"0.7695",
"0.0768",
"0.7695"
],
[
"0.5063",
"0.754",
"0.9224",
"0.754",
"0.9224",
"0.7676",
"0.5063",
"0.7676"
],
[
"0.0768",
"0.7715",
"0.4927",
"0.7715",
"0.4927",
"0.782",
"0.0768",
"0.782"
],
[
"0.505",
"0.768",
"0.9224",
"0.768",
"0.9224",
"0.782",
"0.505",
"0.782"
],
[
"0.0768",
"0.785",
"0.4944",
"0.785",
"0.4944",
"0.799",
"0.0768",
"0.799"
],
[
"0.5063",
"0.784",
"0.9224",
"0.784",
"0.9224",
"0.7974",
"0.5063",
"0.7974"
],
[
"0.076",
"0.8",
"0.2263",
"0.8",
"0.2263",
"0.813",
"0.076",
"0.813"
],
[
"0.5063",
"0.799",
"0.919",
"0.799",
"0.919",
"0.8125",
"0.5063",
"0.8125"
],
[
"0.0972",
"0.8145",
"0.5",
"0.813",
"0.5",
"0.8276",
"0.0972",
"0.829"
],
[
"0.5063",
"0.813",
"0.9224",
"0.813",
"0.9224",
"0.827",
"0.5063",
"0.827"
],
[
"0.076",
"0.83",
"0.4983",
"0.83",
"0.4983",
"0.8433",
"0.076",
"0.8433"
],
[
"0.5063",
"0.827",
"0.9224",
"0.827",
"0.9224",
"0.8403",
"0.5063",
"0.8403"
],
[
"0.5073",
"0.8403",
"0.9233",
"0.8413",
"0.9233",
"0.856",
"0.5073",
"0.855"
],
[
"0.5073",
"0.856",
"0.9233",
"0.856",
"0.9233",
"0.869",
"0.5073",
"0.869"
],
[
"0.09644",
"0.8687",
"0.4983",
"0.8687",
"0.4983",
"0.879",
"0.09644",
"0.879"
],
[
"0.505",
"0.8706",
"0.9224",
"0.8706",
"0.9224",
"0.884",
"0.505",
"0.884"
],
[
"0.0989",
"0.88",
"0.1896",
"0.88",
"0.1896",
"0.89",
"0.0989",
"0.89"
],
[
"0.0989",
"0.89",
"0.2712",
"0.8896",
"0.2712",
"0.901",
"0.0989",
"0.9014"
],
[
"0.5073",
"0.8853",
"0.728",
"0.8853",
"0.728",
"0.8984",
"0.5073",
"0.8984"
],
[
"0.0972",
"0.901",
"0.492",
"0.901",
"0.492",
"0.914",
"0.0972",
"0.914"
],
[
"0.5312",
"0.8994",
"0.9224",
"0.8994",
"0.9224",
"0.9136",
"0.5312",
"0.9136"
],
[
"0.0972",
"0.911",
"0.5",
"0.912",
"0.5",
"0.9253",
"0.0972",
"0.925"
],
[
"0.5024",
"0.914",
"0.921",
"0.914",
"0.921",
"0.928",
"0.5024",
"0.928"
],
[
"0.09644",
"0.9224",
"0.455",
"0.924",
"0.455",
"0.9375",
"0.09644",
"0.936"
],
[
"0.509",
"0.931",
"0.921",
"0.931",
"0.921",
"0.9414",
"0.509",
"0.9414"
]
] | [
"Dynamic Graph Message Passing Networks",
"for Visual Recognition",
"LiZhang, Mohan Chen, Anurag Arnab, Xiangyang Xue,and PhilipH.S.Torr",
"AbstractModelling long-range dependencies is critical for scene understanding tasks in computer vision. Although convolution neural",
"networks (CNNs)have excelled inmanyvision tasks,they are still limited in capturing long-range structured relationshipsas theytypically",
"consist of layers of local kernels.A fully-connected graph, such as the self-attention operation in Transformers, is beneficial for such",
"modelling,however,itscomputationaloverheadisprohibitive.Inthispaper,weproposeadynamicgraphmessagepassingnetwork,that",
"significantlyreducesthecomputationalcomplexitycomparedtorelatedworksmodellingafully-connectedgraph.Thisisachievedby",
" de ",
"adaptively sampling nodes in the graph, conditioned on the input, for message passing. Based on the sampled nodes, we dynamically",
"oredictnode-dependentfilterweiahtsandtheaffinitymatrixforpropagatinginformationbetweenthemThisformulationallowsusto",
"design a self-attention module, and more importantly a new Transformer-based backbone network, that we use for both image",
"classificationpretraining,andforaddressingvariousdownstreamtasks(e.g.objectdetection,instanceandsemanticsegmentation)",
"Usingthismodel,weshowsignificantimprovementswithrespecttostrong,state-of-the-artbaselinesonfourdifferenttasks.Our",
"approach also outperforms fully-connected graphs while using substantially fewer floating-point operations and parameters.Code and",
"modelswillbemadepubliclyavailableathttps://github.com/fudan-zvg/DGMN2.",
"Index TermsDynamic Message Passing, Transformer",
"[']",
"INTRODUCTION",
"I APTURING long-range dependencies is crucial for complex",
" alleviate the computational issues of non-local networks to a",
"scene understanding tasks such as image classification, se-",
"certain extent.However,this stands only if local neighbourhoods",
"manticsegmentation,instancesegmentationandobjectdetection.",
"are considered foreachnode.Employing such local-connected",
"Althoughconvolutionalneuralnetworks(CNNs)haveexcelledin",
"graphs means that the long-range contextual information needed",
"A09L60",
"awiderangeofsceneunderstandingtasks[25].[33],[64l,theyare",
"for complex vision tasks such as segmentation and detection [3],",
"",
"still limited by their ability to capture these long-range interactions.",
"[54l,[57]willonlybepartiallycaptured.Alongthisdirection.",
"To improve the capability of CNNs in this regard, a recent, popular",
"GraphSAGE [23] introduced an efficient graph learning model",
"",
"model Non-local networks [73] proposes a generalisation of the",
" based on graph sampling. However, the proposed sampling method",
"attention model of [69] and achieves significant advance in several",
"considered a uniform sampling strategy along the spatial dimension",
"computer vision tasks.",
"of the input, and was independent of the actual input. Consequently,",
"6027:",
"Non-local networksessentially model pairwise structured",
"the modelling capacity was restricted as it assumed a static input",
"relationships among all feature elements in a feature map to produce",
"graph where the neighbours for each node were fixed and filter",
"the attention weights which are used for feature aggregation.",
"weights were shared among all nodes.",
"Considering each feature element as a node in a graph, Non-",
"To address the aforementioned shortcomings, we propose a",
".AT",
"local networks effectivelymodel a fully-connected featuregraph",
"novel dynamic graph message passing network (DGMN) model,",
"7",
"and thus have a quadratic inference complexity with respect",
"targeting effective and efficient deep representational learning with",
"",
"tothenumber ofthefeatureelements.Thisisinfeasiblefor",
"joint modeling of two key dynamic properties as illustrated in",
"dense prediction tasks on high-resolution imagery, as commonly",
"Fig. I. Our contribution is twofold: (i) We dynamically sample",
"encountered in semantic segmentation [12].Moreover.in dense",
"the neighbourhood of a node from the feature graph, conditioned",
"prediction tasks, capturing relations between all pairs of pixels is",
"on the node features. Intuitively, this learned sampling allows the",
"usuallyunnecessarydueto theredundantinformationcontained",
"network to efficiently gather long-range context by only selecting",
"within the image (Fig. 1). Simply subsampling the feature map to",
"a subset of the most relevant nodes in the graph; (ii) Based on",
"reducethe memory requirements is also suboptimal,as such naive",
"the nodes that have been sampled, we further dynamically predict",
"subsampling would result in smaller objects in the image not being",
"node-dependant, and thus position specific, filter weights and also",
"represented adequately.",
"the affinity matrix, which are used to propagate information among",
"Graph convolution networks (GCNs)[21],[31]-which",
"the feature nodes via message passing. The dynamic weights",
"propagate information along graph-structured input data - can",
"and affinities are especially beneficial to specifically model each",
"sampled feature context, leading to more effective message passing.",
"Both of these dynamic properties are jointly optimised in a single",
"Li Zhang isthe corresponding author with the School of Data Science,",
"model, and we modularise the DGMN as a network layer for simple",
"Fudan Universit",
"E-mail:lizhangfd@fudan.edu.cn",
"deployment into existing networks",
"Anurag Arnab is with Google Research. Mohan Chen and Xiangyang Xue",
"We demonstrate the proposed model on the tasks of image",
"are with School of Computer Science,Fudan University.Philip H.S.Torr is",
"classification, semantic segmentation, object detection and instance",
"with the Department of Engineering Science, University of Oxford.",
"segmentationonthechallengingImageNet[60],Cityscapes[12]"
] | {
"arxiv_id": "2209.09760",
"language": "en",
"page_idx": 0
} |
||
2209.09760-en-1 | 2209.09760-en-1.jpg | 2
<FigureHere>
(a) Fully-connected message passing
(c) Dynamic graph message passing
(b) Locally-connected message passing
Fig. 1. Contextual information is crucial for complex scene understanding tasks. To recognise the "boathouse", one neeeds to consider the“boat" and the“water” next to it. Fully-connected message passing models (a) are able to obtain this information, but are prohibitively expensive. Furthermore, they capture a lot of redundant information $(i.e.“trees”and“sky”).$ Locally-connected models (b) are more efficient, but miss out on important context. Our proposed approach (c), dynamically samples a small subset of relevant feature nodes based on a learned dyniamic sampling scheme, $i.e.$ the learned position-specific random walk (indicated by the white dashed arrow lines), and also dynamically predicts filter weights and affinities (indicated by unique edge and square colors.), which are both conditioned on the sampled feature nodes.
and COCO [44] datasets. We achieve significant performance improvements over the fully-connected model [69],[73], while using substantially fewer floating point operations (FLOPs). Specifically,“plugging” our module into existing networks, we show considerable improvements with respect to strong, state-of-theart baselines on three different vision downstream tasks and of CNNs was to use dilated convolutions[6],[81]. With dilated backbone architectures. Furthermore, we design a new Transformer convolutions, the number of parameters does not change, while backbone network based on the building block of our formulation. It achieves leading performance on Imagenet classification task and the pretrained backbone network also supports varrious vision the convolution operation include deformable convolution [15],
downstream tasks.
A preliminary version of this work was published in[85]. We from which to select input values. However, the weights of the have extended our conference version as follows: (i) We introduce deformable convolution filters do not depend on the selected input, a modified process of dynamic message passing calculation and
present a simplified version of our prior formulation; We further dynamic sampling aims to sample over the whole feature graph to recast our formulation to take a Transformer perspective and obtain a large receptive field, and the predicted affinities and the implement the message updating process into an efficient and weights for message passing are position specific and conditioned effective Transformer layer. (ii) Based on the building block of our
forrmulation, we propose a new backbone network DGMN2 with capture position-based semantic context to enable more effective
multi-scale feature representation which is very compatibility in message passing among feature nodes.
supporting various downstream vision tasks. The experiment on
upstream ImageNet[60] classification task show that alli the variants $_{[30],[37],[53],[61],[74]\text{on the irregular structured and geometricen }}$
of DGMN2 $(i.e.$, DGMN2-Tiny, DGMN2-Small, DGMN2-Medium and DGMN2-Large respectively) can surpass the conventional convolution-based backbones such as ResNet$t$ResNeXt[25],[77], irregular point clouds data. DeepGCN[37] proposes to enlarge and Transformer-based backbone such as DeiT[68] and PVT[72]. the receptive field on graph networks with dilated convolution. $( \textbf{ii\" {n}) With the pretrained backbone network DGMN2, we show DGM[ 30] introduces a differentiable graph module that learnsinger therentiat ares theres theres theres theres theres thes thes thes thes thes thes thes thes thes thes thes thes thes thes thes}$ considerable improvements with respect to strong, state-of-the-art baselines on three different vision downstream tasks such as object detection, instance segmentation and semantic segmentation.
are thus not directly applicable in the feature learning task since feature variables are typically continuous. Coupled with the fact that CRFs are computationally expensive, CRFs are no longer used for most computer vision tasks.
A complementary technique for increasing the receptive field the receptive feld grows exponentially if the dilation rate is linearly increased in successive layers. Other modifications to [90], which learns the offset with respect to a predefined grid and are in fact shared across all different positions. In contrast, our on the dynamically sampled nodes. Our model is thus able to better
Another line of related research targets the graph learning [18]. input. A neural network module EdgeConv is proposed in [74] to learn dynamic graphs with k-nearest neighbor (kNN) on the dynamic adjacency matrices.
The idea of sampling graph nodes has previously been explored in GraphSAGE [23]. Crucially, GraphSAGE simply uniformly samples nodes. In contrast, our sampling strategy is learned based on the node features. Specifically, we first sample the nodes
An early technique for modelling context for computer vision tasks uniformly in the spatial dimension, and then dynamically predict
walks of each node conditioned on the node features. Furthermore,
model[34] was popular as it modelled interactions between all pairs GraphSAGE does not consider our second important property,
$i.e.$,the dynamic prediction of the affinities and the message passing
kernels.
We also note that [29] developed an idea of “dynamic convo-
features. Moreover, they mostly model discrete label spaces, and lution", that is predicting a dynamic convolutional filter for each
# 2 RELATED WORK
involved conditional random fields. In particular, the DenseCRF of pixels in an image. Although such models have been integrated into neural networks [1],[2],[78],[88], they are limited by the fact that the pairwise potentials are based on simple handcrafted | [
[
"0.11926",
"0.2216",
"0.3489",
"0.2216",
"0.3489",
"0.2349",
"0.11926",
"0.2349"
],
[
"0.3816",
"0.2216",
"0.626",
"0.2216",
"0.626",
"0.2349",
"0.3816",
"0.2349"
],
[
"0.6567",
"0.2203",
"0.88",
"0.2203",
"0.88",
"0.2343",
"0.6567",
"0.2343"
],
[
"0.0768",
"0.2507",
"0.9224",
"0.2507",
"0.9224",
"0.2644",
"0.0768",
"0.2644"
],
[
"0.0768",
"0.2644",
"0.921",
"0.2644",
"0.921",
"0.2778",
"0.0768",
"0.2778"
],
[
"0.0768",
"0.2766",
"0.921",
"0.2766",
"0.921",
"0.2903",
"0.0768",
"0.2903"
],
[
"0.0784",
"0.289",
"0.9224",
"0.289",
"0.9224",
"0.303",
"0.0784",
"0.303"
],
[
"0.0768",
"0.3018",
"0.9233",
"0.3018",
"0.9233",
"0.315",
"0.0768",
"0.315"
],
[
"0.0768",
"0.3145",
"0.9233",
"0.3145",
"0.9233",
"0.3276",
"0.0768",
"0.3276"
],
[
"0.0784",
"0.3289",
"0.415",
"0.3289",
"0.415",
"0.339",
"0.0784",
"0.339"
],
[
"0.0768",
"0.3655",
"0.4958",
"0.3655",
"0.4958",
"0.379",
"0.0768",
"0.379"
],
[
"0.5",
"0.3662",
"0.921",
"0.3662",
"0.921",
"0.3801",
"0.5",
"0.3801"
],
[
"0.0784",
"0.3818",
"0.4983",
"0.3818",
"0.4983",
"0.3926",
"0.0784",
"0.3926"
],
[
"0.5034",
"0.3806",
"0.9224",
"0.3806",
"0.9224",
"0.3945",
"0.5034",
"0.3945"
],
[
"0.0768",
"0.3958",
"0.509",
"0.3958",
"0.509",
"0.4092",
"0.0768",
"0.4092"
],
[
"0.5063",
"0.3958",
"0.9233",
"0.3958",
"0.9233",
"0.4092",
"0.5063",
"0.4092"
],
[
"0.0768",
"0.4104",
"0.4944",
"0.4104",
"0.4944",
"0.4236",
"0.0768",
"0.4236"
],
[
"0.4944",
"0.4104",
"0.709",
"0.4104",
"0.709",
"0.4236",
"0.4944",
"0.4236"
],
[
"0.0768",
"0.4248",
"0.4927",
"0.4248",
"0.4927",
"0.4382",
"0.0768",
"0.4382"
],
[
"0.5317",
"0.4268",
"0.9233",
"0.4268",
"0.9233",
"0.44",
"0.5317",
"0.44"
],
[
"0.0768",
"0.4395",
"0.4944",
"0.4395",
"0.4944",
"0.4526",
"0.0768",
"0.4526"
],
[
"0.5034",
"0.4414",
"0.9224",
"0.4414",
"0.9224",
"0.455",
"0.5034",
"0.455"
],
[
"0.0784",
"0.455",
"0.4944",
"0.455",
"0.4944",
"0.4653",
"0.0784",
"0.4653"
],
[
"0.505",
"0.4558",
"0.9233",
"0.4558",
"0.9233",
"0.4697",
"0.505",
"0.4697"
],
[
"0.076",
"0.4673",
"0.4958",
"0.4685",
"0.4958",
"0.4817",
"0.076",
"0.481"
],
[
"0.5073",
"0.4717",
"0.9224",
"0.4717",
"0.9224",
"0.4817",
"0.5073",
"0.4817"
],
[
"0.0768",
"0.483",
"0.4944",
"0.483",
"0.4944",
"0.4968",
"0.0768",
"0.4968"
],
[
"0.5073",
"0.486",
"0.9224",
"0.486",
"0.9224",
"0.4968",
"0.5073",
"0.4968"
],
[
"0.0768",
"0.4976",
"0.4958",
"0.4976",
"0.4958",
"0.511",
"0.0768",
"0.511"
],
[
"0.509",
"0.5005",
"0.921",
"0.5005",
"0.921",
"0.511",
"0.509",
"0.511"
],
[
"0.0784",
"0.513",
"0.1936",
"0.513",
"0.1936",
"0.5234",
"0.0784",
"0.5234"
],
[
"0.509",
"0.5146",
"0.9233",
"0.5146",
"0.9233",
"0.528",
"0.509",
"0.528"
],
[
"0.0989",
"0.5264",
"0.4944",
"0.5264",
"0.4944",
"0.5405",
"0.0989",
"0.5405"
],
[
"0.5063",
"0.529",
"0.9233",
"0.529",
"0.9233",
"0.5425",
"0.5063",
"0.5425"
],
[
"0.0768",
"0.5425",
"0.492",
"0.5425",
"0.492",
"0.553",
"0.0768",
"0.553"
],
[
"0.5063",
"0.5435",
"0.9233",
"0.5435",
"0.9233",
"0.5566",
"0.5063",
"0.5566"
],
[
"0.0768",
"0.556",
"0.4927",
"0.556",
"0.4927",
"0.5693",
"0.0768",
"0.5693"
],
[
"0.5073",
"0.5596",
"0.921",
"0.5596",
"0.921",
"0.5693",
"0.5073",
"0.5693"
],
[
"0.076",
"0.571",
"0.4944",
"0.5693",
"0.4944",
"0.5835",
"0.076",
"0.584"
],
[
"0.5063",
"0.5728",
"0.9233",
"0.5728",
"0.9233",
"0.586",
"0.5063",
"0.586"
],
[
"0.0768",
"0.5854",
"0.4927",
"0.5854",
"0.4927",
"0.5986",
"0.0768",
"0.5986"
],
[
"0.5063",
"0.587",
"0.9224",
"0.587",
"0.9224",
"0.601",
"0.5063",
"0.601"
],
[
"0.0784",
"0.601",
"0.492",
"0.601",
"0.492",
"0.6113",
"0.0784",
"0.6113"
],
[
"0.505",
"0.6016",
"0.9233",
"0.6016",
"0.9233",
"0.6157",
"0.505",
"0.6157"
],
[
"0.0768",
"0.613",
"0.4927",
"0.613",
"0.4927",
"0.6274",
"0.0768",
"0.6274"
],
[
"0.505",
"0.616",
"0.9224",
"0.616",
"0.9224",
"0.63",
"0.505",
"0.63"
],
[
"0.0768",
"0.6274",
"0.492",
"0.6274",
"0.492",
"0.6416",
"0.0768",
"0.6416"
],
[
"0.505",
"0.631",
"0.9233",
"0.631",
"0.9233",
"0.6445",
"0.505",
"0.6445"
],
[
"0.0768",
"0.643",
"0.4927",
"0.643",
"0.4927",
"0.657",
"0.0768",
"0.657"
],
[
"0.5034",
"0.646",
"0.7534",
"0.646",
"0.7534",
"0.659",
"0.5034",
"0.659"
],
[
"0.0768",
"0.6577",
"0.4927",
"0.6577",
"0.4927",
"0.672",
"0.0768",
"0.672"
],
[
"0.5317",
"0.662",
"0.9224",
"0.662",
"0.9224",
"0.6753",
"0.5317",
"0.6753"
],
[
"0.0768",
"0.6724",
"0.4927",
"0.6724",
"0.4927",
"0.686",
"0.0768",
"0.686"
],
[
"0.5063",
"0.677",
"0.9224",
"0.677",
"0.9224",
"0.69",
"0.5063",
"0.69"
],
[
"0.0784",
"0.688",
"0.492",
"0.688",
"0.492",
"0.6987",
"0.0784",
"0.6987"
],
[
"0.5073",
"0.6914",
"0.9224",
"0.6914",
"0.9224",
"0.705",
"0.5073",
"0.705"
],
[
"0.0768",
"0.702",
"0.4927",
"0.702",
"0.4927",
"0.7153",
"0.0768",
"0.7153"
],
[
"0.505",
"0.7056",
"0.9233",
"0.7056",
"0.9233",
"0.7197",
"0.505",
"0.7197"
],
[
"0.0784",
"0.7173",
"0.4927",
"0.7173",
"0.4927",
"0.728",
"0.0784",
"0.728"
],
[
"0.5063",
"0.72",
"0.9224",
"0.72",
"0.9224",
"0.7344",
"0.5063",
"0.7344"
],
[
"0.0768",
"0.73",
"0.4958",
"0.73",
"0.4958",
"0.7437",
"0.0768",
"0.7437"
],
[
"0.5063",
"0.735",
"0.9233",
"0.735",
"0.9233",
"0.7485",
"0.5063",
"0.7485"
],
[
"0.0768",
"0.7437",
"0.492",
"0.7437",
"0.492",
"0.757",
"0.0768",
"0.757"
],
[
"0.5063",
"0.75",
"0.9233",
"0.75",
"0.9233",
"0.763",
"0.5063",
"0.763"
],
[
"0.0768",
"0.759",
"0.4927",
"0.759",
"0.4927",
"0.773",
"0.0768",
"0.773"
],
[
"0.5063",
"0.7646",
"0.6895",
"0.7646",
"0.6895",
"0.778",
"0.5063",
"0.778"
],
[
"0.0768",
"0.7734",
"0.492",
"0.7734",
"0.492",
"0.787",
"0.0768",
"0.787"
],
[
"0.5293",
"0.779",
"0.9233",
"0.78",
"0.9233",
"0.7944",
"0.5293",
"0.7935"
],
[
"0.0768",
"0.7886",
"0.46",
"0.7886",
"0.46",
"0.802",
"0.0768",
"0.802"
],
[
"0.5073",
"0.7954",
"0.9224",
"0.7954",
"0.9224",
"0.8096",
"0.5073",
"0.8096"
],
[
"0.509",
"0.81",
"0.921",
"0.81",
"0.921",
"0.8237",
"0.509",
"0.8237"
],
[
"0.0784",
"0.8237",
"0.0972",
"0.821",
"0.10046",
"0.8306",
"0.0809",
"0.834"
],
[
"0.0989",
"0.8213",
"0.2385",
"0.8213",
"0.2385",
"0.836",
"0.0989",
"0.836"
],
[
"0.509",
"0.8257",
"0.921",
"0.8257",
"0.921",
"0.836",
"0.509",
"0.836"
],
[
"0.076",
"0.839",
"0.4958",
"0.8403",
"0.4958",
"0.855",
"0.076",
"0.854"
],
[
"0.5063",
"0.839",
"0.9233",
"0.839",
"0.9233",
"0.853",
"0.5063",
"0.853"
],
[
"0.07434",
"0.855",
"0.4958",
"0.854",
"0.4958",
"0.8687",
"0.07434",
"0.869"
],
[
"0.505",
"0.855",
"0.9224",
"0.855",
"0.9224",
"0.8657",
"0.505",
"0.8657"
],
[
"0.076",
"0.8706",
"0.4927",
"0.8706",
"0.4927",
"0.884",
"0.076",
"0.884"
],
[
"0.5073",
"0.8687",
"0.9233",
"0.8687",
"0.9233",
"0.882",
"0.5073",
"0.882"
],
[
"0.0768",
"0.8853",
"0.492",
"0.8853",
"0.492",
"0.8984",
"0.0768",
"0.8984"
],
[
"0.505",
"0.882",
"0.9233",
"0.8833",
"0.9233",
"0.898",
"0.505",
"0.8965"
],
[
"0.0768",
"0.8994",
"0.4944",
"0.8994",
"0.4944",
"0.9136",
"0.0768",
"0.9136"
],
[
"0.5063",
"0.8984",
"0.556",
"0.8984",
"0.556",
"0.91",
"0.5063",
"0.91"
],
[
"0.0768",
"0.914",
"0.492",
"0.914",
"0.492",
"0.928",
"0.0768",
"0.928"
],
[
"0.5317",
"0.914",
"0.9224",
"0.914",
"0.9224",
"0.928",
"0.5317",
"0.928"
],
[
"0.0768",
"0.9287",
"0.4968",
"0.9287",
"0.4968",
"0.9424",
"0.0768",
"0.9424"
],
[
"0.5034",
"0.9287",
"0.9224",
"0.9287",
"0.9224",
"0.9424",
"0.5034",
"0.9424"
]
] | [
"(a) Fully-connected message passing",
"(b) Locally-connected message passing",
"(c) Dynamic graph message passing",
"Fig. 1. Contextual information is crucial for complex scene understanding tasks. To recognise the \"boathouse\", one needs to consider",
"the \"boat\" and the \"water\" next to it. Fully-connected message passing models (a) are able to obtain this information, but are",
"prohibitively expensive. Furthermore, they capture a lot of redundant information (i.e.\"trees\" and \"sky\"). Locally-connected models (b)",
"are more efficient, but miss out on important context. Our proposed approach (c), dynamically samples a small subset of relevant",
"feature nodes based on a learned dynamic sampling scheme, i.e.the learned position-specific random walk (indicated by the white",
"dashed arrow lines),and also dynamically predicts filter weights and affinities (indicated by unique edge and square colors.),which",
"arebothconditionedonthesampledfeaturenodes.",
"and COCO [44] datasets. We achieve significant performance",
"are thus not directly applicable in the feature learning task since",
"improvementsoverthefully-connectedmodel[69],[73],while",
"feature variables are typically continuous. Coupled with the fact",
"using substantiallyfewer floating point operations (FLOPs).Specif-",
"that CRFs are computationally expensive, CRFs are no longer used",
"ically,“plugging\"our module into existing networks, we show",
"formostcomputervisiontasks.",
"considerable improvements with respect to strong, state-of-the-",
"A complementary technique for increasing the receptive field",
"art baselines on three different vision downstream tasks and",
"of CNNs was to use dilated convolutions [6], [81]. With dilated",
"backbonearchitectures.Furthermore.wedesignanewTransformer",
"convolutions, the number of parameters does not change, while",
"backbone network based on the building block of our formulation.",
"thereceptivefield growsexponentiallyifthedilationrateis",
"It achieves leading performance on Imagenet classification task",
"linearly increased in successivelayers.Othermodifications to",
"and the pretrained backbone network also supports various vision",
"theconvolutionoperation includedeformableconvolution[15]",
"downstreamtasks.",
"[90], which learns the offset with respect to a predefined grid",
"A preliminary version of this work was published in [85]. We",
"from which to select input values. However, the weights of the",
"haveextendedourconferenceversionasfollows:(i)Weintroduce",
"deformable convolution filters do not depend on the selected input,",
"a modified process of dynamic message passing calculation and",
"andareinfactsharedacrossalldifferentpositions.Incontrast,our",
"present a simplified version of our priorformulation; Wefurther",
"dynamic sampling aims to sample over the whole feature graph to",
"recast our formulation to take a Transformer perspective and",
"obtain a large receptive field, and the predicted affinities and the",
"implementthemessage updatingprocess intoan efficientand",
"weights for message passing are position specific and conditioned",
"effective Transformer layer. (ii) Based on the building block of our",
"on the dynamically sampled nodes. Our model is thus able to better",
"formulation, we propose a new backbone network DGMN2 with",
"capture position-based semantic context to enable more effective",
"multi-scale feature representation which is very compatibility in",
"message passing among feature nodes.",
"supporting various downstream vision tasks. The experiment on",
"Another line of related research targets the graph learning [18],",
"upstream ImageNet [60] classification task show that all the variants",
"[30], [37], [53], [61], [74] on the irregular structured and geometric",
"ofDGMN2(i.e.,DGMN2-Tiny,DGMN2-Small,DGMN2-Medium",
"input. A neural network module EdgeConv is proposed in [74]",
"and DGMN2-Large respectively) can surpass the conventional",
"to learn dynamic graphs with k-nearest neighbor (kNN) on the",
"convolution-basedbackbonessuchasResNet/ResNeXt[25].[77],",
"irregular point clouds data. DeepGCN [37] proposes to enlarge",
"and Transformer-based backbone such as DeiT [68] and PVT [72].",
"the receptive field on graph networks with dilated convolution.",
"(i)With thepretrained backbonenetwork DGMN2, we show",
"DGM [30] introduces a differentiable graph module that learns",
"considerable improvements with respect to strong, state-of-the-art",
"dynamic adjacency matrices.",
"baselines on three different vision downstream tasks such as object",
"The idea of sampling graph nodes has previously been explored",
"detection, instance segmentation and semantic segmentation.",
"in GraphSAGE [23]. Crucially, GraphSAGE simply uniformly",
"samples nodes. In contrast, our sampling strategy is learned based",
"2",
"RELATEDWORK",
"on the node features.Specifically,we first sample the nodes",
"An early technique for modelling context for computer vision tasks",
"uniformly in the spatial dimension, and then dynamically predict",
"involved conditional random fields.In particular,the DenseCRF",
"walksof eachnodeconditioned onthenodefeatures.Furthermore",
"model [34] was popular as it modelled interactions between all pairs",
"GraphSAGE does not consider our second important property,",
"of pixels in an image. Although such models have been integrated",
"i.e., the dynamic prediction of the affinities and the message passing",
"into neural networks [1], [2], [78], [88], they are limited by the",
"kernels",
"fact that the pairwise potentials are based on simple handcrafted",
"We also note that [29] developed an idea of “dynamic convo-",
"features. Moreover, they mostly model discrete label spaces, and",
"lution\", that is predicting a dynamic convolutional filter for each"
] | {
"arxiv_id": "2209.09760",
"language": "en",
"page_idx": 1
} |
||
2209.09760-en-2 | 2209.09760-en-2.jpg | of this operation in the context of natural language processing with a graph-based formulation, and jointly learn dynamic weights and dynamic affinities, which are conditioned on an adaptively sampled neighbourhood for each feature node in the graph using passing.
convolution neural networks. However, there is a bottleneck in such computation and memory usage. PVT [72] uses downsampling operation to shrink the number of key and value tokens. Similarly, MViT [17] also performs pooling to reduce the number of tokens on which self-attention is applied. Swin Transformer [48], on the other hand, predefines a local window region for each query point to linearlize the attention computation. We also instantiate where $A_{i,j}=A[i,j]$ describes the connection relationship $i.e.$, the our formulation into an efficient and effective Transformer layer affinity between latent nodes $\mathbf{h}_i^{(t)}$ and $\mathbf{h}_j^{(t)},\mathcal{N}(i)$ denotes a selfand present a new Transformer backboner enetwork DGMN2 in $\text{included neighborhood of the node h}_i^{( }$ Transformer-based backbone networks, we introduce a graphbased formulation to adaptively sample feature nodes for effective calculation on the hidden node $\mathbf{h}_j^{(t)}$.The message updating function message passing.
3 DYNAMIC
GRAPH
MESSAGE
PASSING
NETWORKS
3.1 Problem definition and notation
of pixels and $C$ is the feature dimension, our goal is to learn a set of refined latent feature vectors $\mathbf{H}=\{\mathbf{h}_{i}\}_{i=1}^{\breve{N}}$ by utilising pixel locations. $\mathbf{H}$ has the same dimension as the observation F. edges and $A$ as its adjacency matrix. Specifically, the nodes of parameers, which, in adatiton to compuational overheaq, resulis the graph represent the latent feature vectors,i.e., $\mathcal{V} = \{ \mathbf{h} _i\} _i^N$, in redundancy in the connections, and also makes t describing the connections between nodes. In this work, we propose $\begin{array}{ll}\text{traınıng data.Theretore,asın Eq.1,alocal node connectıon held}\\\text{a novel dynamic graph message passing network[21] for deep}&\text{is considered in the graph message passing network. How}\end{array}$ representation learning, which refines each graph feature node by
passing messages on the graph $\mathcal{G}$. Different from existing message $\begin{array}{learning}\text{deep representations capturing both local and global}\\\text{recentive felds is imnortant for the model nerformance[26]}\end{array}$ passing neural networks considering a fully- or locally-connected $\begin{array}{c}\text{recceptrve helds is mportant for the model perftormance[26],}\\\text{sta}\end{array}$ with two dynamic properties, $i.e.,dynamic$ sampling of graph nodes to approximate the full graph distribution, and dynamicprediction of node-conditioned filter weights and affinities, in order to achieve more efficient and effective message passing.
# representation learning
Message passing neural networks (MPNNs)[21] present a generalised form of graph neural networks such as graph convolution networks [31], gated graph sequential networks [40] and graph
3
feature position. More recently,[75] further reduced the complexity or directed graph, feed-forward inference is performed through
a message passing phase followed by a readout phase upon the
lightweight grouped convolutions. Unlike[29],[75], we present graph nodes. The message passing phase usually takes $T$ iteration
steps to update feature nodes, while the readout phase is for the final prediction, e.g. , graph classification with updated nodes. In this work, we focus on the message passing phase for learning
the proposed dynamic sampling strategy for effective message efficient and effective feature refinement, since well-represented
features are critical in all downstream tasks. The message passing
Recently, there have been a number of variants of Trans- phase consists of two steps, $i.e.$ a message calculation step $M^t$ former [16],[69] in computer vision. ViT [16] adopts a pure- and a message updating step $U^t$ .Given a latent feature node h$_i^{(t)}$ Transformer architecture with query, key and value self-attention for at an iteration $t$ for computational efficiency, we consider a locally image classification and demonstrates its effectiveness for replacing connected node field with $v_i\subset\mathcal{V}$ and $v_i\in\mathbb{R}^{(K\times C)}$, where
$K\ll N$ is the number of sampled nodes in $v_i.$ Thus we can define
Transformer based models, namely its quadratic complexity in both the message calculation step for node $i$ operated locally as
$$\begin{aligned}\mathbf{m}_{i}^{(t+1)}&=M^t\left(A_{i,j},\{\mathbf{h}_{1}^{(t)},\cdots,\mathbf{h}_{K}^{(t)}\},\mathbf{w}_{j}\right)\\&=\sum_{j\in\mathcal{N}(i)}A_{i,j}\mathbf{h}_{j}^{(t)}\mathbf{w}_{j},\end{aligned}$$
(1)
$U^t$ then updates the node $\mathbf{h}_i^{(t)}$ with a linear combination of the calculated message and the observed feature $\mathbf{f}_i$ at the node position $i$ as:
$$\mathbf{h}_i^{(t+1)}=U^t\left(\mathbf{f}_i,\mathbf{m}_i^{(t+1)}\right)=\sigma\left(\mathbf{f}_i+\alpha_i^m\mathbf{m}_i^{(t+1)}\right),$$
(2)
where $\alpha_i^m$ of a learnable parameter for scaling the message, and
Given an input feature map interpreted as a set of feature vectors, the operation $\sigma(\cdot)$ is a non-linearity function, e.g. , ReLU. By $i.e,\mathbf{F}=\{\mathbf{f}_{i}\}_{i=1}^{N}$ with $\mathbf{f}_i\in\mathbb{R}^{1\times C}$, where $N$ is the number iteratively performing message passing on each node with $T$ steps,
we obtain a refined feature map $\mathbf{H}^{(T)}$ as output.
# hidden structured information among the feature vectors at different 3.3 From a fully-connected graph to a dynamic sam-
parameters, which, in addition to computational overhead, results training data. Therefore, as in Eq. 1, a local node connection field various computer vision tasks, such as detection and segmentation, receptive felds is important for the model performance [26], much fewer parameters than the fully-connected setting, we further explore dynamic sampling strategies in our proposed graph message passing network. We develop a uniform sampling scheme, which we then extend to a predicted random walk sampling scheme, aimed at reducing the redundancy found in a fully-connected graph. This sampling is performed in a dynamic fashion, meaning that for
3.2 Graph message passing neural networks for deep a given node $\mathbf{h}_i$,we aim to sample an optimal subset of $v_i$ from $\mathcal{V}$
to update $\mathbf{h}_i$ via message passing as shown in Fig.2.
Multiple uniform sampling for dynamic receptive fields. Uniform sampling is a commonly used strategy for graph node sampling [36] based on Monte-Carlo estimation. To approximate
attention networks [70]. In order to model structured graph data, in the distribution of $\mathcal{V}$, we consider a set of $S$ uniform sampling which latent variables are represented as nodes on an undirected rates $\varphi$ with $\varphi=\{\rho_q\}_{q=1}^S$, where $\rho_q$ is a sampling rate. Let us | [
[
"0.9126",
"0.03284",
"0.925",
"0.03284",
"0.925",
"0.0442",
"0.9126",
"0.0442"
],
[
"0.0768",
"0.05365",
"0.921",
"0.05365",
"0.921",
"0.06757",
"0.0768",
"0.06757"
],
[
"0.0768",
"0.06946",
"0.4958",
"0.06946",
"0.4958",
"0.0833",
"0.0768",
"0.0833"
],
[
"0.5",
"0.06946",
"0.9224",
"0.0682",
"0.9224",
"0.0821",
"0.5",
"0.0833"
],
[
"0.0768",
"0.084",
"0.5",
"0.084",
"0.5",
"0.09784",
"0.0768",
"0.09784"
],
[
"0.505",
"0.084",
"0.9224",
"0.084",
"0.9224",
"0.09784",
"0.505",
"0.09784"
],
[
"0.0768",
"0.0985",
"0.4944",
"0.0985",
"0.4944",
"0.11237",
"0.0768",
"0.11237"
],
[
"0.5034",
"0.0985",
"0.921",
"0.09784",
"0.921",
"0.1111",
"0.5034",
"0.11237"
],
[
"0.0784",
"0.11426",
"0.5024",
"0.11426",
"0.5024",
"0.125",
"0.0784",
"0.125"
],
[
"0.5034",
"0.113",
"0.9224",
"0.113",
"0.9224",
"0.127",
"0.5034",
"0.127"
],
[
"0.0784",
"0.1288",
"0.4983",
"0.1288",
"0.4983",
"0.1395",
"0.0784",
"0.1395"
],
[
"0.505",
"0.1288",
"0.919",
"0.1288",
"0.919",
"0.1395",
"0.505",
"0.1395"
],
[
"0.0768",
"0.1427",
"0.4944",
"0.1427",
"0.4944",
"0.1559",
"0.0768",
"0.1559"
],
[
"0.5063",
"0.1427",
"0.919",
"0.1427",
"0.919",
"0.1528",
"0.5063",
"0.1528"
],
[
"0.0784",
"0.159",
"0.1299",
"0.159",
"0.1299",
"0.1692",
"0.0784",
"0.1692"
],
[
"0.5063",
"0.1559",
"0.919",
"0.1572",
"0.919",
"0.1704",
"0.5063",
"0.1692"
],
[
"0.0972",
"0.1704",
"0.5034",
"0.1704",
"0.5034",
"0.1843",
"0.0972",
"0.1843"
],
[
"0.5063",
"0.1718",
"0.921",
"0.1718",
"0.921",
"0.1849",
"0.5063",
"0.1849"
],
[
"0.0768",
"0.1849",
"0.4958",
"0.1849",
"0.4958",
"0.1989",
"0.0768",
"0.1989"
],
[
"0.5024",
"0.1863",
"0.919",
"0.1843",
"0.919",
"0.1989",
"0.5024",
"0.2008"
],
[
"0.0768",
"0.2008",
"0.4958",
"0.2008",
"0.4958",
"0.214",
"0.0768",
"0.214"
],
[
"0.5034",
"0.202",
"0.921",
"0.202",
"0.921",
"0.2153",
"0.5034",
"0.2153"
],
[
"0.0768",
"0.2166",
"0.4927",
"0.2166",
"0.4927",
"0.2267",
"0.0768",
"0.2267"
],
[
"0.5034",
"0.214",
"0.921",
"0.214",
"0.921",
"0.2291",
"0.5034",
"0.2291"
],
[
"0.0768",
"0.2311",
"0.4927",
"0.2311",
"0.4927",
"0.2412",
"0.0768",
"0.2412"
],
[
"0.5063",
"0.2299",
"0.9233",
"0.2299",
"0.9233",
"0.2443",
"0.5063",
"0.2443"
],
[
"0.0784",
"0.2456",
"0.4927",
"0.2456",
"0.4927",
"0.2556",
"0.0784",
"0.2556"
],
[
"0.5063",
"0.2456",
"0.8784",
"0.2456",
"0.8784",
"0.2588",
"0.5063",
"0.2588"
],
[
"0.0768",
"0.2588",
"0.492",
"0.2588",
"0.492",
"0.2727",
"0.0768",
"0.2727"
],
[
"0.5703",
"0.2664",
"0.6846",
"0.2651",
"0.6846",
"0.2842",
"0.5703",
"0.2854"
],
[
"0.0768",
"0.2734",
"0.492",
"0.2734",
"0.492",
"0.2874",
"0.0768",
"0.2874"
],
[
"0.6733",
"0.2683",
"0.847",
"0.2683",
"0.847",
"0.286",
"0.6733",
"0.286"
],
[
"0.0784",
"0.289",
"0.4885",
"0.289",
"0.4885",
"0.2998",
"0.0784",
"0.2998"
],
[
"0.648",
"0.2917",
"0.7686",
"0.2935",
"0.7686",
"0.3125",
"0.648",
"0.31"
],
[
"0.899",
"0.2942",
"0.9263",
"0.2942",
"0.9263",
"0.3113",
"0.899",
"0.3113"
],
[
"0.0768",
"0.3018",
"0.492",
"0.3018",
"0.492",
"0.315",
"0.0768",
"0.315"
],
[
"0.6445",
"0.3005",
"0.652",
"0.3005",
"0.652",
"0.3062",
"0.6445",
"0.3062"
],
[
"0.6396",
"0.3113",
"0.689",
"0.3113",
"0.689",
"0.3257",
"0.6396",
"0.3257"
],
[
"0.0768",
"0.3176",
"0.4902",
"0.3176",
"0.4902",
"0.3308",
"0.0768",
"0.3308"
],
[
"0.0784",
"0.3333",
"0.492",
"0.3333",
"0.492",
"0.3435",
"0.0784",
"0.3435"
],
[
"0.5034",
"0.332",
"0.9224",
"0.332",
"0.9224",
"0.3467",
"0.5034",
"0.3467"
],
[
"0.0784",
"0.348",
"0.492",
"0.348",
"0.492",
"0.358",
"0.0784",
"0.358"
],
[
"0.505",
"0.3467",
"0.925",
"0.3467",
"0.925",
"0.3643",
"0.505",
"0.3643"
],
[
"0.0768",
"0.3599",
"0.492",
"0.3599",
"0.492",
"0.3738",
"0.0768",
"0.3738"
],
[
"0.5063",
"0.3674",
"0.9233",
"0.3674",
"0.9233",
"0.3806",
"0.5063",
"0.3806"
],
[
"0.0768",
"0.3757",
"0.4927",
"0.3757",
"0.4927",
"0.3894",
"0.0768",
"0.3894"
],
[
"0.5034",
"0.379",
"0.921",
"0.379",
"0.921",
"0.3965",
"0.5034",
"0.3965"
],
[
"0.0768",
"0.3894",
"0.492",
"0.3901",
"0.492",
"0.404",
"0.0768",
"0.4028"
],
[
"0.5034",
"0.3958",
"0.9233",
"0.3965",
"0.9233",
"0.414",
"0.5034",
"0.4136"
],
[
"0.076",
"0.4048",
"0.4927",
"0.4048",
"0.4927",
"0.4185",
"0.076",
"0.4185"
],
[
"0.076",
"0.4192",
"0.1871",
"0.4204",
"0.1871",
"0.432",
"0.076",
"0.4312"
],
[
"0.5063",
"0.414",
"0.9233",
"0.4136",
"0.9233",
"0.4312",
"0.5063",
"0.432"
],
[
"0.5073",
"0.432",
"0.9224",
"0.432",
"0.9224",
"0.4458",
"0.5073",
"0.4458"
],
[
"0.07434",
"0.4526",
"0.1847",
"0.455",
"0.1847",
"0.4702",
"0.07434",
"0.4685"
],
[
"0.2157",
"0.455",
"0.2778",
"0.455",
"0.2778",
"0.4697",
"0.2157",
"0.4697"
],
[
"0.3064",
"0.455",
"0.3865",
"0.455",
"0.3865",
"0.4697",
"0.3064",
"0.4697"
],
[
"0.4182",
"0.455",
"0.492",
"0.455",
"0.492",
"0.4697",
"0.4182",
"0.4697"
],
[
"0.5107",
"0.4495",
"0.5337",
"0.4495",
"0.5337",
"0.4583",
"0.5107",
"0.4583"
],
[
"0.546",
"0.4653",
"0.869",
"0.464",
"0.869",
"0.4849",
"0.546",
"0.486"
],
[
"0.7466",
"0.473",
"0.7607",
"0.473",
"0.7607",
"0.4766",
"0.7466",
"0.4766"
],
[
"0.7573",
"0.473",
"0.767",
"0.473",
"0.767",
"0.4766",
"0.7573",
"0.4766"
],
[
"0.903",
"0.473",
"0.9165",
"0.473",
"0.9165",
"0.4817",
"0.903",
"0.4817"
],
[
"0.0768",
"0.4912",
"0.3489",
"0.4912",
"0.3489",
"0.505",
"0.0768",
"0.505"
],
[
"0.0784",
"0.4734",
"0.1708",
"0.4734",
"0.1708",
"0.4849",
"0.0784",
"0.4849"
],
[
"0.5073",
"0.4956",
"0.9224",
"0.4956",
"0.9224",
"0.509",
"0.5073",
"0.509"
],
[
"0.0768",
"0.511",
"0.5",
"0.511",
"0.5",
"0.5244",
"0.0768",
"0.5244"
],
[
"0.5073",
"0.5103",
"0.921",
"0.5103",
"0.921",
"0.5234",
"0.5073",
"0.5234"
],
[
"0.076",
"0.523",
"0.4983",
"0.523",
"0.4983",
"0.5405",
"0.076",
"0.5405"
],
[
"0.5063",
"0.5244",
"0.9233",
"0.5244",
"0.9233",
"0.5386",
"0.5063",
"0.5386"
],
[
"0.076",
"0.5405",
"0.4944",
"0.5405",
"0.4944",
"0.5537",
"0.076",
"0.5537"
],
[
"0.5034",
"0.5386",
"0.814",
"0.5386",
"0.814",
"0.553",
"0.5034",
"0.553"
],
[
"0.0768",
"0.5547",
"0.492",
"0.5547",
"0.492",
"0.5693",
"0.0768",
"0.5693"
],
[
"0.0768",
"0.571",
"0.492",
"0.571",
"0.492",
"0.5806",
"0.0768",
"0.5806"
],
[
"0.5073",
"0.5737",
"0.9233",
"0.5737",
"0.9233",
"0.5884",
"0.5073",
"0.5884"
],
[
"0.076",
"0.584",
"0.4944",
"0.5835",
"0.4944",
"0.5967",
"0.076",
"0.5977"
],
[
"0.509",
"0.5903",
"0.591",
"0.5903",
"0.591",
"0.6016",
"0.509",
"0.6016"
],
[
"0.0801",
"0.5986",
"0.4902",
"0.5986",
"0.4902",
"0.6094",
"0.0801",
"0.6094"
],
[
"0.5034",
"0.606",
"0.9224",
"0.607",
"0.9224",
"0.622",
"0.5034",
"0.6206"
],
[
"0.0768",
"0.613",
"0.4927",
"0.613",
"0.4927",
"0.627",
"0.0768",
"0.627"
],
[
"0.0768",
"0.6274",
"0.4944",
"0.6274",
"0.4944",
"0.6416",
"0.0768",
"0.6416"
],
[
"0.505",
"0.6226",
"0.9224",
"0.6226",
"0.9224",
"0.636",
"0.505",
"0.636"
],
[
"0.0768",
"0.6426",
"0.4944",
"0.6426",
"0.4944",
"0.657",
"0.0768",
"0.657"
],
[
"0.5063",
"0.637",
"0.9224",
"0.637",
"0.9224",
"0.651",
"0.5063",
"0.651"
],
[
"0.076",
"0.653",
"0.4944",
"0.653",
"0.4944",
"0.6704",
"0.076",
"0.6704"
],
[
"0.5063",
"0.6514",
"0.9224",
"0.6514",
"0.9224",
"0.6655",
"0.5063",
"0.6655"
],
[
"0.0768",
"0.672",
"0.4927",
"0.672",
"0.4927",
"0.685",
"0.0768",
"0.685"
],
[
"0.5073",
"0.6685",
"0.921",
"0.6685",
"0.921",
"0.6787",
"0.5073",
"0.6787"
],
[
"0.0768",
"0.686",
"0.4927",
"0.686",
"0.4927",
"0.6997",
"0.0768",
"0.6997"
],
[
"0.505",
"0.681",
"0.9224",
"0.681",
"0.9224",
"0.6943",
"0.505",
"0.6943"
],
[
"0.0768",
"0.7007",
"0.492",
"0.7007",
"0.492",
"0.714",
"0.0768",
"0.714"
],
[
"0.509",
"0.6963",
"0.921",
"0.6963",
"0.921",
"0.707",
"0.509",
"0.707"
],
[
"0.076",
"0.7153",
"0.4927",
"0.714",
"0.4927",
"0.728",
"0.076",
"0.729"
],
[
"0.5063",
"0.7104",
"0.9224",
"0.7104",
"0.9224",
"0.7236",
"0.5063",
"0.7236"
],
[
"0.076",
"0.73",
"0.4927",
"0.728",
"0.4927",
"0.741",
"0.076",
"0.7437"
],
[
"0.5063",
"0.7246",
"0.9233",
"0.7246",
"0.9233",
"0.738",
"0.5063",
"0.738"
],
[
"0.076",
"0.7437",
"0.4927",
"0.742",
"0.4927",
"0.757",
"0.076",
"0.7583"
],
[
"0.5073",
"0.7393",
"0.9224",
"0.7393",
"0.9224",
"0.7524",
"0.5073",
"0.7524"
],
[
"0.076",
"0.759",
"0.4927",
"0.759",
"0.4927",
"0.773",
"0.076",
"0.773"
],
[
"0.5073",
"0.755",
"0.921",
"0.755",
"0.921",
"0.765",
"0.5073",
"0.765"
],
[
"0.076",
"0.7734",
"0.4944",
"0.7734",
"0.4944",
"0.787",
"0.076",
"0.787"
],
[
"0.5063",
"0.768",
"0.9224",
"0.768",
"0.9224",
"0.782",
"0.5063",
"0.782"
],
[
"0.5073",
"0.7827",
"0.921",
"0.7827",
"0.921",
"0.797",
"0.5073",
"0.797"
],
[
"0.0768",
"0.789",
"0.4927",
"0.789",
"0.4927",
"0.8",
"0.0768",
"0.8"
],
[
"0.076",
"0.802",
"0.3628",
"0.803",
"0.3628",
"0.8164",
"0.076",
"0.815"
],
[
"0.5073",
"0.7974",
"0.9233",
"0.7974",
"0.9233",
"0.811",
"0.5073",
"0.811"
],
[
"0.5107",
"0.813",
"0.9224",
"0.813",
"0.9224",
"0.8237",
"0.5107",
"0.8237"
],
[
"0.5073",
"0.827",
"0.9233",
"0.827",
"0.9233",
"0.8403",
"0.5073",
"0.8403"
],
[
"0.0768",
"0.836",
"0.492",
"0.836",
"0.492",
"0.8496",
"0.0768",
"0.8496"
],
[
"0.505",
"0.8413",
"0.9233",
"0.8413",
"0.9233",
"0.855",
"0.505",
"0.855"
],
[
"0.0768",
"0.8516",
"0.2532",
"0.8516",
"0.2532",
"0.8657",
"0.0768",
"0.8657"
],
[
"0.5063",
"0.856",
"0.8433",
"0.856",
"0.8433",
"0.869",
"0.5063",
"0.869"
],
[
"0.0768",
"0.869",
"0.5063",
"0.869",
"0.5063",
"0.884",
"0.0768",
"0.884"
],
[
"0.5063",
"0.8706",
"0.925",
"0.8706",
"0.925",
"0.884",
"0.5063",
"0.884"
],
[
"0.0768",
"0.8853",
"0.4944",
"0.8853",
"0.4944",
"0.8984",
"0.0768",
"0.8984"
],
[
"0.5063",
"0.8853",
"0.9224",
"0.8853",
"0.9224",
"0.8984",
"0.5063",
"0.8984"
],
[
"0.0768",
"0.8994",
"0.4944",
"0.8994",
"0.4944",
"0.9136",
"0.0768",
"0.9136"
],
[
"0.5034",
"0.8994",
"0.9224",
"0.8994",
"0.9224",
"0.9136",
"0.5034",
"0.9136"
],
[
"0.0768",
"0.914",
"0.4958",
"0.914",
"0.4958",
"0.928",
"0.0768",
"0.928"
],
[
"0.505",
"0.914",
"0.9224",
"0.914",
"0.9224",
"0.928",
"0.505",
"0.928"
],
[
"0.0768",
"0.9287",
"0.4958",
"0.9297",
"0.4958",
"0.9414",
"0.0768",
"0.94"
],
[
"0.4983",
"0.928",
"0.925",
"0.928",
"0.925",
"0.946",
"0.4983",
"0.946"
]
] | [
"3",
"feature position. More recently, [75] further reduced the complexityor directed graph, feed-forward inference is performed through",
"of this operation in the context of natural language processing with",
"a message passing phase followed by a readout phase upon the",
"lightweight grouped convolutions. Unlike [29], [75], we present",
"graph nodes. The message passing phase usually takes T iteration",
"a graph-based formulation, and jointly learn dynamic weights",
"steps to update feature nodes, while the readout phase is for the",
"and dynamic affinities,which areconditioned on an adaptively",
"final prediction, e.g., graph classification with updated nodes. In",
"sampledneighbourhoodforeachfeaturenodeinthegraphusing",
"this work,wefocus on the messagepassing phaseforlearning",
"the proposed dynamic sampling strategy for effective message",
"efficientandeffectivefeaturerefinement.sincewell-represented",
"passing.",
"features are critical in all downstream tasks. The message passing",
"Recently, there have been a number of variants of Trans-",
"phase consists of two steps, i.e., a message calculation step Mt",
"former [16], [69] in computer vision. ViT [16] adopts a pure-",
"and a message updating step Ut Given a latent feature node h(t)",
"Transformer architecture with query, key and value self-attention for",
"at an iteration t, for computational efficiency, we consider a locally",
"imageclassificationanddemonstratesitseffectivenessforreplacing",
"connected node field with w; C V and vi E R(KxC), where",
"convolution neural networks.However,thereisa bottleneck in such",
"K < N is the number of sampled nodes in vi. Thus we can define",
"Transformerbasedmodels.namelyitsquadraticcomplexityinboth",
"the message calculation step for node i operated locally as",
"computation and memory usage. PVT [72] uses downsampling",
"m(t+1) = M (",
"operation to shrink the number of key and value tokens. Similarly,",
"Ai.j,[ht),...,h?,wj",
"MViT[17alsoperformspoolingtoreducethenumberoftokens",
" Ai,jhft'wj,",
"(1)",
"on which self-attention is applied. Swin Transformer [48], on",
"",
"jEN(i)",
"the other hand, predefines a local window region for each query",
"pointtolinearlize the attention computation.We alsoinstantiate",
"where Ai.j = A[i, j] describes the connection relationship i.e., the",
"ourformulationintoanefficientandeffectiveTransformerlayer",
"affinity between latent nodes h(t) and h(t), N(i) denotes a self-",
"and present a new Transformer backbone network DGMN2 in",
"included neighborhood of the node h,t) which can be derived",
"supporting various vision downstream tasks. Different to above",
"from v; and w, E RCxC is a transformation matrix for message",
"Transformer-based backbonenetworks,we introduce a graph-",
"calculation on the hidden node h(t). The message updating function",
"based formulation to adaptively sample feature nodes for effective",
"message passing.",
"Ut then updates the node h(t) with a linear combination of the",
"calculated message and the observed feature f, at the node position",
"3DYNAMIC",
"GRAPH",
"MESSAGE",
"PASSING",
"zas",
"h(t+1) = Ut (f,m++1) = α (fi + a\"m(t+1)",
"",
"",
"(2",
"3.1 Problem definitionand notation",
"NETWORKS",
"where om of a learnable parameter for scaling the message, and",
"Given an input feature map interpreted as a set of feature vectors,",
"the operation o() is a non-linearity function, e.g. , ReLU. By",
"i.e., F = {f;}i with f, e RlxC, where N is the number",
"iteratively performing message passing on each node with T steps,",
"of pixels and C is the feature dimension, our goal is to learn",
"we obtain a refined feature map H(T) as output.",
"a set of refined latent feature vectors H =- [h,}, by utilising",
"hidden structured informationamongthefeaturevectors atdifferent",
"3.3 From a fully-connected graph to a dynamic sam-",
"pixel locations. H has the same dimension as the observation F.",
"pled graph",
"To learn such structured representations.weconvert thefeature",
"Afully-connected graph typically contains many connections and",
"map into a graph G = [V,&, A) with V as its nodes, as its",
"edges and A as its adjacency matrix. Specifically, the nodes of",
"parameters, which, in addition to computational overhead, results",
"the graph represent the latent feature vectors, i.e., V = [h,}-1,",
"in redundancy in the connections, and also makes the network",
"and Ae RNxN is a binary or learnable matrix with self-loops",
"optimisation more difficult especially when dealing with limited",
"describing the connections between nodes.In this work, we propose",
"training data.Therefore,as inEg.l,alocal nodeconnectionfield",
"a novel dynamic graph message passing network [21] for deep",
"is considered in the graph message passing network. However, in",
"representation learning, which refines each graph feature node by",
"various computer vision tasks, such as detection and segmentation",
"passing messages on the graph G. Different from existing message",
"learning deep representations capturing both local and global",
"passing neural networks considering a fully-or locally-connected",
"receptive fields is important for the model performance [26],",
"static graph [21], [73], we propose a dynamic graph network model",
"[39], [54], [57]. To maintain a large receptive field while utilising",
"with two dynamic properties, i.e., dynamic sampling of graph nodes",
"muchfewerparametersthanthefully-connectedsetting,wefurther",
"to approximate the full graph distribution, and dynamic prediction",
"explore dynamic sampling strategies in our proposed graph message",
"passing network. We develop a uniform sampling scheme, which",
"ofnode-conditionedfilterweightsand affinities,in ordertoachieve",
"more efficient and effective message passing.",
"we then extend to a predicted random walk sampling scheme,",
"aimed atreducingtheredundancyfoundinafully-connectedgraph",
"This sampling is performed in a dynamic fashion, meaning that for",
"3.2 Graph message passing neural networks for deep",
"a given node hi, we aim to sample an optimal subset of u; from V",
"representationlearning",
"to update h; via message passing as shown in Fig. 2.",
"Message passing neural networks (MPNNs) [21] present a gener-",
"Multiple uniform sampling for dynamic receptive fields. Uni-",
"alised form of graph neural networks such as graph convolution",
"form sampling is a commonly used strategy for graph node",
"networks [31], gated graph sequential networks [40] and graph",
"sampling [36] based on Monte-Carlo estimation. To approximate",
"attention networks [70]. In order to model structured graph data, in",
"the distribution of V, we consider a set of S uniform sampling",
"which latent variables arerepresented as nodes on an undirected",
"rates with = (pq)&-1, where Pq is a sampling rate. Let us"
] | {
"arxiv_id": "2209.09760",
"language": "en",
"page_idx": 2
} |
||
2209.09760-en-4 | 2209.09760-en-4.jpg | 5
<FigureHere>
Fig.3. Schematic illustration of the proposed dynamic message passing calculation (DMC) module. The small red square indicates the receiving node whose message is calculated from its neighbourhood, $i.e.$ the sampled $K(e.g.3\times3)$ features nodes. The module accepts a feature map as input and produces its corresponding message map. The symbol * denotes group convolution operation using the dynamically predicted and position specific group kernels and affinities.
where the function soft$\max_c(\cdot)$ denotes a softmax operation along as in existing graph-based feature learning models [9],[41],[73],
the channel axis, which is used to perform a normalisation on [84], we also perform $T=1$ iteration of message updating.the estimated affinity $A_{i,j}^{lq}\in\mathbb{R}^{1}\cdot\mathbf{W}_{i,j}^{k,A}\in\mathbb{R}^{(G\times C+1)\times(K\times C)}$and bi$b_{i,j}^{k,A}\in\mathbb{R}^{(G\times C+1)}$ are matrix
reuuce he number ol ne nler parameers, we consuer groupea Our approach is related to deformable convolution [15],[90] and convolutions [10] with a set of $G$ groups split from the total $C\:\overset{\text{Our approach is related to deformable convolution [15],[90]$
feature channels, and $G\ll C,ie.$, each group of $C/G$ feature Non-local [71], but has several key differences:
channels shares the same set of filter parameters. The predicted dynamic filter weights and the affinities are then used in Eq.5 for dynamic message calculation.
## 3.5 Modular instantiation
A fundamental difference to deformable convolution is that it only learns the offset dependent on the input feature while the filter weights are fixed for all inputs. In contrast, our model learns the random walk, weight and affinity as all being dependent on the input. This property mak specific whereas deformable convolution shares the same weight
Fig.2 and 3 shows how our proposed dynamic graph message across all convolution positions in the feature map. Moreover,[15],
[90] only consider $3\times3$ local neighbours at each convolution
The proposed module accepts
a set of $K$ nodes (where $K\gg9)$ for message passing globally
and F have the same dimension $i. e, \mathbf{F} , \mathbf{H} \in \mathbb{R} ^{H\times W\times C}$,larger receptive field than deformable convolut
Whilst Non-local also learns to refine deep features, it uses a self-attention matrix to guide the message passing between each pair of feature nodes. In contrast, our model learns to sample graph
in Fig. 2 for clarity). The uniform and the random walk sampler feature nodes to capture global feature information efficiently. This
dynamic sampling reduces computational overhead, whilst still
for subsequent dynamic message calculation (DMC) in Fig. 3. The being able to improve upon the accuracy of Non-local across
detail the extension of our model to transformer architectures next.
passing network (DGMN) can be implemented in a neural network. which can be derived from any neural network layer. $\mathbf{H}^{(0)}$ denotes an initial state of the latent feature map, H, and is initialised with where $H,W$ and $C$ are the height, width and the number of feature channels of the feature map respectively. We first define a set of $S$ uniform sampling rates (we show two uniform sampling rates sample the nodes from the full graph and return the node indices other sampling strategies could also be flexibly employed in our framework.
The sampled feature nodes are processed along two paths: one for predicting the node-dependant dynamic affinities $A^{\prime q}\in\mathbb{R}^{\dot{H}\times W\times K}$ and another path for dynamic filters 4.1 Interpretation as a Iranstormer layer$\mathbf{w}^q\in\mathbb{R}^{H\times W\times K\times G}$ where $K(eg.,3\times3)$ is the kernel size for the We further instantiate our formulation into an efficient and effective$\mathbf{w}$ receiving node. The matrix t predict the dynamic filters and affinities in Eq 6 is implemented by a$3\times3$ convolution layer. Message $\mathbf{M}_q\in\mathbb{R}^{H\times W\times C}$ corresponding to the $q$-th sampling rate is then scaled to perform a linear combination with the observed feature map $\mathbf{F}$, to produce a refined feature map $\mathbf{H}^{(1)}$ as output. To balance performance and efficiency, with graph vertexes $\mathcal{V}$.Given the latent node $\mathbf{h}_j^{(t)}$,we can estimate
4 EXTENDING DGMN TO TRANSFORMERS
4.1 Interpretation as a Transformer layer
network DGMN2 to support various vision downstream tasks.
Specifically, we modify the process of dynamic message passing calculation and implement it as follows: With the predicted walk $\triangle\mathbf{d}_j^q$, we can obtain a new set of adaptively sampled nodes | [
[
"0.33",
"0.1559",
"0.3489",
"0.1559",
"0.3489",
"0.1635",
"0.33",
"0.1635"
],
[
"0.4412",
"0.1572",
"0.4558",
"0.1572",
"0.4558",
"0.1604",
"0.4412",
"0.1604"
],
[
"0.0801",
"0.305",
"0.3406",
"0.305",
"0.3406",
"0.315",
"0.0801",
"0.315"
],
[
"0.3381",
"0.305",
"0.4412",
"0.305",
"0.4412",
"0.315",
"0.3381",
"0.315"
],
[
"0.468",
"0.305",
"0.919",
"0.305",
"0.919",
"0.315",
"0.468",
"0.315"
],
[
"0.0768",
"0.3164",
"0.9224",
"0.3164",
"0.9224",
"0.33",
"0.0768",
"0.33"
],
[
"0.0784",
"0.33",
"0.9224",
"0.33",
"0.9224",
"0.3435",
"0.0784",
"0.3435"
],
[
"0.0768",
"0.3403",
"0.607",
"0.3403",
"0.607",
"0.3542",
"0.0768",
"0.3542"
],
[
"0.0768",
"0.3806",
"0.5",
"0.3806",
"0.5",
"0.3945",
"0.0768",
"0.3945"
],
[
"0.4983",
"0.3806",
"0.9233",
"0.3801",
"0.9233",
"0.3945",
"0.4983",
"0.3958"
],
[
"0.0768",
"0.3958",
"0.4944",
"0.3958",
"0.4944",
"0.4092",
"0.0768",
"0.4092"
],
[
"0.5034",
"0.3965",
"0.884",
"0.3965",
"0.884",
"0.4104",
"0.5034",
"0.4104"
],
[
"0.076",
"0.4077",
"0.339",
"0.4077",
"0.339",
"0.4255",
"0.076",
"0.4255"
],
[
"0.3464",
"0.4092",
"0.492",
"0.404",
"0.492",
"0.4185",
"0.3464",
"0.4236"
],
[
"0.07434",
"0.4224",
"0.4958",
"0.4248",
"0.4958",
"0.4426",
"0.07434",
"0.44"
],
[
"0.5063",
"0.4312",
"0.6313",
"0.4312",
"0.6313",
"0.4458",
"0.5063",
"0.4458"
],
[
"0.076",
"0.4426",
"0.492",
"0.4426",
"0.492",
"0.4558",
"0.076",
"0.4558"
],
[
"0.076",
"0.4558",
"0.492",
"0.455",
"0.492",
"0.4697",
"0.076",
"0.4702"
],
[
"0.505",
"0.4507",
"0.9233",
"0.4495",
"0.9233",
"0.464",
"0.505",
"0.4653"
],
[
"0.0768",
"0.4717",
"0.4927",
"0.4717",
"0.4927",
"0.4849",
"0.0768",
"0.4849"
],
[
"0.5073",
"0.4666",
"0.806",
"0.4666",
"0.806",
"0.4797",
"0.5073",
"0.4797"
],
[
"0.0768",
"0.4849",
"0.492",
"0.4849",
"0.492",
"0.4988",
"0.0768",
"0.4988"
],
[
"0.5317",
"0.481",
"0.9233",
"0.481",
"0.9233",
"0.4944",
"0.5317",
"0.4944"
],
[
"0.0768",
"0.5005",
"0.4927",
"0.5005",
"0.4927",
"0.5146",
"0.0768",
"0.5146"
],
[
"0.509",
"0.4956",
"0.9233",
"0.4956",
"0.9233",
"0.509",
"0.509",
"0.509"
],
[
"0.509",
"0.5103",
"0.9224",
"0.5103",
"0.9224",
"0.5234",
"0.509",
"0.5234"
],
[
"0.0768",
"0.5176",
"0.2656",
"0.5176",
"0.2656",
"0.528",
"0.0768",
"0.528"
],
[
"0.509",
"0.5244",
"0.9224",
"0.5244",
"0.9224",
"0.5386",
"0.509",
"0.5386"
],
[
"0.5073",
"0.539",
"0.9233",
"0.539",
"0.9233",
"0.553",
"0.5073",
"0.553"
],
[
"0.076",
"0.5454",
"0.273",
"0.5454",
"0.273",
"0.5596",
"0.076",
"0.5596"
],
[
"0.5073",
"0.5537",
"0.9224",
"0.553",
"0.9224",
"0.5674",
"0.5073",
"0.5684"
],
[
"0.0784",
"0.5664",
"0.492",
"0.5664",
"0.492",
"0.577",
"0.0784",
"0.577"
],
[
"0.5063",
"0.5674",
"0.9224",
"0.5674",
"0.9224",
"0.5806",
"0.5063",
"0.5806"
],
[
"0.0784",
"0.5806",
"0.4927",
"0.5806",
"0.4927",
"0.5913",
"0.0784",
"0.5913"
],
[
"0.5073",
"0.5835",
"0.9224",
"0.5835",
"0.9224",
"0.5967",
"0.5073",
"0.5967"
],
[
"0.0768",
"0.5933",
"0.4927",
"0.5933",
"0.4927",
"0.607",
"0.0768",
"0.607"
],
[
"0.505",
"0.5977",
"0.9224",
"0.5977",
"0.9224",
"0.6113",
"0.505",
"0.6113"
],
[
"0.0768",
"0.606",
"0.4927",
"0.606",
"0.4927",
"0.6206",
"0.0768",
"0.6206"
],
[
"0.5034",
"0.6113",
"0.9224",
"0.613",
"0.9224",
"0.627",
"0.5034",
"0.625"
],
[
"0.0768",
"0.6226",
"0.4927",
"0.6226",
"0.4927",
"0.636",
"0.0768",
"0.636"
],
[
"0.5063",
"0.627",
"0.9233",
"0.627",
"0.9233",
"0.64",
"0.5063",
"0.64"
],
[
"0.076",
"0.636",
"0.4902",
"0.634",
"0.4902",
"0.6484",
"0.076",
"0.651"
],
[
"0.505",
"0.6416",
"0.826",
"0.64",
"0.826",
"0.654",
"0.505",
"0.655"
],
[
"0.0768",
"0.6514",
"0.492",
"0.6514",
"0.492",
"0.6655",
"0.0768",
"0.6655"
],
[
"0.5293",
"0.655",
"0.9224",
"0.656",
"0.9224",
"0.6704",
"0.5293",
"0.6694"
],
[
"0.0768",
"0.6665",
"0.4944",
"0.6665",
"0.4944",
"0.6797",
"0.0768",
"0.6797"
],
[
"0.5073",
"0.6694",
"0.921",
"0.6704",
"0.921",
"0.685",
"0.5073",
"0.6846"
],
[
"0.0768",
"0.681",
"0.4927",
"0.681",
"0.4927",
"0.6943",
"0.0768",
"0.6943"
],
[
"0.509",
"0.6875",
"0.919",
"0.6875",
"0.919",
"0.6978",
"0.509",
"0.6978"
],
[
"0.076",
"0.696",
"0.4927",
"0.696",
"0.4927",
"0.709",
"0.076",
"0.709"
],
[
"0.5063",
"0.7007",
"0.921",
"0.7007",
"0.921",
"0.714",
"0.5063",
"0.714"
],
[
"0.0768",
"0.7104",
"0.4927",
"0.7104",
"0.4927",
"0.7236",
"0.0768",
"0.7236"
],
[
"0.5063",
"0.7153",
"0.919",
"0.7153",
"0.919",
"0.726",
"0.5063",
"0.726"
],
[
"0.0768",
"0.7236",
"0.492",
"0.7236",
"0.492",
"0.7373",
"0.0768",
"0.7373"
],
[
"0.5034",
"0.728",
"0.9224",
"0.729",
"0.9224",
"0.742",
"0.5034",
"0.741"
],
[
"0.0768",
"0.7393",
"0.2443",
"0.7373",
"0.2443",
"0.7505",
"0.0768",
"0.7524"
],
[
"0.25",
"0.7393",
"0.4927",
"0.7393",
"0.4927",
"0.7524",
"0.25",
"0.7524"
],
[
"0.505",
"0.7437",
"0.9224",
"0.7437",
"0.9224",
"0.757",
"0.505",
"0.757"
],
[
"0.0768",
"0.754",
"0.4944",
"0.754",
"0.4944",
"0.7676",
"0.0768",
"0.7676"
],
[
"0.5063",
"0.7583",
"0.9224",
"0.7583",
"0.9224",
"0.7715",
"0.5063",
"0.7715"
],
[
"0.076",
"0.7676",
"0.492",
"0.768",
"0.492",
"0.7827",
"0.076",
"0.782"
],
[
"0.0784",
"0.785",
"0.1487",
"0.785",
"0.1487",
"0.7954",
"0.0784",
"0.7954"
],
[
"0.0989",
"0.7974",
"0.4958",
"0.7974",
"0.4958",
"0.811",
"0.0989",
"0.811"
],
[
"0.509",
"0.8",
"0.523",
"0.8",
"0.523",
"0.808",
"0.509",
"0.808"
],
[
"0.523",
"0.802",
"0.5317",
"0.802",
"0.5317",
"0.8066",
"0.523",
"0.8066"
],
[
"0.536",
"0.797",
"0.864",
"0.797",
"0.864",
"0.811",
"0.536",
"0.811"
],
[
"0.0768",
"0.813",
"0.4958",
"0.813",
"0.4958",
"0.827",
"0.0768",
"0.827"
],
[
"0.5063",
"0.8174",
"0.8164",
"0.8174",
"0.8164",
"0.8306",
"0.5063",
"0.8306"
],
[
"0.1356",
"0.8257",
"0.2361",
"0.8228",
"0.2361",
"0.834",
"0.1356",
"0.837"
],
[
"0.2336",
"0.827",
"0.492",
"0.827",
"0.492",
"0.8403",
"0.2336",
"0.8403"
],
[
"0.0727",
"0.8354",
"0.4958",
"0.837",
"0.4958",
"0.858",
"0.0727",
"0.856"
],
[
"0.0768",
"0.8276",
"0.1405",
"0.8247",
"0.1422",
"0.836",
"0.0784",
"0.839"
],
[
"0.505",
"0.837",
"0.9224",
"0.837",
"0.9224",
"0.851",
"0.505",
"0.851"
],
[
"0.07434",
"0.854",
"0.384",
"0.8516",
"0.384",
"0.869",
"0.07434",
"0.872"
],
[
"0.3857",
"0.856",
"0.4927",
"0.856",
"0.4927",
"0.869",
"0.3857",
"0.869"
],
[
"0.5063",
"0.8516",
"0.9224",
"0.8516",
"0.9224",
"0.8657",
"0.5063",
"0.8657"
],
[
"0.0768",
"0.8706",
"0.4944",
"0.8706",
"0.4944",
"0.884",
"0.0768",
"0.884"
],
[
"0.5024",
"0.8657",
"0.898",
"0.866",
"0.898",
"0.881",
"0.5024",
"0.88"
],
[
"0.07434",
"0.88",
"0.4927",
"0.88",
"0.4927",
"0.901",
"0.07434",
"0.901"
],
[
"0.5293",
"0.881",
"0.921",
"0.882",
"0.921",
"0.896",
"0.5293",
"0.8945"
],
[
"0.0768",
"0.8994",
"0.492",
"0.8994",
"0.492",
"0.9136",
"0.0768",
"0.9136"
],
[
"0.5063",
"0.8965",
"0.9224",
"0.8965",
"0.9224",
"0.91",
"0.5063",
"0.91"
],
[
"0.0768",
"0.914",
"0.4927",
"0.914",
"0.4927",
"0.928",
"0.0768",
"0.928"
],
[
"0.505",
"0.91",
"0.921",
"0.91",
"0.921",
"0.925",
"0.505",
"0.925"
],
[
"0.0768",
"0.928",
"0.4958",
"0.928",
"0.4958",
"0.9424",
"0.0768",
"0.9424"
],
[
"0.505",
"0.928",
"0.8105",
"0.927",
"0.8105",
"0.9414",
"0.505",
"0.9424"
],
[
"0.812",
"0.9287",
"0.921",
"0.9287",
"0.921",
"0.9424",
"0.812",
"0.9424"
]
] | [
"GG",
"",
"Fig.3.Schematicillustrationofthepropo:",
"seddynamicme",
"qepassingcalculation(DMC)module.Thesmallredsguareindicates",
"the receiving node whose message is calculated from its neighbourhood, i.e. the sampled K (e.g. 3 x 3) features nodes. The module",
"accepts a feature map as input and produces its corresponding message map.The symbol * denotes group convolution operation",
"using the dynamically predicted and position specific group kernels and affinities.",
"where the function softmax.() denotes a softmax operation along",
"as in existing graph-based feature learning models [9], [41], [73],",
"the channel axis, which is used to perform a normalisation on",
"[84], we also perform T = 1 iteration of message updating.",
"the estimated affinity A'g., e R'. Wh.",
"E R(G×C+1)x(K×C)",
"and bk,A e R(GxC+1) are matrix transformation parameters. To",
"3.6Discussion",
"reduce the number of the filter parameters, we consider grouped",
"convolutions [10] with a set of G groups split from the total C",
"Our approach is related to deformable convolution [15], [90] and",
"feature channels, and G < C, i.e., each group of C/G feature",
"Non-local [71], but has several key differences:",
"channels shares the same set of filter parameters. The predicted",
"A fundamental difference to deformable convolution is that",
"dynamic filter weights and the affinities are then used in Eq. 5 for",
"it only learns the offset dependent on the input feature while the",
"filter weights are fixed for all inputs. In contrast, our model learns",
"dynamic message calculation.",
"the random walk, weight and affinity as all being dependent on",
"the input. This property makes our weights and affinities position-",
"3.5 Modularinstantiation",
"specific whereas deformable convolution shares the same weight",
"Fig.2and3 showshowourproposeddynamicgraphmessage",
"across all convolution positions in the feature map. Moreover, [15]",
"passingnetwork(DGMN)canbeimplementedinaneuralnetwork.",
"[90] only consider 3 × 3 local neighbours at each convolution",
"The proposed module accepts a single feature map F as input,",
"position. In contrast, our model, for each position, learns to sample",
"which can be derived from any neural network layer. H(o) denotes",
"a set of K nodes (where K 》 9) for message passing globally",
"an initial state of the latent feature map, H, and is initialised with",
"from the whole feature map. This allows our model to capture a",
"F.H and F have the same dimension, i.e., F.H e RHxWxC",
"larger receptive field than deformable convolution.",
"where H, W and C are the height, width and the number of feature",
"Whilst Non-local also learns to refine deep features, it uses a",
"channels of the feature map respectively. We first define a set of",
"self-attention matrix to guide the message passing between each",
"S uniform sampling rates (we show two uniform sampling rates",
"pair offeaturenodes.Incontrast,ourmodel learns tosamplegrapb",
"in Fig. 2 for clarity). The uniform and the random walk sampler",
"feature nodes to capture global feature information efficiently. This",
"sample the nodes from the full graph and return the node indices",
"dynamic sampling reduces computational overhead.whilst still",
"for subsequent dynamic message calculation (DMC) in Fig. 3. The",
"being able to improveupon the accuracy of Non-local across",
"matrixtransformationW",
", to estimate the node random walk in",
"multiple tasks as shown in our experiments (Section 5. We now",
"Eq. 4 is implemented by a 3 × 3 convolution layer [15]. Note that",
"detail the extension of our model to transformer architectures next.",
"other sampling strategies could also be flexibly employed in our",
"framework.",
"The sampled feature nodes are processed along two data",
"4",
"",
"EXTENDINGDGMNTO TRANSFORMERS",
"paths: one for predicting the node-dependant dynamic affini-",
"4.1 Interpretation as a Transformerlayer",
"ERHxWxK",
"and another path for dynamic filters",
"w3 E IRHxwxKxG where K (e.g.,3×3) is the kernel size for the",
"tiesA'q",
"We further instantiate our formulation into an efficient and effective",
"receiving node. The matrix transformation Wk",
"Aused to jointly",
"Transformer layer and present a new Transformer backbone",
"predict the dynamic filters and affinities in Eq 6 is implemented by a",
"network DGMN2 to support various vision downstream tasks.",
"3×3 convolution layer.MessageM,E RixWxC corresponding",
"Specifically,we modify the process of dynamic message",
"to the q-th sampling rate is then scaled to perform a linear",
"passing calculation and implement it as follows: With the predicted",
"combination with the observed feature map F, to produce a refined",
"walk dg, we can obtain a new set of adaptively sampled nodes",
"feature map H(1) as output. To balance performance and efficiency,",
"with graph vertexes V. Given the latent node h(t",
", we can estimate"
] | {
"arxiv_id": "2209.09760",
"language": "en",
"page_idx": 4
} |
||
2204.10290-en-0 | 2204.10290-en-0.jpg | $\textbf{Learning to Revise References}$ for Faithful Summarization
$\textbf{Griffin Adams}^{1* }$ $\textbf{Han- Chin Shing}^2$ $\textbf{Qing Sun}^2$
Christopher Winestock$^2$ $\textbf{Kathleen McKeown}^{1, 2}$ $\text{No\'{e} mie Elhadad}$
{griffin.adams, noemie.elhadad}@columbia.edu
$\{$hanchins, qinsun, winestock, mckeownk}gamazon.com $^{1}$Columbia University, New York, NY$^{2}$Amazon AWS AI, Seattle, WA
## Abstract
Unsupported Reference Sentence
<table>
<tbody>
<tr>
<td> </td>
<td> </td>
<td>She was Thinc and was TPMIPMIN AINP</td>
</tr>
<tr>
<td> </td>
<td> </td>
<td>alinh</td>
</tr>
<tr>
<td> </td>
<td> </td>
<td>Context Alianed ISource</td>
</tr>
</tbody>
</table>
<table>
<tbody>
<tr>
<td> </td>
<td>Startinc COPD flare stahle sternin 1</td>
</tr>
<tr>
<td> </td>
<td>en she arrived on the floor, she was on BiPAP anc had a couqh.</td>
</tr>
</tbody>
</table>
In real-world scenarios with naturally occurring datasets, reference summaries are noisy and may contain information that cannot be inferred from the source text. On large news corpora, removing low quality samples has been shown to reduce model hallucinations. Yet, for smaller, and/or noisier corpora, filtering is detrimental to performance. To improve reference quality while retaining all data, we propose a new approach: to selectively rewrite $un$supported reference sentences to better reflect source data. We automatically generate a synthetic dataset of positive and negative revisions by corrupting supported sentences and learn to revise reference sentences with contrastive learning. The intensity of revisions is treated as a controllable attribute so that, at inference, diverse candidates can be over-generated-then-rescored to balance faithfulness and abstraction. To test our methods, we extract noisy references from publicly available MIMIC-III discharge summaries for the task of hospital-course summarization, and vary the data on which models are trained. According to metrics and human evaluation, models trained on revised clinical references are much more faithful, informative, and fluent than models trained on original or filtered data.
Figure 1: Example of a revised reference sentence. Robitussin and codeine are edited out of the sentence, while cough is correctly kept and a new supported entity BiPAP is added. The model is trained on synthetic data to reconstruct well-supported sentences based on context and diverse model-generated hallucinations.
## 1 Introduction
A common approach to deal with training noise is fltering: to identify and ignore low quality text at the reference (Kang and Hashimoto, 2020; Matsumaru et al., 2020; Nan et al., 2021a; Narayan et al., 2021) or span level (Goyal and Durrett, 2021). Yet, these methods largely work because they are applied to clean, large-scale corpora. For instance, after removing references with “entity hallucinations", Nan et al. (2021a) still have 855k (92% of the original) training examples for Newsroom, 286k (99%) for CNN/DM, 135k (66%) for Xsum.
We consider a noisier, lower resource setting (clinical summarization) and propose a new approach: to revise-not remove-noisy reference content. First, we align each reference sentence to 1-5 sentences in the source text and classify it as supported (to be left alone) or unsupported (to be revised). Our objective is to revise all unsupported reference sentences in such a way that retains faithful content, removes unfaithful content, and, as
The tendency of abstractive systems to produce unfaithful summaries is well-studied (Maynez et al., 2020), yet less attention is paid to the role of the data on which the models are trained. This is problematic for two reasons: (1) many corpora are naturally occurring-not created for training modelsand, as such, are noisy (Kryscinski et al., 2019) and without “inherent quality guarantees”(Bommasani and Cardie, 2020); (2) noisy data is detrimental to training faithful models (Dušek et al., 2019).
$^*$ This project was completed during an NLP research
internship with Amazon Comprehend Medical. | [
[
"0.3435",
"0.09143",
"0.656",
"0.09143",
"0.656",
"0.1057",
"0.3435",
"0.1057"
],
[
"0.357",
"0.111",
"0.642",
"0.111",
"0.642",
"0.1259",
"0.357",
"0.1259"
],
[
"0.3015",
"0.1414",
"0.615",
"0.1426",
"0.615",
"0.1603",
"0.3015",
"0.1592"
],
[
"0.6304",
"0.1437",
"0.7246",
"0.1426",
"0.7246",
"0.158",
"0.6304",
"0.1592"
],
[
"0.2141",
"0.1609",
"0.4307",
"0.1609",
"0.4307",
"0.1758",
"0.2141",
"0.1758"
],
[
"0.434",
"0.1603",
"0.6465",
"0.158",
"0.6465",
"0.1758",
"0.434",
"0.1776"
],
[
"0.656",
"0.1609",
"0.8076",
"0.1609",
"0.8076",
"0.1746",
"0.656",
"0.1746"
],
[
"0.6465",
"0.1663",
"0.6543",
"0.1663",
"0.6543",
"0.1704",
"0.6465",
"0.1704"
],
[
"0.4248",
"0.1633",
"0.4417",
"0.1633",
"0.4417",
"0.1716",
"0.4248",
"0.1716"
],
[
"0.2377",
"0.1776",
"0.769",
"0.1776",
"0.769",
"0.1924",
"0.2377",
"0.1924"
],
[
"0.2007",
"0.1936",
"0.804",
"0.1953",
"0.804",
"0.2102",
"0.2007",
"0.2079"
],
[
"0.2183",
"0.209",
"0.806",
"0.209",
"0.806",
"0.2268",
"0.2183",
"0.2268"
],
[
"0.2627",
"0.253",
"0.3418",
"0.253",
"0.3418",
"0.2673",
"0.2627",
"0.2673"
],
[
"0.567",
"0.2505",
"0.804",
"0.2505",
"0.804",
"0.2642",
"0.567",
"0.2642"
],
[
"0.55",
"0.2666",
"0.843",
"0.2666",
"0.843",
"0.2798",
"0.55",
"0.2798"
],
[
"0.5264",
"0.2778",
"0.864",
"0.279",
"0.864",
"0.2903",
"0.5264",
"0.2893"
],
[
"0.1436",
"0.284",
"0.461",
"0.284",
"0.461",
"0.2976",
"0.1436",
"0.2976"
],
[
"0.5786",
"0.2903",
"0.812",
"0.2903",
"0.812",
"0.304",
"0.5786",
"0.304"
],
[
"0.1453",
"0.2998",
"0.4568",
"0.2998",
"0.4568",
"0.31",
"0.1453",
"0.31"
],
[
"0.02351",
"0.31",
"0.06046",
"0.31",
"0.06046",
"0.4382",
"0.02351",
"0.4382"
],
[
"0.614",
"0.306",
"0.7827",
"0.306",
"0.7827",
"0.3164",
"0.614",
"0.3164"
],
[
"0.147",
"0.3135",
"0.4575",
"0.3135",
"0.4575",
"0.3237",
"0.147",
"0.3237"
],
[
"0.1453",
"0.326",
"0.461",
"0.326",
"0.461",
"0.339",
"0.1453",
"0.339"
],
[
"0.55",
"0.3206",
"0.843",
"0.3206",
"0.843",
"0.3308",
"0.55",
"0.3308"
],
[
"0.531",
"0.3308",
"0.8623",
"0.3308",
"0.8623",
"0.3445",
"0.531",
"0.3445"
],
[
"0.1453",
"0.3403",
"0.4592",
"0.3403",
"0.4592",
"0.354",
"0.1453",
"0.354"
],
[
"0.5205",
"0.3457",
"0.8716",
"0.3457",
"0.8716",
"0.3557",
"0.5205",
"0.3557"
],
[
"0.1436",
"0.354",
"0.4592",
"0.3552",
"0.4592",
"0.3682",
"0.1436",
"0.3677"
],
[
"0.5283",
"0.3552",
"0.864",
"0.3552",
"0.864",
"0.3682",
"0.5283",
"0.3682"
],
[
"0.1453",
"0.3682",
"0.4592",
"0.3682",
"0.4592",
"0.3818",
"0.1453",
"0.3818"
],
[
"0.5225",
"0.3682",
"0.869",
"0.3682",
"0.869",
"0.3818",
"0.5225",
"0.3818"
],
[
"0.1453",
"0.3818",
"0.367",
"0.3818",
"0.367",
"0.3955",
"0.1453",
"0.3955"
],
[
"0.372",
"0.3843",
"0.4575",
"0.3843",
"0.4575",
"0.3955",
"0.372",
"0.3955"
],
[
"0.529",
"0.3813",
"0.8657",
"0.3813",
"0.8657",
"0.3943",
"0.529",
"0.3943"
],
[
"0.1436",
"0.398",
"0.4592",
"0.398",
"0.4592",
"0.411",
"0.1436",
"0.411"
],
[
"0.5264",
"0.3938",
"0.867",
"0.3938",
"0.867",
"0.4067",
"0.5264",
"0.4067"
],
[
"0.517",
"0.4062",
"0.8657",
"0.405",
"0.8657",
"0.4187",
"0.517",
"0.4192"
],
[
"0.1436",
"0.412",
"0.4592",
"0.411",
"0.4592",
"0.4246",
"0.1436",
"0.4258"
],
[
"0.5635",
"0.4192",
"0.8296",
"0.4192",
"0.8296",
"0.4329",
"0.5635",
"0.4329"
],
[
"0.0336",
"0.4312",
"0.05038",
"0.4312",
"0.05038",
"0.4382",
"0.0336",
"0.4382"
],
[
"0.1453",
"0.4258",
"0.4592",
"0.4258",
"0.4592",
"0.4395",
"0.1453",
"0.4395"
],
[
"0.02184",
"0.437",
"0.06464",
"0.437",
"0.06464",
"0.5195",
"0.02184",
"0.5195"
],
[
"0.1453",
"0.4407",
"0.4592",
"0.4407",
"0.4592",
"0.4536",
"0.1453",
"0.4536"
],
[
"0.5156",
"0.4353",
"0.875",
"0.4353",
"0.875",
"0.449",
"0.5156",
"0.449"
],
[
"0.1436",
"0.4531",
"0.4592",
"0.4548",
"0.4592",
"0.4685",
"0.1436",
"0.466"
],
[
"0.6514",
"0.449",
"0.739",
"0.449",
"0.739",
"0.459",
"0.6514",
"0.459"
],
[
"0.1453",
"0.4685",
"0.4592",
"0.4685",
"0.4592",
"0.4822",
"0.1453",
"0.4822"
],
[
"0.5146",
"0.4756",
"0.882",
"0.4756",
"0.882",
"0.4863",
"0.5146",
"0.4863"
],
[
"0.147",
"0.484",
"0.4568",
"0.484",
"0.4568",
"0.4946",
"0.147",
"0.4946"
],
[
"0.511",
"0.4868",
"0.885",
"0.4893",
"0.885",
"0.503",
"0.511",
"0.5005"
],
[
"0.1453",
"0.4963",
"0.4575",
"0.4963",
"0.4575",
"0.5103",
"0.1453",
"0.5103"
],
[
"0.51",
"0.502",
"0.8833",
"0.503",
"0.8833",
"0.5166",
"0.51",
"0.5156"
],
[
"0.147",
"0.5127",
"0.4568",
"0.5127",
"0.4568",
"0.5225",
"0.147",
"0.5225"
],
[
"0.02184",
"0.5195",
"0.06213",
"0.5195",
"0.06213",
"0.73",
"0.02184",
"0.73"
],
[
"0.511",
"0.517",
"0.882",
"0.517",
"0.882",
"0.531",
"0.511",
"0.531"
],
[
"0.1453",
"0.525",
"0.4575",
"0.525",
"0.4575",
"0.538",
"0.1453",
"0.538"
],
[
"0.511",
"0.532",
"0.882",
"0.532",
"0.882",
"0.546",
"0.511",
"0.546"
],
[
"0.1453",
"0.539",
"0.461",
"0.539",
"0.461",
"0.5527",
"0.1453",
"0.5527"
],
[
"0.513",
"0.5474",
"0.8613",
"0.5474",
"0.8613",
"0.558",
"0.513",
"0.558"
],
[
"0.1453",
"0.554",
"0.3225",
"0.554",
"0.3225",
"0.568",
"0.1453",
"0.568"
],
[
"0.3198",
"0.5586",
"0.3267",
"0.5586",
"0.3267",
"0.5625",
"0.3198",
"0.5625"
],
[
"0.3257",
"0.5547",
"0.4575",
"0.5547",
"0.4575",
"0.5654",
"0.3257",
"0.5654"
],
[
"0.1436",
"0.568",
"0.4575",
"0.568",
"0.4575",
"0.5806",
"0.1436",
"0.5806"
],
[
"0.1453",
"0.582",
"0.4592",
"0.582",
"0.4592",
"0.5957",
"0.1453",
"0.5957"
],
[
"0.1453",
"0.5967",
"0.4592",
"0.5967",
"0.4592",
"0.61",
"0.1453",
"0.61"
],
[
"0.531",
"0.594",
"0.882",
"0.594",
"0.882",
"0.6094",
"0.531",
"0.6094"
],
[
"0.147",
"0.6123",
"0.4575",
"0.6123",
"0.4575",
"0.6226",
"0.147",
"0.6226"
],
[
"0.511",
"0.611",
"0.882",
"0.611",
"0.882",
"0.6245",
"0.511",
"0.6245"
],
[
"0.1453",
"0.626",
"0.4575",
"0.626",
"0.4575",
"0.6357",
"0.1453",
"0.6357"
],
[
"0.511",
"0.627",
"0.882",
"0.627",
"0.882",
"0.64",
"0.511",
"0.64"
],
[
"0.1436",
"0.638",
"0.4592",
"0.638",
"0.4592",
"0.652",
"0.1436",
"0.652"
],
[
"0.511",
"0.6426",
"0.882",
"0.6426",
"0.882",
"0.6562",
"0.511",
"0.6562"
],
[
"0.147",
"0.654",
"0.4575",
"0.654",
"0.4575",
"0.6646",
"0.147",
"0.6646"
],
[
"0.511",
"0.659",
"0.882",
"0.659",
"0.882",
"0.673",
"0.511",
"0.673"
],
[
"0.1436",
"0.6675",
"0.4592",
"0.6675",
"0.4592",
"0.681",
"0.1436",
"0.681"
],
[
"0.511",
"0.676",
"0.882",
"0.676",
"0.882",
"0.6895",
"0.511",
"0.6895"
],
[
"0.511",
"0.6914",
"0.8833",
"0.6914",
"0.8833",
"0.706",
"0.511",
"0.706"
],
[
"0.1184",
"0.6997",
"0.1318",
"0.6997",
"0.1318",
"0.708",
"0.1184",
"0.708"
],
[
"0.1351",
"0.7017",
"0.1427",
"0.7017",
"0.1427",
"0.707",
"0.1351",
"0.707"
],
[
"0.1453",
"0.6987",
"0.2603",
"0.6987",
"0.2603",
"0.7114",
"0.1453",
"0.7114"
],
[
"0.511",
"0.708",
"0.8833",
"0.708",
"0.8833",
"0.7217",
"0.511",
"0.7217"
],
[
"0.11755",
"0.7236",
"0.4937",
"0.7236",
"0.4937",
"0.737",
"0.11755",
"0.737"
],
[
"0.511",
"0.7236",
"0.8833",
"0.7236",
"0.8833",
"0.737",
"0.511",
"0.737"
],
[
"0.11755",
"0.7407",
"0.4902",
"0.7407",
"0.4902",
"0.7544",
"0.11755",
"0.7544"
],
[
"0.511",
"0.7407",
"0.8833",
"0.7407",
"0.8833",
"0.7544",
"0.511",
"0.7544"
],
[
"0.11584",
"0.755",
"0.4895",
"0.756",
"0.4895",
"0.771",
"0.11584",
"0.7695"
],
[
"0.51",
"0.7544",
"0.879",
"0.755",
"0.879",
"0.7695",
"0.51",
"0.7686"
],
[
"0.11755",
"0.7725",
"0.4895",
"0.7725",
"0.4895",
"0.786",
"0.11755",
"0.786"
],
[
"0.529",
"0.774",
"0.882",
"0.7754",
"0.882",
"0.7905",
"0.529",
"0.788"
],
[
"0.11755",
"0.789",
"0.4902",
"0.789",
"0.4902",
"0.8027",
"0.11755",
"0.8027"
],
[
"0.509",
"0.7915",
"0.882",
"0.7915",
"0.882",
"0.8047",
"0.509",
"0.8047"
],
[
"0.11584",
"0.8047",
"0.4902",
"0.8047",
"0.4902",
"0.8184",
"0.11584",
"0.8184"
],
[
"0.511",
"0.808",
"0.885",
"0.808",
"0.885",
"0.822",
"0.511",
"0.822"
],
[
"0.11584",
"0.821",
"0.4878",
"0.821",
"0.4878",
"0.8345",
"0.11584",
"0.8345"
],
[
"0.51",
"0.8247",
"0.882",
"0.8247",
"0.882",
"0.8384",
"0.51",
"0.8384"
],
[
"0.11584",
"0.8374",
"0.4878",
"0.8374",
"0.4878",
"0.851",
"0.11584",
"0.851"
],
[
"0.511",
"0.843",
"0.881",
"0.843",
"0.881",
"0.8525",
"0.511",
"0.8525"
],
[
"0.11755",
"0.8525",
"0.4878",
"0.8525",
"0.4878",
"0.866",
"0.11755",
"0.866"
],
[
"0.51",
"0.857",
"0.882",
"0.857",
"0.882",
"0.8706",
"0.51",
"0.8706"
],
[
"0.11755",
"0.8706",
"0.4434",
"0.8706",
"0.4434",
"0.8813",
"0.11755",
"0.8813"
],
[
"0.51",
"0.8726",
"0.882",
"0.8726",
"0.882",
"0.8857",
"0.51",
"0.8857"
],
[
"0.1436",
"0.8936",
"0.4895",
"0.8936",
"0.4895",
"0.9067",
"0.1436",
"0.9067"
],
[
"0.51",
"0.8896",
"0.8833",
"0.8896",
"0.8833",
"0.903",
"0.51",
"0.903"
],
[
"0.1142",
"0.904",
"0.4048",
"0.905",
"0.4048",
"0.9194",
"0.1142",
"0.9185"
],
[
"0.511",
"0.905",
"0.882",
"0.905",
"0.882",
"0.9185",
"0.511",
"0.9185"
]
] | [
"Learning to Revise References",
"for Faithful Summarization",
"Griffin Adams1* Han-Chin Shing?",
"Qing Sun?",
"Christopher Winestock?",
"Kathleen McKeownl,2",
"Noémie Elhadad",
"",
"",
"(griffin.adams, noemie.elhadad)@columbia.edu",
"(hanchins,qinsun, winestock,mckeownk)@amazon.com",
"1Columbia University, New York, NY2Amazon AWS AI, Seattle, WA",
"Abstract",
"Unsupported Reference Sentence",
"She was coughing frequently and was given",
"Robitussin DM as well as Robitussin with codeine",
"In real-world scenarios with naturally occur-",
"for symptomatic reliefof her cough.",
"ring datasets, reference summaries are noisy",
"",
"AlignedSourceContext",
"and may contain informationthat cannotbe in",
"ferred from the source text. On large news cor-",
"Startingsteroidtaper..A.COPD flare...stable",
"24hrs...sats 88-92 P.Antibiotics..steriods.ready for",
"pora, removing low quality samples has been",
"transfertofloor.Presented lasteveto EWwhoputher",
"shown to reduce model hallucinations. Yet,",
"on BIPAP,but transferred her east off of it.SATS in",
"for smaller, and/or noisier corpora, filtering",
"low 90'S.BS'S diminished throughout.+OcC cough.",
"is detrimental to performance.",
"To improve",
"When she arrived here, she told us that she did not",
"reference quality while retaining all data, we",
"feel that shee needed BIPAP any longer, so she has",
"gone all night on 4 liters NC and has done very well.",
"propose a new approach: to selectively re-",
"Supported Model-Generated Revision",
"",
"write unsupported reference sentences to bet-",
"[T']",
"ter reflect source data. We automatically gen-",
"When she arrived on the floor, she was on BiPAP and",
"erate a synthetic dataset of positive and neg-",
"had acough",
"ative revisions by corrupting supported sen-",
"Figure l:Example of arevised reference sentence.Ro-",
"tences and learn to revise reference sentences",
"bitussin and codeine are edited out of the sentence,",
"with contrastive learning. The intensity of re-",
"while cough is correctly kept and a new supported en-",
"visions is treated as a controllable attribute so",
"061102:11",
"tity BiPAP is added. The model is trained on synthetic",
"that, at inference, diverse candidates can be",
"data to reconstruct well-supported sentences based on",
"over-generated-then-rescored to balance faith-",
"contextand diverse model-generated hallucinations.",
"fulness and abstraction.",
"",
"To test our meth",
"ods, we extract noisy references from publicly",
"available MIMIC-III discharge summaries for",
"the task of hospital-course summarization, and",
"A common approach to deal with training noise",
"vary the data on which models are trained. Ac",
"is filtering: to identify and ignore low quality text",
"cordingtometricsandhumanevaluation,mod",
"at the reference (Kang and Hashimoto, 2020; Mat",
"els trained on revised clinical references are",
"sumaru et al., 2020; Nan et al., 2021a; Narayan",
"much more faithful,informative, and fluent",
"et al., 2021) or span level (Goyal and Durrett, 2021)",
"than models trained on original or filtered data.",
"Yet, these methods largely work because they are",
"applied to clean, large-scale corpora. For instance,",
"1",
"",
"Introduction",
"after removing references with “entity hallucina-",
"The tendency of abstractive systems to produce un-",
"tions\", Nan et al. (2021a) still have 855k (92%",
"faithful summaries is well-studied (Maynez et al.,",
"of the original) training examples for Newsroom,",
"2020), yet less attention is paid to the role of the",
"286k (99%) for CNN/DM, 135k (66%) for Xsum.",
"data on which the models are trained. This is prob-",
"We consider a noisier, lower resource setting",
"lematic for two reasons: (1) many corpora are nat-",
"(clinical summarization) and propose a new ap-",
"urally occurring-not created for training models--",
"proach: to revise-not remove-noisy reference con-",
"and, as such, are noisy (Kryscinski et al., 2019) and",
"tent. First, we align each reference sentence to",
"without “inherent quality guarantees\" (Bommasani",
"1-5sentencesinthesourcetextandclassifyitas",
"and Cardie, 2020); (2) noisy data is detrimental to",
"supported (to be left alone) or unsupported (to be",
"trainingfaithful models (Dusek et al.,2019)",
"revised). Our objective is to revise all unsupported",
"* This project was completed during an NLP research",
"reference sentences in such a way that retains faith-",
"internship with Amazon Comprehend Medical.",
"ful content, removes unfaithful content, and, as"
] | {
"arxiv_id": "2204.10290",
"language": "en",
"page_idx": 0
} |
||
2204.10290-en-1 | 2204.10290-en-1.jpg | needed to preserve length, adds relevant context. An example output is shown in Figure 1. In a coherent sentence, the model removes unsupported entities (Robitussin, codeine) and introduces a twice mentioned concept from the context (BiPAP).
To learn this revision task, we need examples of supported and unsupported reference sentences for the same context. Without directly observing it, we generate synthetic data. At a high-level, we take each supported sentence, corrupt it to form a diverse set of unsupported alternatives, and use this mix of real and synthetic data to create examples of (un)faithful revisions for contrastive learning.
As a test case, we consider a task of real-world significance-summarizing a hospital admissionand extract a corpus from a noisy source-notes from the Electronic Health Record (EHR). We experiment with the publicly available MIMIC-III dataset (Johnson et al., 2016). As in Adams et al. (2021), we treat the Brief Hospital Course (BHC) section of the discharge summary as a reference summary and all notes prior to discharge as source. Data coverage is a huge issue as only 60% of reference summary entities can be found in the source.
The contributions of this work are: (1) Proposing a new method to address variable reference quality: reference revision, which, as a data preprocessing step, is model agnostic and complementary to other faithfulness approaches; (2) Showing that training on revised references can improve faithfulness while also improving informativeness and fluency; (3) Providing code$^{1}$, pre-processed datasets, and models for alignment, corruption, revision, post-hoc editing, and generation of clinical summaries$^2;$ (4) Beyond its primary use case of data pre-processing, demonstrating that reference revision can have standalone value: as a post-hoc editor and a pre-training objective for faithfulness.
## Related Work
Faithfulness. Efforts to address faithfulness have focused on smarter models (Huang et al., 2020), more suitable metrics (Durmus et al.,2020), content-plan editing (Narayan et al., 2021), and post-hoc interventions: ranking (Falke et al., 2019) and editing (Cao et al., 2020; Dong et al., 2020; Zhu et al., 2021). Synthetic errors (Kryscinski
$^{\mathrm{l}}$https://github.com/amazon-research/
summary-reference-revision
The datasets are accessible via PhysioNet (Goldberger et al., 2000) and models via HuggingFace (Wolf et al., 2020). Please refer to our GitHub README for further details.
et al., 2020) are useful for optimizing (Cao and Wang, 2021) and evaluating (Goyal and Durrett, 2020) faithfulness, yet are best supplemented with fine-grained annotations (Goyal and Durrett, 2021). The impact of training data noise on faithfulness
is less studied. The most common proposal is to identify low quality samples-with entailment (Matsumaru et al., 2020; Goyal and Durrett, 2021) or entity overlap (Nan et al., 2021a; Narayan et al., 2021)-and drop them. These filtering methods tend to improve faithfulness yet can degrade informativeness. Kang and Hashimoto (2020) address the data hunger issue by first training on all samples before implementing Loss Truncation-ignoring high $\log\log$loss datapoints. This is effective on a relatively clean Gigaword corpus yet untested on noisier corpora in which hallucination behavior from the unfiltered data may be difficult to unlearn. Filtering can be“insufficient to train a competitive model” when high-quality data is limited (Filippova, 2020). Our proposed method takes advantage of all available data while seeking to redress the underlying issue.
Clinical Summarization. Faithfulness is less studied for clinical summarization because most proposed methods are extractive (Pivovarov and Elhadad, 2015; Moen et al., 2016; Alsentzer and Kim, 2018). Abstractive approaches tend to focus on finer temporal granularities, e.g., synthesizing a single radiology report (MacAvaney et al., 2019; Sotudeh Gharebagh et al., 2020; Zhang et al., 2020b) and doctor-patient conversations (Krishna et al., 2020; Joshi et al., 2020; Zhang et al., 2021a).
Most similar, Shing et al. (2021) survey extractthen-abstract approaches to section-specific discharge summarization on MIMIC-III. They measure factuality with entity overlap and remove poorly supported references. When analyzing a proprietary EHR-derived, hospital-course summarization corpus, Adams et al. (2021) perform oracle extractive analysis and confirm, as we do, that EHR-derived summary references are highly noisy.
## 2 Data
The publicly available MIMIC-III dataset contains de-identified clinical records from patients admitted to Beth Israel Deaconess Medical Center (Johnson et al., 2016). Hospital-admission summarization is a challenging task: in at most a paragraph, a summary must discuss what happened to the patient during the admission, why it happened, and what needs to happen next. To create a dataset for | [
[
"0.11755",
"0.0861",
"0.4963",
"0.0861",
"0.4963",
"0.0998",
"0.11755",
"0.0998"
],
[
"0.509",
"0.08435",
"0.882",
"0.0849",
"0.882",
"0.0998",
"0.509",
"0.0986"
],
[
"0.4954",
"0.09143",
"0.5054",
"0.09143",
"0.5054",
"0.0944",
"0.4954",
"0.0944"
],
[
"0.1184",
"0.10394",
"0.4895",
"0.10394",
"0.4895",
"0.114",
"0.1184",
"0.114"
],
[
"0.509",
"0.10156",
"0.881",
"0.10156",
"0.881",
"0.1152",
"0.509",
"0.1152"
],
[
"0.11584",
"0.11816",
"0.4895",
"0.11816",
"0.4895",
"0.1318",
"0.11584",
"0.1318"
],
[
"0.51",
"0.11816",
"0.881",
"0.11816",
"0.881",
"0.1318",
"0.51",
"0.1318"
],
[
"0.11755",
"0.1342",
"0.4895",
"0.1342",
"0.4895",
"0.1478",
"0.11755",
"0.1478"
],
[
"0.513",
"0.1342",
"0.882",
"0.1342",
"0.882",
"0.1478",
"0.513",
"0.1478"
],
[
"0.11584",
"0.1497",
"0.4534",
"0.1497",
"0.4534",
"0.1633",
"0.11584",
"0.1633"
],
[
"0.529",
"0.1484",
"0.882",
"0.1509",
"0.882",
"0.1644",
"0.529",
"0.1621"
],
[
"0.1368",
"0.1663",
"0.4878",
"0.1663",
"0.4878",
"0.1799",
"0.1368",
"0.1799"
],
[
"0.511",
"0.1663",
"0.882",
"0.1663",
"0.882",
"0.1799",
"0.511",
"0.1799"
],
[
"0.11755",
"0.1829",
"0.4895",
"0.1829",
"0.4895",
"0.1965",
"0.11755",
"0.1965"
],
[
"0.513",
"0.1841",
"0.881",
"0.1841",
"0.881",
"0.1948",
"0.513",
"0.1948"
],
[
"0.11584",
"0.1978",
"0.4878",
"0.199",
"0.4878",
"0.2131",
"0.11584",
"0.212"
],
[
"0.51",
"0.199",
"0.8833",
"0.199",
"0.8833",
"0.212",
"0.51",
"0.212"
],
[
"0.11584",
"0.2156",
"0.4878",
"0.2156",
"0.4878",
"0.2286",
"0.11584",
"0.2286"
],
[
"0.511",
"0.2144",
"0.8833",
"0.2144",
"0.8833",
"0.228",
"0.511",
"0.228"
],
[
"0.11755",
"0.2322",
"0.4895",
"0.2322",
"0.4895",
"0.2452",
"0.11755",
"0.2452"
],
[
"0.51",
"0.231",
"0.882",
"0.231",
"0.882",
"0.2446",
"0.51",
"0.2446"
],
[
"0.11584",
"0.2477",
"0.4895",
"0.2477",
"0.4895",
"0.2612",
"0.11584",
"0.2612"
],
[
"0.509",
"0.2477",
"0.8833",
"0.2477",
"0.8833",
"0.2612",
"0.509",
"0.2612"
],
[
"0.1142",
"0.2625",
"0.4895",
"0.2625",
"0.4895",
"0.2756",
"0.1142",
"0.2756"
],
[
"0.5073",
"0.2625",
"0.881",
"0.2625",
"0.881",
"0.2756",
"0.5073",
"0.2756"
],
[
"0.11584",
"0.2778",
"0.4727",
"0.279",
"0.4727",
"0.2935",
"0.11584",
"0.2922"
],
[
"0.513",
"0.279",
"0.882",
"0.279",
"0.882",
"0.2935",
"0.513",
"0.2935"
],
[
"0.1377",
"0.2976",
"0.4878",
"0.2976",
"0.4878",
"0.308",
"0.1377",
"0.308"
],
[
"0.513",
"0.2957",
"0.882",
"0.2957",
"0.882",
"0.3088",
"0.513",
"0.3088"
],
[
"0.11755",
"0.3123",
"0.492",
"0.3123",
"0.492",
"0.326",
"0.11755",
"0.326"
],
[
"0.513",
"0.3113",
"0.882",
"0.3113",
"0.882",
"0.3247",
"0.513",
"0.3247"
],
[
"0.1184",
"0.33",
"0.4895",
"0.33",
"0.4895",
"0.3403",
"0.1184",
"0.3403"
],
[
"0.513",
"0.328",
"0.882",
"0.328",
"0.882",
"0.3416",
"0.513",
"0.3416"
],
[
"0.11755",
"0.3433",
"0.4902",
"0.3433",
"0.4902",
"0.357",
"0.11755",
"0.357"
],
[
"0.513",
"0.3433",
"0.882",
"0.3433",
"0.882",
"0.357",
"0.513",
"0.357"
],
[
"0.11755",
"0.3599",
"0.486",
"0.3599",
"0.486",
"0.3735",
"0.11755",
"0.3735"
],
[
"0.511",
"0.3594",
"0.882",
"0.3594",
"0.882",
"0.3723",
"0.511",
"0.3723"
],
[
"0.11584",
"0.3748",
"0.486",
"0.3748",
"0.486",
"0.3884",
"0.11584",
"0.3884"
],
[
"0.51",
"0.376",
"0.882",
"0.376",
"0.882",
"0.3896",
"0.51",
"0.3896"
],
[
"0.1184",
"0.3914",
"0.4895",
"0.3914",
"0.4895",
"0.405",
"0.1184",
"0.405"
],
[
"0.51",
"0.3914",
"0.882",
"0.3914",
"0.882",
"0.405",
"0.51",
"0.405"
],
[
"0.11755",
"0.408",
"0.4878",
"0.408",
"0.4878",
"0.4216",
"0.11755",
"0.4216"
],
[
"0.511",
"0.408",
"0.882",
"0.4067",
"0.882",
"0.4216",
"0.511",
"0.4229"
],
[
"0.11584",
"0.4246",
"0.4895",
"0.4246",
"0.4895",
"0.4382",
"0.11584",
"0.4382"
],
[
"0.513",
"0.4246",
"0.882",
"0.4246",
"0.882",
"0.4382",
"0.513",
"0.4382"
],
[
"0.11755",
"0.4407",
"0.4895",
"0.4407",
"0.4895",
"0.4536",
"0.11755",
"0.4536"
],
[
"0.513",
"0.449",
"0.882",
"0.449",
"0.882",
"0.462",
"0.513",
"0.462"
],
[
"0.11584",
"0.4573",
"0.4844",
"0.4573",
"0.4844",
"0.4702",
"0.11584",
"0.4702"
],
[
"0.513",
"0.4656",
"0.882",
"0.4656",
"0.882",
"0.4785",
"0.513",
"0.4785"
],
[
"0.1351",
"0.4727",
"0.4902",
"0.4727",
"0.4902",
"0.4863",
"0.1351",
"0.4863"
],
[
"0.511",
"0.481",
"0.882",
"0.4797",
"0.882",
"0.4946",
"0.511",
"0.4954"
],
[
"0.1142",
"0.4893",
"0.4878",
"0.4893",
"0.4878",
"0.503",
"0.1142",
"0.503"
],
[
"0.511",
"0.4954",
"0.882",
"0.4963",
"0.882",
"0.5103",
"0.511",
"0.509"
],
[
"0.1142",
"0.505",
"0.4895",
"0.505",
"0.4895",
"0.5186",
"0.1142",
"0.5186"
],
[
"0.513",
"0.513",
"0.8833",
"0.513",
"0.8833",
"0.527",
"0.513",
"0.527"
],
[
"0.11584",
"0.521",
"0.4902",
"0.5195",
"0.4902",
"0.5337",
"0.11584",
"0.535"
],
[
"0.51",
"0.5293",
"0.882",
"0.5293",
"0.882",
"0.542",
"0.51",
"0.542"
],
[
"0.11755",
"0.5376",
"0.4878",
"0.5376",
"0.4878",
"0.5503",
"0.11755",
"0.5503"
],
[
"0.509",
"0.5444",
"0.882",
"0.5444",
"0.882",
"0.558",
"0.509",
"0.558"
],
[
"0.11755",
"0.5527",
"0.4878",
"0.5527",
"0.4878",
"0.568",
"0.11755",
"0.568"
],
[
"0.513",
"0.561",
"0.8833",
"0.561",
"0.8833",
"0.5747",
"0.513",
"0.5747"
],
[
"0.1184",
"0.5693",
"0.4878",
"0.5693",
"0.4878",
"0.583",
"0.1184",
"0.583"
],
[
"0.513",
"0.5776",
"0.882",
"0.5776",
"0.882",
"0.5913",
"0.513",
"0.5913"
],
[
"0.11584",
"0.5845",
"0.4878",
"0.585",
"0.4878",
"0.5996",
"0.11584",
"0.5986"
],
[
"0.513",
"0.5933",
"0.882",
"0.5933",
"0.882",
"0.607",
"0.513",
"0.607"
],
[
"0.11755",
"0.6016",
"0.4878",
"0.6016",
"0.4878",
"0.615",
"0.11755",
"0.615"
],
[
"0.529",
"0.61",
"0.8833",
"0.61",
"0.8833",
"0.6235",
"0.529",
"0.6235"
],
[
"0.11584",
"0.6177",
"0.4844",
"0.616",
"0.4844",
"0.6313",
"0.11584",
"0.632"
],
[
"0.511",
"0.627",
"0.8833",
"0.627",
"0.8833",
"0.64",
"0.511",
"0.64"
],
[
"0.1184",
"0.6353",
"0.486",
"0.6353",
"0.486",
"0.6455",
"0.1184",
"0.6455"
],
[
"0.513",
"0.641",
"0.882",
"0.641",
"0.882",
"0.655",
"0.513",
"0.655"
],
[
"0.11755",
"0.6494",
"0.4878",
"0.6494",
"0.4878",
"0.663",
"0.11755",
"0.663"
],
[
"0.513",
"0.6577",
"0.882",
"0.6577",
"0.882",
"0.6714",
"0.513",
"0.6714"
],
[
"0.11755",
"0.6665",
"0.4878",
"0.6665",
"0.4878",
"0.6797",
"0.11755",
"0.6797"
],
[
"0.513",
"0.675",
"0.882",
"0.675",
"0.882",
"0.6885",
"0.513",
"0.6885"
],
[
"0.11584",
"0.6816",
"0.4895",
"0.6816",
"0.4895",
"0.6953",
"0.11584",
"0.6953"
],
[
"0.513",
"0.6904",
"0.882",
"0.6904",
"0.882",
"0.7036",
"0.513",
"0.7036"
],
[
"0.11755",
"0.7104",
"0.241",
"0.7104",
"0.241",
"0.7236",
"0.11755",
"0.7236"
],
[
"0.511",
"0.706",
"0.882",
"0.706",
"0.882",
"0.7197",
"0.511",
"0.7197"
],
[
"0.51",
"0.7217",
"0.882",
"0.7227",
"0.882",
"0.737",
"0.51",
"0.7363"
],
[
"0.11755",
"0.7354",
"0.4878",
"0.7354",
"0.4878",
"0.749",
"0.11755",
"0.749"
],
[
"0.513",
"0.7393",
"0.882",
"0.7393",
"0.882",
"0.753",
"0.513",
"0.753"
],
[
"0.11755",
"0.752",
"0.4878",
"0.752",
"0.4878",
"0.7656",
"0.11755",
"0.7656"
],
[
"0.11755",
"0.7686",
"0.4895",
"0.7686",
"0.4895",
"0.782",
"0.11755",
"0.782"
],
[
"0.5156",
"0.7686",
"0.5366",
"0.7686",
"0.5366",
"0.777",
"0.5156",
"0.777"
],
[
"0.538",
"0.764",
"0.588",
"0.7666",
"0.586",
"0.781",
"0.5366",
"0.7793"
],
[
"0.11584",
"0.7837",
"0.4878",
"0.783",
"0.4878",
"0.7974",
"0.11584",
"0.799"
],
[
"0.51",
"0.7915",
"0.882",
"0.7915",
"0.882",
"0.8047",
"0.51",
"0.8047"
],
[
"0.11755",
"0.8003",
"0.4878",
"0.8003",
"0.4878",
"0.814",
"0.11755",
"0.814"
],
[
"0.513",
"0.808",
"0.8833",
"0.808",
"0.8833",
"0.822",
"0.513",
"0.822"
],
[
"0.11755",
"0.8164",
"0.4878",
"0.8164",
"0.4878",
"0.83",
"0.11755",
"0.83"
],
[
"0.5146",
"0.826",
"0.8833",
"0.826",
"0.8833",
"0.836",
"0.5146",
"0.836"
],
[
"0.11584",
"0.832",
"0.486",
"0.832",
"0.486",
"0.8467",
"0.11584",
"0.8467"
],
[
"0.513",
"0.8403",
"0.8833",
"0.8403",
"0.8833",
"0.854",
"0.513",
"0.854"
],
[
"0.1368",
"0.857",
"0.4626",
"0.857",
"0.4626",
"0.8706",
"0.1368",
"0.8706"
],
[
"0.511",
"0.857",
"0.8833",
"0.857",
"0.8833",
"0.8706",
"0.511",
"0.8706"
],
[
"0.11755",
"0.8706",
"0.357",
"0.8706",
"0.357",
"0.8813",
"0.11755",
"0.8813"
],
[
"0.51",
"0.8726",
"0.882",
"0.8726",
"0.882",
"0.8857",
"0.51",
"0.8857"
],
[
"0.1351",
"0.879",
"0.486",
"0.88",
"0.486",
"0.8945",
"0.1351",
"0.8936"
],
[
"0.11755",
"0.8936",
"0.4895",
"0.8936",
"0.4895",
"0.9067",
"0.11755",
"0.9067"
],
[
"0.51",
"0.8896",
"0.882",
"0.8896",
"0.882",
"0.903",
"0.51",
"0.903"
],
[
"0.11584",
"0.905",
"0.4592",
"0.905",
"0.4592",
"0.9185",
"0.11584",
"0.9185"
],
[
"0.511",
"0.905",
"0.882",
"0.905",
"0.882",
"0.9185",
"0.511",
"0.9185"
]
] | [
"needed to preserve length, adds relevant context.",
"et al., 2020) are useful for optimizing (Cao and",
"",
"Anexampleoutputis showninFigure1.Ina coher-",
"Wang, 2021) and evaluating (Goyal and Durrett",
"ent sentence, the model removes unsupported enti-",
"2020) faithfulness, yet are best supplemented with",
"ties (Robitussin, codeine) and introduces a twice",
"fine-grained annotations (Goyal and Durrett, 2021)",
"mentioned concept from the context (BiPAP)",
"The impact of training data noise on faithfulness",
"To learn this revision task, we need examples",
"is less studied. The most common proposal is to",
"of supported and unsupported reference sentences",
"identifylowquality samples-withentailment (Mat-",
"for the same context. Without directly observing",
"sumaru et al., 2020; Goyal and Durrett, 2021) or",
"it, we generate synthetic data. At a high-level, we",
"entity overlap (Nan et al., 2021a; Narayan et al.",
"take each supported sentence, corrupt it to form a",
"2021)-and drop them. These filtering methods tend",
"diverse set of unsupported alternatives, and use this",
"to improve faithfulness yet can degrade informa-",
"mix of real and synthetic data to create examples",
"tiveness. Kang and Hashimoto (2020) address the",
"of (un)faithful revisions for contrastive learning.",
"data hunger issue by first training on all samples be-",
"As a test case, we consider a task of real-world",
"fore implementing Loss Truncation-ignoring high",
"significance-summarizing a hospital admission-",
"log loss datapoints. This is effective on a relatively",
"and extract a corpus from anoisy source-notes",
"clean Gigaword corpus yet untested on noisier cor",
"from the Electronic Health Record (EHR). We ex-",
"pora in which hallucination behavior from the unfil",
"periment with the publicly available MIMIC-III",
"tered data may be difficult to unlearn. Filtering can",
"dataset (Johnson et al., 2016). As in Adams et al",
"be \"insufficient to train a competitive model\"when",
"(2021), we treat the Brief Hospital Course (BHC)",
"high-quality data is limited (Filippova, 2020). Our",
"section of the discharge summary as a reference",
"proposed method takes advantage of all available",
" summary and all notes prior to discharge as source.",
"data while seeking to redress the underlying issue",
"Data coverage is a huge issue as only 60% of refer-",
"Clinical Summarization. Faithfulness is less",
"ence summary entities can be found in the source.",
"studied for clinical summarization because most",
"The contributions of this work are: (1) Propos-",
"proposed methods are extractive (Pivovarov and",
"ing a new method to address variable reference",
"Elhadad. 2015; Moen et al., 2016; Alsentzer and",
"quality: reference revision, which, as a data pre-",
"Kim, 2018). Abstractive approaches tend to fo-",
"processing step, is model agnostic and complemen-",
"cus on finer temporal granularities, e.g., synthesiz-",
"tary to other faithfulness approaches; (2) Showing",
"ing a single radiology report (MacAvaney et al..",
"that training on revised references can improve",
"2019; Sotudeh Gharebagh et al., 2020; Zhang et al.,",
"faithfulness while also improving informativeness",
"2020b) and doctor-patient conversations (Krishna",
"and fluency;(3)Providing code',pre-processed",
"et al., 2020; Joshi et al., 2020; Zhang et al., 2021a)",
"datasets, and models for alignment, corruption, re-",
"Most similar, Shing et al. (2021) survey extract-",
"vision, post-hoc editing, and generation of clinical",
"then-abstract approaches to section-specific dis-",
"summaries:(4)Beyond its primary use case of",
"charge summarization on MIMIC-III. They mea",
"data pre-processing, demonstrating that reference",
"sure factuality with entity overlap and remove",
"revision can have standalone value: as a post-hoc",
"poorly supported references. When analyzing a",
"editor and a pre-training objective for faithfulness.",
"proprietary EHR-derived, hospital-course summa-",
"Related Work",
"rization corpus, Adams et al. (2021) perform ora-",
"cle extractive analysis and confirm, as we do, that",
"Faithfulness.Efforts to address faithfulness have",
"EHR-derived summary references are highly noisy.",
"focused on smarter models (Huang et al., 2020),",
"more suitable metrics (Durmus et al., 2020),",
"2",
"Data",
"content-plan editing (Narayan et al., 2021), and",
"The publicly available MIMIC-III dataset contains",
"post-hoc interventions: ranking (Falke et al., 2019)",
"de-identified clinical records from patients admit-",
"and editing (Cao et al., 2020; Dong et al., 2020;",
"tedtoBethIsraelDeaconessMedicalCenter(John-",
"Zhu et al., 2021). Synthetic errors (Kryscinski",
"son et al., 2016). Hospital-admission summariza-",
"'https://github.com/amazon-research/",
"tion is a challenging task: in at most a paragraph,",
"summary-reference-revision",
"a summary must discuss what happened to the pa",
"2The datasets are accessible via PhysioNet (Goldberger",
"et al., 2000) and models via HuggingFace (Wolf et al., 2020).",
"tient during the admission, why it happened, and",
"Please refer to our GitHub README for further details.",
"what needs to happen next. To create a dataset for"
] | {
"arxiv_id": "2204.10290",
"language": "en",
"page_idx": 1
} |
||
2204.10290-en-3 | 2204.10290-en-3.jpg | unsupported, yet realistic, alternatives (§4.1), and then use this mix of real and synthetic data to create a supervised set of positive and negative revisions to support a contrastive learning objective (§4.2)^4.
4.1 Generating Synthetic Hallucinations
(D)esiderata. Based on Figure 2, unsupported generation. Each component $(k,\boldsymbol{e}_{\boldsymbol{s}-k}^{+k},f(s)_{-k}^{+k})$ is
sentences look normal (unfaithful only in the context of available data) (D1); contain many hallucinated entities (D2); exhibit a wide range of semantic divergence from the aligned source context $(\mathbf{D}3);$ and in spite of clear differences, are topically similar to aligned context (D4). To illustrate D3, we note that the second sentence could be revised by simply removing the bigram “without complications", yet the first and third sentences require more substantial re-writing to remove unfaithful content while still remaining informative and coherent.
High-Level. The simplest way to construct, and control for, hallucinations is to perform entity swaps (Kryscinski et al., 2020; Zhao et al., 2020; Zhang et al., 2021b; Chen et al., 2021). Yet, this can produce disfluent text which is easily detectable (Goyal and Durrett, 2021). In contrast, generating from a LLM produces fluent (Zhou et al., 2021), more diverse text (Cao and Wang, 2021), yet without as much control over hallucinations. Given our desiderata, we combine entity swaps (D2) into a generative framework (D1,D3) and incorporate a set of topical entities to avoid excessive semantic drift (D4). Our proposed method is called ReDRESS: reference distractor entity set swapping.
Training Objective. The ReDRESS backbone is a BART encoder-decoder model (Lewis et al., 2020a). BART is trained as a denoising autoencoder to reconstruct a corrupted sentence: $p(s|f(s))$,where $f$ is an arbitrary noise function(s). Across natural language tasks, the BART authors find span deletion to be the most effective noise. ReDRESS also uses span deletion but adds an extra noise function: entity swaps. Specifically, for each sentence $s$, we extract a set of topically related entities $e_s$ and then exchange entities between $s$ and $e_s.$ Let us denote span deletion as $f$ and the swap transformation as $g(s,\boldsymbol e_{\mathbf{s}},k)\to$ $e_{s_{-k}^{+k},s_{-k}^{+k},\text{where }k}$ represents the number of entities exchanged. To vary the level of corruption
$^{4}$We do not rely on unsupported sentences during training because they are set aside for inference. To use them, we would need to synthetically construct supported alternatives, which is not possible without first knowing how to revise.
$(\mathbf{D}3)$, we sample a different $k$ for each example such that, on average, half of the entity mentions in $s$ are swapped out for entities in $e_s.$ The final pre-training objective is $p(s|k,\boldsymbol{e}_{\boldsymbol{s}-k}^{+k},f(s)_{-k}^{+k}).k$ is represented by a special token and added to the input to make entity swaps a controllable aspect of separated by a special <sep> token and passed to the BART encoder. During training, the decoder learns to reconstruct $s$ by re-writing $f(s)_-k^{+k}$ such that it reverses the $k$ entity swaps performed between $e_s$ and $s$ and fills in a deleted text span$^5.$
Inference. To use ReDRESS to generate a plausible, corrupted version of a sentence $s$, we apply $f$ to $s$ and sample $k$ to apply $g$ to both $s$ and its distractor set $e_s.$ Two key modifications are introduced to discourage the model from reconstructing $s{:}(\mathbf{m1})$ entities removed from $s$ are not added to $e_s$,and (m2) $k$ swaps are implemented, yet the model is provided $k+1$ as the swap code to trick the model into performing an additional swap than is required for reconstruction. Using the notation above, ReDRESS generates $(k+1,\boldsymbol{e}_{\boldsymbol{s}-k},f(s)_{-k}^{+k})\to\hat{s}$ using standard beam search decoding. Without access to the original entities from $s$ in $e_s$ (m1), the model looks for plausible, hallucinated alternatives. In Appendix C, we show that (m1, m2) increase the diversity of synthetic hallucinations $\hat{s}$ to mirror the variable level of coverage observed in the data.
Implementation Details. We train ReDRESS from bart-base on a large set of unlabeled sentences extracted from MIMIC-II discharge summaries. To find the distractor set $e_s$ specific to each $s$,we first retrieve the sentences most similar to $s$ (using BioSentVec (Chen et al., 2019) embeddings to create a Faiss index (Johnson et al., 2017)). The first 25 unique concepts extracted from the set of nearest neighbors form the distractor set $e_s$ for $s.$
# 4.2 Learning a Revision Model
We apply ReDRESS to the set of supported reference summary sentences to generate positive and negative revision examples for contrastive learning.
Notation. Let $r$ represent a reference sentence with aligned source sentences $S.\hat{r}_n$ is a corrupted version of $r$ generated by ReDRESS with random seed $n$. Given $N$ diverse outputs, each generated
$^5e_s$ is provided as input yet is not part of the target output. In other words, the output of the model is a natural sentence. | [
[
"0.11584",
"0.0861",
"0.4895",
"0.0849",
"0.4895",
"0.0998",
"0.11584",
"0.10095"
],
[
"0.51",
"0.0861",
"0.881",
"0.0861",
"0.881",
"0.0998",
"0.51",
"0.0998"
],
[
"0.1184",
"0.10394",
"0.4878",
"0.10394",
"0.4878",
"0.114",
"0.1184",
"0.114"
],
[
"0.51",
"0.1027",
"0.881",
"0.1027",
"0.881",
"0.1164",
"0.51",
"0.1164"
],
[
"0.11584",
"0.11816",
"0.4895",
"0.11816",
"0.4895",
"0.1318",
"0.11584",
"0.1318"
],
[
"0.51",
"0.11816",
"0.881",
"0.11816",
"0.881",
"0.1318",
"0.51",
"0.1318"
],
[
"0.1142",
"0.1342",
"0.4844",
"0.133",
"0.4844",
"0.1478",
"0.1142",
"0.1484"
],
[
"0.511",
"0.1342",
"0.881",
"0.1342",
"0.881",
"0.1484",
"0.511",
"0.1484"
],
[
"0.511",
"0.1509",
"0.882",
"0.1509",
"0.882",
"0.1644",
"0.511",
"0.1644"
],
[
"0.11755",
"0.1609",
"0.444",
"0.1609",
"0.444",
"0.1758",
"0.11755",
"0.1758"
],
[
"0.51",
"0.1663",
"0.882",
"0.1663",
"0.882",
"0.1799",
"0.51",
"0.1799"
],
[
"0.11755",
"0.1829",
"0.4895",
"0.1829",
"0.4895",
"0.1965",
"0.11755",
"0.1965"
],
[
"0.51",
"0.1829",
"0.881",
"0.1829",
"0.881",
"0.1965",
"0.51",
"0.1965"
],
[
"0.11584",
"0.2007",
"0.4878",
"0.2007",
"0.4878",
"0.2114",
"0.11584",
"0.2114"
],
[
"0.509",
"0.199",
"0.879",
"0.199",
"0.879",
"0.212",
"0.509",
"0.212"
],
[
"0.1142",
"0.2156",
"0.492",
"0.2156",
"0.492",
"0.2286",
"0.1142",
"0.2286"
],
[
"0.509",
"0.2131",
"0.882",
"0.2144",
"0.882",
"0.2286",
"0.509",
"0.228"
],
[
"0.11755",
"0.2322",
"0.4937",
"0.2322",
"0.4937",
"0.2452",
"0.11755",
"0.2452"
],
[
"0.51",
"0.231",
"0.882",
"0.231",
"0.882",
"0.2452",
"0.51",
"0.2452"
],
[
"0.11755",
"0.2477",
"0.4902",
"0.2477",
"0.4902",
"0.2612",
"0.11755",
"0.2612"
],
[
"0.509",
"0.2465",
"0.882",
"0.2452",
"0.882",
"0.26",
"0.509",
"0.2612"
],
[
"0.11755",
"0.2632",
"0.4895",
"0.2632",
"0.4895",
"0.2766",
"0.11755",
"0.2766"
],
[
"0.509",
"0.2632",
"0.8564",
"0.2625",
"0.8564",
"0.2766",
"0.509",
"0.2778"
],
[
"0.11755",
"0.2808",
"0.486",
"0.2808",
"0.486",
"0.2915",
"0.11755",
"0.2915"
],
[
"0.511",
"0.2893",
"0.8833",
"0.2903",
"0.8833",
"0.3047",
"0.511",
"0.304"
],
[
"0.1184",
"0.2964",
"0.486",
"0.2964",
"0.486",
"0.3071",
"0.1184",
"0.3071"
],
[
"0.11755",
"0.31",
"0.4902",
"0.31",
"0.4902",
"0.3237",
"0.11755",
"0.3237"
],
[
"0.509",
"0.306",
"0.882",
"0.3071",
"0.882",
"0.3218",
"0.509",
"0.3206"
],
[
"0.11755",
"0.328",
"0.4878",
"0.328",
"0.4878",
"0.3416",
"0.11755",
"0.3416"
],
[
"0.51",
"0.3237",
"0.8833",
"0.3237",
"0.8833",
"0.3374",
"0.51",
"0.3374"
],
[
"0.11584",
"0.3433",
"0.4878",
"0.3433",
"0.4878",
"0.357",
"0.11584",
"0.357"
],
[
"0.509",
"0.339",
"0.882",
"0.3384",
"0.882",
"0.3528",
"0.509",
"0.354"
],
[
"0.1142",
"0.3594",
"0.4626",
"0.3594",
"0.4626",
"0.3723",
"0.1142",
"0.3723"
],
[
"0.511",
"0.3557",
"0.8833",
"0.3557",
"0.8833",
"0.3694",
"0.511",
"0.3694"
],
[
"0.511",
"0.3718",
"0.882",
"0.3718",
"0.882",
"0.3855",
"0.511",
"0.3855"
],
[
"0.11584",
"0.3843",
"0.486",
"0.3843",
"0.486",
"0.398",
"0.11584",
"0.398"
],
[
"0.51",
"0.3884",
"0.882",
"0.3884",
"0.882",
"0.402",
"0.51",
"0.402"
],
[
"0.11584",
"0.402",
"0.4878",
"0.402",
"0.4878",
"0.415",
"0.11584",
"0.415"
],
[
"0.509",
"0.4038",
"0.881",
"0.4038",
"0.881",
"0.4175",
"0.509",
"0.4175"
],
[
"0.11584",
"0.4175",
"0.4878",
"0.4175",
"0.4878",
"0.4312",
"0.11584",
"0.4312"
],
[
"0.509",
"0.4192",
"0.879",
"0.4192",
"0.879",
"0.4329",
"0.509",
"0.4329"
],
[
"0.11755",
"0.4329",
"0.4895",
"0.4329",
"0.4895",
"0.4465",
"0.11755",
"0.4465"
],
[
"0.509",
"0.4358",
"0.881",
"0.4358",
"0.881",
"0.4495",
"0.509",
"0.4495"
],
[
"0.11755",
"0.4495",
"0.4895",
"0.4495",
"0.4895",
"0.4631",
"0.11755",
"0.4631"
],
[
"0.51",
"0.452",
"0.882",
"0.452",
"0.882",
"0.466",
"0.51",
"0.466"
],
[
"0.11755",
"0.4644",
"0.4878",
"0.4656",
"0.4878",
"0.4797",
"0.11755",
"0.4785"
],
[
"0.51",
"0.4685",
"0.882",
"0.4685",
"0.882",
"0.4822",
"0.51",
"0.4822"
],
[
"0.11755",
"0.4822",
"0.4902",
"0.4822",
"0.4902",
"0.4954",
"0.11755",
"0.4954"
],
[
"0.509",
"0.484",
"0.879",
"0.484",
"0.879",
"0.4976",
"0.509",
"0.4976"
],
[
"0.1142",
"0.4963",
"0.4937",
"0.4976",
"0.4937",
"0.5127",
"0.1142",
"0.511"
],
[
"0.51",
"0.5005",
"0.882",
"0.5005",
"0.882",
"0.514",
"0.51",
"0.514"
],
[
"0.11584",
"0.514",
"0.4895",
"0.514",
"0.4895",
"0.528",
"0.11584",
"0.528"
],
[
"0.51",
"0.5166",
"0.882",
"0.5166",
"0.882",
"0.53",
"0.51",
"0.53"
],
[
"0.11584",
"0.531",
"0.4895",
"0.531",
"0.4895",
"0.5444",
"0.11584",
"0.5444"
],
[
"0.511",
"0.531",
"0.882",
"0.532",
"0.882",
"0.5464",
"0.511",
"0.546"
],
[
"0.11584",
"0.5464",
"0.4895",
"0.5464",
"0.4895",
"0.56",
"0.11584",
"0.56"
],
[
"0.5054",
"0.5474",
"0.8564",
"0.549",
"0.8564",
"0.5635",
"0.5054",
"0.5625"
],
[
"0.11584",
"0.5625",
"0.4902",
"0.5625",
"0.4902",
"0.576",
"0.11584",
"0.576"
],
[
"0.11755",
"0.579",
"0.4902",
"0.579",
"0.4902",
"0.593",
"0.11755",
"0.593"
],
[
"0.51",
"0.576",
"0.882",
"0.5747",
"0.882",
"0.589",
"0.51",
"0.5903"
],
[
"0.11584",
"0.5933",
"0.4844",
"0.5957",
"0.4844",
"0.6094",
"0.11584",
"0.607"
],
[
"0.51",
"0.593",
"0.881",
"0.593",
"0.881",
"0.6055",
"0.51",
"0.6055"
],
[
"0.51",
"0.608",
"0.8833",
"0.608",
"0.8833",
"0.6216",
"0.51",
"0.6216"
],
[
"0.11584",
"0.6177",
"0.486",
"0.6177",
"0.486",
"0.632",
"0.11584",
"0.632"
],
[
"0.509",
"0.626",
"0.882",
"0.626",
"0.882",
"0.6396",
"0.509",
"0.6396"
],
[
"0.1142",
"0.6353",
"0.4878",
"0.6353",
"0.4878",
"0.6484",
"0.1142",
"0.6484"
],
[
"0.5073",
"0.641",
"0.8833",
"0.641",
"0.8833",
"0.655",
"0.5073",
"0.655"
],
[
"0.11584",
"0.651",
"0.4878",
"0.651",
"0.4878",
"0.6646",
"0.11584",
"0.6646"
],
[
"0.51",
"0.6567",
"0.881",
"0.6567",
"0.881",
"0.6704",
"0.51",
"0.6704"
],
[
"0.11755",
"0.6685",
"0.4902",
"0.6685",
"0.4902",
"0.6816",
"0.11755",
"0.6816"
],
[
"0.511",
"0.675",
"0.881",
"0.675",
"0.881",
"0.685",
"0.511",
"0.685"
],
[
"0.11584",
"0.684",
"0.4895",
"0.683",
"0.4895",
"0.6978",
"0.11584",
"0.6987"
],
[
"0.51",
"0.6895",
"0.882",
"0.6895",
"0.882",
"0.703",
"0.51",
"0.703"
],
[
"0.1184",
"0.7007",
"0.4878",
"0.7007",
"0.4878",
"0.7144",
"0.1184",
"0.7144"
],
[
"0.51",
"0.706",
"0.876",
"0.706",
"0.876",
"0.7197",
"0.51",
"0.7197"
],
[
"0.1184",
"0.7188",
"0.4878",
"0.7188",
"0.4878",
"0.7285",
"0.1184",
"0.7285"
],
[
"0.11755",
"0.733",
"0.4895",
"0.733",
"0.4895",
"0.7466",
"0.11755",
"0.7466"
],
[
"0.513",
"0.7354",
"0.769",
"0.7354",
"0.769",
"0.749",
"0.513",
"0.749"
],
[
"0.11755",
"0.75",
"0.4895",
"0.75",
"0.4895",
"0.763",
"0.11755",
"0.763"
],
[
"0.513",
"0.7573",
"0.8833",
"0.7573",
"0.8833",
"0.7715",
"0.513",
"0.7715"
],
[
"0.1142",
"0.764",
"0.4844",
"0.764",
"0.4844",
"0.778",
"0.1142",
"0.778"
],
[
"0.511",
"0.774",
"0.881",
"0.774",
"0.881",
"0.7876",
"0.511",
"0.7876"
],
[
"0.11584",
"0.782",
"0.4902",
"0.782",
"0.4902",
"0.796",
"0.11584",
"0.796"
],
[
"0.51",
"0.7905",
"0.882",
"0.7905",
"0.882",
"0.804",
"0.51",
"0.804"
],
[
"0.1142",
"0.7974",
"0.486",
"0.7974",
"0.486",
"0.811",
"0.1142",
"0.811"
],
[
"0.11755",
"0.8125",
"0.4878",
"0.8135",
"0.4878",
"0.8276",
"0.11755",
"0.8267"
],
[
"0.51",
"0.8164",
"0.882",
"0.818",
"0.882",
"0.832",
"0.51",
"0.8306"
],
[
"0.1125",
"0.826",
"0.4937",
"0.8237",
"0.4937",
"0.844",
"0.1125",
"0.8467"
],
[
"0.51",
"0.8345",
"0.882",
"0.8345",
"0.882",
"0.847",
"0.51",
"0.847"
],
[
"0.11755",
"0.8457",
"0.4895",
"0.8457",
"0.4895",
"0.8594",
"0.11755",
"0.8594"
],
[
"0.5073",
"0.8496",
"0.882",
"0.8496",
"0.882",
"0.8633",
"0.5073",
"0.8633"
],
[
"0.1368",
"0.869",
"0.4895",
"0.869",
"0.4895",
"0.883",
"0.1368",
"0.883"
],
[
"0.5073",
"0.866",
"0.882",
"0.866",
"0.882",
"0.88",
"0.5073",
"0.88"
],
[
"0.1142",
"0.88",
"0.4878",
"0.8813",
"0.4878",
"0.8945",
"0.1142",
"0.8936"
],
[
"0.11755",
"0.8936",
"0.4895",
"0.8936",
"0.4895",
"0.9067",
"0.11755",
"0.9067"
],
[
"0.531",
"0.891",
"0.8833",
"0.8936",
"0.8833",
"0.9077",
"0.531",
"0.9062"
],
[
"0.11584",
"0.9067",
"0.4634",
"0.9067",
"0.4634",
"0.9175",
"0.11584",
"0.9175"
],
[
"0.509",
"0.905",
"0.879",
"0.905",
"0.879",
"0.9185",
"0.509",
"0.9185"
]
] | [
"unsupported, yet realistic, alternatives (s4.1), and",
"(D3), we sample a different k for each example",
"then usethis mix of real and syntheticdatato create",
" such that, on average, half of the entity mentions",
"a supervised set of positive and negative revisions",
"in s are swapped out for entities in es. The final",
"to support a contrastive learning objective (s4.2)4",
"pre-training objective is p(s|k,eg-k, f(s)+k). k",
"is represented by a special token and added to the",
"4.1 Generating Synthetic Hallucinations",
"input to make entity swaps a controllable aspect of",
"(D)esiderata. Based on Figure 2, unsupported",
"generation. Each component (k, es-k, f(s)-k) is",
"sentenceslooknormal(unfaithfulonlyinthecon-",
" separated by a special <sep> token and passed to",
"text of available data) (D1); contain many hallu-",
"the BART encoder. During training, the decoder",
"cinated entities (D2); exhibit a wide range of se-",
"learns to reconstruct s by re-writing f(s)-k such",
"mantic divergence from the aligned source context",
"that it reverses the k entity swaps performed be",
"(D3); and in spite of clear differences, are topically",
"tween eg and s and fills in a deleted text span?",
"similar to aligned context (D4).To illustrateD3",
"Inference.To use ReDRESS to generate a plau-",
"wenotethatthesecondsentencecouldberevised",
"by simply removing the bigram without complica-",
" sible, corrupted version of a sentence s, we apply f",
"tions\", yet the first and third sentences require more",
"to s and sample k to apply g to both s and its distrac-",
"substantial re-writing to remove unfaithful content",
"tor set eg. Two key modifications are introduced to",
" while still remaining informative and coherent.",
"discourage the model from reconstructing s: (ml)",
"entities removed from s are not added to es, and",
"High-Level.The simplest way to construct, and",
"(m2) k swaps are implemented, yet the model is",
"control for, hallucinations is to perform entity",
"provided k + 1 as the swap code to trick the model",
"swaps (Kryscinski et al., 2020; Zhao et al., 2020;",
"into performing an additional swap than is required",
"Zhang et al., 2021b; Chen et al., 2021). Yet, this",
"for reconstruction. Using the notation above, Re",
"can produce disfluent text which is easily detectable",
"DRESS generates (k + 1,es-k, f(s)-k) → $ us-",
"(Goyal and Durrett, 2021). In contrast, generating",
"ing standard beam search decoding. Without access",
"from a LLM produces fluent (Zhou et al., 2021),",
"to the original entities from s in es (ml), the model",
"more diverse text (Cao and Wang, 2021), yet with-",
"looks for plausible, hallucinated alternatives. In",
"out as much control over hallucinations. Given",
"Appendix C, we show that (ml, m2) increase the",
"our desiderata, we combine entity swaps (D2) into",
"diversity of synthetic hallucinations s to mirror the",
"a generative framework (D1, D3) and incorporate",
"variable level of coverage observed in the data.",
"a set of topical entities to avoid excessive seman-",
"tic drift (D4). Our proposed method is called Re-",
"Implementation Details.We train ReDRESS",
"DRESS: reference distractor entity set swapping.",
"from bart -base on a large set of unlabeled sen-",
"tences extracted from MIMIC-III discharge sum-",
"Training Objective.The ReDRESS backbone",
"maries. To find the distractor set es specific to each",
"is a BART encoder-decoder model (Lewis et al.,",
" s, we first retrieve the sentences most similar to s",
"2020a).BART is trained as a denoising au-",
"(using BioSentVec (Chen et al., 2019) embeddings",
"toencoder to reconstruct a corrupted sentence:",
"to create a Faiss index (Johnson et al., 2017)). The",
"p(slf(s)), where f is an arbitrary noise function(s).",
"first 25 unique concepts extracted from the set of",
"Across natural language tasks, the BART authors",
"nearest neighbors form the distractor set e for s.",
"findspandeletiontobethemosteffectivenoise.",
"ReDRESS also uses span deletion but adds an",
"4.2Learning a Revision Model",
"extra noise function: entity swaps. Specifically,",
"We apply ReDRESS to the set of supported refer-",
"for each sentence s, we extract a set of topically",
"ence summary sentences to generate positive and",
"related entities eg and then exchange entities be-",
"negative revision examples for contrastive learning",
"tween s and es. Let us denote span deletion as",
"f and the swap transformation as g(s, es, k) ",
"Notation.Let r represent a reference sentence",
"esk, s+, where k represents the number of en-",
"with aligned source sentences S. rn is a corrupted",
"tities exchanged. To vary the level of corruption",
"version of r generated by ReDRESS with random",
"4We do not rely on unsupported sentences during training",
" seed n. Given N diverse outputs, each generated",
"because they are set aside for inference. To use them, we",
"would need to synthetically construct supported alternatives,",
"eg is provided as input yet is not part of the target output.",
"which is not possible withoutfirstknowing how to revise",
"In other words, the output of the model is a natural sentence."
] | {
"arxiv_id": "2204.10290",
"language": "en",
"page_idx": 3
} |
||
1802.10062-en-1 | 1802.10062-en-1.jpg | deeper, regular network in Table 1. The main reason of using MCNN in [18] is the flexible receptive fields provided by convolutional filters with different sizes across the column. Intuitively, each column of MCNN is dedicated to a certain level of congested scene. However, the effectiveness of using MCNN may not be prominent. We present Fig. 2 to illustrate the features learned by three separated columns (representing large, medium, and small receptive fields) in MCNN and evaluate them with ShanghaiTech Part\_A[18] dataset. The three curves in this figure share very similar patterns (estimated error rate) for 50 test cases with different congest densities meaning that each column in such branch structure learn nearly identical features. It performs against the original intention of the MCNN design for learning different features for each column.
In this paper, we design a deeper network called CSRNet for counting crowd and generating high-quality density maps. Unlike the latest works such as [4,5] which use the deep CNN for ancillary, we focus on designing a CNN-based density map generator. Our model uses pure convolutional layers as the backbone to support input images with flexible resolutions. To limit the network complexity, we use the small size of convolution filters (like $3\times3)$ in all layers. We deploy the first 10 layers from VGG-16[21] as the front-end and dilated convolution layers as the back-end to enlarge receptive fields and extract deeper features without losing resolutions (since pooling layers are not used). By taking advantage of such innovative structure, we outperform the state-of-the-art crowd counting solutions (a MCNN based solution called CP-CNN [5]) with $7\%,47.3\%,10.0\%$, and $2.9\%$ lower Mean Absolute Error (MAE) in ShanghaiTech [18] Part.A, Part\_B, $UCF\_CC\_50[22], and WorldExpo'10[3] datasets respec-$ tively. Also, we achieve high performance on the UCSD dataset [23] with 1.16 MAE. After extending this work to vehicle counting on TRANCOS dataset [20], we achieve 15.4% lower MAE than the current best approach, called FCN- HA[ 24] .
The rest of the paper is structured as follows. Sec.2 presents the previous works for crowd counting and density map generation. Sec.3 introduces the architecture and configuration of our model while Sec. 4 presents the experimental results on several datasets. In Sec. 5, we conclude the paper.
## $2. \textbf{ Related work}$
Following the idea proposed by Loy et $al.[25]$, the potential solutions for crowd scenes analysis can be classified into three categories: detection-based methods, regressionbased methods, and density estimation-based methods. By combining the deep learning, the CNN-based solutions show even stronger ability in this task and outperform the traditional methods.
<FigureHere>
Figure 2. The estimated error of 50 samples from the testing set in ShanghaiTech Part\_A [18] generated by the three pre-trained columns of MCNN. Small, Medium, Large respectively stand for the columns with small, medium or large kernels in the MCNN.
<table>
<tbody>
<tr>
<th>Method</th>
<th>Parameters</th>
<th>$MAE$</th>
<th>$MSE$</th>
</tr>
<tr>
<td>Col.1 of MCNN</td>
<td>57.75k</td>
<td>141.2</td>
<td>206.8</td>
</tr>
<tr>
<td>Col. 2 of MCNN</td>
<td>45.99k</td>
<td>160.5</td>
<td>239.0</td>
</tr>
<tr>
<td>Col. 3 of MCNN</td>
<td>25.14k</td>
<td>153.7</td>
<td>230.2</td>
</tr>
<tr>
<td>MCNNTotal</td>
<td>127.68k</td>
<td>110.2</td>
<td>185.9</td>
</tr>
<tr>
<td>$A($ deeper $CNN$</td>
<td>83.84k</td>
<td>93.0</td>
<td>142.2</td>
</tr>
</tbody>
</table>
Table 1. To demonstrate that MCNN [18] may not be the best choice, we design a deeper, single-column network with fewer parameters compared to MCNN. The architecture of the proposed small network is: $CR(32,3)-M-CR(64,3)-\dot{M}-$ $CR(64,3)-M-CR(32,3)-CR(32,3)-CR(1,1).CR(m,n)$ represents the convolutional layer with m filters whose size is $n\times n$ followed by the ReLu layer. $\dot{M}$ is the max-pooling layer. Results show that the single-column version achieves higher performance on ShanghaiTech Part\_A dataset [18] with the lowest MAE and Mean Squared Error (MSE)
# 2.1. Detection-based approaches
Most of the early researches focus on detection-based approaches using a moving-window-like detector to detect people and count their number [26]. These methods require well-trained classifiers to extract low-level features from the whole human body (like Haar wavelets [27] and HOG (histogram oriented gradients) [28]). However, they perform poorly on highly congested scenes since most of the targeted objects are obscured. To tackle this problem, researchers detect particular body parts instead of the whole body to complete crowd scenes analysis [29].
# 2.2. Regression-based approaches
Since detection-based approaches can not be adapted to highly congested scenes, researchers try to deploy regression-based approaches to learn the relations among extracted features from cropped image patches, and then calculate the number of particular objects. More features, | [
[
"0.0801",
"0.09216",
"0.468",
"0.09216",
"0.468",
"0.1061",
"0.0801",
"0.1061"
],
[
"0.0801",
"0.108",
"0.468",
"0.108",
"0.468",
"0.1218",
"0.0801",
"0.1218"
],
[
"0.5293",
"0.1225",
"0.544",
"0.1225",
"0.544",
"0.13",
"0.5293",
"0.13"
],
[
"0.7812",
"0.11743",
"0.8364",
"0.11743",
"0.8364",
"0.1276",
"0.7812",
"0.1276"
],
[
"0.0809",
"0.125",
"0.4673",
"0.125",
"0.4673",
"0.1351",
"0.0809",
"0.1351"
],
[
"0.7983",
"0.127",
"0.846",
"0.1288",
"0.846",
"0.1395",
"0.7983",
"0.137"
],
[
"0.0784",
"0.1383",
"0.4714",
"0.1383",
"0.4714",
"0.1515",
"0.0784",
"0.1515"
],
[
"0.5312",
"0.1433",
"0.544",
"0.1433",
"0.544",
"0.1515",
"0.5312",
"0.1515"
],
[
"0.0809",
"0.1547",
"0.468",
"0.1547",
"0.468",
"0.1654",
"0.0809",
"0.1654"
],
[
"0.5664",
"0.1528",
"0.841",
"0.1459",
"0.8433",
"0.2102",
"0.569",
"0.2166"
],
[
"0.505",
"0.161",
"0.5186",
"0.161",
"0.5186",
"0.1963",
"0.505",
"0.1963"
],
[
"0.0801",
"0.1674",
"0.468",
"0.1674",
"0.468",
"0.1812",
"0.0801",
"0.1812"
],
[
"0.0801",
"0.1831",
"0.4714",
"0.1831",
"0.4714",
"0.1963",
"0.0801",
"0.1963"
],
[
"0.524",
"0.1895",
"0.544",
"0.1895",
"0.544",
"0.1976",
"0.524",
"0.1976"
],
[
"0.0809",
"0.1989",
"0.4697",
"0.1989",
"0.4697",
"0.2122",
"0.0809",
"0.2122"
],
[
"0.5073",
"0.2026",
"0.517",
"0.2026",
"0.517",
"0.2166",
"0.5073",
"0.2166"
],
[
"0.0801",
"0.2134",
"0.468",
"0.2134",
"0.468",
"0.2267",
"0.0801",
"0.2267"
],
[
"0.5215",
"0.2102",
"0.5474",
"0.2102",
"0.5474",
"0.2184",
"0.5215",
"0.2184"
],
[
"0.0809",
"0.2299",
"0.468",
"0.2299",
"0.468",
"0.2405",
"0.0809",
"0.2405"
],
[
"0.5215",
"0.2311",
"0.55",
"0.2311",
"0.55",
"0.2437",
"0.5215",
"0.2437"
],
[
"0.0801",
"0.2437",
"0.4697",
"0.2437",
"0.4697",
"0.2568",
"0.0801",
"0.2568"
],
[
"0.0784",
"0.2588",
"0.468",
"0.2583",
"0.468",
"0.2715",
"0.0784",
"0.2727"
],
[
"0.523",
"0.254",
"0.5513",
"0.254",
"0.5513",
"0.2651",
"0.523",
"0.2651"
],
[
"0.613",
"0.2651",
"0.6313",
"0.2651",
"0.6313",
"0.2778",
"0.613",
"0.2778"
],
[
"0.671",
"0.2664",
"0.6855",
"0.2664",
"0.6855",
"0.276",
"0.671",
"0.276"
],
[
"0.7256",
"0.2664",
"0.74",
"0.2664",
"0.74",
"0.2747",
"0.7256",
"0.2747"
],
[
"0.789",
"0.2615",
"0.8013",
"0.2708",
"0.789",
"0.2798",
"0.777",
"0.2708"
],
[
"0.5537",
"0.2683",
"0.568",
"0.26",
"0.5786",
"0.2708",
"0.5664",
"0.279"
],
[
"0.8364",
"0.2664",
"0.851",
"0.2664",
"0.851",
"0.276",
"0.8364",
"0.276"
],
[
"0.0801",
"0.2734",
"0.4697",
"0.2734",
"0.4697",
"0.2874",
"0.0801",
"0.2874"
],
[
"0.6685",
"0.2766",
"0.739",
"0.2766",
"0.739",
"0.2874",
"0.6685",
"0.2874"
],
[
"0.0801",
"0.289",
"0.468",
"0.289",
"0.468",
"0.303",
"0.0801",
"0.303"
],
[
"0.5024",
"0.2922",
"0.892",
"0.2922",
"0.892",
"0.3062",
"0.5024",
"0.3062"
],
[
"0.0801",
"0.303",
"0.2917",
"0.303",
"0.2917",
"0.3164",
"0.0801",
"0.3164"
],
[
"0.5024",
"0.307",
"0.892",
"0.307",
"0.892",
"0.3208",
"0.5024",
"0.3208"
],
[
"0.0972",
"0.318",
"0.468",
"0.3176",
"0.468",
"0.332",
"0.0972",
"0.3333"
],
[
"0.501",
"0.3193",
"0.8906",
"0.3193",
"0.8906",
"0.3333",
"0.501",
"0.3333"
],
[
"0.0768",
"0.3333",
"0.468",
"0.334",
"0.468",
"0.3484",
"0.0768",
"0.348"
],
[
"0.5024",
"0.3352",
"0.88",
"0.3352",
"0.88",
"0.3484",
"0.5024",
"0.3484"
],
[
"0.0801",
"0.3499",
"0.468",
"0.3499",
"0.468",
"0.363",
"0.0801",
"0.363"
],
[
"0.6626",
"0.3623",
"0.7466",
"0.3643",
"0.7466",
"0.379",
"0.6626",
"0.377"
],
[
"0.814",
"0.363",
"0.859",
"0.363",
"0.859",
"0.3774",
"0.814",
"0.3774"
],
[
"0.5312",
"0.3643",
"0.5933",
"0.3643",
"0.5933",
"0.3757",
"0.5312",
"0.3757"
],
[
"0.0801",
"0.3643",
"0.4697",
"0.3643",
"0.4697",
"0.3774",
"0.0801",
"0.3774"
],
[
"0.755",
"0.3643",
"0.8037",
"0.3643",
"0.8037",
"0.3757",
"0.755",
"0.3757"
],
[
"0.0784",
"0.3774",
"0.4697",
"0.379",
"0.4697",
"0.3933",
"0.0784",
"0.3926"
],
[
"0.5312",
"0.3806",
"0.6543",
"0.3806",
"0.6543",
"0.3945",
"0.5312",
"0.3945"
],
[
"0.6807",
"0.3806",
"0.732",
"0.3806",
"0.732",
"0.3958",
"0.6807",
"0.3958"
],
[
"0.7573",
"0.3806",
"0.8037",
"0.3806",
"0.8037",
"0.3958",
"0.7573",
"0.3958"
],
[
"0.8145",
"0.3806",
"0.8613",
"0.3806",
"0.8613",
"0.3958",
"0.8145",
"0.3958"
],
[
"0.0801",
"0.3945",
"0.468",
"0.3945",
"0.468",
"0.4077",
"0.0801",
"0.4077"
],
[
"0.5317",
"0.3965",
"0.653",
"0.3965",
"0.653",
"0.4104",
"0.5317",
"0.4104"
],
[
"0.7573",
"0.3965",
"0.8037",
"0.3965",
"0.8037",
"0.411",
"0.7573",
"0.411"
],
[
"0.814",
"0.3965",
"0.8613",
"0.3965",
"0.8613",
"0.411",
"0.814",
"0.411"
],
[
"0.0801",
"0.4104",
"0.4673",
"0.4104",
"0.4673",
"0.4236",
"0.0801",
"0.4236"
],
[
"0.681",
"0.3977",
"0.732",
"0.3977",
"0.732",
"0.4092",
"0.681",
"0.4092"
],
[
"0.535",
"0.4124",
"0.6543",
"0.4124",
"0.6543",
"0.4255",
"0.535",
"0.4255"
],
[
"0.759",
"0.4124",
"0.8013",
"0.4124",
"0.8013",
"0.4268",
"0.759",
"0.4268"
],
[
"0.814",
"0.4124",
"0.8594",
"0.4124",
"0.8594",
"0.4268",
"0.814",
"0.4268"
],
[
"0.0801",
"0.4248",
"0.468",
"0.4248",
"0.468",
"0.4382",
"0.0801",
"0.4382"
],
[
"0.683",
"0.4136",
"0.7305",
"0.4136",
"0.7305",
"0.4248",
"0.683",
"0.4248"
],
[
"0.5337",
"0.428",
"0.6274",
"0.428",
"0.6274",
"0.4426",
"0.5337",
"0.4426"
],
[
"0.6787",
"0.428",
"0.7344",
"0.428",
"0.7344",
"0.4426",
"0.6787",
"0.4426"
],
[
"0.7607",
"0.428",
"0.8013",
"0.428",
"0.8013",
"0.4426",
"0.7607",
"0.4426"
],
[
"0.814",
"0.428",
"0.8613",
"0.428",
"0.8613",
"0.4426",
"0.814",
"0.4426"
],
[
"0.0784",
"0.44",
"0.468",
"0.44",
"0.468",
"0.4539",
"0.0784",
"0.4539"
],
[
"0.6807",
"0.443",
"0.732",
"0.443",
"0.732",
"0.4583",
"0.6807",
"0.4583"
],
[
"0.814",
"0.443",
"0.8613",
"0.443",
"0.8613",
"0.459",
"0.814",
"0.459"
],
[
"0.7646",
"0.4443",
"0.7974",
"0.4443",
"0.7974",
"0.457",
"0.7646",
"0.457"
],
[
"0.5317",
"0.4443",
"0.6396",
"0.4443",
"0.6396",
"0.4583",
"0.5317",
"0.4583"
],
[
"0.0784",
"0.4526",
"0.468",
"0.455",
"0.468",
"0.4697",
"0.0784",
"0.4673"
],
[
"0.0801",
"0.4702",
"0.4697",
"0.4702",
"0.4697",
"0.4841",
"0.0801",
"0.4841"
],
[
"0.5034",
"0.4702",
"0.892",
"0.4702",
"0.892",
"0.4841",
"0.5034",
"0.4841"
],
[
"0.0784",
"0.4841",
"0.4697",
"0.4849",
"0.4697",
"0.5",
"0.0784",
"0.4988"
],
[
"0.5024",
"0.4849",
"0.892",
"0.4849",
"0.892",
"0.4988",
"0.5024",
"0.4988"
],
[
"0.0801",
"0.5005",
"0.4697",
"0.5005",
"0.4697",
"0.5146",
"0.0801",
"0.5146"
],
[
"0.5024",
"0.4988",
"0.8906",
"0.4988",
"0.8906",
"0.512",
"0.5024",
"0.512"
],
[
"0.0809",
"0.5176",
"0.4673",
"0.5176",
"0.4673",
"0.528",
"0.0809",
"0.528"
],
[
"0.501",
"0.512",
"0.876",
"0.5103",
"0.876",
"0.5234",
"0.501",
"0.526"
],
[
"0.0784",
"0.53",
"0.4697",
"0.53",
"0.4697",
"0.5435",
"0.0784",
"0.5435"
],
[
"0.5024",
"0.526",
"0.8906",
"0.526",
"0.8906",
"0.539",
"0.5024",
"0.539"
],
[
"0.0801",
"0.5444",
"0.4697",
"0.5444",
"0.4697",
"0.558",
"0.0801",
"0.558"
],
[
"0.501",
"0.5405",
"0.892",
"0.539",
"0.892",
"0.553",
"0.501",
"0.5537"
],
[
"0.5024",
"0.5537",
"0.892",
"0.5537",
"0.892",
"0.5674",
"0.5024",
"0.5674"
],
[
"0.0801",
"0.56",
"0.468",
"0.56",
"0.468",
"0.5737",
"0.0801",
"0.5737"
],
[
"0.5024",
"0.5674",
"0.892",
"0.5674",
"0.892",
"0.5806",
"0.5024",
"0.5806"
],
[
"0.0801",
"0.5757",
"0.4697",
"0.5757",
"0.4697",
"0.589",
"0.0801",
"0.589"
],
[
"0.5024",
"0.5806",
"0.8936",
"0.5806",
"0.8936",
"0.5947",
"0.5024",
"0.5947"
],
[
"0.0784",
"0.5903",
"0.468",
"0.5903",
"0.468",
"0.604",
"0.0784",
"0.604"
],
[
"0.5024",
"0.595",
"0.671",
"0.595",
"0.671",
"0.6094",
"0.5024",
"0.6094"
],
[
"0.0801",
"0.606",
"0.4714",
"0.606",
"0.4714",
"0.619",
"0.0801",
"0.619"
],
[
"0.0768",
"0.619",
"0.4697",
"0.6206",
"0.4697",
"0.6353",
"0.0768",
"0.634"
],
[
"0.5024",
"0.6157",
"0.751",
"0.617",
"0.751",
"0.632",
"0.5024",
"0.63"
],
[
"0.0809",
"0.636",
"0.4697",
"0.636",
"0.4697",
"0.6494",
"0.0809",
"0.6494"
],
[
"0.523",
"0.64",
"0.892",
"0.64",
"0.892",
"0.654",
"0.523",
"0.654"
],
[
"0.0809",
"0.6514",
"0.1765",
"0.6514",
"0.1765",
"0.6636",
"0.0809",
"0.6636"
],
[
"0.5024",
"0.657",
"0.892",
"0.657",
"0.892",
"0.6704",
"0.5024",
"0.6704"
],
[
"0.0989",
"0.6665",
"0.4697",
"0.6665",
"0.4697",
"0.681",
"0.0989",
"0.681"
],
[
"0.5024",
"0.672",
"0.892",
"0.672",
"0.892",
"0.685",
"0.5024",
"0.685"
],
[
"0.0801",
"0.6816",
"0.468",
"0.6816",
"0.468",
"0.696",
"0.0801",
"0.696"
],
[
"0.501",
"0.6875",
"0.892",
"0.686",
"0.892",
"0.6997",
"0.501",
"0.7007"
],
[
"0.0784",
"0.6963",
"0.4697",
"0.696",
"0.4697",
"0.7104",
"0.0784",
"0.711"
],
[
"0.5024",
"0.702",
"0.892",
"0.702",
"0.892",
"0.7153",
"0.5024",
"0.7153"
],
[
"0.0801",
"0.712",
"0.4697",
"0.712",
"0.4697",
"0.726",
"0.0801",
"0.726"
],
[
"0.5024",
"0.7173",
"0.8906",
"0.7173",
"0.8906",
"0.731",
"0.5024",
"0.731"
],
[
"0.0784",
"0.726",
"0.468",
"0.726",
"0.468",
"0.7393",
"0.0784",
"0.7393"
],
[
"0.505",
"0.7344",
"0.8906",
"0.7344",
"0.8906",
"0.744",
"0.505",
"0.744"
],
[
"0.0801",
"0.7437",
"0.1462",
"0.7437",
"0.1462",
"0.755",
"0.0801",
"0.755"
],
[
"0.5024",
"0.7476",
"0.892",
"0.7476",
"0.892",
"0.761",
"0.5024",
"0.761"
],
[
"0.5024",
"0.763",
"0.892",
"0.763",
"0.892",
"0.7764",
"0.5024",
"0.7764"
],
[
"0.0784",
"0.7695",
"0.2141",
"0.771",
"0.2141",
"0.785",
"0.0784",
"0.784"
],
[
"0.5024",
"0.7764",
"0.8013",
"0.7764",
"0.8013",
"0.7905",
"0.5024",
"0.7905"
],
[
"0.09644",
"0.7944",
"0.468",
"0.7954",
"0.468",
"0.81",
"0.09644",
"0.8096"
],
[
"0.5034",
"0.803",
"0.761",
"0.803",
"0.761",
"0.8174",
"0.5034",
"0.8174"
],
[
"0.0801",
"0.811",
"0.468",
"0.811",
"0.468",
"0.8247",
"0.0801",
"0.8247"
],
[
"0.0801",
"0.827",
"0.4673",
"0.827",
"0.4673",
"0.8403",
"0.0801",
"0.8403"
],
[
"0.5195",
"0.8247",
"0.892",
"0.8257",
"0.892",
"0.8403",
"0.5195",
"0.839"
],
[
"0.0784",
"0.8413",
"0.468",
"0.8413",
"0.468",
"0.855",
"0.0784",
"0.855"
],
[
"0.501",
"0.8413",
"0.892",
"0.8413",
"0.892",
"0.855",
"0.501",
"0.855"
],
[
"0.0801",
"0.857",
"0.4697",
"0.857",
"0.4697",
"0.8706",
"0.0801",
"0.8706"
],
[
"0.5024",
"0.857",
"0.892",
"0.857",
"0.892",
"0.8706",
"0.5024",
"0.8706"
],
[
"0.0801",
"0.872",
"0.4697",
"0.872",
"0.4697",
"0.8853",
"0.0801",
"0.8853"
],
[
"0.501",
"0.872",
"0.8936",
"0.872",
"0.8936",
"0.886",
"0.501",
"0.886"
],
[
"0.0784",
"0.887",
"0.2133",
"0.887",
"0.2133",
"0.901",
"0.0784",
"0.901"
],
[
"0.5024",
"0.887",
"0.892",
"0.887",
"0.892",
"0.901",
"0.5024",
"0.901"
]
] | [
"deeper, regular network in Table 1. The main reason of us-",
"ing MCNN in [18]is the flexible receptive fields provided",
"400",
"-0-Small",
"byconvolutional filters withdifferent sizes acrossthecol-",
"Medium",
"umn. Intuitively, each column of MCNN is dedicated to a",
"200",
"certainlevel ofcongested scene.However,theeffectiveness",
"MM",
"",
"of using MCNN may not be prominent.We present Fig.2",
"to illustrate the features learned by three separated columns",
"-200",
"(representing large, medium, and small receptive fields) in",
"SE",
"MCNN and evaluate them with ShanghaiTech Part_A [18]",
"400",
"dataset.The three curves in this figure share very similar",
"600",
"patterns (estimated error rate) for 50 test cases with different",
"congest densities meaning that each column in such branch",
"800",
"10",
"20",
"30",
"10",
"O",
"50",
"structure learn nearly identical features. It performs against",
"TestingSample",
"the original intention of the MCNN design for learning dif-",
"Figure 2. The estimated error of 50 samples from the testing set",
"ferent features for each column.",
"in ShanghaiTech Part_A [18] generated by the three pre-trained",
"In this paper, we design a deeper network called CSR-",
"columns of MCNN. Small, Medium, Large respectively stand for",
"Net for counting crowd and generating high-quality den-",
"the columns with small, medium or large kernels in the MCNN.",
"sity maps. Unlike the latest works such as [4, 5] which",
"Parameters",
"MSE",
"Method",
"use the deep CNN for ancillary, we focus on designing a",
"MAE",
"CNN-based density map generator. Our model uses pure",
"Col.1 of MCNN",
"57.75k",
"141.2",
"206.8",
"convolutional layers as the backbone to support input im-",
"Col.2of MCNN",
"160.5",
"239.0",
"ages with flexible resolutions. To limit the network com-",
"45.99k",
"Col. 3 of MCNN",
"153.7",
"230.2",
"plexity, we use the small size of convolution filters (like",
"25.14k",
"MCNN Total",
"127.68k",
"110.2",
"185.9",
"3 × 3) in all layers. We deploy the first 10 layers from",
"83.84k",
"142.2",
"93.0",
"Adeeper CNN",
"VGG-16 [21] as the front-end and dilated convolution lay-",
"ers as the back-end to enlarge receptive fields and extract",
"Table 1. To demonstrate that MCNN [18] may not be the best",
"deeper features without losing resolutions (since pooling",
"choice, we design a deeper, single-column network with fewer",
"layers are not used). By taking advantage of such innovative",
"parameters compared to MCNN. The architecture of the pro-",
"structure,we outperform the state-of-the-art crowd count-",
"posed small network is:CR(32,3)-M-CR(64.3)=M",
"ing solutions (a MCNN based solution called CP-CNN [5])",
"CR(64,3)-M-CR(32,3)-CR(32,3)-CR(1,1).CR(m, n)",
"with 7%, 47.3%, 10.0%, and 2.9% lower Mean Abso-",
"represents the convolutional layer with mfilters whose size is n Xn",
"followed by the ReLu layer. M is the max-pooling layer. Results",
"lute Error (MAE) in ShanghaiTech [18] Part_A, PartB,",
"show that the single-column version achieves higher performance",
"UCF_CC_50 [22], and WorldExpo'10 [3] datasets respec-",
"on ShanghaiTech Part._A dataset [18] with the lowest MAE and",
"tively. Also, we achieve high performance on the UCSD",
"Mean Squared Error (MSE)",
"dataset [23] with 1.16 MAE. After extending this work to",
"vehicle counting on TRANCOS dataset [20], we achieve",
"2.1.Detection-based approaches",
"15.4% lower MAE than the current best approach, called",
"Most of the early researches focus on detection-based",
"FCN-HA [24]",
"approaches using a moving-window-like detector to detect",
"The rest of the paper is structured as follows. Sec. 2",
"people and count their number [26]. These methods re-",
"presents the previous works for crowd counting and den-",
"quire well-trained classifiers to extract low-level features",
"sity map generation. Sec. 3 introduces the architecture and",
"from the whole human body (like Haar wavelets [27] and",
"configuration of our model while Sec. 4 presents the exper-",
"HOG (histogram oriented gradients) [28]). However, they",
"imental results on several datasets. In Sec. 5, we conclude",
"performpoorlyonhighlycongestedscenessincemostof",
"the paper.",
"the targeted objects are obscured. To tackle this problem,",
"researchers detect particular body parts instead of the whole",
"2.Related work",
"body to complete crowd scenes analysis [29].",
"Following the idea proposed by Loy et al. [25], the po-",
"2.2.Regression-based approaches",
"tential solutions for crowd scenes analysis can be classified",
"into three categories: detection-based methods, regression-",
"Since detection-based approaches can not be adapted",
"based methods, and density estimation-based methods. By",
"to highly congested scenes, researchers try to deploy",
"combining the deep learning, the CNN-based solutions",
"regression-based approaches to learn the relations among",
"show even stronger ability in this task and outperform the",
"extracted features from cropped image patches, and then",
"traditional methods.",
"calculate the number of particular objects. More features,"
] | {
"arxiv_id": "1802.10062",
"language": "en",
"page_idx": 1
} |
||
1802.10062-en-2 | 1802.10062-en-2.jpg | such as foreground and texture features, have been used for generating low-level information [30]. Following similar approaches, Idrees et $al.[22]$ propose a model to extract features by employing Fourier analysis and SIFT (Scaleinvariant feature transform) [31] interest-point based count- $ing.$
2.3. Density estimation-based approaches
When executing the regression-based solution, one critical feature, called saliency, is overlooked which causes inaccurate results in local regions. Lempitsky et al.[32] propose a method to solve this problem by learning a linear mapping between features in the local region and its object density maps. It integrates the information of saliency during the learning process. Since the ideal linear mapping is hard to obtain, Pham $etal.[33]$ use random forest regression to learn a non-linear mapping instead of the linear one.
### 2.4. CNN-based approaches
Literature also focuses on the CNN-based approaches to predict the density map because of its success in classification and recognition [34,21,35]. In the work presented by Walach and Wolf [36], a method is demonstrated with selective sampling and layered boosting. Instead of using patch-based training, Shang et al.[37] try an end-toend regression method using CNNs which takes the entire image as input and directly outputs the final crowd count. Boominathan et al.[19] present the frst work purely using convolutional networks and dual-column architecture for generating density map. Marsden et al.[38] explore single-column fully convolutional networks while Sindagi $etal.[39]$ propose a CNN which uses the high-level prior information to boost the density prediction performance. An improved structure is proposed by Zhang et al.[18] who introduce a multi-column based architecture (MCNN) for crowd counting. Similar idea is shown in Onoro and Sastre[20] where a scale-aware, multi-column counting model called Hydra CNN is presented for object density estimation. It is clear that the CNN-based solutions outperform the previous works mentioned in Sec.2.1 to 2.3.
$2. 5. \textbf{ Limitations of the state- of- the- art approaches}$
Most recently, Sam et $al.[4]$ propose the Switch-CNN using a density level classifier to choose different regressors for particular input patches. Sindagi $et$ al.[5] present a Contextual Pyramid CNN, which uses CNN networks to estimate context at various levels for achieving lower count error and better quality density maps. These two solutions achieve the state-of-the-art performance, and both of them used multi-column based architecture (MCNN) and density level classifier. However, we observe several disadvantages in these approaches: (1) Multi-column CNNs are hard to train according to the training method described in
work[18]. Such bloated network structure requires more time to train. (2) Multi-column CNNs introduce redundant structure as we mentioned in Sec. 1. Different columns seem to perform similarly without obvious differences. (3) Both solutions require density level classifier before sending pictures in the MCNN. However, the granularity of density level is hard to define in real-time congested scene analysis since the number of objects keeps changing with a large range. Also, using a fine-grained classifier means more columns need to be implemented which makes the design more complicated and causes more redundancy. (4) These works spend a large portion of parameters for density level classification to label the input regions instead of allocating parameters to the final density map generation. Since the branch structure in MCNN is not efficient, the lack of parameters for generating density map lowers the final accuracy. Taking all above disadvantages into consideration, we propose a novel approach to concentrate on encoding the deeper features in congested scenes and generating highquality density map.
## $3. \textbf{Proposed Solution}$
The fundamental idea of the proposed design is to deploy a deeper CNN for capturing high-level features with larger receptive felds and generating high-quality density maps without brutally expanding network complexity. In this section, we first introduce the architecture we proposed, and then we present the corresponding training methods.
## 3.1. CSRNet architecture
Following the similar idea in [19,4,5], we choose VGG- 16[21] as the front-end of CSRNet because of its strong transfer learning ability and its flexible architecture for easily concatenating the back-end for density map generation. In CrowdNet [19], the authors directly carve the first 13 layers from VGG-16 and add a $1\times1$ convolutional layer as output layer. The absence of modifications results in very weak performance. Other architectures, such as [4], uses VGG-16 as the density level classiffer for labeling input images before sending them to the most suitable column of the MCNN, while the CP-CNN [5] incorporates the result of classification with the features from density map generator. In these cases, the VGG-16 performs as an ancillary without significantly boosting the final accuracy. In this paper, we first remove the classification part of VGG- 16 (fully-connected layers) and build the proposed CSRNet with convolutional layers in VGG-16. The output size of this front-end network is 1/8 of the original input size. If we continue to stack more convolutional layers and pooling layers (basic components in VGG-16), output size would be further shrunken, and it is hard to generate high-quality density maps. Inspired by the works [10,11,40], we try to deploy dilated convolutional layers as the back-end for | [
[
"0.0801",
"0.09155",
"0.4697",
"0.09155",
"0.4697",
"0.1061",
"0.0801",
"0.1061"
],
[
"0.5024",
"0.09155",
"0.892",
"0.09155",
"0.892",
"0.1061",
"0.5024",
"0.1061"
],
[
"0.0809",
"0.108",
"0.4697",
"0.108",
"0.4697",
"0.1218",
"0.0809",
"0.1218"
],
[
"0.5034",
"0.1092",
"0.8906",
"0.1092",
"0.8906",
"0.1193",
"0.5034",
"0.1193"
],
[
"0.0809",
"0.1237",
"0.4714",
"0.1237",
"0.4714",
"0.137",
"0.0809",
"0.137"
],
[
"0.5024",
"0.1225",
"0.892",
"0.1225",
"0.892",
"0.1364",
"0.5024",
"0.1364"
],
[
"0.0801",
"0.1383",
"0.4697",
"0.1383",
"0.4697",
"0.1515",
"0.0801",
"0.1515"
],
[
"0.5024",
"0.1383",
"0.892",
"0.1383",
"0.892",
"0.1515",
"0.5024",
"0.1515"
],
[
"0.0809",
"0.1547",
"0.4673",
"0.1547",
"0.4673",
"0.1654",
"0.0809",
"0.1654"
],
[
"0.5034",
"0.1528",
"0.8906",
"0.1528",
"0.8906",
"0.1666",
"0.5034",
"0.1666"
],
[
"0.0801",
"0.1692",
"0.10864",
"0.1692",
"0.10864",
"0.1818",
"0.0801",
"0.1818"
],
[
"0.5034",
"0.1674",
"0.8906",
"0.1674",
"0.8906",
"0.1818",
"0.5034",
"0.1818"
],
[
"0.5024",
"0.1831",
"0.8906",
"0.1831",
"0.8906",
"0.1963",
"0.5024",
"0.1963"
],
[
"0.0801",
"0.1925",
"0.3962",
"0.1925",
"0.3962",
"0.207",
"0.0801",
"0.207"
],
[
"0.5034",
"0.1989",
"0.892",
"0.1989",
"0.892",
"0.2122",
"0.5034",
"0.2122"
],
[
"0.10046",
"0.2166",
"0.468",
"0.2166",
"0.468",
"0.2299",
"0.10046",
"0.2299"
],
[
"0.5024",
"0.2134",
"0.892",
"0.2134",
"0.892",
"0.2267",
"0.5024",
"0.2267"
],
[
"0.0784",
"0.2299",
"0.4673",
"0.2311",
"0.4673",
"0.2456",
"0.0784",
"0.2443"
],
[
"0.501",
"0.2267",
"0.892",
"0.2279",
"0.892",
"0.2424",
"0.501",
"0.2412"
],
[
"0.0801",
"0.2468",
"0.468",
"0.2468",
"0.468",
"0.26",
"0.0801",
"0.26"
],
[
"0.5024",
"0.2437",
"0.892",
"0.2437",
"0.892",
"0.2568",
"0.5024",
"0.2568"
],
[
"0.0801",
"0.262",
"0.468",
"0.262",
"0.468",
"0.276",
"0.0801",
"0.276"
],
[
"0.5034",
"0.26",
"0.8906",
"0.26",
"0.8906",
"0.2708",
"0.5034",
"0.2708"
],
[
"0.0801",
"0.2766",
"0.4697",
"0.2766",
"0.4697",
"0.2903",
"0.0801",
"0.2903"
],
[
"0.501",
"0.2727",
"0.8906",
"0.2734",
"0.8906",
"0.2886",
"0.501",
"0.2874"
],
[
"0.0801",
"0.2922",
"0.468",
"0.2922",
"0.468",
"0.3062",
"0.0801",
"0.3062"
],
[
"0.5034",
"0.289",
"0.892",
"0.289",
"0.892",
"0.303",
"0.5034",
"0.303"
],
[
"0.0784",
"0.3062",
"0.468",
"0.307",
"0.468",
"0.3213",
"0.0784",
"0.3208"
],
[
"0.501",
"0.303",
"0.892",
"0.3037",
"0.892",
"0.318",
"0.501",
"0.3176"
],
[
"0.0768",
"0.3208",
"0.4697",
"0.3213",
"0.4697",
"0.336",
"0.0768",
"0.3352"
],
[
"0.5034",
"0.3208",
"0.8896",
"0.3208",
"0.8896",
"0.3308",
"0.5034",
"0.3308"
],
[
"0.0801",
"0.3372",
"0.4412",
"0.3372",
"0.4412",
"0.351",
"0.0801",
"0.351"
],
[
"0.5034",
"0.334",
"0.892",
"0.334",
"0.892",
"0.348",
"0.5034",
"0.348"
],
[
"0.5034",
"0.3499",
"0.892",
"0.3499",
"0.892",
"0.363",
"0.5034",
"0.363"
],
[
"0.0784",
"0.3591",
"0.2974",
"0.3599",
"0.2974",
"0.3774",
"0.0784",
"0.377"
],
[
"0.5024",
"0.3643",
"0.8906",
"0.3643",
"0.8906",
"0.3774",
"0.5024",
"0.3774"
],
[
"0.0989",
"0.385",
"0.468",
"0.385",
"0.468",
"0.3984",
"0.0989",
"0.3984"
],
[
"0.5034",
"0.3801",
"0.6396",
"0.3801",
"0.6396",
"0.3945",
"0.5034",
"0.3945"
],
[
"0.0801",
"0.401",
"0.468",
"0.401",
"0.468",
"0.414",
"0.0801",
"0.414"
],
[
"0.5034",
"0.4072",
"0.6763",
"0.4072",
"0.6763",
"0.4216",
"0.5034",
"0.4216"
],
[
"0.0801",
"0.4167",
"0.4673",
"0.4167",
"0.4673",
"0.43",
"0.0801",
"0.43"
],
[
"0.0801",
"0.4312",
"0.4697",
"0.4312",
"0.4697",
"0.4443",
"0.0801",
"0.4443"
],
[
"0.5215",
"0.432",
"0.8906",
"0.432",
"0.8906",
"0.4458",
"0.5215",
"0.4458"
],
[
"0.0801",
"0.4463",
"0.4714",
"0.4463",
"0.4714",
"0.4602",
"0.0801",
"0.4602"
],
[
"0.501",
"0.4475",
"0.8906",
"0.4475",
"0.8906",
"0.461",
"0.501",
"0.461"
],
[
"0.0801",
"0.461",
"0.4697",
"0.461",
"0.4697",
"0.4749",
"0.0801",
"0.4749"
],
[
"0.501",
"0.461",
"0.892",
"0.4622",
"0.892",
"0.4766",
"0.501",
"0.476"
],
[
"0.0809",
"0.4766",
"0.4697",
"0.4766",
"0.4697",
"0.4905",
"0.0809",
"0.4905"
],
[
"0.501",
"0.4778",
"0.8906",
"0.4778",
"0.8906",
"0.4912",
"0.501",
"0.4912"
],
[
"0.0801",
"0.4912",
"0.468",
"0.4912",
"0.468",
"0.505",
"0.0801",
"0.505"
],
[
"0.5024",
"0.4937",
"0.8906",
"0.4937",
"0.8906",
"0.507",
"0.5024",
"0.507"
],
[
"0.0801",
"0.506",
"0.4697",
"0.506",
"0.4697",
"0.5195",
"0.0801",
"0.5195"
],
[
"0.5024",
"0.5083",
"0.875",
"0.5083",
"0.875",
"0.5215",
"0.5024",
"0.5215"
],
[
"0.0801",
"0.5215",
"0.4697",
"0.5215",
"0.4697",
"0.535",
"0.0801",
"0.535"
],
[
"0.0801",
"0.536",
"0.468",
"0.536",
"0.468",
"0.55",
"0.0801",
"0.55"
],
[
"0.5034",
"0.531",
"0.6978",
"0.531",
"0.6978",
"0.5444",
"0.5034",
"0.5444"
],
[
"0.0809",
"0.552",
"0.468",
"0.552",
"0.468",
"0.565",
"0.0809",
"0.565"
],
[
"0.5215",
"0.553",
"0.8906",
"0.553",
"0.8906",
"0.5664",
"0.5215",
"0.5664"
],
[
"0.0801",
"0.5664",
"0.468",
"0.565",
"0.468",
"0.58",
"0.0801",
"0.5806"
],
[
"0.5024",
"0.5684",
"0.892",
"0.5693",
"0.892",
"0.584",
"0.5024",
"0.5835"
],
[
"0.0801",
"0.582",
"0.4697",
"0.582",
"0.4697",
"0.595",
"0.0801",
"0.595"
],
[
"0.5024",
"0.5854",
"0.8906",
"0.5854",
"0.8906",
"0.5986",
"0.5024",
"0.5986"
],
[
"0.0801",
"0.5977",
"0.4697",
"0.5977",
"0.4697",
"0.6113",
"0.0801",
"0.6113"
],
[
"0.5024",
"0.601",
"0.892",
"0.601",
"0.892",
"0.6143",
"0.5024",
"0.6143"
],
[
"0.0809",
"0.613",
"0.4673",
"0.613",
"0.4673",
"0.6235",
"0.0809",
"0.6235"
],
[
"0.5024",
"0.6157",
"0.8906",
"0.6157",
"0.8906",
"0.629",
"0.5024",
"0.629"
],
[
"0.0809",
"0.629",
"0.4656",
"0.629",
"0.4656",
"0.6396",
"0.0809",
"0.6396"
],
[
"0.5024",
"0.63",
"0.892",
"0.63",
"0.892",
"0.643",
"0.5024",
"0.643"
],
[
"0.0801",
"0.6426",
"0.468",
"0.6426",
"0.468",
"0.656",
"0.0801",
"0.656"
],
[
"0.5034",
"0.6465",
"0.8906",
"0.6465",
"0.8906",
"0.657",
"0.5034",
"0.657"
],
[
"0.0801",
"0.657",
"0.468",
"0.657",
"0.468",
"0.6704",
"0.0801",
"0.6704"
],
[
"0.5034",
"0.66",
"0.892",
"0.66",
"0.892",
"0.674",
"0.5034",
"0.674"
],
[
"0.0784",
"0.6704",
"0.468",
"0.672",
"0.468",
"0.686",
"0.0784",
"0.685"
],
[
"0.5034",
"0.675",
"0.8906",
"0.675",
"0.8906",
"0.688",
"0.5034",
"0.688"
],
[
"0.0801",
"0.686",
"0.398",
"0.686",
"0.398",
"0.6997",
"0.0801",
"0.6997"
],
[
"0.5034",
"0.69",
"0.8936",
"0.69",
"0.8936",
"0.704",
"0.5034",
"0.704"
],
[
"0.0801",
"0.709",
"0.4631",
"0.7104",
"0.4631",
"0.728",
"0.0801",
"0.7266"
],
[
"0.501",
"0.704",
"0.8936",
"0.705",
"0.8936",
"0.7197",
"0.501",
"0.7183"
],
[
"0.5024",
"0.7197",
"0.8936",
"0.72",
"0.8936",
"0.735",
"0.5024",
"0.7344"
],
[
"0.0989",
"0.735",
"0.468",
"0.735",
"0.468",
"0.75",
"0.0989",
"0.75"
],
[
"0.5024",
"0.735",
"0.8896",
"0.735",
"0.8896",
"0.7485",
"0.5024",
"0.7485"
],
[
"0.0784",
"0.75",
"0.468",
"0.7505",
"0.468",
"0.765",
"0.0784",
"0.7646"
],
[
"0.5024",
"0.7505",
"0.892",
"0.7505",
"0.892",
"0.7646",
"0.5024",
"0.7646"
],
[
"0.0801",
"0.765",
"0.4673",
"0.765",
"0.4673",
"0.779",
"0.0801",
"0.779"
],
[
"0.5034",
"0.7666",
"0.8906",
"0.7666",
"0.8906",
"0.78",
"0.5034",
"0.78"
],
[
"0.0809",
"0.781",
"0.4697",
"0.781",
"0.4697",
"0.7944",
"0.0809",
"0.7944"
],
[
"0.5034",
"0.781",
"0.892",
"0.781",
"0.892",
"0.7944",
"0.5034",
"0.7944"
],
[
"0.0801",
"0.797",
"0.4697",
"0.797",
"0.4697",
"0.81",
"0.0801",
"0.81"
],
[
"0.5034",
"0.797",
"0.892",
"0.797",
"0.892",
"0.81",
"0.5034",
"0.81"
],
[
"0.0801",
"0.811",
"0.4697",
"0.811",
"0.4697",
"0.8247",
"0.0801",
"0.8247"
],
[
"0.5034",
"0.811",
"0.892",
"0.811",
"0.892",
"0.8247",
"0.5034",
"0.8247"
],
[
"0.0801",
"0.827",
"0.4697",
"0.827",
"0.4697",
"0.8403",
"0.0801",
"0.8403"
],
[
"0.5024",
"0.827",
"0.892",
"0.827",
"0.892",
"0.8403",
"0.5024",
"0.8403"
],
[
"0.0784",
"0.8413",
"0.468",
"0.8403",
"0.468",
"0.854",
"0.0784",
"0.855"
],
[
"0.5024",
"0.8413",
"0.892",
"0.8413",
"0.892",
"0.855",
"0.5024",
"0.855"
],
[
"0.0809",
"0.858",
"0.4673",
"0.858",
"0.4673",
"0.8687",
"0.0809",
"0.8687"
],
[
"0.5024",
"0.857",
"0.8906",
"0.857",
"0.8906",
"0.8706",
"0.5024",
"0.8706"
],
[
"0.0809",
"0.872",
"0.4697",
"0.872",
"0.4697",
"0.8853",
"0.0809",
"0.8853"
],
[
"0.5024",
"0.872",
"0.8906",
"0.872",
"0.8906",
"0.8853",
"0.5024",
"0.8853"
],
[
"0.0784",
"0.887",
"0.4697",
"0.887",
"0.4697",
"0.901",
"0.0784",
"0.901"
],
[
"0.5024",
"0.887",
"0.892",
"0.887",
"0.892",
"0.901",
"0.5024",
"0.901"
]
] | [
"such as foreground and texture features, have been used for",
"work [18]. Such bloated network structure requires more",
"generating low-level information [30].Following similar",
"timeto train.(2)Multi-column CNNs introduceredundant",
"approaches, Idrees et al. [22] propose a model to extract",
"structure as we mentioned in Sec. l. Different columns",
"features by employing Fourier analysis and SIFT (Scale-",
"seem to perform similarly without obvious differences. (3)",
"invariantfeaturetransform)[311interest-pointbasedcount-",
"Both solutions require density level classifier before send-",
"ing.",
"ing pictures in the MCNN. However, the granularity of den-",
"sity level is hard to define in real-time congested scene anal-",
"2.3.Density estimation-based approaches",
"ysis since the number of objects keeps changing with a large",
"When executing the regression-based solution, one crit-",
"range. Also, using a fine-grained classifier means more",
"ical feature, called saliency, is overlooked which causes in-",
"columns need to be implemented which makes the design",
"accurate results in local regions. Lempitsky et al. [32] pro-",
"more complicated and causes more redundancy. (4) These",
"pose a method to solve this problem by learning a linear",
"worksspendalargeportionofparametersfordensitylevel",
"mapping between features in the local region and its object",
"classification to label the input regions instead of allocating",
"density maps. It integrates the information of saliency dur-",
"parameters to the final density map generation. Since the",
"ing the learning process. Since the ideal linear mapping is",
"branch structure in MCNN is not efficient, the lack of pa-",
"hard to obtain, Pham et al. [33] use random forest regression",
"rametersforgeneratingdensitymaplowersthefinalaccu-",
"to learn a non-linear mapping instead of the linear one.",
"racy. Taking all above disadvantages into consideration, we",
"propose a novel approach to concentrate on encoding the",
"2.4. CNN-based approaches",
"deeper features in congested scenes and generating high-",
"Literature also focuses on the CNN-based approaches",
"quality density map.",
"to predict the density map because of its success in clas-",
"3.Proposed Solution",
"sification and recognition [34, 21, 35]. In the work pre-",
"sented by Walach and Wolf [36], a method is demonstrated",
"The fundamental idea of the proposed design is to de-",
"with selective sampling and layered boosting.Instead of",
"ploy a deeper CNN for capturing high-level features with",
"using patch-based training, Shang et al. [37] try an end-to-",
"larger receptive fields and generating high-quality density",
"end regression method using CNNs which takes the entire",
"maps without brutally expanding network complexity.In",
"image as input and directly outputs the final crowd count.",
"this section, we first introduce the architecture we proposed,",
"Boominathan et al. [19] present the first work purely us-",
"and then we present the corresponding training methods.",
"ing convolutional networks and dual-column architecture",
"for generating density map. Marsden et al. [38] explore",
"3.1. CSRNet architecture",
"single-column fully convolutional networks while Sindagi",
"Following the similar idea in [19, 4. 5], we choose VGG-",
"et al. [39] propose a CNN which uses the high-level prior in-",
"16 [21] as the front-end of CSRNet because of its strong",
"formation to boost the density prediction performance.An",
"transfer learning ability and its flexible architecture for eas-",
"improved structure is proposed by Zhang et al. [18] who",
"ily concatenating the back-end for density map generation.",
"introduceamulti-columnbased architecture(MCNN)for",
"In CrowdNet [19], the authors directly carve the first 13 lay-",
"crowd counting.Similarideaisshown in Onoro and Sas-",
"ers from VGG-16 and add a 1 x 1 convolutional layer as",
"tre [20] where a scale-aware, multi-column counting model",
"output layer. The absence of modifications results in very",
"called Hydra CNN is presented for object density estima-",
"weak performance. Other architectures, such as [4], uses",
"tion. It is clear that the CNN-based solutions outperform",
"VGG-16 as the density level classifier for labeling input im-",
"the previous works mentioned in Sec. 2.1 to 2.3.",
"ages before sending them to the most suitable column of",
"2.5. Limitations of the state-of-the-art approaches",
"the MCNN, while the CP-CNN [5] incorporates the result",
"of classification with the features from density map gen-",
"Most recently, Sam et al. [4] propose the Switch-CNN",
"erator. In these cases, the VGG-16 performs as an ancil-",
"using a density level classifier to choose different regres-",
"lary without significantly boosting the final accuracy.In",
"sors for particular input patches. Sindagi et al. [5] present",
"this paper, we first remove the classification part of VGG-",
"a Contextual Pyramid CNN, which uses CNN networks to",
"16 (fully-connected layers) and build the proposed CSRNet",
"estimate context at various levels for achieving lower count",
"with convolutional layers in VGG-16. The output size of",
"error and better quality density maps. These two solutions",
"this front-end network is 1/8 of the original input size. If",
"achieve the state-of-the-art performance, and both of them",
"we continue to stack more convolutional layers and pooling",
"used multi-column based architecture (MCNN)and den-",
"layers (basic components in VGG-16), output size would",
"sity level classifier.However, we observe several disad-",
"be further shrunken, and it is hard to generate high-quality",
"vantages in these approaches: (1) Multi-column CNNs are",
"density maps. Inspired by the works [10, 11, 40], we try",
"hard to train according to the training method described in",
"to deploy dilated convolutional layers as the back-end for"
] | {
"arxiv_id": "1802.10062",
"language": "en",
"page_idx": 2
} |
||
1802.10062-en-4 | 1802.10062-en-4.jpg | <table>
<tbody>
<tr>
<th>Dataset</th>
<th>Generating method</th>
</tr>
<tr>
<td>ShanghaiTech Part$\_$A [ 18] UCF$\_$CC$\_$50 [22]</td>
<td>Geometry-adaptive kernels</td>
</tr>
<tr>
<td>ShanghaiTech Part$\_$B [ 18] </td>
<td>Fixed kernel: $\sigma = 15$</td>
</tr>
<tr>
<td>TRANCOS$\left [ 44\right ]$</td>
<td>${\mathrm{Fixed~kernel:~}}\sigma=10$</td>
</tr>
<tr>
<td>${\mathrm{The~WorldExpo'10~[3]}}$ The UCSD [ 23] </td>
<td>Fixed kernel: $\sigma$= 3</td>
</tr>
</tbody>
</table>
Table 2. The ground truth generating methods for different datasets
resolution as the input image. With the same size, CSRNet generated results are comparable with the ground truth results using the PSNR (Peak Signal-to-Noise Ratio) and SSIM (Structural Similarity in Image [43]).
## 3.2. Training method
In this section, we provide specific details of CSRNet training. By taking advantage of the regular CNN network (without branch structures), CSRNet is easy to implement and fast to deploy.
## 3.2.1 Ground truth generation
Following the method of generating density maps in [18], we use the geometry-adaptive kernels to tackle the highly congested scenes. By blurring each head annotation using a Gaussian kernel (which is normalized to 1), we generate the ground truth considering the spatial distribution of all images from each dataset. The geometry-adaptive kernel is defined as:
$$F(\boldsymbol{x})=\sum_{i=1}^N\delta(\boldsymbol{x}-\boldsymbol{x}_i)\times G_{\sigma_i}(\boldsymbol{x}),\:with\:\sigma_i=\beta\overline{d_i}$$
: (2)
For each targeted object $x_i$ in the ground truth $\delta$,we use $\overline{d_i}$ to indicate the average distance of $k$ nearest neighbors. To generate the density map, we convolve $\delta(x-x_i)$ with a Gaussian kernel with parameter $\sigma_i$ (standard deviation), where $x$ is the position of pixel in the image. In experiment, we follow the confguration in [18] where $\beta=0.3$ and $k=3.$ For input with sparse crowd, we adapt the Gaussian kernel to the average head size to blur all the annotations. The setups for different datasets are shown in Table 2.
# 3.2.2 Data augmentation
We crop 9 patches from each image at different locations with 1/4 size of the original image. The first four patches contain four quarters of the image without overlapping while the other five patches are randomly cropped from the input image. After that, we mirror the patches so that we double the training set.
<table>
<tbody>
<tr>
<th colspan="4">$\overline{{\mathrm{Configurations~of~CSRNet}}}$</th>
</tr>
<tr>
<th>$A$</th>
<th>B</th>
<th>$C$</th>
<th>D</th>
</tr>
<tr>
<td>$in$</td>
<td>$ut(unfixed-reso$</td>
<td>${\mathrm{ution~color}}$ ima</td>
<td>ge)</td>
</tr>
<tr>
<td> </td>
<td>froni</td>
<td>end</td>
<td> </td>
</tr>
<tr>
<td> </td>
<td>(fine-tuned 1 $\textbf{fi}$</td>
<td>VGG- 16) $)m$</td>
<td> </td>
</tr>
<tr>
<td> </td>
<td>conv</td>
<td>.64-1</td>
<td> </td>
</tr>
<tr>
<td> </td>
<td>conv</td>
<td>.64-1</td>
<td> </td>
</tr>
<tr>
<td> </td>
<td>$max-F$</td>
<td>oling</td>
<td> </td>
</tr>
<tr>
<td> </td>
<td>$conv3$</td>
<td>128-1</td>
<td> </td>
</tr>
<tr>
<td> </td>
<td>conv3</td>
<td>128-1</td>
<td> </td>
</tr>
<tr>
<td> </td>
<td>$max-P$</td>
<td>oling</td>
<td> </td>
</tr>
<tr>
<td> </td>
<td>conv3</td>
<td>256-1</td>
<td> </td>
</tr>
<tr>
<td> </td>
<td>conv3</td>
<td>256-1</td>
<td> </td>
</tr>
<tr>
<td> </td>
<td>conv3</td>
<td>256-1</td>
<td> </td>
</tr>
<tr>
<td> </td>
<td>$max-p$</td>
<td>oling</td>
<td> </td>
</tr>
<tr>
<td> </td>
<td>conv3</td>
<td>512-1</td>
<td> </td>
</tr>
<tr>
<td> </td>
<td>conv3</td>
<td>512-1</td>
<td> </td>
</tr>
<tr>
<td> </td>
<td>conv3</td>
<td>512-1</td>
<td> </td>
</tr>
<tr>
<td>bac</td>
<td>k- end ( four $r$diff$\epsilon$</td>
<td>rent configurati</td>
<td>$0ns)$</td>
</tr>
<tr>
<td>conv3- 512- 1</td>
<td>conv3- 512- 2</td>
<td>conv3- 512- 2</td>
<td>conv3- 512- 4</td>
</tr>
<tr>
<td>conv3- 512- 1</td>
<td>conv3- 512- 2</td>
<td>conv3- 512- 2</td>
<td>conv3- 512- 4</td>
</tr>
<tr>
<td>conv3- 512- 1</td>
<td>conv3- 512- 2</td>
<td>conv3- 512- 2</td>
<td>conv3- 512- 4</td>
</tr>
<tr>
<td>conv3- 256- 1</td>
<td>conv3- 256- 2</td>
<td>conv3- 256- 4</td>
<td>conv3- 256- 4</td>
</tr>
<tr>
<td>conv3- 128- 1</td>
<td>conv3- 128- 2</td>
<td>conv3- 128- 4</td>
<td>conv3- 128- 4</td>
</tr>
<tr>
<td>conv3- 64- 1</td>
<td>conv3- 64- 2</td>
<td>conv3- 64- 4</td>
<td>conv3- 64- 4</td>
</tr>
<tr>
<td> </td>
<td>conv</td>
<td>-1-1</td>
<td> </td>
</tr>
</tbody>
</table>
Table 3. Configuration of CSRNet. All convolutional layers use padding to maintain the previous size. The convolutional layers' parameters are denoted as “conv-(kernel size)-(number of filters)- (dilation rate)", max-pooling layers are conducted over a $2\times2$ pixel window with stride 2.
### 3.2.3 Training details
We use a straightforward way to train the CSRNet as an end-to-end structure. The first 10 convolutional layers are fine-tuned from a well-trained VGG-16[21]. For the other layers, the initial values come from a Gaussian initialization with 0.01 standard deviation. Stochastic gradient descent (SGD) is applied with frxed learning rate at le-6 during training. Also, we choose the Euclidean distance to measure the difference between the ground truth and the estimated density map we generated which is similar to other works [19, 18,4]. The loss function is given as follow:
$$L(\Theta)=\frac{1}{2N}\sum_{i=1}^{N}\parallel\:Z(X_i;\Theta)-Z_i^{GT}\parallel_2^2$$
(3)
where $N$ is the size of training batch and $Z(X_{i};\Theta)$ is the output generated by CSRNet with parameters shown as $\Theta.X_{i}$ represents the input image while $Z_i^GT$ is the ground truth result of the input image $X_i.$ | [
[
"0.10046",
"0.0884",
"0.156",
"0.0903",
"0.1544",
"0.1048",
"0.0989",
"0.1029"
],
[
"0.304",
"0.09216",
"0.4182",
"0.09216",
"0.4182",
"0.1029",
"0.304",
"0.1029"
],
[
"0.6177",
"0.0903",
"0.777",
"0.0903",
"0.777",
"0.1042",
"0.6177",
"0.1042"
],
[
"0.649",
"0.1016",
"0.6626",
"0.11237",
"0.6504",
"0.1206",
"0.638",
"0.1111"
],
[
"0.7427",
"0.1016",
"0.7534",
"0.1092",
"0.7427",
"0.1193",
"0.732",
"0.11237"
],
[
"0.8286",
"0.1048",
"0.847",
"0.1048",
"0.847",
"0.1187",
"0.8286",
"0.1187"
],
[
"0.5474",
"0.1048",
"0.5645",
"0.1048",
"0.5645",
"0.1187",
"0.5474",
"0.1187"
],
[
"0.10046",
"0.1061",
"0.2646",
"0.1067",
"0.2646",
"0.1218",
"0.10046",
"0.1206"
],
[
"0.281",
"0.11426",
"0.4453",
"0.11426",
"0.4453",
"0.1276",
"0.281",
"0.1276"
],
[
"0.1013",
"0.1218",
"0.2133",
"0.1218",
"0.2133",
"0.1351",
"0.1013",
"0.1351"
],
[
"0.584",
"0.1206",
"0.8105",
"0.1218",
"0.8105",
"0.1364",
"0.584",
"0.1351"
],
[
"0.10046",
"0.1338",
"0.2615",
"0.1338",
"0.2615",
"0.1477",
"0.10046",
"0.1477"
],
[
"0.2957",
"0.1351",
"0.4265",
"0.1351",
"0.4265",
"0.1483",
"0.2957",
"0.1483"
],
[
"0.6685",
"0.137",
"0.726",
"0.137",
"0.726",
"0.1483",
"0.6685",
"0.1483"
],
[
"0.2957",
"0.1483",
"0.4265",
"0.1483",
"0.4265",
"0.1622",
"0.2957",
"0.1622"
],
[
"0.6177",
"0.1497",
"0.777",
"0.1497",
"0.777",
"0.1635",
"0.6177",
"0.1635"
],
[
"0.10297",
"0.1642",
"0.2443",
"0.1642",
"0.2443",
"0.178",
"0.10297",
"0.178"
],
[
"0.10455",
"0.1509",
"0.2051",
"0.1509",
"0.2051",
"0.1622",
"0.10455",
"0.1622"
],
[
"0.2974",
"0.1692",
"0.4224",
"0.1704",
"0.4224",
"0.1849",
"0.2974",
"0.1843"
],
[
"0.661",
"0.1654",
"0.732",
"0.1654",
"0.732",
"0.1768",
"0.661",
"0.1768"
],
[
"0.10455",
"0.1755",
"0.2051",
"0.178",
"0.2051",
"0.1925",
"0.10455",
"0.1901"
],
[
"0.6626",
"0.1787",
"0.732",
"0.1787",
"0.732",
"0.1901",
"0.6626",
"0.1901"
],
[
"0.0801",
"0.1925",
"0.4697",
"0.1932",
"0.4697",
"0.2084",
"0.0801",
"0.207"
],
[
"0.6567",
"0.1913",
"0.738",
"0.1913",
"0.738",
"0.2058",
"0.6567",
"0.2058"
],
[
"0.6567",
"0.2058",
"0.7363",
"0.2058",
"0.7363",
"0.2203",
"0.6567",
"0.2203"
],
[
"0.0784",
"0.2247",
"0.4673",
"0.2229",
"0.4673",
"0.2361",
"0.0784",
"0.238"
],
[
"0.6567",
"0.2197",
"0.7363",
"0.2197",
"0.7363",
"0.2343",
"0.6567",
"0.2343"
],
[
"0.6562",
"0.2329",
"0.738",
"0.2349",
"0.738",
"0.25",
"0.6562",
"0.2474"
],
[
"0.0801",
"0.2393",
"0.4697",
"0.2393",
"0.4697",
"0.2524",
"0.0801",
"0.2524"
],
[
"0.0801",
"0.2551",
"0.4697",
"0.2551",
"0.4697",
"0.2683",
"0.0801",
"0.2683"
],
[
"0.6587",
"0.2507",
"0.7344",
"0.2507",
"0.7344",
"0.262",
"0.6587",
"0.262"
],
[
"0.6567",
"0.2632",
"0.7344",
"0.2632",
"0.7344",
"0.2747",
"0.6567",
"0.2747"
],
[
"0.0801",
"0.2695",
"0.3635",
"0.2695",
"0.3635",
"0.2827",
"0.0801",
"0.2827"
],
[
"0.6567",
"0.276",
"0.7363",
"0.276",
"0.7363",
"0.2903",
"0.6567",
"0.2903"
],
[
"0.0801",
"0.2974",
"0.2427",
"0.2974",
"0.2427",
"0.3125",
"0.0801",
"0.3125"
],
[
"0.6587",
"0.2922",
"0.7363",
"0.2922",
"0.7363",
"0.3037",
"0.6587",
"0.3037"
],
[
"0.6567",
"0.305",
"0.7363",
"0.305",
"0.7363",
"0.3193",
"0.6567",
"0.3193"
],
[
"0.0972",
"0.3225",
"0.4697",
"0.3225",
"0.4697",
"0.336",
"0.0972",
"0.336"
],
[
"0.6587",
"0.3193",
"0.7344",
"0.3193",
"0.7344",
"0.3308",
"0.6587",
"0.3308"
],
[
"0.0784",
"0.3372",
"0.468",
"0.336",
"0.468",
"0.351",
"0.0784",
"0.3516"
],
[
"0.6587",
"0.334",
"0.7344",
"0.334",
"0.7344",
"0.3452",
"0.6587",
"0.3452"
],
[
"0.0801",
"0.3528",
"0.4697",
"0.3528",
"0.4697",
"0.3662",
"0.0801",
"0.3662"
],
[
"0.5786",
"0.348",
"0.818",
"0.348",
"0.818",
"0.361",
"0.5786",
"0.361"
],
[
"0.6055",
"0.361",
"0.6914",
"0.3591",
"0.6914",
"0.3738",
"0.6055",
"0.3757"
],
[
"0.794",
"0.3623",
"0.88",
"0.3599",
"0.88",
"0.3738",
"0.794",
"0.3757"
],
[
"0.7036",
"0.363",
"0.7837",
"0.363",
"0.7837",
"0.375",
"0.7036",
"0.375"
],
[
"0.0784",
"0.3687",
"0.2009",
"0.3687",
"0.2009",
"0.3818",
"0.0784",
"0.3818"
],
[
"0.5146",
"0.363",
"0.5947",
"0.363",
"0.5947",
"0.375",
"0.5146",
"0.375"
],
[
"0.5146",
"0.3757",
"0.5933",
"0.3757",
"0.5933",
"0.387",
"0.5146",
"0.387"
],
[
"0.609",
"0.3757",
"0.6914",
"0.3738",
"0.6914",
"0.387",
"0.609",
"0.3894"
],
[
"0.7036",
"0.377",
"0.7817",
"0.377",
"0.7817",
"0.3882",
"0.7036",
"0.3882"
],
[
"0.7983",
"0.377",
"0.8774",
"0.377",
"0.8774",
"0.3882",
"0.7983",
"0.3882"
],
[
"0.5156",
"0.3901",
"0.5947",
"0.3901",
"0.5947",
"0.4016",
"0.5156",
"0.4016"
],
[
"0.6113",
"0.3901",
"0.6895",
"0.3901",
"0.6895",
"0.4016",
"0.6113",
"0.4016"
],
[
"0.7036",
"0.3901",
"0.7837",
"0.3901",
"0.7837",
"0.4016",
"0.7036",
"0.4016"
],
[
"0.7974",
"0.3901",
"0.8774",
"0.3901",
"0.8774",
"0.4016",
"0.7974",
"0.4016"
],
[
"0.609",
"0.404",
"0.6914",
"0.4016",
"0.6914",
"0.4153",
"0.609",
"0.4172"
],
[
"0.794",
"0.4028",
"0.8784",
"0.401",
"0.8784",
"0.4153",
"0.794",
"0.4172"
],
[
"0.7036",
"0.4048",
"0.7837",
"0.4048",
"0.7837",
"0.4167",
"0.7036",
"0.4167"
],
[
"0.1274",
"0.4077",
"0.3057",
"0.4077",
"0.3057",
"0.4216",
"0.1274",
"0.4216"
],
[
"0.5156",
"0.4048",
"0.5933",
"0.4048",
"0.5933",
"0.4167",
"0.5156",
"0.4167"
],
[
"0.5156",
"0.4185",
"0.5947",
"0.4185",
"0.5947",
"0.43",
"0.5156",
"0.43"
],
[
"0.0801",
"0.4092",
"0.1299",
"0.4092",
"0.1299",
"0.4204",
"0.0801",
"0.4204"
],
[
"0.6113",
"0.4185",
"0.6895",
"0.4185",
"0.6895",
"0.43",
"0.6113",
"0.43"
],
[
"0.7036",
"0.4185",
"0.7837",
"0.4185",
"0.7837",
"0.43",
"0.7036",
"0.43"
],
[
"0.7974",
"0.4185",
"0.8774",
"0.4185",
"0.8774",
"0.43",
"0.7974",
"0.43"
],
[
"0.5215",
"0.432",
"0.591",
"0.432",
"0.591",
"0.443",
"0.5215",
"0.443"
],
[
"0.615",
"0.432",
"0.687",
"0.432",
"0.687",
"0.443",
"0.615",
"0.443"
],
[
"0.71",
"0.432",
"0.7793",
"0.432",
"0.7793",
"0.443",
"0.71",
"0.443"
],
[
"0.8022",
"0.432",
"0.8735",
"0.432",
"0.8735",
"0.443",
"0.8022",
"0.443"
],
[
"0.6665",
"0.4463",
"0.7295",
"0.4463",
"0.7295",
"0.4583",
"0.6665",
"0.4583"
],
[
"0.0989",
"0.435",
"0.4697",
"0.435",
"0.4697",
"0.449",
"0.0989",
"0.449"
],
[
"0.0801",
"0.449",
"0.4697",
"0.4495",
"0.4697",
"0.464",
"0.0801",
"0.4634"
],
[
"0.0801",
"0.4653",
"0.468",
"0.4653",
"0.468",
"0.4792",
"0.0801",
"0.4792"
],
[
"0.5034",
"0.4717",
"0.892",
"0.4717",
"0.892",
"0.4849",
"0.5034",
"0.4849"
],
[
"0.0801",
"0.4797",
"0.468",
"0.4797",
"0.468",
"0.4937",
"0.0801",
"0.4937"
],
[
"0.5034",
"0.4873",
"0.888",
"0.4873",
"0.888",
"0.4976",
"0.5034",
"0.4976"
],
[
"0.0801",
"0.4956",
"0.4697",
"0.4956",
"0.4697",
"0.509",
"0.0801",
"0.509"
],
[
"0.5024",
"0.5",
"0.8906",
"0.5",
"0.8906",
"0.513",
"0.5024",
"0.513"
],
[
"0.0801",
"0.5103",
"0.468",
"0.5103",
"0.468",
"0.5234",
"0.0801",
"0.5234"
],
[
"0.5024",
"0.513",
"0.892",
"0.513",
"0.892",
"0.5264",
"0.5024",
"0.5264"
],
[
"0.0801",
"0.5264",
"0.1691",
"0.5264",
"0.1691",
"0.5386",
"0.0801",
"0.5386"
],
[
"0.5024",
"0.528",
"0.665",
"0.528",
"0.665",
"0.5415",
"0.5024",
"0.5415"
],
[
"0.1666",
"0.5645",
"0.1805",
"0.5693",
"0.1732",
"0.5806",
"0.1602",
"0.5757"
],
[
"0.5024",
"0.5625",
"0.665",
"0.5625",
"0.665",
"0.5757",
"0.5024",
"0.5757"
],
[
"0.10046",
"0.5806",
"0.1584",
"0.5806",
"0.1584",
"0.595",
"0.10046",
"0.595"
],
[
"0.1528",
"0.577",
"0.4697",
"0.579",
"0.4697",
"0.5977",
"0.1528",
"0.595"
],
[
"0.1569",
"0.5977",
"0.1814",
"0.5977",
"0.1814",
"0.6094",
"0.1569",
"0.6094"
],
[
"0.523",
"0.595",
"0.892",
"0.595",
"0.892",
"0.6094",
"0.523",
"0.6094"
],
[
"0.5034",
"0.6113",
"0.892",
"0.6113",
"0.892",
"0.625",
"0.5034",
"0.625"
],
[
"0.10046",
"0.6226",
"0.468",
"0.6226",
"0.468",
"0.637",
"0.10046",
"0.637"
],
[
"0.5024",
"0.6255",
"0.892",
"0.6255",
"0.892",
"0.6396",
"0.5024",
"0.6396"
],
[
"0.0784",
"0.637",
"0.4673",
"0.638",
"0.4673",
"0.653",
"0.0784",
"0.6514"
],
[
"0.5024",
"0.6416",
"0.8906",
"0.6416",
"0.8906",
"0.655",
"0.5024",
"0.655"
],
[
"0.0801",
"0.654",
"0.468",
"0.654",
"0.468",
"0.6685",
"0.0801",
"0.6685"
],
[
"0.5024",
"0.656",
"0.892",
"0.656",
"0.892",
"0.6694",
"0.5024",
"0.6694"
],
[
"0.0801",
"0.6694",
"0.468",
"0.6694",
"0.468",
"0.683",
"0.0801",
"0.683"
],
[
"0.5024",
"0.672",
"0.892",
"0.672",
"0.892",
"0.685",
"0.5024",
"0.685"
],
[
"0.0801",
"0.6846",
"0.468",
"0.6846",
"0.468",
"0.6978",
"0.0801",
"0.6978"
],
[
"0.5024",
"0.686",
"0.892",
"0.686",
"0.892",
"0.6997",
"0.5024",
"0.6997"
],
[
"0.0809",
"0.6997",
"0.4697",
"0.6997",
"0.4697",
"0.7134",
"0.0809",
"0.7134"
],
[
"0.5024",
"0.702",
"0.8896",
"0.702",
"0.8896",
"0.7153",
"0.5024",
"0.7153"
],
[
"0.0784",
"0.7134",
"0.468",
"0.714",
"0.468",
"0.728",
"0.0784",
"0.7266"
],
[
"0.5024",
"0.7163",
"0.892",
"0.7163",
"0.892",
"0.73",
"0.5024",
"0.73"
],
[
"0.0801",
"0.73",
"0.468",
"0.73",
"0.468",
"0.7437",
"0.0801",
"0.7437"
],
[
"0.5024",
"0.732",
"0.863",
"0.732",
"0.863",
"0.7456",
"0.5024",
"0.7456"
],
[
"0.0809",
"0.744",
"0.437",
"0.744",
"0.437",
"0.759",
"0.0809",
"0.759"
],
[
"0.6284",
"0.781",
"0.642",
"0.781",
"0.642",
"0.7925",
"0.6284",
"0.7925"
],
[
"0.6562",
"0.776",
"0.671",
"0.776",
"0.671",
"0.785",
"0.6562",
"0.785"
],
[
"0.0801",
"0.784",
"0.12335",
"0.784",
"0.12335",
"0.799",
"0.0801",
"0.799"
],
[
"0.1168",
"0.785",
"0.2646",
"0.785",
"0.2646",
"0.799",
"0.1168",
"0.799"
],
[
"0.564",
"0.7886",
"0.6167",
"0.7886",
"0.6167",
"0.803",
"0.564",
"0.803"
],
[
"0.6504",
"0.7886",
"0.671",
"0.7886",
"0.671",
"0.802",
"0.6504",
"0.802"
],
[
"0.687",
"0.786",
"0.834",
"0.782",
"0.835",
"0.8037",
"0.689",
"0.808"
],
[
"0.869",
"0.789",
"0.892",
"0.789",
"0.892",
"0.802",
"0.869",
"0.802"
],
[
"0.687",
"0.7954",
"0.6953",
"0.7935",
"0.697",
"0.797",
"0.689",
"0.799"
],
[
"0.819",
"0.797",
"0.8286",
"0.797",
"0.8286",
"0.802",
"0.819",
"0.802"
],
[
"0.652",
"0.806",
"0.675",
"0.806",
"0.675",
"0.8174",
"0.652",
"0.8174"
],
[
"0.6216",
"0.7974",
"0.653",
"0.7974",
"0.653",
"0.81",
"0.6216",
"0.81"
],
[
"0.0989",
"0.811",
"0.4673",
"0.811",
"0.4673",
"0.8247",
"0.0989",
"0.8247"
],
[
"0.0801",
"0.827",
"0.4697",
"0.827",
"0.4697",
"0.8403",
"0.0801",
"0.8403"
],
[
"0.0801",
"0.8413",
"0.468",
"0.8413",
"0.468",
"0.855",
"0.0801",
"0.855"
],
[
"0.523",
"0.8413",
"0.892",
"0.8413",
"0.892",
"0.855",
"0.523",
"0.855"
],
[
"0.0801",
"0.857",
"0.468",
"0.857",
"0.468",
"0.8706",
"0.0801",
"0.8706"
],
[
"0.501",
"0.857",
"0.8896",
"0.857",
"0.8896",
"0.8706",
"0.501",
"0.8706"
],
[
"0.0784",
"0.8706",
"0.4697",
"0.872",
"0.4697",
"0.886",
"0.0784",
"0.8853"
],
[
"0.5024",
"0.869",
"0.8936",
"0.869",
"0.8936",
"0.887",
"0.5024",
"0.887"
],
[
"0.0801",
"0.886",
"0.232",
"0.887",
"0.232",
"0.901",
"0.0801",
"0.8994"
],
[
"0.5024",
"0.887",
"0.728",
"0.887",
"0.728",
"0.901",
"0.5024",
"0.901"
]
] | [
"Dataset",
"Generating method",
"Configurations of CSRNet",
"S",
"C",
"D",
"A",
"ShanghaiTech Part.A [18]",
"Geometry-adaptive kernels",
"UCFLCC-50[22]",
"input(unfixed-resolution color image)",
"ShanghaiTech Part B [18]",
"Fixed kernel: α = 15",
"front-end",
"Fixed kernel: g = 10",
"(fine-tuned from VGG-16)",
"The WorldExpo'10 [3]",
"TRANCOS[44]",
"Fixed kernel: = 3",
"conv3-64-1",
"The UCSD [23]",
"conv3-64-1",
"Table 2. The ground truth generating methods for different datasets",
"max-pooling",
"conv3-128-1",
"resolution as the input image. With the same size, CSR-",
"conv3-128-1",
"max-pooling",
"Net generated results are comparable with the ground truth",
"results using the PSNR (Peak Signal-to-Noise Ratio) and",
"conv3-256-1",
"conv3-256-1",
"SSIM (Structural Similarity in Image [43])",
"conv3-256-1",
"3.2. Training method",
"max-pooling",
"conv3-512-1",
"In this section, we provide specific details of CSRNet",
"conv3-512-1",
"training. By taking advantage of the regular CNN network",
"conv3-512-1",
"(without branch structures), CSRNet is easy to implement",
"back-end (four different configurations)",
"conv3-512-2",
"conv3-512-4",
"conv3-512-2",
"and fast to deploy.",
"conv3-512-1",
"conv3-512-1",
"conv3-512-2",
"conv3-512-2",
"conv3-512-4",
"conv3-512-1",
"conv3-512-2",
"conv3-512-2",
"conv3-512-4",
"conv3-256-2",
"conv3-256-4",
"conv3-256-4",
"Ground truth generation",
"conv3-256-1",
"conv3-128-1",
"3.2.1",
"conv3-128-2",
"conv3-128-4",
"conv3-128-4",
"conv3-64-1",
"conv3-64-2",
"conv3-64-4",
"conv3-64-4",
"convl-1-1",
"Following the method of generating density maps in",
"[18], we use the geometry-adaptive kernels to tackle the",
"highly congested scenes. By blurring each head annotation",
"Table 3. Configuration of CSRNet. All convolutional layers use",
"using a Gaussian kernel (which is normalized to 1), we gen-",
"padding to maintain the previous size.The convolutional layers",
"erate the ground truth considering the spatial distribution of",
"parameters are denoted as “conv-(kernel size)-(number of filters)-",
"all images from each dataset. The geometry-adaptive kernel",
"(dilation rate)\", max-pooling layers are conducted over a 2 × 2",
"is defined as:",
"pixel window with stride 2.",
"N",
"3.2.3Training details",
"F(α) =",
"o(α -i) × Ga.(α), with oi=βd, (2)",
"i=1",
"We use a straightforward way to train the CSRNet as",
"an end-to-end structure. The first 10 convolutional layers",
"For each targeted object c, in the ground truth d, we use",
"are fine-tuned from a well-trained VGG-16 [21]. For the",
"d, to indicate the average distance of k nearest neighbors.",
"other layers, the initial values come from a Gaussian ini-",
"To generate the density map, we convolve d(aα - i) with",
"tialization with 0.01 standard deviation. Stochastic gradient",
"a Gaussian kernel with parameter ; (standard deviation),",
"descent (SGD) is applied with fixed learning rate at le-6",
"where is the position of pixel in the image. In experiment,",
"during training. Also, we choose the Euclidean distance to",
"we follow the configuration in [18] where β = 0.3 and",
"measure the difference between the ground truth and the es-",
"k=3.Forinputwith sparsecrowd,weadapt theGaussian",
"timated density map we generated which is similar to other",
"kernel to the average head size to blur all the annotations.",
"works [19, 18, 4]. The loss function is given as follow:",
"The setups for different datasets are shown in Table 2.",
"1",
"N",
"3.2.2",
"Data augmentation",
"L(O) =",
">",
"Z(X;0) - ZST l2",
"(3)",
"",
"15",
"2=1",
"2N",
"We crop 9 patches from each image at different loca-",
"tions with 1/4 size of the original image. The first four",
"patches contain four quarters of the image without overlap-",
"where N is the size of training batch and Z(Xi; O) is",
"ping while the other five patches are randomly cropped from",
"the output generated by CSRNet with parameters shown as",
"the input image. After that, we mirror the patches so that we",
"O. X, represents the input image while zGT is the ground",
"double the training set.",
"truth result of the input image X,."
] | {
"arxiv_id": "1802.10062",
"language": "en",
"page_idx": 4
} |
||
2003.00189-en-0 | 2003.00189-en-0.jpg | Logarithmic Regret for Adversarial Online Control
Dylan J. Foster
Max Simchowitz
MIT
UC Berkeley
dylanf@mit.edu
msimchow@berkeley.edu
$\mathbf{Abstract}$
We introduce a new algorithm for online linear-quadratic control in a known system subject to adversarial disturbances. Existing regret bounds for this setting scale as $\sqrt{T}$ unless strong stochastic assumptions are imposed on the disturbance process. We give the first algorithm with $\log$arithmic regret for arbitrary adversarial disturbance sequences, provided the state and control costs are given by known quadratic functions. Our algorithm and analysis use a characterization for the optimal offline control law to reduce the online control problem to (delayed) online learning with approximate $advantage\textit{functions. Compared to previous techniques, our approach}$ does not need to control movement costs for the iterates, leading to logarithmic regret.
## 1 Introduction
$\begin{aligned}\text{Reinforcement learning and control consider the behavior of an agent making decisions in a dynamic}\\\text{Reinforcement learning and control consider the behavior of an agent making decisions in a dyn}\end{aligned}$ environment in order to suffer minimal loss. In light of recent practical breakthroughs in datadriven approaches to continuous RL and control (Lillicrap et al., 2016; Mnih et al., 2015; Silver et al., 2017), there is great interest in applying these techniques in real-world decision making applications. However, to reliably deploy data-driven RL and control in physical systems such as self-driving cars, it is critical to develop principled algorithms with provable safety and robustness guarantees. At the same time, algorithms should not be overly pessimistic, and should be able to take advantage of benign environments whenever possible.
In this paper we develop algorithms for online linear-quadratic control which ensure robust worst-case performance while optimally adapting to the environment at hand. Linear control has traditionally been studied in settings where the dynamics of the environment are either governed by a well-behaved stochastic process or driven by a worst-case process to which the learner must remain robust in the $\mathcal{H}_\infty$ sense. We consider an intermediate approach introduced by Agarwal et al. (2019a) in which disturbances are non-stochastic but performance is evaluated in terms of $regret.$ This benchmark forces the learner's control policy to achieve near optimal performance on any specific disturbance process encountered.
Concretely, we consider a setting in which the state evolves according to linear dynamics:
$$x_{t+1}=Ax_{t}+Bu_{t}+w_{t},$$
(1)
where $x_t\in\mathbb{R}^d\mathbf{x}$ are states, $u_t\in\mathbb{R}^d\mathbf{u}$ are inputs, and $A\in\mathbb{R}^d\mathbf{x}\times d\mathbf{x}$ and $B\in\mathbb{R}^d\mathbf{x}\times d\mathbf{u}$ are system matrices known to the learner. We refer to $w_t\in\mathbb{R}^{d_{\mathbf{x}}}$ as the disturbance (or,“noise”), which we assume is selected by an adaptive adversary and satisfies $\|w_t\|\leq1;$ we let $\boldsymbol{w}$ refer to the entire sequence $w_{1:T}.$ We consider fixed quadratic costs of the form $\ell(x,u):=x^{\intercal}R_{x}x+u^{\intercal}R_{u}u$, where $R_x,R_u\geq0$ are given. This model encompasses noise which is uncorrelated $(\mathcal{H}_2)$, worst-case $(\mathcal{H}_\infty)$, or governed by some non-stationary stochastic process. The model also approximates control techniques such
1 | [
[
"0.2157",
"0.1433",
"0.785",
"0.1427",
"0.785",
"0.161",
"0.2157",
"0.1622"
],
[
"0.3096",
"0.1863",
"0.4412",
"0.1863",
"0.4412",
"0.2008",
"0.3096",
"0.2008"
],
[
"0.528",
"0.1863",
"0.669",
"0.1863",
"0.669",
"0.2008",
"0.528",
"0.2008"
],
[
"0.353",
"0.2026",
"0.3962",
"0.2026",
"0.3962",
"0.2184",
"0.353",
"0.2184"
],
[
"0.546",
"0.2039",
"0.6543",
"0.2039",
"0.6543",
"0.2184",
"0.546",
"0.2184"
],
[
"0.3145",
"0.2229",
"0.4387",
"0.2229",
"0.4387",
"0.2374",
"0.3145",
"0.2374"
],
[
"0.5063",
"0.2229",
"0.6914",
"0.2229",
"0.6914",
"0.2374",
"0.5063",
"0.2374"
],
[
"0.02287",
"0.2424",
"0.06046",
"0.2424",
"0.06046",
"0.375",
"0.02287",
"0.375"
],
[
"0.4631",
"0.2727",
"0.5376",
"0.2727",
"0.5376",
"0.2874",
"0.4631",
"0.2874"
],
[
"0.1854",
"0.2954",
"0.839",
"0.2954",
"0.839",
"0.3093",
"0.1854",
"0.3093"
],
[
"0.1602",
"0.3113",
"0.8384",
"0.3113",
"0.8384",
"0.3245",
"0.1602",
"0.3245"
],
[
"0.1609",
"0.3257",
"0.8384",
"0.3257",
"0.8384",
"0.3403",
"0.1609",
"0.3403"
],
[
"0.1602",
"0.3416",
"0.839",
"0.3416",
"0.839",
"0.3547",
"0.1602",
"0.3547"
],
[
"0.1602",
"0.3567",
"0.839",
"0.3567",
"0.839",
"0.3706",
"0.1602",
"0.3706"
],
[
"0.1602",
"0.3706",
"0.8384",
"0.3706",
"0.8384",
"0.3838",
"0.1602",
"0.3838"
],
[
"0.0196",
"0.377",
"0.058",
"0.3757",
"0.06046",
"0.4634",
"0.02287",
"0.464"
],
[
"0.1602",
"0.387",
"0.839",
"0.387",
"0.839",
"0.401",
"0.1602",
"0.401"
],
[
"0.1584",
"0.401",
"0.7817",
"0.4016",
"0.7817",
"0.4167",
"0.1584",
"0.4153"
],
[
"0.1177",
"0.4443",
"0.1316",
"0.4443",
"0.1316",
"0.455",
"0.1177",
"0.455"
],
[
"0.1544",
"0.44",
"0.3022",
"0.44",
"0.3022",
"0.4583",
"0.1544",
"0.4583"
],
[
"0.04166",
"0.457",
"0.05148",
"0.457",
"0.05148",
"0.461",
"0.04166",
"0.461"
],
[
"0.02124",
"0.464",
"0.058",
"0.464",
"0.06046",
"0.6875",
"0.0237",
"0.6875"
],
[
"0.1168",
"0.4749",
"0.8813",
"0.4749",
"0.8813",
"0.4893",
"0.1168",
"0.4893"
],
[
"0.1136",
"0.4924",
"0.8813",
"0.4912",
"0.8813",
"0.506",
"0.1136",
"0.507"
],
[
"0.1152",
"0.509",
"0.8813",
"0.509",
"0.8813",
"0.5234",
"0.1152",
"0.5234"
],
[
"0.1136",
"0.526",
"0.8823",
"0.5264",
"0.8823",
"0.5415",
"0.1136",
"0.5405"
],
[
"0.1168",
"0.5444",
"0.8823",
"0.5444",
"0.8823",
"0.558",
"0.1168",
"0.558"
],
[
"0.1168",
"0.561",
"0.8823",
"0.561",
"0.8823",
"0.5757",
"0.1168",
"0.5757"
],
[
"0.1168",
"0.5776",
"0.8823",
"0.5776",
"0.8823",
"0.5923",
"0.1168",
"0.5923"
],
[
"0.1168",
"0.595",
"0.576",
"0.595",
"0.576",
"0.6094",
"0.1168",
"0.6094"
],
[
"0.1438",
"0.6123",
"0.884",
"0.6123",
"0.884",
"0.627",
"0.1438",
"0.627"
],
[
"0.1168",
"0.63",
"0.8823",
"0.63",
"0.8823",
"0.643",
"0.1168",
"0.643"
],
[
"0.1177",
"0.6465",
"0.8823",
"0.6465",
"0.8823",
"0.661",
"0.1177",
"0.661"
],
[
"0.1168",
"0.6636",
"0.8823",
"0.6636",
"0.8823",
"0.678",
"0.1168",
"0.678"
],
[
"0.1168",
"0.6797",
"0.8823",
"0.6797",
"0.8823",
"0.6943",
"0.1168",
"0.6943"
],
[
"0.1152",
"0.6978",
"0.884",
"0.6978",
"0.884",
"0.711",
"0.1152",
"0.711"
],
[
"0.1136",
"0.7134",
"0.884",
"0.7134",
"0.884",
"0.731",
"0.1136",
"0.731"
],
[
"0.1152",
"0.733",
"0.468",
"0.733",
"0.468",
"0.747",
"0.1152",
"0.747"
],
[
"0.1447",
"0.75",
"0.8433",
"0.75",
"0.8433",
"0.763",
"0.1447",
"0.763"
],
[
"0.4126",
"0.78",
"0.5884",
"0.78",
"0.5884",
"0.7944",
"0.4126",
"0.7944"
],
[
"0.857",
"0.778",
"0.884",
"0.778",
"0.884",
"0.7954",
"0.857",
"0.7954"
],
[
"0.1136",
"0.8066",
"0.8857",
"0.8066",
"0.8857",
"0.8276",
"0.1136",
"0.8276"
],
[
"0.1152",
"0.829",
"0.8823",
"0.829",
"0.8823",
"0.8433",
"0.1152",
"0.8433"
],
[
"0.1152",
"0.8467",
"0.8813",
"0.8467",
"0.8813",
"0.8613",
"0.1152",
"0.8613"
],
[
"0.1136",
"0.8623",
"0.8823",
"0.8613",
"0.8823",
"0.879",
"0.1136",
"0.88"
],
[
"0.1152",
"0.881",
"0.8823",
"0.881",
"0.8823",
"0.896",
"0.1152",
"0.896"
],
[
"0.1168",
"0.898",
"0.8823",
"0.898",
"0.8823",
"0.912",
"0.1168",
"0.912"
],
[
"0.4927",
"0.9375",
"0.5063",
"0.9375",
"0.5063",
"0.9478",
"0.4927",
"0.9478"
]
] | [
"Logarithmic Regret for Adversarial Online Control",
"Dylan J. Foster",
"MaxSimchowitz",
"MIT",
"UC Berkeley",
"dylanf@mit.edu",
"msimchow@berkeley.edu",
" ",
"Abstract",
"We introduce a new algorithm for online linear-quadratic control in a known system subject",
"to adversarial disturbances. Existing regret bounds for this setting scale as VT unless strong",
"stochastic assumptions are imposed on the disturbance process. We give the first algorithm with",
"logarithmic regret for arbitrary adversarial disturbance sequences, provided the state and control",
"costs are given by known quadratic functions. Our algorithm and analysis use a characterization",
"for the optimal offline control law to reduce the online control problem to (delayed) online",
"[DT's]",
"learning with approximate adoantage functions. Compared to previous techniques, our approach",
"does not need to control movement costs for the iterates, leading to logarithmic regret.",
"1",
"Introduction",
"",
"06800500010",
"Reinforcement learning and control consider the behavior of an agent making decisions in a dynamic",
"environment in order to suffer minimal loss. In light of recent practical breakthroughs in data",
"driven approaches to continuous RL and control (Lillicrap et al., 2016; Mnih et al., 2015; Silver",
"et al.,2017),there is great interest in applying these techniques in real-world decision making",
"applications. However, to reliably deploy data-driven RL and control in physical systems such as",
"self-driving cars, it is critical to develop principled algorithms with provable safety and robustness",
"guarantees. At the same time, algorithms should not be overly pessimistic, and should be able to",
"take advantage of benign environments whenever possible.",
"In this paper we develop algorithms for online linear-quadratic control which ensure robust",
"worst-case performance while optimally adapting to the environment at hand. Linear control has",
"traditionally been studied in settings where the dynamics of the environment are either governed",
"by a well-behaved stochastic process or driven by a worst-case process to which the learner must",
"remain robust in the Ho sense. We consider an intermediate approach introduced by Agarwal",
"et al. (2019a) in which disturbances are non-stochastic but performance is evaluated in terms of",
"regret. This benchmark forces the learner's control policy to achieve near optimal performance on",
"any specific disturbance process encountered.",
"Concretely, we consider a setting in which the state evolves according to linear dynamics:",
"Ct+1 = Act + But + Wt,",
"(1)",
"where at e Rdx are states, ut e Rdu are inputs, and Ae Rdxxdx and B e Rdxdu are system matrices",
"known to the learner.We refer to wt e Rdx as the disturbance (or, “noise\"), which we assume",
"is selected by an adaptive adversary and satisfies wtl ≤ l; we let w refer to the entire sequence",
"Wi:T. We consider fixed quadratic costs of the form l(c, u) := c' Rr&+ u Ruu, where Ra, Ru ≥ 0 are",
"given. This model encompasses noise which is uncorrelated (H2), worst-case (H), or governed",
"by some non-stationary stochastic process. The model also approximates control techniques such",
"1"
] | {
"arxiv_id": "2003.00189",
"language": "en",
"page_idx": 0
} |
||
2003.00189-en-1 | 2003.00189-en-1.jpg | as feedback linearization and trajectory tracking (Slotine and Li, 1991), where $A$ and $B$ are the result of linearizing a known nonlinear system and the disturbances arise due to systematic errors in linearization rather than from a benign noise process.
For any policy $\pi$ that selects controls based on the current state and disturbances observed so
far, we measure its performance over a time horizon $T$ by
$$J_T(\pi;\boldsymbol{w})=\sum_{t=1}^T\ell(x_t^\pi,u_t^\pi),$$
the total cost incurred by following $u_t=\pi_t(x_t,w_{1:t-1}).$ Letting $\pi^K$ denote a state-feedback control
law of the form $\pi_{t}^{K}(x)=-Kx$ for all $t$,the learning algorithm's goal is to minimize
$$\mathrm{Reg}_{T}=J_{T}(\pi^{\mathrm{alg}};\boldsymbol{w})-\inf_{K\epsilon\mathcal{K}}J_{T}(\pi^{K};\boldsymbol{w}),$$
where $\pi^\mathrm{alg}$ denotes the learner's policy and $\mathcal{K}$ is an appropriately defined set of stabilizing controllers. Thus, $\pi^\mathrm{alg}$ has low regret when its performance nearly matches the optimal controller $K\in\mathcal{K}$ on the specific, realized noise sequence. While the class $\mathcal{K}$ contains the optimal $\mathcal{H}_\infty$ and $\mathcal{H}_2$ control policies, we also develop algorithms to compete with a more general class of stabilizing linear controllers, which may fare better for certain noise sequences (Appendix A).
Achieving logarithmic regret in adversarial online control. Agarwal et al. (2019a) introduced the adversarial LQR setting we study and provided an efficient algorithm with $\sqrt{T}$-regret. Subsequent works (Agarwal et al., 2019b; Simchowitz et al., 2020) have shown that logarithmic regret is possible when the disturbances follow a semi-adversarial process with persistent excitation. Our main result is to achieve logarithmic regret for fully adversarial disturbances, provided that costs are known and quadratic.
## 1.1 Contributions
We introduce Riccatitron (Algorithm 1), a new algorithm for online linear control with adversarial
disturbances which attains polylogarithmic regret.
Theorem 1 (informal). Riccatitron $attains\textit{regret O}\left ( \log ^3T\right ) , where\mathcal{O} \textit{hides factors polynomial in}$
$relevant~problem~parameters.$
Riccatitron has comparable computational efficiency to previous methods. We show in Appendix A that the algorithm also extends to a more general benchmark class of linear controllers with internal state, and to“tracking” loss functions of the form $\ell_t(x,u):=\ell(x-a_t,u-b_t).$ Some conceptual contributions are as follows.
When is logarithmic regret possible in online control? Simchowitz and Foster (2020) and Cassel et al. (2020) independently show that logarithmic regret is impossible in a minimax sense if the system matrices $(A,B)$ are unknown, even when disturbances are i.i.d. gaussian. Conversely, our result shows that if $A$ and $B$ are known, logarithmic regret is possible even when disturbances are adversarial. Together, these results paint a clear picture of when logarithmic regret is achievable in online linear control. We note, however, that our approach heavily leverages the structure of linear control with strongly convex, quadratic costs. We refer the reader to the related work section for discussion of further structural assumptions that facilitate logarithmic regret.
2 | [
[
"0.1152",
"0.09216",
"0.8813",
"0.09216",
"0.8813",
"0.1067",
"0.1152",
"0.1067"
],
[
"0.1168",
"0.1111",
"0.8823",
"0.1111",
"0.8823",
"0.125",
"0.1168",
"0.125"
],
[
"0.1136",
"0.127",
"0.556",
"0.1276",
"0.556",
"0.1427",
"0.1136",
"0.1414"
],
[
"0.1438",
"0.1433",
"0.88",
"0.1433",
"0.88",
"0.1578",
"0.1438",
"0.1578"
],
[
"0.1168",
"0.1622",
"0.5664",
"0.1622",
"0.5664",
"0.1755",
"0.1168",
"0.1755"
],
[
"0.5",
"0.1945",
"0.5107",
"0.1945",
"0.5107",
"0.1989",
"0.5",
"0.1989"
],
[
"0.406",
"0.2008",
"0.5923",
"0.2008",
"0.5923",
"0.2184",
"0.406",
"0.2184"
],
[
"0.4968",
"0.2172",
"0.513",
"0.2172",
"0.513",
"0.2267",
"0.4968",
"0.2267"
],
[
"0.1136",
"0.2405",
"0.884",
"0.2393",
"0.884",
"0.26",
"0.1136",
"0.2615"
],
[
"0.1152",
"0.2615",
"0.7656",
"0.2615",
"0.7656",
"0.276",
"0.1152",
"0.276"
],
[
"0.3513",
"0.2886",
"0.6445",
"0.2886",
"0.6445",
"0.3093",
"0.3513",
"0.3093"
],
[
"0.523",
"0.3062",
"0.5513",
"0.3062",
"0.5513",
"0.3132",
"0.523",
"0.3132"
],
[
"0.11273",
"0.327",
"0.884",
"0.3276",
"0.884",
"0.3452",
"0.11273",
"0.3447"
],
[
"0.1168",
"0.3467",
"0.8813",
"0.3467",
"0.8813",
"0.361",
"0.1168",
"0.361"
],
[
"0.1168",
"0.3643",
"0.8823",
"0.3643",
"0.8823",
"0.379",
"0.1168",
"0.379"
],
[
"0.1168",
"0.3806",
"0.8813",
"0.3806",
"0.8813",
"0.3958",
"0.1168",
"0.3958"
],
[
"0.1136",
"0.3984",
"0.763",
"0.3984",
"0.763",
"0.4136",
"0.1136",
"0.4136"
],
[
"0.1168",
"0.432",
"0.8823",
"0.432",
"0.8823",
"0.4463",
"0.1168",
"0.4463"
],
[
"0.1168",
"0.4495",
"0.88",
"0.4495",
"0.88",
"0.464",
"0.1168",
"0.464"
],
[
"0.1152",
"0.4673",
"0.8813",
"0.4673",
"0.8813",
"0.481",
"0.1152",
"0.481"
],
[
"0.1168",
"0.4841",
"0.8813",
"0.4841",
"0.8813",
"0.4988",
"0.1168",
"0.4988"
],
[
"0.1168",
"0.502",
"0.88",
"0.502",
"0.88",
"0.515",
"0.1168",
"0.515"
],
[
"0.1152",
"0.518",
"0.3594",
"0.518",
"0.3594",
"0.532",
"0.1152",
"0.532"
],
[
"0.1177",
"0.5547",
"0.152",
"0.5547",
"0.152",
"0.5674",
"0.1177",
"0.5674"
],
[
"0.1666",
"0.5547",
"0.3015",
"0.5547",
"0.3015",
"0.5693",
"0.1666",
"0.5693"
],
[
"0.1152",
"0.5806",
"0.8823",
"0.5806",
"0.8823",
"0.595",
"0.1152",
"0.595"
],
[
"0.1136",
"0.5977",
"0.5107",
"0.5986",
"0.5107",
"0.613",
"0.1136",
"0.6123"
],
[
"0.1136",
"0.6235",
"0.884",
"0.625",
"0.884",
"0.6426",
"0.1136",
"0.6416"
],
[
"0.1152",
"0.6445",
"0.3406",
"0.6445",
"0.3406",
"0.6577",
"0.1152",
"0.6577"
],
[
"0.1438",
"0.672",
"0.8813",
"0.672",
"0.8813",
"0.686",
"0.1438",
"0.686"
],
[
"0.1168",
"0.6895",
"0.8823",
"0.6895",
"0.8823",
"0.704",
"0.1168",
"0.704"
],
[
"0.1136",
"0.704",
"0.8823",
"0.705",
"0.8823",
"0.7227",
"0.1136",
"0.7217"
],
[
"0.1152",
"0.7236",
"0.4265",
"0.7236",
"0.4265",
"0.738",
"0.1152",
"0.738"
],
[
"0.1168",
"0.757",
"0.8823",
"0.757",
"0.8823",
"0.7715",
"0.1168",
"0.7715"
],
[
"0.1168",
"0.7744",
"0.884",
"0.7744",
"0.884",
"0.789",
"0.1168",
"0.789"
],
[
"0.1168",
"0.7925",
"0.8823",
"0.7925",
"0.8823",
"0.8066",
"0.1168",
"0.8066"
],
[
"0.1152",
"0.8096",
"0.8823",
"0.8096",
"0.8823",
"0.8237",
"0.1152",
"0.8237"
],
[
"0.1168",
"0.827",
"0.8813",
"0.827",
"0.8813",
"0.8403",
"0.1168",
"0.8403"
],
[
"0.1152",
"0.8433",
"0.8823",
"0.8433",
"0.8823",
"0.858",
"0.1152",
"0.858"
],
[
"0.1168",
"0.86",
"0.8823",
"0.86",
"0.8823",
"0.875",
"0.1168",
"0.875"
],
[
"0.1168",
"0.8774",
"0.749",
"0.8774",
"0.749",
"0.8916",
"0.1168",
"0.8916"
],
[
"0.492",
"0.9346",
"0.509",
"0.9346",
"0.509",
"0.9497",
"0.492",
"0.9497"
]
] | [
"as feedback linearization and trajectory tracking (Slotine and Li, 1991), where A and B are the",
"result of linearizing a known nonlinear system and the disturbances arise due to systematic errors",
"in linearizationratherthanfromabenignnoiseprocess.",
"For any policy π that selects controls based on the current state and disturbances observed sc",
"far,wemeasureitsperformanceoveratimehorizonTby",
"",
"Jr(π;w) = El(cT,ut),",
"t=1",
"the total cost incurred by following ut = nt(at, Wit-1). Letting nK denote a state-feedback control",
"law of the form π (α) = -Kr for all t, the learning algorithm's goal is to minimize",
"RegT = Jr(πalg; w)- inf Jr(K;w),",
"KEK",
"where πalg denotes the learner's policy and K is an appropriately defined set of stabilizing con-",
"trollers. Thus, πalg has low regret when its performance nearly matches the optimal controller",
"K e K on the specific, realized noise sequence. While the class K contains the optimal Ho and",
"H2 control policies, we also develop algorithms to compete with a more general class of stabilizing",
"linear controllers, which may fare better for certain noise sequences (Appendix A).",
"Achieving logarithmic regret in adversarial online control. Agarwal et al. (2019a) intro-",
"duced the adversarial LQR setting we study and provided an efficient algorithm with VT-regret",
"Subsequent works (Agarwal et al., 2019b; Simchowitz et al., 2020) have shown that logarithmic re",
"gret is possible when the disturbances follow a semi-adversarial process with persistent excitation",
"Our main result is to achieve logarithmic regret for fully adversarial disturbances, provided that",
"costs are known and quadratic.",
"1.1.",
"Contributions",
"We introduce Riccatitron (Algorithm I), a new algorithm for online linear control with adversarial",
"disturbances which attains polylogarithmic regret.",
"Theorem 1 (informal). Riccatitron attains regret O(log3 T), where O hides factors polynomial in",
"relevant problem parameters.",
"Riccatitron has comparable computational efficiency to previous methods.We show in Ap",
"pendix A that the algorithm also extends to a more general benchmark class of linear controllers",
"with internal state, and to “tracking\" loss functions of the form et(c, u) := l(α - at,u -bt). Some",
"conceptual contributions are as follows.",
"When is logarithmic regret possible in online control? Simchowitz and Foster (2020) and",
"Cassel et al. (2020) independently show that logarithmic regret is impossible in a minimax sense if",
"the system matrices (A, B) are unknown, even when disturbances are i.i.d. gaussian. Conversely",
"our result shows that if A and B are known, logarithmic regret is possible even when disturbances",
"are adversarial. Together, these results paint a clear picture of when logarithmic regret is achievable",
"in online linear control. We note, however, that our approach heavily leverages the structure of",
"linear control with strongly convex, quadratic costs. We refer the reader to the related work section",
"for discussion of further structural assumptions that facilitate logarithmic regret.",
"2"
] | {
"arxiv_id": "2003.00189",
"language": "en",
"page_idx": 1
} |
||
2003.00189-en-2 | 2003.00189-en-2.jpg | Addressing trajectory mismatch. Riccatitron represents a new approach to a problem we call $trajectory\textit{ mismatch that arises when considering policy regret in online learning problems with}$ state. In dynamic environments, different policies inevitably visit different state trajectories. Lowregret algorithms must address the mismatch between the performance of the learner's policy $\pi^\mathrm{alg}$ on its own realized trajectory and the performance of each benchmark policy $\pi$ on the alternative trajectories it would induce. Most algorithms with policy regret guarantees (Even-Dar et al., 2009; Zimin and Neu, 2013; Abbasi-Yadkori et al., 2013; Arora et al., 2012; Anava et al., 2015; Abbasi-Yadkori et al., 2014; Cohen et al., 2018; Agarwal et al., 2019a,b; Simchowitz et al., 2020) adopt an approach to addressing this trajectory mismatch that we refer to as“online learning with stationary costs", or OLwS. At each round $t$, the learner's adaptive policy $\pi^\mathrm{alg}$ commits to a policy $\pi^{(t)}$, typically from a benchmark class $\Pi .$ The goal is to ensure that the iterates $\pi ^{( t) }$ attain low regret on a proxy sequence of stationary cost functions $\pi\mapsto\lambda_t(\pi)$ that describes the loss the learner would suffer at stage $t$ under the fictional trajectory that would arise if she had played the policy $\pi$ at all stages up to time $t$ (or in some cases, on the corresponding steady-state trajectory as $t\to\infty).$ Since the stationary cost does not depend on the learner's state, low regret on the sequence $\{\lambda_t\}$ can be obtained by feeding these losses directly into a standard online learning algorithm. To relate regret on the proxy sequence back to regret on the true sequence, most approaches use that the iterates produced by the online learner are sufficiently slow-moving. Appendix B explains both the general OLwS paradigm, and its instantiation for online control, in further detail.
The main technical challenge Riccatitron overcomes is that for the stationary costs that arise in our setting, no known algorithm produces iterates which move sufficiently slowly to yield logarithmic regret via OLwS (Appendix B.4). We adopt a new approach for online control we call $online\textit{learning}$ $with~advantages,~$or OLwA, which abandons stationary costs, and instead considers the controltheoretic advantages of actions relative to the unconstrained offline optimal policy $\pi^\star.$ Somewhat miraculously, we find that these advantages remove the explicit dependence on the learner's state, thereby eliminating the issue of trajectory mismatch described above. In particular, unlike OLwS, we $do\textit{not need to verify that the iterates produced by our algorithm change slowly}.$
1.2 Our approach: Online learning with advantages
In this section we sketch the $online\textit{ learning with advantages ( OLwA) technique underlying Ric- 10001}$ catitron. Let $\pi^\star$ denote the optimal unconstrained policy given knowledge of the entire disturbance sequence $\boldsymbol{w}$, and let $\mathbf{Q}_t^\star(x,u;\boldsymbol{w})$ be the associated Q-function (this quantity is formally defined in $\begin{aligned}\text{Definition 3). The }advantage~\mathrm{with~respect~to~\pi^{\star},~}\mathbf{A}_{t}^{\star}(u;x,\boldsymbol{w}):=\mathbf{Q}_{t}^{\star}(x,u;\boldsymbol{w})-\mathbf{Q}_{t}^{\star}(w,u,\pi^{\star}(x);\boldsymbol{w}),\end{aligned}$ describes the difference between the total cost accumulated by selecting action $u$ in state $x$ at time $t$ and subsequently playing according to the optimal policy $\pi^\star$,versus choosing $u=\pi_t^\star(x;\boldsymbol{w})$ as well.$^1$ By the well-known performance difference lemma (Kakade, 2003), the relative cost of a policy is equal the sum of the advantages under the states visited by said policy:$^2$
$$J_T(\pi;\boldsymbol{w})-J_T(\pi^\star;\boldsymbol{w})=\sum_{t=1}^T\mathbf{A}_t^\star(u_t^\pi;x_t^\pi,\boldsymbol{w}).$$
(2)
$^{1}$Since we use losses rather than rewards,“advantage” refers to the advantage of $\pi^{\star}$ over $u$ rather than the
advantage of $u$ over $\pi^\star;$ the latter terminology is more common in reinforcement learning.
$^2$See Lemma C.12 in Appendix D for a general statement of the performance difference lemma. The invocation of the performance difference lemma here is slightly different from other results on online learning in MDPs such as Even-Dar et al. (2009), in that the role of $\pi$ and $\pi^{\star}$ is swapped.
3 | [
[
"0.1177",
"0.09216",
"0.8813",
"0.09216",
"0.8813",
"0.1067",
"0.1177",
"0.1067"
],
[
"0.1136",
"0.1092",
"0.8813",
"0.10986",
"0.8813",
"0.125",
"0.1136",
"0.1237"
],
[
"0.1152",
"0.1276",
"0.8823",
"0.1276",
"0.8823",
"0.1414",
"0.1152",
"0.1414"
],
[
"0.1136",
"0.1445",
"0.8823",
"0.1427",
"0.8823",
"0.1572",
"0.1136",
"0.159"
],
[
"0.1152",
"0.161",
"0.8823",
"0.161",
"0.8823",
"0.1755",
"0.1152",
"0.1755"
],
[
"0.1168",
"0.1787",
"0.8823",
"0.1787",
"0.8823",
"0.1925",
"0.1168",
"0.1925"
],
[
"0.1168",
"0.1957",
"0.8857",
"0.1957",
"0.8857",
"0.2102",
"0.1168",
"0.2102"
],
[
"0.1168",
"0.2134",
"0.8823",
"0.2134",
"0.8823",
"0.2267",
"0.1168",
"0.2267"
],
[
"0.1152",
"0.2299",
"0.8823",
"0.2299",
"0.8823",
"0.2443",
"0.1152",
"0.2443"
],
[
"0.1152",
"0.2468",
"0.8813",
"0.2468",
"0.8813",
"0.2615",
"0.1152",
"0.2615"
],
[
"0.1168",
"0.2644",
"0.4436",
"0.2644",
"0.4436",
"0.279",
"0.1168",
"0.279"
],
[
"0.4453",
"0.2632",
"0.8813",
"0.2632",
"0.8813",
"0.2778",
"0.4453",
"0.2778"
],
[
"0.1152",
"0.2822",
"0.4272",
"0.2822",
"0.4272",
"0.2954",
"0.1152",
"0.2954"
],
[
"0.4207",
"0.2827",
"0.556",
"0.2827",
"0.556",
"0.2935",
"0.4207",
"0.2935"
],
[
"0.576",
"0.281",
"0.8823",
"0.2798",
"0.8823",
"0.2942",
"0.576",
"0.2954"
],
[
"0.1152",
"0.2986",
"0.4453",
"0.2986",
"0.4453",
"0.3125",
"0.1152",
"0.3125"
],
[
"0.4395",
"0.303",
"0.4756",
"0.303",
"0.4756",
"0.31",
"0.4395",
"0.31"
],
[
"0.4795",
"0.3037",
"0.5107",
"0.3037",
"0.5107",
"0.308",
"0.4795",
"0.308"
],
[
"0.5254",
"0.3037",
"0.5537",
"0.3037",
"0.5537",
"0.308",
"0.5254",
"0.308"
],
[
"0.5596",
"0.2986",
"0.8813",
"0.2986",
"0.8813",
"0.3125",
"0.5596",
"0.3125"
],
[
"0.1177",
"0.3176",
"0.29",
"0.3176",
"0.29",
"0.3276",
"0.1177",
"0.3276"
],
[
"0.607",
"0.3176",
"0.8813",
"0.3176",
"0.8813",
"0.3276",
"0.607",
"0.3276"
],
[
"0.1168",
"0.334",
"0.2632",
"0.334",
"0.2632",
"0.3447",
"0.1168",
"0.3447"
],
[
"0.6323",
"0.336",
"0.6543",
"0.336",
"0.6543",
"0.3435",
"0.6323",
"0.3435"
],
[
"0.6567",
"0.3333",
"0.8813",
"0.3333",
"0.8813",
"0.3467",
"0.6567",
"0.3467"
],
[
"0.1152",
"0.3499",
"0.3218",
"0.3499",
"0.3218",
"0.363",
"0.1152",
"0.363"
],
[
"0.58",
"0.3384",
"0.613",
"0.3384",
"0.613",
"0.3423",
"0.58",
"0.3423"
],
[
"0.3284",
"0.3528",
"0.357",
"0.3528",
"0.357",
"0.3599",
"0.3284",
"0.3599"
],
[
"0.609",
"0.3499",
"0.8813",
"0.3499",
"0.8813",
"0.363",
"0.609",
"0.363"
],
[
"0.3782",
"0.356",
"0.3896",
"0.356",
"0.3896",
"0.3591",
"0.3782",
"0.3591"
],
[
"0.1168",
"0.3694",
"0.2255",
"0.3694",
"0.2255",
"0.3801",
"0.1168",
"0.3801"
],
[
"0.6978",
"0.3674",
"0.8823",
"0.3674",
"0.8823",
"0.3806",
"0.6978",
"0.3806"
],
[
"0.1168",
"0.3865",
"0.2573",
"0.3865",
"0.2573",
"0.3965",
"0.1168",
"0.3965"
],
[
"0.6846",
"0.3838",
"0.8823",
"0.3838",
"0.8823",
"0.3977",
"0.6846",
"0.3977"
],
[
"0.1152",
"0.4016",
"0.246",
"0.3997",
"0.246",
"0.414",
"0.1152",
"0.4167"
],
[
"0.634",
"0.4028",
"0.7466",
"0.4028",
"0.7466",
"0.4136",
"0.634",
"0.4136"
],
[
"0.1478",
"0.4192",
"0.2255",
"0.4192",
"0.2255",
"0.43",
"0.1478",
"0.43"
],
[
"0.687",
"0.4216",
"0.7236",
"0.4216",
"0.7236",
"0.4287",
"0.687",
"0.4287"
],
[
"0.689",
"0.4248",
"0.699",
"0.4248",
"0.699",
"0.428",
"0.689",
"0.428"
],
[
"0.728",
"0.4204",
"0.88",
"0.4204",
"0.88",
"0.4312",
"0.728",
"0.4312"
],
[
"0.1177",
"0.4375",
"0.2303",
"0.4375",
"0.2303",
"0.4475",
"0.1177",
"0.4475"
],
[
"0.638",
"0.435",
"0.8813",
"0.435",
"0.8813",
"0.449",
"0.638",
"0.449"
],
[
"0.1152",
"0.4526",
"0.3145",
"0.4526",
"0.3145",
"0.4666",
"0.1152",
"0.4666"
],
[
"0.649",
"0.452",
"0.8813",
"0.4526",
"0.8813",
"0.4666",
"0.649",
"0.4653"
],
[
"0.4084",
"0.4583",
"0.429",
"0.4583",
"0.429",
"0.4622",
"0.4084",
"0.4622"
],
[
"0.539",
"0.4583",
"0.5605",
"0.4583",
"0.5605",
"0.4622",
"0.539",
"0.4622"
],
[
"0.1152",
"0.4697",
"0.46",
"0.4685",
"0.46",
"0.483",
"0.1152",
"0.4841"
],
[
"0.4722",
"0.4749",
"0.527",
"0.4749",
"0.527",
"0.4792",
"0.4722",
"0.4792"
],
[
"0.5703",
"0.4749",
"0.5864",
"0.4749",
"0.5864",
"0.4792",
"0.5703",
"0.4792"
],
[
"0.6094",
"0.4697",
"0.8823",
"0.4697",
"0.8823",
"0.483",
"0.6094",
"0.483"
],
[
"0.1177",
"0.488",
"0.5186",
"0.488",
"0.5186",
"0.4988",
"0.1177",
"0.4988"
],
[
"0.539",
"0.4912",
"0.5596",
"0.4912",
"0.5596",
"0.4956",
"0.539",
"0.4956"
],
[
"0.5923",
"0.4873",
"0.884",
"0.4873",
"0.884",
"0.5005",
"0.5923",
"0.5005"
],
[
"0.1152",
"0.504",
"0.8857",
"0.504",
"0.8857",
"0.5176",
"0.1152",
"0.5176"
],
[
"0.1168",
"0.521",
"0.886",
"0.521",
"0.886",
"0.535",
"0.1168",
"0.535"
],
[
"0.1136",
"0.5386",
"0.7715",
"0.5386",
"0.7715",
"0.553",
"0.1136",
"0.553"
],
[
"0.1168",
"0.5747",
"0.1503",
"0.5747",
"0.1503",
"0.587",
"0.1168",
"0.587"
],
[
"0.1626",
"0.5713",
"0.6274",
"0.5728",
"0.6274",
"0.5903",
"0.1626",
"0.589"
],
[
"0.1152",
"0.601",
"0.8823",
"0.601",
"0.8823",
"0.6157",
"0.1152",
"0.6157"
],
[
"0.1152",
"0.617",
"0.8813",
"0.617",
"0.8813",
"0.632",
"0.1152",
"0.632"
],
[
"0.1152",
"0.6353",
"0.8813",
"0.6353",
"0.8813",
"0.6484",
"0.1152",
"0.6484"
],
[
"0.1136",
"0.651",
"0.4983",
"0.6514",
"0.4983",
"0.6665",
"0.1136",
"0.6655"
],
[
"0.4944",
"0.653",
"0.8823",
"0.653",
"0.8823",
"0.6665",
"0.4944",
"0.6665"
],
[
"0.1168",
"0.6694",
"0.4983",
"0.6694",
"0.4983",
"0.683",
"0.1168",
"0.683"
],
[
"0.5107",
"0.6704",
"0.8784",
"0.6704",
"0.8784",
"0.681",
"0.5107",
"0.681"
],
[
"0.1136",
"0.686",
"0.8813",
"0.685",
"0.8813",
"0.6997",
"0.1136",
"0.7007"
],
[
"0.1152",
"0.704",
"0.8823",
"0.704",
"0.8823",
"0.7183",
"0.1152",
"0.7183"
],
[
"0.11273",
"0.72",
"0.683",
"0.7197",
"0.683",
"0.7344",
"0.11273",
"0.735"
],
[
"0.5254",
"0.75",
"0.5415",
"0.75",
"0.5415",
"0.759",
"0.5254",
"0.759"
],
[
"0.3381",
"0.761",
"0.66",
"0.761",
"0.66",
"0.776",
"0.3381",
"0.776"
],
[
"0.857",
"0.7603",
"0.884",
"0.7603",
"0.884",
"0.779",
"0.857",
"0.779"
],
[
"0.523",
"0.7764",
"0.546",
"0.7764",
"0.546",
"0.7886",
"0.523",
"0.7886"
],
[
"0.1381",
"0.799",
"0.8823",
"0.799",
"0.8823",
"0.813",
"0.1381",
"0.813"
],
[
"0.1136",
"0.813",
"0.7114",
"0.813",
"0.7114",
"0.827",
"0.1136",
"0.827"
],
[
"0.1365",
"0.8257",
"0.8823",
"0.8257",
"0.8823",
"0.8403",
"0.1365",
"0.8403"
],
[
"0.1152",
"0.8413",
"0.884",
"0.8413",
"0.884",
"0.855",
"0.1152",
"0.855"
],
[
"0.1152",
"0.855",
"0.5415",
"0.855",
"0.5415",
"0.8687",
"0.1152",
"0.8687"
],
[
"0.492",
"0.9346",
"0.509",
"0.9346",
"0.509",
"0.9487",
"0.492",
"0.9487"
]
] | [
"Addressing trajectory mismatch.Riccatitron represents a new approach to a problem we call",
"trajectory mismatch that arises when considering policy regret in online learning problems with",
"state. In dynamic environments, different policies inevitably visit different state trajectories.Low-",
"regret algorithms must address the mismatch between the performance of the learner's policy alg",
"on its own realized trajectory and the performance of each benchmark policy on the alternative",
"trajectories it would induce. Most algorithms with policy regret guarantees (Even-Dar et al.",
"2009; Zimin and Neu, 2013; Abbasi-Yadkori et al., 2013; Arora et al., 2012; Anava et al., 2015;",
"Abbasi-Yadkori et al., 2014: Cohen et al., 2018: Agarwal et al., 2019a,b; Simchowitz et al., 2020)",
"adopt an approach to addressing this trajectory mismatch that we refer to as “online learning with",
"stationary costs\", or OLws. At each round t, the learner's adaptive policy πalg commits to a policy",
"(t), typically from a benchmark class II",
"The goal is to ensure that the iterates (t) attain low",
"regret on a proxy sequence of stationary",
"/costfunctions",
"At() that describes the loss the learner",
"would suffer at stage t under the fictional ",
"trale",
"OOr",
"a",
"would arise if she had played the policy T",
"at all stages uptotim",
"igsteady-statetrajectoryast→oo)",
"Since the stationar",
"low",
" regret on the sequence {t]",
"can be obtained by feeding",
"ST",
"thec",
"online learning algorithm. To relate",
"T5",
"regret on the",
"approaches use that the",
"iteratesproduced",
"endix B explains both the",
"general OLwS pa",
"further detail",
"The main",
"ionar",
"",
"y costs that arisein",
"oursetting.no",
"ently slowlytoyield logarithmic",
"regret via OLwS (Appendi",
"control wecall online learning",
"",
"",
"with advantages, or OLwA, which abandons",
"sao0",
"OS",
"and instead considers the control",
"theoreticadvantages of actionsrelativetotheunco",
"FYa",
"offline optimal policy *.Somewhat",
"miraculously, we find that these advantages remove the explicit dependence on the learner's state,",
"thereby eliminating the issue of trajectory mismatch described above. In particular, unlike OLws,",
"we do not need to verify that the iterates produced by our algorithm change slowly.",
"1.2",
"Our approach: Online learning with advantages",
"In this section we sketch the online learning with advantages (OLwA) technique underlying Ric-",
"catitron. Let * denote the optimal unconstrained policy given knowledge of the entire disturbance",
"sequence w, and let Q*(, u; w) be the associated Q-function (this quantity is formally defined in",
"Definition 3).The advantage with respect to π*",
", A*(u;c,w) := Qt(α, u; w) -Q*(w, u, π*(); w)",
"describes the difference between the total cost acc",
"mulated byselecting actionu in stater at time",
"and subsequently playing according to the optimal policy *, versus choosing u = *(c; w) as well.1",
"By the well-known performance difference lemma (Kakade, 2003), the relative cost of a policy is",
"equal the sum of the advantages under the states visited by said policy:2",
"T",
"JT(π;w) - JT(π*;w) =/ At(ut ;t,w).",
"(2)",
"t=1",
"ISince we use losses rather than rewards, “advantage\" refers to the advantage of * over u rather than the",
"advantage of u over *; the latter terminology is more common in reinforcement learning.",
"2See Lemma C.12 in Appendix D for a general statement of the performance difference lemma. The invocation",
"of the performance difference lemma here is slightly different from other results on online learning in MDPs such as",
"Even-Dar et al. (2009), in that the role of π and π* is swapped.",
"3"
] | {
"arxiv_id": "2003.00189",
"language": "en",
"page_idx": 2
} |
||
2003.00189-en-3 | 2003.00189-en-3.jpg | With this observation, the regret of any algorithm $\pi^\mathrm{alg}$ to a policy class $\Pi$ can be expressed as:
$$\mathrm{Reg}_{T}(\pi^{\mathrm{alg}};\Pi,\boldsymbol{w})=\sum_{t=1}^{T}\mathbf{A}_{t}^{\star}(u_{t}^{\mathrm{alg}};x_{t}^{\mathrm{alg}},\boldsymbol{w})-\inf_{\pi\epsilon\Pi}\sum_{t=1}^{T}\mathbf{A}_{t}^{\star}(u_{t}^{\pi};x_{t}^{\pi},\boldsymbol{w}).$$
(3)
The expression (3) suggests that a reasonable approach might be to run an online learner on the functions $\pi \mapsto \mathbf{A} _{t}^{\star }( u_{t}^{\pi }; x_{t}^{\pi }, \boldsymbol{w}) .$ However, there are two issues. First, the advantages in the first sum are evaluated on the states $x_{t}^{\mathrm{alg}}$under $\pi ^{\mathrm{alg}}$, and in the second sum under the comparator trajectories $x^\pi$ (trajectory mismatch). Second, like $\pi^\star$ itself, the advantages require knowledge of all future disturbances, which are not yet known to the learner at time $t.$ We show—somewhat miraculously—that if the control policies are parametrized using a particular optimal control law, the advantages do not depend on the state, and can be approximated using only finite lookahead.
Theorem 2 (informal). For control policies π with a suitable parametrization, the mapping π→ $\mathbf{A}_{t}^{\star}(u_{t}^{\pi};x_{t}^{\pi},\boldsymbol{w})$ can be arbitrarilily-well approximated by $a$ function $\pi\mapsto\widehat{\mathbf{A}}_{t;h}(\pi;w_{1:t+h})$ which (1) does not depend $on$ the $state$, (2) can $be$ determined $by$ the learner $at$ time $t+ h$, and (3) has $a$
$simple\textit{ quadratic structure. }$
The“magic”behind this theorem is that the functional dependence of the unconstrained optimal policy $\pi^\star(x;\boldsymbol{w})$ on the state $x$ is linear, and does not depend $\boldsymbol w$ (Theorem 3). As a consequence, the state-dependent portion of $\pi^\star$ can be built into the controller parametrization, leaving only the $\boldsymbol{w}$-dependent portion up to the online learner. In light of this result, we use online learning to $\hat{\text{ensure low regret on the sequence of loss functions }}f_{t}(\pi):=\widehat{\mathbf{A}}_{t;h}(\pi;w_{1:t+h});$ we address the fact that $f_t$ is only revealed to the learner after a delay of $h$ steps via a standard reduction (Joulani et al., 2013). We then show that for an appropriate controller parameterization $f_t(\pi)$ is exp-concave with respective to the learner's policy and hence second-order online learning algorithms attain regret (Hazan et al., 2007).
We refer the reader to Appendix B for an in-depth overview of the OLwS framework, its relationship to OLwA, and challenges associated with using these techniques to achieve logarithmic regret.
## 1.3 Related work
Linear control for known systems. Cohen et al. (2018) establish $\sqrt T$ regret for online control of known linear systems under stochastic noise and time varying quadratic cost. Agarwal et al. (2019a) achieve $\sqrt{T}$-regret with both adversarial disturbances and time varying, adversarially chosen loss functions $\ell_t$ via a reduction to online convex optimization with memory (Anava et al., 2015). Their approach adopts a“disturbance-action” policy parameterization (or, DAP), which we utilize as well (Definition 2). Certain previous results achieve logarithmic regret by making assumptions that ensure stationary costs are strongly convex, allowing for logarithmic regret and movement cost via Anava et al. (2015) or similar arguments. Abbasi-Yadkori et al. (2014) consider an online tracking problem with known system parameters zero exogenous noise. The absence of noise enables an approach based on MDP-E (see Appendix B.5), for which the relevant Q-functions in this setting are strongly convex, leading to logarithmic regret. More recently Agarwal et al. (2019b) showed that in the noisy setting the stationary costs $\lambda_t$ themselves are strongly convex in a disturbance-action parametrization, provided that the loss functions $\ell_t$ are strongly convex and the noise covariance is well-conditioned, which also leads to logarithmic regret. Simchowitz et al. (2020) show that this approach extends to“semi-adversarial” disturbances with a well-conditioned stochastic component
4 | [
[
"0.1152",
"0.0903",
"0.8594",
"0.0903",
"0.8594",
"0.108",
"0.1152",
"0.108"
],
[
"0.2566",
"0.1288",
"0.4019",
"0.1288",
"0.4019",
"0.1497",
"0.2566",
"0.1497"
],
[
"0.5073",
"0.13",
"0.74",
"0.1307",
"0.74",
"0.1497",
"0.5073",
"0.1483"
],
[
"0.857",
"0.132",
"0.884",
"0.132",
"0.884",
"0.1509",
"0.857",
"0.1509"
],
[
"0.4902",
"0.1338",
"0.5073",
"0.1338",
"0.5073",
"0.1427",
"0.4902",
"0.1427"
],
[
"0.3945",
"0.137",
"0.4004",
"0.137",
"0.4004",
"0.1445",
"0.3945",
"0.1445"
],
[
"0.4004",
"0.1383",
"0.4043",
"0.1383",
"0.4043",
"0.1433",
"0.4004",
"0.1433"
],
[
"0.406",
"0.1383",
"0.4126",
"0.1383",
"0.4126",
"0.1414",
"0.406",
"0.1414"
],
[
"0.423",
"0.1338",
"0.4958",
"0.1338",
"0.4958",
"0.1483",
"0.423",
"0.1483"
],
[
"0.4084",
"0.1483",
"0.4248",
"0.1483",
"0.4248",
"0.1578",
"0.4084",
"0.1578"
],
[
"0.569",
"0.1445",
"0.6235",
"0.1445",
"0.6235",
"0.159",
"0.569",
"0.159"
],
[
"0.1422",
"0.1704",
"0.8823",
"0.1718",
"0.8823",
"0.1895",
"0.1422",
"0.1881"
],
[
"0.1168",
"0.1913",
"0.8823",
"0.1913",
"0.8823",
"0.2058",
"0.1168",
"0.2058"
],
[
"0.1152",
"0.2084",
"0.398",
"0.2084",
"0.398",
"0.2216",
"0.1152",
"0.2216"
],
[
"0.392",
"0.2039",
"0.8823",
"0.207",
"0.8823",
"0.2247",
"0.392",
"0.2216"
],
[
"0.1152",
"0.2247",
"0.884",
"0.2247",
"0.884",
"0.2393",
"0.1152",
"0.2393"
],
[
"0.1168",
"0.2424",
"0.8823",
"0.2424",
"0.8823",
"0.2556",
"0.1168",
"0.2556"
],
[
"0.1168",
"0.2588",
"0.8823",
"0.2588",
"0.8823",
"0.2734",
"0.1168",
"0.2734"
],
[
"0.1152",
"0.2747",
"0.8857",
"0.276",
"0.8857",
"0.2903",
"0.1152",
"0.289"
],
[
"0.1152",
"0.3193",
"0.8813",
"0.3213",
"0.8813",
"0.336",
"0.1152",
"0.334"
],
[
"0.1152",
"0.336",
"0.884",
"0.3372",
"0.884",
"0.3547",
"0.1152",
"0.3542"
],
[
"0.1152",
"0.356",
"0.8823",
"0.356",
"0.8823",
"0.3706",
"0.1152",
"0.3706"
],
[
"0.1136",
"0.3738",
"0.3228",
"0.3738",
"0.3228",
"0.3882",
"0.1136",
"0.3882"
],
[
"0.1447",
"0.4016",
"0.8813",
"0.4016",
"0.8813",
"0.4153",
"0.1447",
"0.4153"
],
[
"0.1168",
"0.4192",
"0.884",
"0.4192",
"0.884",
"0.433",
"0.1168",
"0.433"
],
[
"0.1168",
"0.4363",
"0.8813",
"0.4363",
"0.8813",
"0.4495",
"0.1168",
"0.4495"
],
[
"0.1168",
"0.4526",
"0.8823",
"0.4526",
"0.8823",
"0.4673",
"0.1168",
"0.4673"
],
[
"0.1152",
"0.4702",
"0.8813",
"0.4702",
"0.8813",
"0.4849",
"0.1152",
"0.4849"
],
[
"0.1177",
"0.4873",
"0.8823",
"0.4873",
"0.8823",
"0.5005",
"0.1177",
"0.5005"
],
[
"0.1152",
"0.504",
"0.8823",
"0.504",
"0.8823",
"0.518",
"0.1152",
"0.518"
],
[
"0.1152",
"0.5215",
"0.884",
"0.5215",
"0.884",
"0.536",
"0.1152",
"0.536"
],
[
"0.1152",
"0.5386",
"0.2754",
"0.5386",
"0.2754",
"0.553",
"0.1152",
"0.553"
],
[
"0.1438",
"0.556",
"0.8813",
"0.556",
"0.8813",
"0.5693",
"0.1438",
"0.5693"
],
[
"0.1152",
"0.5728",
"0.8813",
"0.5728",
"0.8813",
"0.587",
"0.1152",
"0.587"
],
[
"0.1152",
"0.5923",
"0.1691",
"0.5923",
"0.1691",
"0.604",
"0.1152",
"0.604"
],
[
"0.1136",
"0.6235",
"0.2942",
"0.6235",
"0.2942",
"0.638",
"0.1136",
"0.638"
],
[
"0.1152",
"0.6514",
"0.884",
"0.6514",
"0.884",
"0.6665",
"0.1152",
"0.6665"
],
[
"0.1152",
"0.6704",
"0.8823",
"0.6704",
"0.8823",
"0.6846",
"0.1152",
"0.6846"
],
[
"0.1152",
"0.686",
"0.884",
"0.686",
"0.884",
"0.7007",
"0.1152",
"0.7007"
],
[
"0.1136",
"0.704",
"0.8813",
"0.7026",
"0.8813",
"0.7173",
"0.1136",
"0.7183"
],
[
"0.1152",
"0.72",
"0.8813",
"0.72",
"0.8813",
"0.735",
"0.1152",
"0.735"
],
[
"0.1152",
"0.738",
"0.8823",
"0.738",
"0.8823",
"0.7524",
"0.1152",
"0.7524"
],
[
"0.1152",
"0.756",
"0.8823",
"0.756",
"0.8823",
"0.771",
"0.1152",
"0.771"
],
[
"0.1136",
"0.7715",
"0.8823",
"0.773",
"0.8823",
"0.787",
"0.1136",
"0.786"
],
[
"0.1152",
"0.789",
"0.8813",
"0.789",
"0.8813",
"0.8037",
"0.1152",
"0.8037"
],
[
"0.1136",
"0.8066",
"0.8823",
"0.806",
"0.8823",
"0.821",
"0.1136",
"0.8213"
],
[
"0.1152",
"0.8237",
"0.8823",
"0.8237",
"0.8823",
"0.8384",
"0.1152",
"0.8384"
],
[
"0.11273",
"0.839",
"0.8823",
"0.8403",
"0.8823",
"0.855",
"0.11273",
"0.854"
],
[
"0.1168",
"0.858",
"0.88",
"0.858",
"0.88",
"0.8726",
"0.1168",
"0.8726"
],
[
"0.1152",
"0.875",
"0.8823",
"0.875",
"0.8823",
"0.8896",
"0.1152",
"0.8896"
],
[
"0.1168",
"0.8926",
"0.8823",
"0.8926",
"0.8823",
"0.907",
"0.1168",
"0.907"
],
[
"0.492",
"0.935",
"0.509",
"0.935",
"0.509",
"0.9497",
"0.492",
"0.9497"
]
] | [
"With this observation, the regret of any algorithm ralg to a policy class II can be expressed as:",
"Regr(πalg; II, w) -",
"alg,w)-infA;(ut;at,w).",
"(3)",
".o",
"",
"1",
"",
"At(ualg;",
"t",
"元ll台",
"The expression (3) suggests that a reasonable approach might be to run an online learner on",
"the functions π → A*(ut; αt, w). However, there are two issues. First, the advantages in the first",
"sum are evaluated on the states ",
"alg under xalg, and in the second sum under the comparator",
"trajectories α\" (trajectory mismatch). Second, like * itself, the advantages require knowledge of",
"all future disturbances, which are not yet known to the learner at time t.We show-somewhat",
"miraculouslythat if the control policies are parametrized using a particular optimal control law",
"the advantages do not depend on the state, and can be approximated using only finite lookahead.",
"Theorem 2 (informal). For control policies with a suitable parametrization, the mapping ",
"A,(uf ; rt, w) can be arbitrarilily-well approximated by a function π - At;h(π; Wi:t+h) which (1)",
"does not depend on the state, (2) can be determined by the learner at time t + h, and (3) has a",
"simplequadraticstructure.",
"The “magic\" behind this theorem is that the functional dependence of the unconstrained optimal",
"policy *(a; w) on the state is linear, and does not depend w (Theorem 3). As a consequence,",
"the state-dependent portion of * can be built into the controller parametrization, leaving only",
"the w-dependent portion up to the online learner. In light of this result, we use online learning to",
"ensure low regret on the sequence of loss functions ft(π) := At:h(π; wi1:t+h); we address the fact that",
"ft is only revealed to the learner after a delay of h steps via a standard reduction (Joulani et al.",
"2013). We then show that for an appropriate controller parameterization ft() is exp-concave with",
"respective to the learner's policy and hence second-order online learning algorithms attain regret",
"(Hazan et al., 2007).",
"We refer the reader to Appendix B for an in-depth overview of the oLws framework, its re",
"lationship to OLwA, and challenges associated with using these techniques to achieve logarithmic",
"regret.",
"1.3Related work",
"Linear control for known systems. Cohen et al. (2018) establish VT regret for online control of",
"known linear systems under stochastic noise and time varying quadratic cost. Agarwal et al. (2019a)",
"achieve VT-regret with both adversarial disturbances and time varying, adversarially chosen loss",
"functions lt via a reduction to online convex optimization with memory (Anava et al., 2015). Their",
"approach adopts a “disturbance-action\" policy parameterization (or, DAp), which we utilize as well",
"(Definition 2). Certain previous results achieve logarithmic regret by making assumptions that",
"ensure stationary costs are strongly convex, allowing for logarithmic regret and movement cost via",
"Anava et al. (2015) or similar arguments: Abbasi-Yadkori et al. (2014) consider an online tracking",
"problem with known system parameters zero exogenous noise.The absence of noise enables an",
"approach based on MDP-E (see Appendix B.5), for which the relevant Q-functions in this setting",
"are strongly convex, leading to logarithmic regret. More recently Agarwal et al. (2019b) showed that",
"in the noisy setting the stationary costs At themselves are strongly convex in a disturbance-action",
"parametrization, provided that the loss functions lt are strongly convex and the noise covariance",
"is well-conditioned, which also leads to logarithmic regret. Simchowitz et al. (2020) show that this",
"approach extends to “semi-adversarial\" disturbances with a well-conditioned stochastic component",
"4"
] | {
"arxiv_id": "2003.00189",
"language": "en",
"page_idx": 3
} |
||
2003.00189-en-4 | 2003.00189-en-4.jpg | and a possibly adversarial component. Our results (with the restriction that costs are quadratic) give the first logarithmic regret bounds for the fully adversarial setting and, to the best of our knowledge, give the first instance in online control where an exp-concave but not strongly convex parametrization attains logarithmic regret.
Linear control for unknown systems. For unknown systems, various works (Abbasi-Yadkori and Szepesvári, 2011; Faradonbeh et al., 2018; Cohen et al., 2019; Mania et al., 2019) establish $\sqrt{T}$-regret for fixed quadratic losses and stationary stochastic noise, which is optimal for this setting (Simchowitz and Foster, 2020; Cassel et al., 2020). Because of the stochastic nature of these problems, purely statistical techniques suffice. By combining these techniques with OCO with memory (Anava et al., 2015), other recent works have addressed both unknown dynamics and adversarial noise (Hazan et al., 2020; Simchowitz et al., 2020). (Cassel et al., 2020) show that logarithmic regret $is$ achievable under stochastic noise for systems $(A,B)$ where only $A$ is unknown, or where only $B$ is unknown and the optimal controller satisfes a non-degeneracy assumption
Online reinforcement learning. Online linear control belongs to a broader line of work on online reinforcement learning in (known or unknown) Markov decision processes with adversarial costs or transitions. Given the staggering breadth of work in this direction from the online learning, control, and RL communities, we focus on past contributions which are most closely related to our setting. As discussed earlier, essentially all prior approaches to online RL abide by the OLwS paradigm. Perhaps the first result in this direction is the MDP-E algorithm of Even-Dar et al. (2009), which attains $\sqrt T$ policy regret in a tabular MDP with known stationary dynamics and adversarially chosen rewards. Subsequent works (Abbasi-Yadkori et al., 2013) achieves $\sqrt{T}$-regret in a tabular setting where the both the rewards and transition kernels are selected by an adversary. A parallel line of work on adversarial tabular MDPs considers the episodic setting (Zimin and Neu, 2013; Rosenberg and Mansour, 2019), which alleviates the need to bound the movement costs between iterates.
$\textbf{Policy regret. All of the approaches described so far can be viewed as special cases of the general}$ problem of minimizing $policy\textit{ regret in online learning. A finite- memory formulation of the policy}$ regret benchmark was popularized by Arora et al.(2012). Anava et al.(2015) generalize this result to the $online\textit{ conve}x\textit{ optimization with memory setting and demonstrate that many popular}$ online learning algorithms naturally produce slow-moving iterates, yielding near-optimal policy regret bounds (see Appendix B.3.1 for detailed discussion). These results have found immediate application in online linear control (Agarwal et al., 2019a; Hazan et al., 2020; Simchowitz et al., 2020). However, the analysis of Anava et al. (2015) does not extend to give fast rates for the exp-concave loss functions which arise in our setting.
# 1.4 Preliminaries
We consider the linear control setting in (1). For normalization, we assume $\|w_t\|\leq1$ for all $t.$ We also assume $x_1=0.$ A comprehensive summary of all notation used throughout the paper is provided in Table 1 in the appendix.
Policies and trajectories. We consider policies $\pi$ parameterized as functions of $x_t$ and $\boldsymbol{w}$ via $u_t=\pi_t(x_t;\boldsymbol{w}).$ We assume that, when selecting action $u_t$ at time $t$, the learner has access to all states $x_1:t,u_{1:t-1}$, as well as $w_1: t- 1$ ( the latter assumption is without loss of generality by the identity
5 | [
[
"0.1168",
"0.09216",
"0.8823",
"0.09216",
"0.8823",
"0.1067",
"0.1168",
"0.1067"
],
[
"0.1152",
"0.10986",
"0.8813",
"0.1092",
"0.8813",
"0.1237",
"0.1152",
"0.125"
],
[
"0.1168",
"0.1276",
"0.8823",
"0.1276",
"0.8823",
"0.1414",
"0.1168",
"0.1414"
],
[
"0.1168",
"0.1445",
"0.4534",
"0.1445",
"0.4534",
"0.159",
"0.1168",
"0.159"
],
[
"0.1168",
"0.1799",
"0.8823",
"0.1799",
"0.8823",
"0.1945",
"0.1168",
"0.1945"
],
[
"0.1168",
"0.1976",
"0.8823",
"0.1976",
"0.8823",
"0.2122",
"0.1168",
"0.2122"
],
[
"0.1177",
"0.214",
"0.8823",
"0.214",
"0.8823",
"0.2291",
"0.1177",
"0.2291"
],
[
"0.1136",
"0.2317",
"0.8823",
"0.2311",
"0.8823",
"0.2456",
"0.1136",
"0.2468"
],
[
"0.1152",
"0.2488",
"0.8813",
"0.2474",
"0.8813",
"0.262",
"0.1152",
"0.2632"
],
[
"0.1152",
"0.2664",
"0.8813",
"0.2664",
"0.8813",
"0.2798",
"0.1152",
"0.2798"
],
[
"0.1168",
"0.2842",
"0.8813",
"0.2842",
"0.8813",
"0.2974",
"0.1168",
"0.2974"
],
[
"0.1152",
"0.3005",
"0.8823",
"0.3005",
"0.8823",
"0.315",
"0.1152",
"0.315"
],
[
"0.1152",
"0.318",
"0.853",
"0.318",
"0.853",
"0.332",
"0.1152",
"0.332"
],
[
"0.1168",
"0.3542",
"0.8813",
"0.3542",
"0.8813",
"0.3674",
"0.1168",
"0.3674"
],
[
"0.1152",
"0.3706",
"0.8823",
"0.3706",
"0.8823",
"0.385",
"0.1152",
"0.385"
],
[
"0.1136",
"0.3882",
"0.884",
"0.3882",
"0.884",
"0.4028",
"0.1136",
"0.4028"
],
[
"0.1136",
"0.4048",
"0.8823",
"0.4048",
"0.8823",
"0.4192",
"0.1136",
"0.4192"
],
[
"0.1152",
"0.4224",
"0.8823",
"0.4224",
"0.8823",
"0.4375",
"0.1152",
"0.4375"
],
[
"0.1168",
"0.4395",
"0.884",
"0.4395",
"0.884",
"0.4539",
"0.1168",
"0.4539"
],
[
"0.1168",
"0.4558",
"0.8823",
"0.4558",
"0.8823",
"0.4702",
"0.1168",
"0.4702"
],
[
"0.1152",
"0.4734",
"0.8823",
"0.4734",
"0.8823",
"0.488",
"0.1152",
"0.488"
],
[
"0.1136",
"0.4905",
"0.884",
"0.4912",
"0.884",
"0.506",
"0.1136",
"0.505"
],
[
"0.1152",
"0.5083",
"0.8823",
"0.5083",
"0.8823",
"0.5215",
"0.1152",
"0.5215"
],
[
"0.1168",
"0.5244",
"0.8823",
"0.5244",
"0.8823",
"0.539",
"0.1168",
"0.539"
],
[
"0.1152",
"0.5415",
"0.2507",
"0.5415",
"0.2507",
"0.5547",
"0.1152",
"0.5547"
],
[
"0.1136",
"0.5757",
"0.884",
"0.577",
"0.884",
"0.5947",
"0.1136",
"0.5933"
],
[
"0.1136",
"0.5947",
"0.884",
"0.5947",
"0.884",
"0.6123",
"0.1136",
"0.6123"
],
[
"0.1152",
"0.613",
"0.8823",
"0.613",
"0.8823",
"0.6274",
"0.1152",
"0.6274"
],
[
"0.1152",
"0.63",
"0.8813",
"0.63",
"0.8813",
"0.6445",
"0.1152",
"0.6445"
],
[
"0.1152",
"0.6465",
"0.8813",
"0.6465",
"0.8813",
"0.661",
"0.1152",
"0.661"
],
[
"0.1136",
"0.664",
"0.8823",
"0.6636",
"0.8823",
"0.678",
"0.1136",
"0.6787"
],
[
"0.1152",
"0.6816",
"0.8823",
"0.681",
"0.8823",
"0.696",
"0.1152",
"0.6963"
],
[
"0.1152",
"0.6987",
"0.8823",
"0.6987",
"0.8823",
"0.7134",
"0.1152",
"0.7134"
],
[
"0.1136",
"0.7153",
"0.5312",
"0.7163",
"0.5312",
"0.731",
"0.1136",
"0.73"
],
[
"0.1136",
"0.7505",
"0.2942",
"0.7505",
"0.2942",
"0.7646",
"0.1136",
"0.7646"
],
[
"0.1152",
"0.778",
"0.8857",
"0.778",
"0.8857",
"0.7925",
"0.1152",
"0.7925"
],
[
"0.1168",
"0.7944",
"0.8813",
"0.7944",
"0.8813",
"0.8096",
"0.1168",
"0.8096"
],
[
"0.1152",
"0.813",
"0.407",
"0.813",
"0.407",
"0.8276",
"0.1152",
"0.8276"
],
[
"0.1152",
"0.8486",
"0.8813",
"0.8486",
"0.8813",
"0.863",
"0.1152",
"0.863"
],
[
"0.1136",
"0.866",
"0.8813",
"0.8643",
"0.8813",
"0.879",
"0.1136",
"0.881"
],
[
"0.1111",
"0.882",
"0.8823",
"0.88",
"0.8823",
"0.898",
"0.1111",
"0.8994"
],
[
"0.492",
"0.9346",
"0.5073",
"0.9346",
"0.5073",
"0.9497",
"0.492",
"0.9497"
]
] | [
"and a possibly adversarial component. Our results (with the restriction that costs are quadratic)",
"give the first logarithmic regret bounds for the fully adversarial setting and, to the best of our",
"knowledge, give the firstinstance in online control where an exp-concave but not strongly convex",
"parametrization attains logarithmic regret.",
"Linear control for unknown systems. For unknown systems, various works (Abbasi-Yadkori",
"and Szepesvari, 2011; Faradonbeh et al., 2018; Cohen et al., 2019; Mania et al., 2019) establish",
"VT-regret for fixed quadratic losses and stationary stochastic noise, which is optimal for this",
"setting (Simchowitz and Foster, 2020; Cassel et al., 2020). Because of the stochastic nature of",
"these problems, purely statistical techniques suffice. By combining these technigues with OCO",
"with memory (Anava et al., 2015), other recent works have addressed both unknown dynamics",
"and adversarial noise (Hazan et al., 2020: Simchowitz et al., 2020). (Cassel et al., 2020) show that",
"logarithmic regret is achievable under stochastic noise for systems (A, B) where only A is unknown.",
"or where only B is unknown and the optimal controller satisfies a non-degeneracy assumption.",
"Online reinforcement learning.Online linear control belongs to a broader line of work on",
"online reinforcement learning in (known or unknown) Markov decision processes with adversarial",
"costs or transitions. Given the staggering breadth of work in this direction from the online learning,",
"control, and RL communities, we focus on past contributions which are most closely related to our",
"setting. As discussed earlier, essentially all prior approaches to online RL abide by the OLws",
"paradigm. Perhaps the first result in this direction is the MDP-E algorithm of Even-Dar et al.",
"(2009),which attains VT policy regret in a tabular MDP with known stationary dynamics and",
"adversarially chosen rewards. Subsequent works (Abbasi-Yadkori et al., 2013) achieves VT-regret",
"in a tabular setting where the both the rewards and transition kernels are selected by an adversary.",
"A parallel line of work on adversarial tabular MDPs considers the episodic setting (Zimin and",
"Neu, 2013; Rosenberg and Mansour, 2019), which alleviates the need to bound the movement costs",
"between iterates.",
"Policy regret.All of the approaches described so far can be viewed as special cases of the general",
"problem of minimizing policy regret in online learning: A finite-memory formulation of the policy",
"regret benchmark was popularized by Arora et al. (2012). Anava et al. (2015) generalize this",
"result to the online conver optimization with memory setting and demonstrate that many popularn",
"online learning algorithms naturally produce slow-moving iterates, yielding near-optimal policy",
"regret bounds (see Appendix B.3.1 for detailed discussion). These results have found immediate",
"application in online linear control (Agarwal et al., 2019a; Hazan et al., 2020; Simchowitz et al..",
"2020). However, the analysis of Anava et al. (2015) does not extend to give fast rates for the",
"exp-concave loss functions which arise in our setting.",
"1.4Preliminaries",
"We consider the linear control setting in (1).For normalization, we assumewt 1 for all t.",
"We also assume ci = 0. A comprehensive summary of all notation used throughout the paper is",
"provided in Table 1 in the appendix.",
"Policies and trajectories. We consider policies T parameterized as functions of &t and w via",
"ut = t(ct;w). We assume that, when selecting action ut at time t, the learner has access to all",
"states C1:t, u1:t-1, as well as wit-1 (the latter assumption is without loss of generality by the identity",
"5"
] | {
"arxiv_id": "2003.00189",
"language": "en",
"page_idx": 4
} |
||
2310.05295-en-0 | 2310.05295-en-0.jpg | Visual Storytelling with Question-Answer Plans
$\textbf{Danyang Liu, Mirella Lapata, Frank Keller}$
Institute for Language, Cognition and Computation
School of Informatics, University of Edinburgh
10 Crichton Street, Edinburgh EH8 9AB
danyang.liu@ed.ac.uk, {mlap, keller}@inf.ed.ac.uk
## Abstract
Visual storytelling aims to generate compelling narratives from image sequences. Existing models often focus on enhancing the representation of the image sequence, e.g., with external knowledge sources or advanced graph structures. Despite recent progress, the stories are often repetitive, illogical, and lacking in detail. To mitigate these issues, we present a novel framework which integrates visual representations with pretrained language models and planning. Our model translates the image sequence into a visual prefix, a sequence of continuous embeddings which language models can interpret. It also leverages a sequence of question-answer pairs as a blueprint plan for selecting salient visual concepts and determining how they should be assembled into a narrative. Automatic and human evaluation on the VIST benchmark (Huang et al., 2016) demonstrates that blueprint-based models generate stories that are more coherent, interesting, and natural compared to competitive baselines and state-of-the-art systems.
## 1 Introduction
Visual storytelling involves narrating an engaging and logically coherent story based on a sequence of images (see the example in Figure 1). The task lies at the intersection of natural language processing and computer vision and has recently attracted increasing interest from both communities (Wang et al., 2022; Hsu et al., 2021; Xu et al., 2021; Chen et al., 2021; Hsu et al., 2020; Wang et al., 2020; Huang et al., 2016). Visual storytelling differs from image captioning, which typically focuses on generating descriptive text, e.g., by identifying and depicting objects within an image. It requires a deeper understanding of how images and the events they illustrate relate to each other in order to create a convincing narrative.
Visual storytelling is commonly modeled as a two-stage process. The image sequence is frst
<FigureHere>
Figure 1: Blueprint annotation for a visual story. Colorcoded answers are extracted from the gold story. Questions are generated by feeding the answers and the gold story as context to a pretrained question generator.
encoded into a representation which typically includes image embeddings and detected objects. Subsequently, a decoder generates a story token by token based on the encoding of the image sequence. Recent work has mainly focused on enhancing the first stage of the generation process e.g., by leveraging external knowledge sources (Hsu et al., 2021; Chen et al., 2021; Hsu et al., 2020; Yang et al., 2019). Advanced representations for image sequences have also been explored, such as scene graphs (Hong et al., 2020) and story graphs (Hsu et al., 2021). Despite recent progress, these methods struggle to produce meaningful narratives, are prone to hallucination and repetition, often generate vague sentences, and have difficulty identifying salient visual concepts.
We attribute the lack of story quality to at least two reasons. Previous work on text-based genera- | [
[
"0.2551",
"0.0932",
"0.7446",
"0.0932",
"0.7446",
"0.1081",
"0.2551",
"0.1081"
],
[
"0.3132",
"0.133",
"0.692",
"0.133",
"0.692",
"0.1478",
"0.3132",
"0.1478"
],
[
"0.293",
"0.1509",
"0.7095",
"0.1509",
"0.7095",
"0.1644",
"0.293",
"0.1644"
],
[
"0.309",
"0.165",
"0.693",
"0.1663",
"0.693",
"0.1812",
"0.309",
"0.1799"
],
[
"0.339",
"0.1829",
"0.6665",
"0.1829",
"0.6665",
"0.1965",
"0.339",
"0.1965"
],
[
"0.2527",
"0.2007",
"0.748",
"0.2007",
"0.748",
"0.2144",
"0.2527",
"0.2144"
],
[
"0.2612",
"0.253",
"0.34",
"0.253",
"0.34",
"0.2673",
"0.2612",
"0.2673"
],
[
"0.1436",
"0.2798",
"0.4575",
"0.2808",
"0.4575",
"0.2957",
"0.1436",
"0.2944"
],
[
"0.3955",
"0.2944",
"0.4575",
"0.2976",
"0.4568",
"0.3088",
"0.3938",
"0.306"
],
[
"0.517",
"0.2935",
"0.8696",
"0.2935",
"0.8696",
"0.3071",
"0.517",
"0.3071"
],
[
"0.5205",
"0.306",
"0.869",
"0.306",
"0.869",
"0.3196",
"0.5205",
"0.3196"
],
[
"0.1453",
"0.2957",
"0.403",
"0.2957",
"0.403",
"0.3088",
"0.1453",
"0.3088"
],
[
"0.02351",
"0.31",
"0.06046",
"0.31",
"0.06046",
"0.437",
"0.02351",
"0.437"
],
[
"0.1436",
"0.3088",
"0.4592",
"0.3088",
"0.4592",
"0.3225",
"0.1436",
"0.3225"
],
[
"0.5205",
"0.3176",
"0.869",
"0.3176",
"0.869",
"0.3308",
"0.5205",
"0.3308"
],
[
"0.1453",
"0.3237",
"0.4592",
"0.3237",
"0.4592",
"0.3374",
"0.1453",
"0.3374"
],
[
"0.5205",
"0.33",
"0.8696",
"0.33",
"0.8696",
"0.3433",
"0.5205",
"0.3433"
],
[
"0.1453",
"0.3384",
"0.4592",
"0.3384",
"0.4592",
"0.3516",
"0.1453",
"0.3516"
],
[
"0.517",
"0.3425",
"0.8447",
"0.3425",
"0.8447",
"0.3557",
"0.517",
"0.3557"
],
[
"0.147",
"0.354",
"0.4575",
"0.354",
"0.4575",
"0.364",
"0.147",
"0.364"
],
[
"0.1453",
"0.3652",
"0.4575",
"0.3652",
"0.4575",
"0.379",
"0.1453",
"0.379"
],
[
"0.517",
"0.3652",
"0.8564",
"0.3652",
"0.8564",
"0.379",
"0.517",
"0.379"
],
[
"0.1453",
"0.3813",
"0.4592",
"0.3813",
"0.4592",
"0.3943",
"0.1453",
"0.3943"
],
[
"0.5225",
"0.379",
"0.693",
"0.379",
"0.693",
"0.3926",
"0.5225",
"0.3926"
],
[
"0.1486",
"0.3967",
"0.4575",
"0.3967",
"0.4575",
"0.4067",
"0.1486",
"0.4067"
],
[
"0.517",
"0.3938",
"0.7656",
"0.3938",
"0.7656",
"0.4067",
"0.517",
"0.4067"
],
[
"0.1453",
"0.4092",
"0.4575",
"0.4092",
"0.4575",
"0.4229",
"0.1453",
"0.4229"
],
[
"0.5225",
"0.4067",
"0.6094",
"0.4067",
"0.6094",
"0.4187",
"0.5225",
"0.4187"
],
[
"0.1453",
"0.4233",
"0.4575",
"0.4233",
"0.4575",
"0.437",
"0.1453",
"0.437"
],
[
"0.5205",
"0.4187",
"0.7676",
"0.4187",
"0.7676",
"0.4316",
"0.5205",
"0.4316"
],
[
"0.04114",
"0.4316",
"0.0487",
"0.4316",
"0.0487",
"0.437",
"0.04114",
"0.437"
],
[
"0.5225",
"0.4316",
"0.5977",
"0.4316",
"0.5977",
"0.4436",
"0.5225",
"0.4436"
],
[
"0.02016",
"0.4382",
"0.05878",
"0.437",
"0.06213",
"0.514",
"0.02351",
"0.5156"
],
[
"0.1453",
"0.4382",
"0.4592",
"0.4382",
"0.4592",
"0.452",
"0.1453",
"0.452"
],
[
"0.519",
"0.4453",
"0.791",
"0.4453",
"0.791",
"0.459",
"0.519",
"0.459"
],
[
"0.1453",
"0.452",
"0.4592",
"0.452",
"0.4592",
"0.4656",
"0.1453",
"0.4656"
],
[
"0.5225",
"0.459",
"0.636",
"0.459",
"0.636",
"0.4714",
"0.5225",
"0.4714"
],
[
"0.1453",
"0.4656",
"0.4592",
"0.4656",
"0.4592",
"0.4785",
"0.1453",
"0.4785"
],
[
"0.519",
"0.4727",
"0.7466",
"0.4727",
"0.7466",
"0.4863",
"0.519",
"0.4863"
],
[
"0.1453",
"0.4797",
"0.4592",
"0.4797",
"0.4592",
"0.4934",
"0.1453",
"0.4934"
],
[
"0.5225",
"0.4868",
"0.665",
"0.4868",
"0.665",
"0.4976",
"0.5225",
"0.4976"
],
[
"0.1453",
"0.4946",
"0.4575",
"0.4946",
"0.4575",
"0.5083",
"0.1453",
"0.5083"
],
[
"0.5205",
"0.5005",
"0.867",
"0.5005",
"0.867",
"0.514",
"0.5205",
"0.514"
],
[
"0.1453",
"0.5083",
"0.4592",
"0.5083",
"0.4592",
"0.5215",
"0.1453",
"0.5215"
],
[
"0.5156",
"0.513",
"0.693",
"0.513",
"0.693",
"0.527",
"0.5156",
"0.527"
],
[
"0.02184",
"0.521",
"0.06213",
"0.521",
"0.06213",
"0.73",
"0.02184",
"0.73"
],
[
"0.1453",
"0.5225",
"0.4575",
"0.5225",
"0.4575",
"0.536",
"0.1453",
"0.536"
],
[
"0.147",
"0.538",
"0.4575",
"0.538",
"0.4575",
"0.549",
"0.147",
"0.549"
],
[
"0.511",
"0.5435",
"0.882",
"0.5435",
"0.882",
"0.557",
"0.511",
"0.557"
],
[
"0.1427",
"0.55",
"0.4592",
"0.5503",
"0.4592",
"0.5654",
"0.1427",
"0.564"
],
[
"0.511",
"0.558",
"0.8833",
"0.558",
"0.8833",
"0.572",
"0.511",
"0.572"
],
[
"0.1427",
"0.564",
"0.4592",
"0.5654",
"0.4592",
"0.58",
"0.1427",
"0.579"
],
[
"0.511",
"0.5723",
"0.882",
"0.5723",
"0.882",
"0.586",
"0.511",
"0.586"
],
[
"0.1453",
"0.58",
"0.4592",
"0.58",
"0.4592",
"0.5933",
"0.1453",
"0.5933"
],
[
"0.1453",
"0.5933",
"0.3376",
"0.594",
"0.3376",
"0.608",
"0.1453",
"0.607"
],
[
"0.513",
"0.5884",
"0.8506",
"0.5884",
"0.8506",
"0.5986",
"0.513",
"0.5986"
],
[
"0.12177",
"0.6245",
"0.1293",
"0.6245",
"0.1293",
"0.6313",
"0.12177",
"0.6313"
],
[
"0.1453",
"0.6226",
"0.2603",
"0.6226",
"0.2603",
"0.6353",
"0.1453",
"0.6353"
],
[
"0.511",
"0.63",
"0.882",
"0.63",
"0.882",
"0.6436",
"0.511",
"0.6436"
],
[
"0.11584",
"0.6455",
"0.4878",
"0.648",
"0.4878",
"0.662",
"0.11584",
"0.66"
],
[
"0.51",
"0.6455",
"0.8833",
"0.6445",
"0.8833",
"0.659",
"0.51",
"0.66"
],
[
"0.11584",
"0.663",
"0.4895",
"0.663",
"0.4895",
"0.677",
"0.11584",
"0.677"
],
[
"0.513",
"0.662",
"0.882",
"0.662",
"0.882",
"0.676",
"0.513",
"0.676"
],
[
"0.11755",
"0.6797",
"0.4895",
"0.6797",
"0.4895",
"0.6934",
"0.11755",
"0.6934"
],
[
"0.511",
"0.6777",
"0.882",
"0.6777",
"0.882",
"0.6914",
"0.511",
"0.6914"
],
[
"0.11584",
"0.695",
"0.492",
"0.6953",
"0.492",
"0.7104",
"0.11584",
"0.709"
],
[
"0.513",
"0.695",
"0.882",
"0.695",
"0.882",
"0.708",
"0.513",
"0.708"
],
[
"0.11755",
"0.712",
"0.4878",
"0.712",
"0.4878",
"0.7256",
"0.11755",
"0.7256"
],
[
"0.513",
"0.7114",
"0.882",
"0.7114",
"0.882",
"0.7246",
"0.513",
"0.7246"
],
[
"0.11584",
"0.727",
"0.486",
"0.727",
"0.486",
"0.7407",
"0.11584",
"0.7407"
],
[
"0.511",
"0.727",
"0.885",
"0.727",
"0.885",
"0.7407",
"0.511",
"0.7407"
],
[
"0.11755",
"0.7437",
"0.4878",
"0.7437",
"0.4878",
"0.7573",
"0.11755",
"0.7573"
],
[
"0.513",
"0.742",
"0.8833",
"0.742",
"0.8833",
"0.756",
"0.513",
"0.756"
],
[
"0.11755",
"0.759",
"0.4895",
"0.759",
"0.4895",
"0.774",
"0.11755",
"0.774"
],
[
"0.511",
"0.7573",
"0.8833",
"0.7583",
"0.8833",
"0.7725",
"0.511",
"0.7715"
],
[
"0.11755",
"0.7754",
"0.4878",
"0.7754",
"0.4878",
"0.789",
"0.11755",
"0.789"
],
[
"0.51",
"0.774",
"0.881",
"0.774",
"0.881",
"0.7876",
"0.51",
"0.7876"
],
[
"0.11755",
"0.792",
"0.4895",
"0.792",
"0.4895",
"0.8057",
"0.11755",
"0.8057"
],
[
"0.511",
"0.7905",
"0.882",
"0.7905",
"0.882",
"0.8047",
"0.511",
"0.8047"
],
[
"0.11755",
"0.808",
"0.4878",
"0.808",
"0.4878",
"0.822",
"0.11755",
"0.822"
],
[
"0.511",
"0.807",
"0.882",
"0.807",
"0.882",
"0.821",
"0.511",
"0.821"
],
[
"0.11584",
"0.8237",
"0.4895",
"0.8223",
"0.4895",
"0.8374",
"0.11584",
"0.8384"
],
[
"0.51",
"0.8237",
"0.881",
"0.8237",
"0.881",
"0.8374",
"0.51",
"0.8374"
],
[
"0.1184",
"0.843",
"0.4878",
"0.843",
"0.4878",
"0.8525",
"0.1184",
"0.8525"
],
[
"0.511",
"0.839",
"0.8833",
"0.839",
"0.8833",
"0.8525",
"0.511",
"0.8525"
],
[
"0.11584",
"0.857",
"0.4895",
"0.857",
"0.4895",
"0.8706",
"0.11584",
"0.8706"
],
[
"0.509",
"0.854",
"0.882",
"0.855",
"0.882",
"0.869",
"0.509",
"0.868"
],
[
"0.11584",
"0.8726",
"0.287",
"0.8726",
"0.287",
"0.8857",
"0.11584",
"0.8857"
],
[
"0.511",
"0.8716",
"0.684",
"0.8716",
"0.684",
"0.8853",
"0.511",
"0.8853"
],
[
"0.1335",
"0.887",
"0.4895",
"0.888",
"0.4895",
"0.903",
"0.1335",
"0.902"
],
[
"0.5283",
"0.887",
"0.882",
"0.888",
"0.882",
"0.903",
"0.5283",
"0.902"
],
[
"0.11584",
"0.905",
"0.4895",
"0.905",
"0.4895",
"0.9185",
"0.11584",
"0.9185"
],
[
"0.5073",
"0.905",
"0.882",
"0.905",
"0.882",
"0.9185",
"0.5073",
"0.9185"
]
] | [
"Visual Storytelling with Question-Answer Plans",
"Danyang Liu, Mirella Lapata, Frank Keller",
"Institute for Language, Cognition and Computation",
"School of Informatics, University of Edinburgh",
"10 Crichton Street, Edinburgh EH8 9AB",
"danyang.liu@ed.ac.uk,(mlap,kellerj@inf.ed.ac.uk",
"Abstract",
"Visual storytelling aims to generate compelling",
"Existing",
"A small indie rock group performed for a music store in",
"our city. The girls doing harmonies in the songs were",
"narratives from image sequences.",
"",
"models often focus on enhancing the represen-",
"phenomenal. The band had three guitarists. One of the",
"tation of the image sequence, e.g., with external",
"guitarists resembled James Morrison, More independent",
"knowledge sources or advanced graph struc-",
"bands played for the grand opening of the music store.",
"tures.Despite recent progress,the stories are",
"often repetitive, illogical, and lacking in de-",
"Qo: What group performed for a music store in our city?",
"tail. To mitigate these issues, we present a",
"Ao: a small indie rock group",
"novelframeworkwhichintegratesvisualrep-",
"Q: What kind of music did the girls do?",
"resentations with pretrained language models",
"A,: harmonies",
"and planning. Our model translates the image",
"Q: Who did the harmonies in the songs?",
"",
"A,: the girls",
"['S]",
"sequence into a visual prefix, a sequence of",
"Q,: How many guitarists did the group have?",
"continuous embeddings which language mod-",
"A,: three guitarists",
"els can interpret. It also leverages a sequence",
"Q:Who resembled James Morrison?",
"of question-answer pairs as a blueprint plan",
"A:one of the guitarists",
"for selecting salient visual concepts and deter-",
"Q,: Who played for the grand opening of the music store?",
"mining how they should be assembled into a",
"A,:more independent bands",
"5251EA",
"narrative. Automatic and human evaluation",
"on the VIST benchmark (Huang et al.,2016)",
"Figure 1: Blueprint annotation for a visual story. Color-",
"demonstrates that blueprint-based models gen-",
"coded answers are extracted from the gold story. Ques-",
"erate stories that are more coherent, interesting.",
"tions are generated by feeding the answers and the gold",
"and natural compared to competitive baselines",
"and state-of-the-art systems.",
"story as context to a pretrained question generator.",
"1",
"Introduction",
"encoded into a representation which typically in-",
"Visual storytelling involves narrating an engaging",
"cludes image embeddings and detected objects.",
"and logically coherent story based on a sequence",
"Subsequently, a decoder generates a story token by",
"of images (see the example in Figure 1). The task",
"token based on the encoding of the image sequence.",
"lies at the intersection of natural language process-",
"Recent work has mainly focused on enhancing the",
"ing and computer vision and has recently attracted",
"first stage of the generation process e.g., by leverag.",
"increasing interest from both communities (Wang",
"ing external knowledge sources (Hsu et al., 2021;",
"et al., 2022; Hsu et al., 2021; Xu et al., 2021; Chen",
"Chen et al., 2021; Hsu et al., 2020; Yang et al..",
"et al., 2021; Hsu et al., 2020; Wang et al., 2020;",
"2019). Advanced representations for image se-",
"Huang et al., 2016). Visual storytelling differs from",
"quences have also been explored, such as scene",
"image captioning, which typically focuses on gen-",
"graphs (Hong et al., 2020) and story graphs (Hsu",
"erating descriptive text, e.g., by identifying and",
"et al., 2021). Despite recent progress, these meth-",
"depicting objects within an image. It requires a",
" ods struggle to produce meaningful narratives, are",
"deeperunderstandingofhowimagesandtheevents",
"prone to hallucination and repetition, often gener-",
"they illustrate relate to each other in order to create",
"ate vague sentences, and have difficulty identifying",
"a convincing narrative.",
"salient visual concepts.",
"Visual storytelling is commonly modeled as a",
"We attribute the lack of story quality to at least",
"two-stage process. The image sequence is first",
" two reasons. Previous work on text-based genera-"
] | {
"arxiv_id": "2310.05295",
"language": "en",
"page_idx": 0
} |
||
2310.05295-en-1 | 2310.05295-en-1.jpg | tion has demonstrated that planning can improve story coherence, allowing to control the trajectory of events, the characters described and their actions (Yao et al., 2019; Xu et al., 2018; Rashkin et al., 2020; Goldfarb-Tarrant et al., 2020a; Fan et al., 2019; Yang et al., 2022). However, planning has not been considered in the context of visual storytelling, existing models adopt black-box architectures which are not particularly controllable or interpretable. Another limitation concerns the nature of current models which are essentially trained from scratch, and as a result have limited language modelling and generalization capabilities (they only see multimodal training samples; see top of Figure 1). Although pretrained language models (Raffel et al., 2020; Lewis et al., 2020; Brown et al., 2020) have been widely adopted for generalpurpose story generation, their potential for visual storytelling remains unexplored.
In this work we propose an approach to visual
storytelling which integrates pretrained language
models with visual representations and incorpo-
rates an intermediate planning step before gener-
ating the full story. Our encoder translates the im-
age sequence into a visual prefix, a sequence of
continuous embeddings which language models
can interpret. Following Narayan et al. (2022), we
represent plans as a sequence of question-answer
pairs, called blueprints, which serve as a proxy for
content selection (i.e., what to say) and planning
(i.e., in what order). Blueprints are loosely related
to the Question-under-Discussion (QUD) theory of
discourse (Larsson, 2002; Roberts, 2012; De Kuthy
et al., 2020), which posits that text structure can be
analyzed by identifying implicit questions raised
and answered by subsequent spans of text. We
augment visual storytelling training data with story
blueprints (see Figure 1), which we obtain automat-
ically thanks to state-of-the-art question generation
technology.
We fine-tune pretrained language models to gen-
erate blueprints from image sequences and the sto-
ries based on them. We showcase two types of
storytelling models, which vary in how the plan-
ning mechanism is implemented. A top-down
model generates the blueprint first and then contin-
ues to generate the corresponding story in one go,
whereas an integrated model interleaves planning
with text generation rather than determining a plan
in advance; generation is iteratively conditioned on
the image input, the blueprint and the story gener-
ated so far. Experiments on the VIST benchmark (Huang et al., 2016) show that blueprint-based models generate more coherent, interesting, and humanlike stories compared to the state of the art and large language models (LLMs) like GPT-3.5, according to automatic and human evaluation.
### 2 $\textbf{Related Work}$
Visual Storytelling Huang et al.(2016) introduced visual storytelling as a vehicle for developing AI tools with human-like understanding of grounded event structure and linguistic abilities that go beyond descriptive language. While earlier work (Gonzalez-Rico and Fuentes-Pineda, 2018; Kim et al., 2018) employed simple encoderdecoder architectures (using CNNs to extract visual features and RNNs to generate text), more recent methods (Xu et al., 2021; Chen et al., 2021; Hsu et al., 2020; Yang et al., 2019) leverage external resources (e.g., ConceptNet) as a way of instilling commonsense reasoning skills. Sometimes, scene graphs are also used to model relations between objects (Lu et al., 2016; Hong et al., 2020; Wang et al., 2020). To our knowledge, none of these approaches make use of plan-based decoding. Hsu et al. (2021) construct a graph representing the image sequence (based on training data and external resources) and identify the highest scoring path as the best storyline encapsulated therein. The storyline can be viewed as a form of planning, however, on the encoder side.
Most existing approaches (Xu et al., 2021; Hsu et al., 2020; Wang et al., 2020; Yang et al., 2019) train Transformer models from scratch, with the exception of Chen et al. (2021), who employ a vanilla BART model as a baseline without task-specific adaptation. In contrast, our work leverages the language modeling and generalization capabilities of pretrained language models for visual storytelling. Planning and Generation In the domain of automatic story generation, planning has been effec-
tive at capturing the content and structure of stories. The generation process is often decomposed into two stages, namely planning an outline and then elaborating on it, e.g., by filling in specific details of a story. Plans have been represented as a sequence of event or phrase keywords (Yao et al., 2019; Xu et al., 2018; Rashkin et al., 2020), character actions (Liu et al., 2020), plot structures (Goldfarb-Tarrant et al., 2020a), and more elaborate descriptions including details about the setting | [
[
"0.11755",
"0.0861",
"0.4895",
"0.0861",
"0.4895",
"0.0998",
"0.11755",
"0.0998"
],
[
"0.511",
"0.0861",
"0.881",
"0.0861",
"0.881",
"0.0998",
"0.511",
"0.0998"
],
[
"0.1142",
"0.10156",
"0.4895",
"0.1027",
"0.4895",
"0.1164",
"0.1142",
"0.1152"
],
[
"0.511",
"0.10156",
"0.882",
"0.10156",
"0.882",
"0.1152",
"0.511",
"0.1152"
],
[
"0.11584",
"0.11816",
"0.4895",
"0.11816",
"0.4895",
"0.1318",
"0.11584",
"0.1318"
],
[
"0.513",
"0.11816",
"0.882",
"0.11816",
"0.882",
"0.1318",
"0.513",
"0.1318"
],
[
"0.11584",
"0.133",
"0.486",
"0.133",
"0.486",
"0.1467",
"0.11584",
"0.1467"
],
[
"0.51",
"0.133",
"0.881",
"0.1342",
"0.881",
"0.1484",
"0.51",
"0.1478"
],
[
"0.11584",
"0.1484",
"0.4878",
"0.1484",
"0.4878",
"0.1621",
"0.11584",
"0.1621"
],
[
"0.51",
"0.1497",
"0.882",
"0.1497",
"0.882",
"0.1644",
"0.51",
"0.1644"
],
[
"0.11755",
"0.165",
"0.4895",
"0.165",
"0.4895",
"0.1787",
"0.11755",
"0.1787"
],
[
"0.51",
"0.1663",
"0.7773",
"0.1663",
"0.7773",
"0.1799",
"0.51",
"0.1799"
],
[
"0.11584",
"0.1812",
"0.486",
"0.1812",
"0.486",
"0.1948",
"0.11584",
"0.1948"
],
[
"0.5156",
"0.1965",
"0.5425",
"0.1965",
"0.5425",
"0.2048",
"0.5156",
"0.2048"
],
[
"0.5366",
"0.1936",
"0.665",
"0.1936",
"0.665",
"0.2079",
"0.5366",
"0.2079"
],
[
"0.11584",
"0.2144",
"0.4902",
"0.2131",
"0.4902",
"0.228",
"0.11584",
"0.2286"
],
[
"0.11755",
"0.199",
"0.4895",
"0.199",
"0.4895",
"0.212",
"0.11755",
"0.212"
],
[
"0.511",
"0.2173",
"0.882",
"0.2185",
"0.882",
"0.2328",
"0.511",
"0.2322"
],
[
"0.11755",
"0.231",
"0.4895",
"0.231",
"0.4895",
"0.2446",
"0.11755",
"0.2446"
],
[
"0.511",
"0.2351",
"0.882",
"0.2351",
"0.882",
"0.2488",
"0.511",
"0.2488"
],
[
"0.11755",
"0.2488",
"0.4844",
"0.2488",
"0.4844",
"0.2588",
"0.11755",
"0.2588"
],
[
"0.511",
"0.2517",
"0.881",
"0.2517",
"0.881",
"0.2654",
"0.511",
"0.2654"
],
[
"0.11584",
"0.2632",
"0.4878",
"0.2632",
"0.4878",
"0.2766",
"0.11584",
"0.2766"
],
[
"0.511",
"0.2673",
"0.882",
"0.2673",
"0.882",
"0.2808",
"0.511",
"0.2808"
],
[
"0.11584",
"0.2798",
"0.4895",
"0.2798",
"0.4895",
"0.2935",
"0.11584",
"0.2935"
],
[
"0.51",
"0.284",
"0.882",
"0.284",
"0.882",
"0.2976",
"0.51",
"0.2976"
],
[
"0.11755",
"0.2957",
"0.4878",
"0.2957",
"0.4878",
"0.3088",
"0.11755",
"0.3088"
],
[
"0.511",
"0.2986",
"0.881",
"0.2986",
"0.881",
"0.3123",
"0.511",
"0.3123"
],
[
"0.11584",
"0.31",
"0.4895",
"0.31",
"0.4895",
"0.3237",
"0.11584",
"0.3237"
],
[
"0.511",
"0.3142",
"0.882",
"0.3154",
"0.882",
"0.33",
"0.511",
"0.3289"
],
[
"0.11755",
"0.3267",
"0.4878",
"0.3267",
"0.4878",
"0.3403",
"0.11755",
"0.3403"
],
[
"0.513",
"0.3333",
"0.882",
"0.3333",
"0.882",
"0.3467",
"0.513",
"0.3467"
],
[
"0.11755",
"0.3433",
"0.4895",
"0.3433",
"0.4895",
"0.357",
"0.11755",
"0.357"
],
[
"0.51",
"0.3474",
"0.881",
"0.3474",
"0.881",
"0.361",
"0.51",
"0.361"
],
[
"0.11755",
"0.3599",
"0.4878",
"0.3599",
"0.4878",
"0.3735",
"0.11755",
"0.3735"
],
[
"0.51",
"0.364",
"0.882",
"0.364",
"0.882",
"0.3777",
"0.51",
"0.3777"
],
[
"0.11755",
"0.376",
"0.357",
"0.376",
"0.357",
"0.3896",
"0.11755",
"0.3896"
],
[
"0.511",
"0.3813",
"0.882",
"0.3813",
"0.882",
"0.3943",
"0.511",
"0.3943"
],
[
"0.1351",
"0.3984",
"0.4878",
"0.3984",
"0.4878",
"0.412",
"0.1351",
"0.412"
],
[
"0.509",
"0.3967",
"0.881",
"0.3967",
"0.881",
"0.411",
"0.509",
"0.411"
],
[
"0.1142",
"0.4133",
"0.4895",
"0.4146",
"0.4895",
"0.4287",
"0.1142",
"0.4275"
],
[
"0.51",
"0.4133",
"0.882",
"0.4133",
"0.882",
"0.427",
"0.51",
"0.427"
],
[
"0.11584",
"0.4312",
"0.492",
"0.4312",
"0.492",
"0.4448",
"0.11584",
"0.4448"
],
[
"0.51",
"0.4287",
"0.882",
"0.4275",
"0.882",
"0.4424",
"0.51",
"0.4436"
],
[
"0.11584",
"0.4478",
"0.4937",
"0.4478",
"0.4937",
"0.4614",
"0.11584",
"0.4614"
],
[
"0.513",
"0.4448",
"0.882",
"0.4448",
"0.882",
"0.459",
"0.513",
"0.459"
],
[
"0.11584",
"0.4631",
"0.492",
"0.462",
"0.492",
"0.4768",
"0.11584",
"0.478"
],
[
"0.51",
"0.459",
"0.882",
"0.4602",
"0.882",
"0.4744",
"0.51",
"0.4739"
],
[
"0.11755",
"0.4797",
"0.4895",
"0.4797",
"0.4895",
"0.4934",
"0.11755",
"0.4934"
],
[
"0.511",
"0.4768",
"0.882",
"0.4768",
"0.882",
"0.4905",
"0.511",
"0.4905"
],
[
"0.11584",
"0.4954",
"0.4878",
"0.4954",
"0.4878",
"0.509",
"0.11584",
"0.509"
],
[
"0.509",
"0.4922",
"0.882",
"0.4934",
"0.882",
"0.5083",
"0.509",
"0.5073"
],
[
"0.11755",
"0.511",
"0.4878",
"0.511",
"0.4878",
"0.525",
"0.11755",
"0.525"
],
[
"0.51",
"0.5103",
"0.879",
"0.5103",
"0.879",
"0.524",
"0.51",
"0.524"
],
[
"0.11755",
"0.528",
"0.4878",
"0.528",
"0.4878",
"0.5415",
"0.11755",
"0.5415"
],
[
"0.51",
"0.5254",
"0.882",
"0.5254",
"0.882",
"0.539",
"0.51",
"0.539"
],
[
"0.11584",
"0.5435",
"0.4895",
"0.542",
"0.4895",
"0.557",
"0.11584",
"0.558"
],
[
"0.509",
"0.5405",
"0.881",
"0.5405",
"0.881",
"0.554",
"0.509",
"0.554"
],
[
"0.11755",
"0.56",
"0.486",
"0.56",
"0.486",
"0.5737",
"0.11755",
"0.5737"
],
[
"0.513",
"0.558",
"0.8833",
"0.558",
"0.8833",
"0.572",
"0.513",
"0.572"
],
[
"0.11755",
"0.576",
"0.486",
"0.576",
"0.486",
"0.589",
"0.11755",
"0.589"
],
[
"0.51",
"0.5737",
"0.663",
"0.5737",
"0.663",
"0.5874",
"0.51",
"0.5874"
],
[
"0.11584",
"0.5913",
"0.4878",
"0.5913",
"0.4878",
"0.605",
"0.11584",
"0.605"
],
[
"0.5283",
"0.589",
"0.882",
"0.5884",
"0.882",
"0.6025",
"0.5283",
"0.604"
],
[
"0.11584",
"0.607",
"0.486",
"0.607",
"0.486",
"0.6206",
"0.11584",
"0.6206"
],
[
"0.51",
"0.604",
"0.881",
"0.604",
"0.881",
"0.618",
"0.51",
"0.618"
],
[
"0.11755",
"0.6235",
"0.4878",
"0.6235",
"0.4878",
"0.637",
"0.11755",
"0.637"
],
[
"0.511",
"0.6216",
"0.882",
"0.6216",
"0.882",
"0.6353",
"0.511",
"0.6353"
],
[
"0.11584",
"0.64",
"0.4878",
"0.64",
"0.4878",
"0.654",
"0.11584",
"0.654"
],
[
"0.51",
"0.638",
"0.882",
"0.638",
"0.882",
"0.652",
"0.51",
"0.652"
],
[
"0.1142",
"0.6562",
"0.4895",
"0.655",
"0.4895",
"0.6694",
"0.1142",
"0.6704"
],
[
"0.51",
"0.655",
"0.882",
"0.655",
"0.882",
"0.6685",
"0.51",
"0.6685"
],
[
"0.1142",
"0.6714",
"0.4878",
"0.6714",
"0.4878",
"0.685",
"0.1142",
"0.685"
],
[
"0.509",
"0.6694",
"0.882",
"0.6694",
"0.882",
"0.683",
"0.509",
"0.683"
],
[
"0.11584",
"0.687",
"0.4937",
"0.6885",
"0.4937",
"0.703",
"0.11584",
"0.7017"
],
[
"0.51",
"0.687",
"0.8833",
"0.685",
"0.8833",
"0.6997",
"0.51",
"0.7017"
],
[
"0.1142",
"0.703",
"0.4895",
"0.7036",
"0.4895",
"0.7188",
"0.1142",
"0.7173"
],
[
"0.511",
"0.7017",
"0.881",
"0.7017",
"0.881",
"0.7163",
"0.511",
"0.7163"
],
[
"0.1142",
"0.7188",
"0.2065",
"0.72",
"0.2065",
"0.7354",
"0.1142",
"0.733"
],
[
"0.513",
"0.727",
"0.882",
"0.727",
"0.882",
"0.7407",
"0.513",
"0.7407"
],
[
"0.1368",
"0.7437",
"0.4902",
"0.7437",
"0.4902",
"0.7583",
"0.1368",
"0.7583"
],
[
"0.51",
"0.7446",
"0.882",
"0.7446",
"0.882",
"0.7583",
"0.51",
"0.7583"
],
[
"0.11584",
"0.759",
"0.4895",
"0.759",
"0.4895",
"0.7725",
"0.11584",
"0.7725"
],
[
"0.511",
"0.7603",
"0.882",
"0.7603",
"0.882",
"0.774",
"0.511",
"0.774"
],
[
"0.11755",
"0.7754",
"0.4895",
"0.7754",
"0.4895",
"0.789",
"0.11755",
"0.789"
],
[
"0.51",
"0.7754",
"0.882",
"0.7754",
"0.882",
"0.789",
"0.51",
"0.789"
],
[
"0.11755",
"0.7915",
"0.4937",
"0.7915",
"0.4937",
"0.8047",
"0.11755",
"0.8047"
],
[
"0.51",
"0.792",
"0.882",
"0.792",
"0.882",
"0.8057",
"0.51",
"0.8057"
],
[
"0.11755",
"0.808",
"0.4895",
"0.808",
"0.4895",
"0.822",
"0.11755",
"0.822"
],
[
"0.51",
"0.808",
"0.882",
"0.808",
"0.882",
"0.822",
"0.51",
"0.822"
],
[
"0.1142",
"0.8247",
"0.4902",
"0.8247",
"0.4902",
"0.8384",
"0.1142",
"0.8384"
],
[
"0.513",
"0.8237",
"0.882",
"0.8237",
"0.882",
"0.8384",
"0.513",
"0.8384"
],
[
"0.11755",
"0.8413",
"0.4902",
"0.8413",
"0.4902",
"0.855",
"0.11755",
"0.855"
],
[
"0.51",
"0.8403",
"0.882",
"0.8403",
"0.882",
"0.854",
"0.51",
"0.854"
],
[
"0.11584",
"0.857",
"0.4878",
"0.857",
"0.4878",
"0.8706",
"0.11584",
"0.8706"
],
[
"0.51",
"0.8555",
"0.882",
"0.855",
"0.882",
"0.869",
"0.51",
"0.8706"
],
[
"0.11755",
"0.8726",
"0.4895",
"0.8726",
"0.4895",
"0.8857",
"0.11755",
"0.8857"
],
[
"0.511",
"0.8726",
"0.8833",
"0.8726",
"0.8833",
"0.8857",
"0.511",
"0.8857"
],
[
"0.11584",
"0.8896",
"0.4895",
"0.8896",
"0.4895",
"0.903",
"0.11584",
"0.903"
],
[
"0.51",
"0.888",
"0.8833",
"0.888",
"0.8833",
"0.902",
"0.51",
"0.902"
],
[
"0.11584",
"0.905",
"0.4937",
"0.905",
"0.4937",
"0.9194",
"0.11584",
"0.9194"
],
[
"0.509",
"0.905",
"0.881",
"0.905",
"0.881",
"0.9185",
"0.509",
"0.9185"
]
] | [
"tion has demonstrated that planning can improve",
"ated so far. Experiments on the VIST benchmark",
" story coherence, allowing to control the trajectory",
"(Huang et al., 2016) show that blueprint-based mod-",
"of events, the characters described and their ac-",
"els generate more coherent, interesting, and human-",
"tions (Yao et al., 2019: Xu et al., 2018; Rashkin",
"like stories compared to the state of the art and large",
"et al.. 2020: Goldfarb-Tarrant et al.. 2020a: Fan",
"language models (LLMs) like GPT-3.5, according",
"et al., 2019: Yang et al., 2022). However, plan-",
" to automatic and human evaluation.",
"ning has not been considered in the context of vi-",
"2",
"Related Work",
"architectures which are not particularly control-",
"sual storytelling, existing models adopt black-box",
"Visual Storytelling Huang et al. (2016) intro-",
"lable or interpretable. Another limitation concerns",
"duced visual storytelling as a vehicle for devel-",
"thenature of currentmodels whichareessentially",
"oping AI tools with human-like understanding",
"trained from scratch, and as a result have limited",
"of grounded event structure and linguistic abili-",
"language modelling and generalization capabilities",
"ties that go beyond descriptive language. While",
"(they only see multimodal training samples; see top",
"earlier work (Gonzalez-Rico and Fuentes-Pineda.",
"of Figure 1). Although pretrained language mod-",
"2018; Kim et al., 2018) employed simple encoder-",
"els (Raffel et al., 2020; Lewis et al., 2020; Brown",
"decoder architectures (using CNNs to extract visual",
"et al., 2020) have been widely adopted for general-",
"features and RNNs to generate text), more recent",
"purpose story generation, their potential for visual",
"methods (Xu et al., 2021; Chen et al., 2021; Hsu",
"storytelling remains unexplored.",
"et al., 2020; Yang et al., 2019)leverage external",
"In this work we propose an approach to visual",
"resources (e.g., ConceptNet) as a way of instilling",
"storytelling which integrates pretrained language",
"commonsense reasoning skills. Sometimes, scene",
"models with visual representations and incorpo-",
"graphs are also used to model relations between",
"rates an intermediate planning step before gener-",
"objects (Lu et al., 2016; Hong et al., 2020; Wang",
"ating the full story. Our encoder translates the im-",
"et al., 2020). To our knowledge, none of these ap",
"age sequence into a visual prefix, a sequence of",
"proaches make use of plan-based decoding. Hsu",
"continuous embeddings which language models",
"et al. (2021) construct a graph representing the im-",
"can interpret. Following Narayan et al. (2022), we",
"age sequence (based on training data and external",
"represent plans as a sequence of question-answer",
"resources) and identify the highest scoring path as",
"pairs, called blueprints, which serve as a proxy for",
" the best storyline encapsulated therein. The story",
"content selection (i.e., what to say) and planning",
"line can be viewed as a form of planning, however,",
"(i.e., in what order). Blueprints are loosely related",
"on the encoder side.",
"to the Question-under-Discussion (QUD) theory of",
"Most existing approaches (Xu et al., 2021; Hsu",
"discourse (Larsson, 2002: Roberts, 2012; De Kuthy",
"et al., 2020; Wang et al., 2020: Yang et al., 2019)",
"et al., 2020), which posits that text structure can be",
"train Transformer models from scratch, with the ex-",
"analyzed by identifying implicit questions raised",
"ception of Chen et al. (2021), who employ a vanilla",
"and answered by subsequent spans of text. We",
"BART model as a baseline without task-specific",
" augment visual storytelling training data with story",
" adaptation. In contrast, our work leverages the lan-",
"blueprints (see Figure I), which we obtain automat-",
"guage modeling and generalization capabilities of",
"ically thanks to state-of-the-art question generation",
"pretrained language models for visual storytelling",
"technology.",
"Planning and GenerationIn the domain of au-",
"We fine-tune pretrained language models to gen-",
"tomatic story generation, planning has been effec",
"erate blueprints from image sequences and the sto-",
"tive at capturing the content and structure of sto-",
"ries based on them. We showcase two types of",
"ries. The generation process is often decomposed",
"storytelling models, which vary in how the plan-",
" into two stages, namely planning an outline and",
"ning mechanism is implemented. A top-down",
"then elaborating on it, e.g., by filling in specific",
" model generates the blueprint first and then contin-",
"details of a story. Plans have been represented",
"ues to generate the corresponding story in one go,",
"as a sequence of event or phrase keywords (Yao",
"whereas an integrated model interleaves planning",
"et al., 2019; Xu et al., 2018; Rashkin et al., 2020),",
"with text generation rather than determining a plan",
"character actions (Liu et al., 2020), plot structures",
"in advance; generation is iteratively conditioned on",
"(Goldfarb-Tarrant et al., 2020a), and more elabo-",
"the image input, the blueprint and the story gener-",
"rate descriptions including details about the setting"
] | {
"arxiv_id": "2310.05295",
"language": "en",
"page_idx": 1
} |
||
2310.05295-en-2 | 2310.05295-en-2.jpg | of the story, its characters, and main plot points
(Yang et al., 2022).
The idea of having a separate planning stage has also been explored for other text generation tasks including summarization (Narayan et al., 2021; Liu and Chen, 2021) and data-to-text generation (Moryossef et al., 2019; Puduppully et al., 2022). Our work is closest to Narayan et al. (2022) who propose the use of question-answer pairs as intermediate plans for summarization. However, their approach is designed for descriptive text. Our work extends their framework to a multimodal setting, where the input consists of image sequences, and the output are narratives characterized by more abstract and figurative language.
## 3 Blueprint-based Visual Storytelling
Let $I$ represent a sequence of $k$ images, denoted as $\{v_1,v_2...,v_k\}.$ Given this input, our goal is to generate a blueprint plan $B$ (i.e., an ordered set of question-answer pairs) and a story $S$ based on it. Most generation datasets do not include blueprint annotations, and visual storytelling is no exception. We first describe how we automatically obtain $\{I_i,B_i,S_i\}_{i=1}^N$ training data samples (Section 3.1), and then introduce our story generation models (Section 3.2).
## 3.1 Blueprint Annotation
Let $\{I_i,S_i\}_{i=1}^N$ denote a dataset consisting of pairs of image sequences and their corresponding stories. We automatically create blueprint $B_i$ based on story $S_i$ using state-of-the-art question generators (Romero, 2021; Raffel et al., 2020), coupled with a filtering procedure to remove repetitions and ill-formed questions.
The generation of question-answer pairs involves two steps, namely answer extraction and question generation. In the context of storytelling, capturing key events is crucial for a compelling narrative. While noun phrases and named entities are commonly recognized as significant content units in other tasks such as summarization (Narayan et al., 2022; Deutsch and Roth, 2023), verb phrases also play a vital role in conveying story dynamics, actions, and relationships (Trabasso et al., 1989; Eisenberg and Finlayson, 2017; Liu et al., 2020). Therefore, in addition to named entities and noun phrases, we also extract verb phrases as answer candidates using the spaCy library.
We then generate questions for answer candi-
<FigureHere>
Figure 2: Visual prefix construction. The pretrained image encoder and concept detector are frozen. $FF$ refers to a feed-forward layer, $CD$ and $EL$ denote a concept detector and embedding layer, respectively
dates with a T5 model (Raffel et al., 2020; Romero, 2021) fine-tuned on the SQuAD reading comprehension dataset (Rajpurkar et al., 2018). The answer and the story are provided as context to predict the corresponding question. We decontextualize stories by replacing pronouns with their corresponding head mentions, using a state-of-the-art coreference resolution model (Dobrovolskii, 2021).
Question-answer pairs are subsequently filtered to eliminate noise which is unavoidable due to the automatic preprocessing steps mentioned above. We thus remove any question-answer pairs where the answer is already present in the question. We also employ a round-trip consistency check ( Alberti et al., 2019) which discards questions if they yield answers different from those used to generate them.
## 3.2 Blueprint Models
Our approach leverages the generation capabilities of pre-trained sequence-to-sequence models. As our backbone model, we employ BART-base (Lewis et al., 2020) which has been fine-tuned for text generation. We adapt this model to our visual storytelling task in two ways. Aside from enabling the generation of blueprints, we convert the image sequence to a visual prefx which the pretrained language model can interpret. The pretrained language model is prompted with this prefix to generate the blueprint, and eventually the story.
Visual Prefix Construction Our model needs to grasp what the image sequence is about, e.g., the depicted objects, actions, and their associations. Drawing inspiration from recent advances in vision and language research (Mokady et al., 2021; Tsimpoukelli et al., 2021; Alayrac et al., 2022; Zhai et al., 2022; Liu et al., 2023; Huang et al., 2023), we translate the input sequence of images into a | [
[
"0.11755",
"0.0861",
"0.486",
"0.0861",
"0.486",
"0.0998",
"0.11755",
"0.0998"
],
[
"0.8286",
"0.0885",
"0.867",
"0.0885",
"0.867",
"0.0998",
"0.8286",
"0.0998"
],
[
"0.7407",
"0.0944",
"0.7886",
"0.0968",
"0.7886",
"0.10394",
"0.7407",
"0.10156"
],
[
"0.1184",
"0.10156",
"0.2603",
"0.10156",
"0.2603",
"0.1152",
"0.1184",
"0.1152"
],
[
"0.7446",
"0.1051",
"0.785",
"0.1051",
"0.785",
"0.11224",
"0.7446",
"0.11224"
],
[
"0.827",
"0.10095",
"0.8657",
"0.10095",
"0.8657",
"0.1134",
"0.827",
"0.1134"
],
[
"0.1368",
"0.11816",
"0.4844",
"0.11816",
"0.4844",
"0.1318",
"0.1368",
"0.1318"
],
[
"0.11755",
"0.1354",
"0.4878",
"0.1354",
"0.4878",
"0.1484",
"0.11755",
"0.1484"
],
[
"0.11755",
"0.1509",
"0.4878",
"0.1509",
"0.4878",
"0.1644",
"0.11755",
"0.1644"
],
[
"0.644",
"0.1467",
"0.7",
"0.1467",
"0.7",
"0.1567",
"0.644",
"0.1567"
],
[
"0.65",
"0.1592",
"0.6987",
"0.1592",
"0.6987",
"0.1704",
"0.65",
"0.1704"
],
[
"0.11755",
"0.1663",
"0.486",
"0.1663",
"0.486",
"0.1799",
"0.11755",
"0.1799"
],
[
"0.7466",
"0.1609",
"0.785",
"0.1609",
"0.785",
"0.165",
"0.7466",
"0.165"
],
[
"0.7466",
"0.1704",
"0.784",
"0.1704",
"0.784",
"0.1746",
"0.7466",
"0.1746"
],
[
"0.11755",
"0.1829",
"0.486",
"0.1829",
"0.486",
"0.1965",
"0.11755",
"0.1965"
],
[
"0.11755",
"0.1996",
"0.4878",
"0.1996",
"0.4878",
"0.2131",
"0.11755",
"0.2131"
],
[
"0.11584",
"0.2162",
"0.4895",
"0.2156",
"0.4895",
"0.2286",
"0.11584",
"0.2299"
],
[
"0.11755",
"0.2322",
"0.4895",
"0.2322",
"0.4895",
"0.2452",
"0.11755",
"0.2452"
],
[
"0.511",
"0.2299",
"0.8833",
"0.2299",
"0.8833",
"0.2446",
"0.511",
"0.2446"
],
[
"0.1142",
"0.2477",
"0.4895",
"0.2465",
"0.4895",
"0.2612",
"0.1142",
"0.2625"
],
[
"0.509",
"0.2446",
"0.882",
"0.2434",
"0.882",
"0.257",
"0.509",
"0.2583"
],
[
"0.11584",
"0.2642",
"0.4895",
"0.2642",
"0.4895",
"0.2778",
"0.11584",
"0.2778"
],
[
"0.511",
"0.2588",
"0.882",
"0.2588",
"0.882",
"0.2725",
"0.511",
"0.2725"
],
[
"0.51",
"0.2737",
"0.858",
"0.2737",
"0.858",
"0.2874",
"0.51",
"0.2874"
],
[
"0.11755",
"0.2798",
"0.4878",
"0.2798",
"0.4878",
"0.2935",
"0.11755",
"0.2935"
],
[
"0.11584",
"0.2944",
"0.4895",
"0.2957",
"0.4895",
"0.31",
"0.11584",
"0.3088"
],
[
"0.1142",
"0.31",
"0.34",
"0.3123",
"0.34",
"0.3267",
"0.1142",
"0.3247"
],
[
"0.511",
"0.3135",
"0.882",
"0.3142",
"0.882",
"0.3289",
"0.511",
"0.328"
],
[
"0.513",
"0.3308",
"0.882",
"0.3308",
"0.882",
"0.3445",
"0.513",
"0.3445"
],
[
"0.1142",
"0.3403",
"0.451",
"0.3403",
"0.451",
"0.3552",
"0.1142",
"0.3552"
],
[
"0.51",
"0.3467",
"0.8833",
"0.3457",
"0.8833",
"0.3599",
"0.51",
"0.361"
],
[
"0.511",
"0.3635",
"0.8833",
"0.3635",
"0.8833",
"0.3772",
"0.511",
"0.3772"
],
[
"0.11584",
"0.3694",
"0.486",
"0.3694",
"0.486",
"0.383",
"0.11584",
"0.383"
],
[
"0.513",
"0.379",
"0.885",
"0.379",
"0.885",
"0.3938",
"0.513",
"0.3938"
],
[
"0.11584",
"0.3855",
"0.4878",
"0.3855",
"0.4878",
"0.3997",
"0.11584",
"0.3997"
],
[
"0.511",
"0.3955",
"0.8833",
"0.3955",
"0.8833",
"0.4092",
"0.511",
"0.4092"
],
[
"0.11584",
"0.402",
"0.486",
"0.402",
"0.486",
"0.415",
"0.11584",
"0.415"
],
[
"0.511",
"0.411",
"0.882",
"0.411",
"0.882",
"0.4246",
"0.511",
"0.4246"
],
[
"0.11755",
"0.4175",
"0.4878",
"0.4175",
"0.4878",
"0.4312",
"0.11755",
"0.4312"
],
[
"0.511",
"0.427",
"0.881",
"0.427",
"0.881",
"0.4407",
"0.511",
"0.4407"
],
[
"0.1184",
"0.4353",
"0.486",
"0.4353",
"0.486",
"0.4453",
"0.1184",
"0.4453"
],
[
"0.531",
"0.4436",
"0.882",
"0.4436",
"0.882",
"0.4573",
"0.531",
"0.4573"
],
[
"0.11755",
"0.4495",
"0.4895",
"0.4495",
"0.4895",
"0.4631",
"0.11755",
"0.4631"
],
[
"0.11584",
"0.4656",
"0.486",
"0.4656",
"0.486",
"0.4785",
"0.11584",
"0.4785"
],
[
"0.511",
"0.4602",
"0.882",
"0.4602",
"0.882",
"0.4739",
"0.511",
"0.4739"
],
[
"0.12006",
"0.481",
"0.4895",
"0.481",
"0.4895",
"0.4954",
"0.12006",
"0.4954"
],
[
"0.51",
"0.4756",
"0.8833",
"0.4744",
"0.8833",
"0.4893",
"0.51",
"0.4905"
],
[
"0.511",
"0.4905",
"0.882",
"0.4912",
"0.882",
"0.506",
"0.511",
"0.505"
],
[
"0.11755",
"0.4976",
"0.4895",
"0.4976",
"0.4895",
"0.5127",
"0.11755",
"0.5127"
],
[
"0.11584",
"0.511",
"0.2183",
"0.513",
"0.2183",
"0.528",
"0.11584",
"0.5254"
],
[
"0.513",
"0.5083",
"0.882",
"0.5083",
"0.882",
"0.5215",
"0.513",
"0.5215"
],
[
"0.513",
"0.525",
"0.882",
"0.525",
"0.882",
"0.538",
"0.513",
"0.538"
],
[
"0.11755",
"0.5415",
"0.33",
"0.5415",
"0.33",
"0.5557",
"0.11755",
"0.5557"
],
[
"0.511",
"0.539",
"0.881",
"0.539",
"0.881",
"0.5527",
"0.511",
"0.5527"
],
[
"0.511",
"0.557",
"0.882",
"0.557",
"0.882",
"0.571",
"0.511",
"0.571"
],
[
"0.11584",
"0.564",
"0.486",
"0.564",
"0.486",
"0.5776",
"0.11584",
"0.5776"
],
[
"0.11584",
"0.582",
"0.4895",
"0.582",
"0.4895",
"0.5957",
"0.11584",
"0.5957"
],
[
"0.513",
"0.583",
"0.6943",
"0.583",
"0.6943",
"0.5967",
"0.513",
"0.5967"
],
[
"0.11755",
"0.597",
"0.4878",
"0.597",
"0.4878",
"0.611",
"0.11755",
"0.611"
],
[
"0.513",
"0.605",
"0.882",
"0.605",
"0.882",
"0.619",
"0.513",
"0.619"
],
[
"0.1142",
"0.6123",
"0.4895",
"0.6123",
"0.4895",
"0.627",
"0.1142",
"0.627"
],
[
"0.511",
"0.6216",
"0.8833",
"0.6216",
"0.8833",
"0.6353",
"0.511",
"0.6353"
],
[
"0.1142",
"0.6274",
"0.486",
"0.6274",
"0.486",
"0.6426",
"0.1142",
"0.6426"
],
[
"0.51",
"0.6357",
"0.881",
"0.6357",
"0.881",
"0.6494",
"0.51",
"0.6494"
],
[
"0.11755",
"0.6455",
"0.4878",
"0.6455",
"0.4878",
"0.66",
"0.11755",
"0.66"
],
[
"0.51",
"0.652",
"0.881",
"0.652",
"0.881",
"0.665",
"0.51",
"0.665"
],
[
"0.11584",
"0.66",
"0.2737",
"0.661",
"0.2737",
"0.676",
"0.11584",
"0.675"
],
[
"0.511",
"0.6694",
"0.882",
"0.6694",
"0.882",
"0.683",
"0.511",
"0.683"
],
[
"0.1368",
"0.6787",
"0.4878",
"0.6787",
"0.4878",
"0.6924",
"0.1368",
"0.6924"
],
[
"0.51",
"0.684",
"0.879",
"0.684",
"0.879",
"0.6978",
"0.51",
"0.6978"
],
[
"0.11584",
"0.6953",
"0.4878",
"0.6953",
"0.4878",
"0.709",
"0.11584",
"0.709"
],
[
"0.511",
"0.7007",
"0.881",
"0.7007",
"0.881",
"0.7144",
"0.511",
"0.7144"
],
[
"0.11755",
"0.7104",
"0.4895",
"0.7104",
"0.4895",
"0.7246",
"0.11755",
"0.7246"
],
[
"0.513",
"0.7188",
"0.882",
"0.7188",
"0.882",
"0.7324",
"0.513",
"0.7324"
],
[
"0.11584",
"0.7256",
"0.486",
"0.7256",
"0.486",
"0.7393",
"0.11584",
"0.7393"
],
[
"0.513",
"0.734",
"0.882",
"0.734",
"0.882",
"0.7476",
"0.513",
"0.7476"
],
[
"0.1142",
"0.741",
"0.4878",
"0.742",
"0.4878",
"0.7573",
"0.1142",
"0.756"
],
[
"0.513",
"0.75",
"0.882",
"0.75",
"0.882",
"0.763",
"0.513",
"0.763"
],
[
"0.11584",
"0.7583",
"0.486",
"0.7583",
"0.486",
"0.7715",
"0.11584",
"0.7715"
],
[
"0.513",
"0.7666",
"0.769",
"0.7666",
"0.769",
"0.78",
"0.513",
"0.78"
],
[
"0.1142",
"0.774",
"0.4844",
"0.774",
"0.4844",
"0.7876",
"0.1142",
"0.7876"
],
[
"0.11755",
"0.7905",
"0.4895",
"0.7905",
"0.4895",
"0.804",
"0.11755",
"0.804"
],
[
"0.513",
"0.7915",
"0.882",
"0.7915",
"0.882",
"0.8047",
"0.513",
"0.8047"
],
[
"0.11755",
"0.807",
"0.4895",
"0.807",
"0.4895",
"0.821",
"0.11755",
"0.821"
],
[
"0.51",
"0.808",
"0.881",
"0.808",
"0.881",
"0.822",
"0.51",
"0.822"
],
[
"0.1142",
"0.8223",
"0.4895",
"0.822",
"0.4895",
"0.836",
"0.1142",
"0.8374"
],
[
"0.51",
"0.8237",
"0.882",
"0.8237",
"0.882",
"0.8374",
"0.51",
"0.8374"
],
[
"0.11755",
"0.839",
"0.4902",
"0.839",
"0.4902",
"0.8525",
"0.11755",
"0.8525"
],
[
"0.51",
"0.839",
"0.882",
"0.8403",
"0.882",
"0.855",
"0.51",
"0.854"
],
[
"0.1142",
"0.8555",
"0.4878",
"0.8555",
"0.4878",
"0.869",
"0.1142",
"0.869"
],
[
"0.51",
"0.8555",
"0.8833",
"0.8555",
"0.8833",
"0.869",
"0.51",
"0.869"
],
[
"0.11755",
"0.8726",
"0.4895",
"0.8726",
"0.4895",
"0.8857",
"0.11755",
"0.8857"
],
[
"0.51",
"0.8726",
"0.882",
"0.8726",
"0.882",
"0.8857",
"0.51",
"0.8857"
],
[
"0.11755",
"0.888",
"0.3762",
"0.888",
"0.3762",
"0.902",
"0.11755",
"0.902"
],
[
"0.511",
"0.888",
"0.882",
"0.888",
"0.882",
"0.902",
"0.511",
"0.902"
],
[
"0.1368",
"0.904",
"0.4902",
"0.904",
"0.4902",
"0.9185",
"0.1368",
"0.9185"
],
[
"0.51",
"0.905",
"0.8833",
"0.905",
"0.8833",
"0.9185",
"0.51",
"0.9185"
]
] | [
"of the story, its characters, and main plot points",
"Visual",
"Mapping",
"(Yang et al., 2022).",
"Network",
"Prefix",
"The idea of having a separate planning stage has",
"also been explored for other text generation tasks",
"including summarization (Narayan et al., 2021;",
"*Image",
"Encoder",
"Liu and Chen, 2021) and data-to-text generation",
"Concent",
"mhadda",
"(Moryossef et al., 2019; Puduppully et al., 2022)",
"Our work is closest to Narayan et al. (2022) who",
"propose the use of question-answer pairs as inter-",
"mediate plans for summarization. However, their",
"Figure 2: Visual prefix construction. The pretrained",
"approach is designed for descriptive text. Our work",
"image encoder and concept detector are frozen.FF",
"extends their framework to a multimodal setting,",
"refers to a feed-forward layer, CD and EL denote a",
"concept detector and embedding layer, respectively.",
"where the input consists of image sequences, and",
"the output are narratives characterized by more ab-",
"stract and figurative language.",
"dates with a T5 model (Raffel et al., 2020; Romero",
"2021) fine-tuned on the SQuAD reading compre-",
"3 Blueprint-based Visual Storytelling",
"hension dataset (Rajpurkar et al., 2018). The an-",
"swer and the story are provided as context to pre-",
"Let I represent a sequence of k images, denoted",
"dict the corresponding question. We decontextu-",
"as [u1, v2..., Uk]. Given this input, our goal is to",
"alize stories by replacing pronouns with their cor-",
"generate a blueprint plan B (i.e., an ordered set of",
"responding head mentions, using a state-of-the-art",
"question-answer pairs) and a story S based on it.",
"coreference resolution model (Dobrovolski, 2021)",
"Most generationdatasetsdonotincludeblueprint",
"Question-answer pairs are subsequently filtered",
"annotations, and visual storytelling is no excep-",
"tion. We first describe how we automatically obtain",
"to eliminate noise which is unavoidable due to the",
"[I, B, S,, training data samples (Section 3.1),",
"automatic preprocessing steps mentioned above.",
"We thus remove any question-answer pairs where",
"and then introduce our story generation models",
"(Section 3.2).",
"the answer is already present in the question. We",
"also employ a round-trip consistency check (Alberti",
"3.1 BlueprintAnnotation",
"et al., 2019) which discards questions if they yield",
"answers different from those used to generate them.",
"Let { I;. S,, denote a dataset consisting of pairs",
"of image sequences and their corresponding sto-",
"3.2Blueprint Models",
"ries. We automatically create blueprint B, based",
"Our approach leverages the generation capabili-",
"on story S; using state-of-the-art question genera-",
"ties of pre-trained sequence-to-sequence models",
"tors (Romero, 2021; Raffel et al., 2020), coupled",
"As our backbone model, we employ BART-base",
"with a filtering procedure to remove repetitions and",
"(Lewis et al.. 2020) which has been fine-tuned for",
"ill-formed questions.",
"text generation. We adapt this model to our visual",
"The generation of question-answer pairs involves",
"storytelling task in two ways. Aside from enabling",
"two steps, namely answer extraction and question",
"the generation of blueprints, we convert the image",
"generation. In the context of storytelling, captur-",
"sequence to a visual prefix which the pretrained lan-",
"ing key events is crucial for a compelling narra-",
"guage model can interpret. The pretrained language",
"tive. While noun phrases and named entities are",
"model is prompted with this prefix to generate the",
"commonly recognized as significant content units",
"blueprint, and eventually the story.",
"in other tasks such as summarization (Narayan",
"et al., 2022; Deutsch and Roth, 2023), verb phrases",
"Visual Prefix ConstructionOur model needs to",
"also play a vital role in conveying story dynamics,",
" grasp what the image sequence is about, e.g., the",
"actions, and relationships (Trabasso et al., 1989;",
"depicted objects, actions, and their associations",
"Eisenberg and Finlayson, 2017; Liu et al., 2020).",
"Drawing inspiration from recent advances in vision",
"Therefore, in addition to named entities and noun",
"and language research (Mokady et al., 2021; Tsim-",
"phrases, we also extract verb phrases as answer",
"poukelli et al., 2021; Alayrac et al., 2022; Zhai",
"candidates using the spaCy library.",
"et al., 2022; Liu et al., 2023; Huang et al., 2023),",
"We then generate questions for answer candi-",
"we translate the input sequence of images into a"
] | {
"arxiv_id": "2310.05295",
"language": "en",
"page_idx": 2
} |
||
2310.05295-en-4 | 2310.05295-en-4.jpg | <table>
<tbody>
<tr>
<td>Length of image sequences</td>
<td>5.0</td>
</tr>
<tr>
<td>Number of sentences in the story</td>
<td>5.0</td>
</tr>
<tr>
<td>Number of tokens per story</td>
<td>52.3</td>
</tr>
<tr>
<td>Number of QA pairs per story</td>
<td>11.1</td>
</tr>
<tr>
<td>Number of tokens per QA pair</td>
<td>10.3</td>
</tr>
<tr>
<td>Number of tokens per story plus QA pair</td>
<td>166.2</td>
</tr>
</tbody>
</table>
Table 1: VIST dataset Statistics (average values).
$s_i$ represents the $i$-th sentence in the story, and $b_i$ its associated blueprint. Each $b_i$ consists of answerblueprint. The training objective for the iterative model is defined as follows:
$$\max_{\theta,\phi}\sum_{j=1}^N\sum_{i=1}^k\log p_{\theta,\phi}\left(b_{i+1},s_{i+1}|b_{1:i},s_{1:i},I_j\right)$$
where $b_{1:i}$ and $s_{1:i}$ refer to the blueprint and sen-
tences generated so far, from step 1 to step $i.$
At each time step $i$, the encoder takes image sequence $I$ as input, and the decoder takes the context (i.e., blueprint and sentences generated so far $\{b_1,b_2,\ldots,b_i;s_1,s_2,\ldots,s_i\})$ as a prompt to predict the next blueprint $b_{i+1}$ and sentence $s_{i+1}.$ Therefore, the iterative model is trained on samples $\{ I, ( b_{1: i}, s_{1: i}) , b_{i+ 1}, s_{i+ 1}\} .$ We prefix$(b_1:i,s_{1:i})$, $b_{i+1}$, and $s_{i+1}$ with $Context:,Plan:$, and Next Sentence, respectively. To handle the first time step, we introduce special token $\langle START\rangle$ as context to predict $b_1$ and $s_1.$ We also use $\langle END\rangle$ to indicate the completion of an iteration (see Figure 3 for an illustration). It is important to note that $(b_1:i,s_{1:i})$ are masked out when computing the loss because they serve as prompts to the decoder. We want to avoid the model repeatedly predicting and overly optimizing the blueprints and sentences that appear at the beginning of the output.
# 4 Experimental Setting
## 4.1 Dataset
We performed experiments on the widely used VIST dataset (Huang et al., 2016), which contains 10,117 Flickr albums and 210,819 unique photos. Each training sample consists of $k=5$ images and a corresponding story of $k=5$ sentences. As described in Section 3.1, we augment each story with an automatically generated blueprint.
## 4.2 Implementation Details
Our models are built on top of BART-base (Lewis et al., 2020) and finetuned with a learning rate of 3e-5, batch size of 64, and warm-up ratio of 0.05. We select the best checkpoint on the validation set using a QA-based metric which quantifies the extent to which the output story follows its blueprint (see Section 4.4). During inference, we employ beam search (size 5). For our visual prefix, we employed the Clarifai Concept Detector which was trained on a dataset containing 9,098 concepts and
question pairs, denoted as $\{ a_{1}^{i}, q_{1}^{i}, \ldots , a_{l( i) }^{i}, q_{l( i) }^{i}\} $, 20 million images ( multiple concepts are assigned where $l_{( i) }$ is the number of pairs in the $i- \textbf{th}$ to each image) , and is integrated with the Incep- to each i
tionV2 architecture (Szegedy et al., 2016) .
## 4.3 Comparison Systems
We compared our models against several baselines and state-of-the-art visual storytelling models. These included a vanilla BART-base model with the same encoder and visual prefix as ours but no planning (VP-BART; it generates the story directly in an autoregressive manner without the blueprint). KG-Story (Hsu et al., 2020) predicts a set of words representative of the image sequence, enriches them using external knowledge graphs, and generates stories based on the enriched word set. PR-VIST (Hsu et al., 2021) is a state-of-the-art model which constructs a graph representing the relations between elements in the image sequence, identifies the best storyline captured therein, and proceeds to generate a story based on it. The process of constructing the story graph can be viewed as a form of planning. Along similar lines, Chen et al.(2021) build a common sense knowledge graph capturing concepts in the image sequence, and use MCSM, a Maximal Clique Selection Module to identify which ones to write a story about. They use BART-large to generate the story based on selected concepts (and image features).
We also compared against an LLM that generates stories via prompting. We provide GPT-3.5 with a visual prefix, namely the concepts identified in the image sequence, and a prompt which explains how to create the blueprint and generate the story together with examples (in-context learning). Details on the prompt can be found in Appendix A.
## 4.4 Automatic Evaluation
We evaluated our stories using BLEU, ROUGE, METEOR, and CIDER, mainly to compare to previous work. Several studies (Hsu et al., 2022, 2021, | [
[
"0.44",
"0.0873",
"0.4685",
"0.0873",
"0.4685",
"0.10095",
"0.44",
"0.10095"
],
[
"0.51",
"0.08435",
"0.7407",
"0.0849",
"0.7407",
"0.0998",
"0.51",
"0.0986"
],
[
"0.1368",
"0.10156",
"0.357",
"0.10156",
"0.357",
"0.1152",
"0.1368",
"0.1152"
],
[
"0.1368",
"0.0885",
"0.3225",
"0.0885",
"0.3225",
"0.10156",
"0.1368",
"0.10156"
],
[
"0.44",
"0.10156",
"0.4685",
"0.10156",
"0.4685",
"0.1152",
"0.44",
"0.1152"
],
[
"0.513",
"0.1081",
"0.882",
"0.1081",
"0.882",
"0.1217",
"0.513",
"0.1217"
],
[
"0.1377",
"0.1164",
"0.3208",
"0.1164",
"0.3208",
"0.13",
"0.1377",
"0.13"
],
[
"0.4307",
"0.114",
"0.4685",
"0.114",
"0.4685",
"0.13",
"0.4307",
"0.13"
],
[
"0.51",
"0.1223",
"0.882",
"0.12354",
"0.882",
"0.1383",
"0.51",
"0.1372"
],
[
"0.1351",
"0.13",
"0.3376",
"0.13",
"0.3376",
"0.1437",
"0.1351",
"0.1437"
],
[
"0.4324",
"0.1312",
"0.465",
"0.1312",
"0.465",
"0.1437",
"0.4324",
"0.1437"
],
[
"0.1351",
"0.1443",
"0.34",
"0.1455",
"0.34",
"0.1592",
"0.1351",
"0.158"
],
[
"0.4324",
"0.1443",
"0.4668",
"0.1443",
"0.4668",
"0.1567",
"0.4324",
"0.1567"
],
[
"0.511",
"0.1401",
"0.8833",
"0.1401",
"0.8833",
"0.1538",
"0.511",
"0.1538"
],
[
"0.1351",
"0.1592",
"0.4114",
"0.1592",
"0.4114",
"0.1729",
"0.1351",
"0.1729"
],
[
"0.4207",
"0.1592",
"0.4685",
"0.1592",
"0.4685",
"0.1733",
"0.4207",
"0.1733"
],
[
"0.511",
"0.1567",
"0.882",
"0.1567",
"0.882",
"0.1704",
"0.511",
"0.1704"
],
[
"0.511",
"0.1729",
"0.882",
"0.1729",
"0.882",
"0.1859",
"0.511",
"0.1859"
],
[
"0.1351",
"0.1882",
"0.4668",
"0.1882",
"0.4668",
"0.2019",
"0.1351",
"0.2019"
],
[
"0.511",
"0.1895",
"0.882",
"0.1895",
"0.882",
"0.2031",
"0.511",
"0.2031"
],
[
"0.51",
"0.2037",
"0.882",
"0.2048",
"0.882",
"0.2197",
"0.51",
"0.2185"
],
[
"0.511",
"0.2197",
"0.882",
"0.2203",
"0.882",
"0.2351",
"0.511",
"0.234"
],
[
"0.11584",
"0.234",
"0.4836",
"0.234",
"0.4836",
"0.2477",
"0.11584",
"0.2477"
],
[
"0.511",
"0.2369",
"0.882",
"0.2369",
"0.882",
"0.2505",
"0.511",
"0.2505"
],
[
"0.11584",
"0.2494",
"0.4878",
"0.2494",
"0.4878",
"0.2632",
"0.11584",
"0.2632"
],
[
"0.513",
"0.253",
"0.882",
"0.253",
"0.882",
"0.2666",
"0.513",
"0.2666"
],
[
"0.1142",
"0.2632",
"0.4902",
"0.2654",
"0.4902",
"0.2832",
"0.1142",
"0.2808"
],
[
"0.509",
"0.2673",
"0.881",
"0.2673",
"0.881",
"0.2808",
"0.509",
"0.2808"
],
[
"0.11755",
"0.282",
"0.4878",
"0.282",
"0.4878",
"0.2964",
"0.11755",
"0.2964"
],
[
"0.509",
"0.284",
"0.881",
"0.2852",
"0.881",
"0.2998",
"0.509",
"0.2986"
],
[
"0.11755",
"0.2976",
"0.4895",
"0.2976",
"0.4895",
"0.3113",
"0.11755",
"0.3113"
],
[
"0.51",
"0.3005",
"0.8286",
"0.3005",
"0.8286",
"0.3154",
"0.51",
"0.3154"
],
[
"0.11584",
"0.3142",
"0.3257",
"0.3142",
"0.3257",
"0.328",
"0.11584",
"0.328"
],
[
"0.511",
"0.3289",
"0.723",
"0.33",
"0.723",
"0.3445",
"0.511",
"0.3433"
],
[
"0.1704",
"0.3433",
"0.215",
"0.3433",
"0.215",
"0.3552",
"0.1704",
"0.3552"
],
[
"0.12006",
"0.3499",
"0.492",
"0.354",
"0.492",
"0.379",
"0.12006",
"0.3748"
],
[
"0.511",
"0.3516",
"0.882",
"0.354",
"0.882",
"0.3677",
"0.511",
"0.3652"
],
[
"0.1335",
"0.3723",
"0.1587",
"0.3723",
"0.1587",
"0.3801",
"0.1335",
"0.3801"
],
[
"0.513",
"0.3694",
"0.882",
"0.3694",
"0.882",
"0.383",
"0.513",
"0.383"
],
[
"0.1621",
"0.3772",
"0.22",
"0.3772",
"0.22",
"0.3884",
"0.1621",
"0.3884"
],
[
"0.511",
"0.3843",
"0.879",
"0.3843",
"0.879",
"0.398",
"0.511",
"0.398"
],
[
"0.11755",
"0.4067",
"0.4902",
"0.4067",
"0.4902",
"0.4204",
"0.11755",
"0.4204"
],
[
"0.511",
"0.402",
"0.879",
"0.402",
"0.879",
"0.415",
"0.511",
"0.415"
],
[
"0.511",
"0.4163",
"0.881",
"0.4175",
"0.881",
"0.4316",
"0.511",
"0.4312"
],
[
"0.11584",
"0.4233",
"0.4475",
"0.4233",
"0.4475",
"0.437",
"0.11584",
"0.437"
],
[
"0.1351",
"0.4407",
"0.486",
"0.4412",
"0.486",
"0.456",
"0.1351",
"0.4548"
],
[
"0.5146",
"0.4353",
"0.879",
"0.4353",
"0.879",
"0.4453",
"0.5146",
"0.4453"
],
[
"0.513",
"0.4495",
"0.882",
"0.4495",
"0.882",
"0.4631",
"0.513",
"0.4631"
],
[
"0.11755",
"0.4578",
"0.4878",
"0.4578",
"0.4878",
"0.4714",
"0.11755",
"0.4714"
],
[
"0.511",
"0.4656",
"0.882",
"0.466",
"0.882",
"0.4797",
"0.511",
"0.4785"
],
[
"0.11755",
"0.4744",
"0.4878",
"0.4744",
"0.4878",
"0.488",
"0.11755",
"0.488"
],
[
"0.511",
"0.4822",
"0.882",
"0.4822",
"0.882",
"0.4954",
"0.511",
"0.4954"
],
[
"0.11755",
"0.4905",
"0.4878",
"0.4905",
"0.4878",
"0.505",
"0.11755",
"0.505"
],
[
"0.511",
"0.4976",
"0.882",
"0.4976",
"0.882",
"0.511",
"0.511",
"0.511"
],
[
"0.1184",
"0.506",
"0.4902",
"0.506",
"0.4902",
"0.521",
"0.1184",
"0.521"
],
[
"0.51",
"0.514",
"0.882",
"0.514",
"0.882",
"0.528",
"0.51",
"0.528"
],
[
"0.11755",
"0.5225",
"0.4895",
"0.5225",
"0.4895",
"0.536",
"0.11755",
"0.536"
],
[
"0.509",
"0.5293",
"0.882",
"0.53",
"0.882",
"0.5444",
"0.509",
"0.5435"
],
[
"0.12006",
"0.539",
"0.4895",
"0.539",
"0.4895",
"0.554",
"0.12006",
"0.554"
],
[
"0.51",
"0.5464",
"0.882",
"0.5464",
"0.882",
"0.56",
"0.51",
"0.56"
],
[
"0.11584",
"0.5547",
"0.4902",
"0.554",
"0.4902",
"0.5684",
"0.11584",
"0.5693"
],
[
"0.51",
"0.561",
"0.881",
"0.561",
"0.881",
"0.5747",
"0.51",
"0.5747"
],
[
"0.11584",
"0.571",
"0.4895",
"0.571",
"0.4895",
"0.585",
"0.11584",
"0.585"
],
[
"0.511",
"0.579",
"0.8833",
"0.579",
"0.8833",
"0.593",
"0.511",
"0.593"
],
[
"0.1142",
"0.586",
"0.4878",
"0.585",
"0.4878",
"0.5996",
"0.1142",
"0.601"
],
[
"0.51",
"0.594",
"0.882",
"0.5933",
"0.882",
"0.608",
"0.51",
"0.6094"
],
[
"0.11755",
"0.6025",
"0.486",
"0.6025",
"0.486",
"0.616",
"0.11755",
"0.616"
],
[
"0.513",
"0.611",
"0.882",
"0.611",
"0.882",
"0.6245",
"0.513",
"0.6245"
],
[
"0.11755",
"0.618",
"0.486",
"0.618",
"0.486",
"0.632",
"0.11755",
"0.632"
],
[
"0.513",
"0.627",
"0.882",
"0.627",
"0.882",
"0.64",
"0.513",
"0.64"
],
[
"0.1142",
"0.633",
"0.4878",
"0.6353",
"0.4878",
"0.6494",
"0.1142",
"0.648"
],
[
"0.5146",
"0.6436",
"0.8833",
"0.6436",
"0.8833",
"0.6567",
"0.5146",
"0.6567"
],
[
"0.11755",
"0.652",
"0.4836",
"0.652",
"0.4836",
"0.662",
"0.11755",
"0.662"
],
[
"0.5146",
"0.66",
"0.882",
"0.66",
"0.882",
"0.6704",
"0.5146",
"0.6704"
],
[
"0.11584",
"0.6665",
"0.486",
"0.6665",
"0.486",
"0.6797",
"0.11584",
"0.6797"
],
[
"0.51",
"0.6733",
"0.881",
"0.6733",
"0.881",
"0.687",
"0.51",
"0.687"
],
[
"0.11584",
"0.6816",
"0.4844",
"0.6816",
"0.4844",
"0.6953",
"0.11584",
"0.6953"
],
[
"0.511",
"0.6904",
"0.881",
"0.6904",
"0.881",
"0.7036",
"0.511",
"0.7036"
],
[
"0.11584",
"0.6997",
"0.486",
"0.6997",
"0.486",
"0.7134",
"0.11584",
"0.7134"
],
[
"0.513",
"0.708",
"0.8237",
"0.708",
"0.8237",
"0.7217",
"0.513",
"0.7217"
],
[
"0.11755",
"0.7173",
"0.339",
"0.7173",
"0.339",
"0.728",
"0.11755",
"0.728"
],
[
"0.529",
"0.7236",
"0.8833",
"0.7236",
"0.8833",
"0.7383",
"0.529",
"0.7383"
],
[
"0.51",
"0.7393",
"0.882",
"0.7383",
"0.882",
"0.753",
"0.51",
"0.7544"
],
[
"0.1184",
"0.752",
"0.1351",
"0.752",
"0.1351",
"0.761",
"0.1184",
"0.761"
],
[
"0.126",
"0.749",
"0.3333",
"0.7505",
"0.3333",
"0.7656",
"0.126",
"0.763"
],
[
"0.511",
"0.756",
"0.882",
"0.756",
"0.882",
"0.771",
"0.511",
"0.771"
],
[
"0.511",
"0.7725",
"0.8833",
"0.7725",
"0.8833",
"0.786",
"0.511",
"0.786"
],
[
"0.1184",
"0.783",
"0.22",
"0.783",
"0.22",
"0.7944",
"0.1184",
"0.7944"
],
[
"0.511",
"0.789",
"0.882",
"0.789",
"0.882",
"0.8027",
"0.511",
"0.8027"
],
[
"0.11584",
"0.807",
"0.486",
"0.807",
"0.486",
"0.821",
"0.11584",
"0.821"
],
[
"0.51",
"0.804",
"0.882",
"0.8047",
"0.882",
"0.8193",
"0.51",
"0.8184"
],
[
"0.11755",
"0.8237",
"0.4895",
"0.8237",
"0.4895",
"0.8384",
"0.11755",
"0.8384"
],
[
"0.511",
"0.822",
"0.8833",
"0.822",
"0.8833",
"0.835",
"0.511",
"0.835"
],
[
"0.11584",
"0.8403",
"0.486",
"0.8403",
"0.486",
"0.854",
"0.11584",
"0.854"
],
[
"0.513",
"0.8496",
"0.729",
"0.8496",
"0.729",
"0.8633",
"0.513",
"0.8633"
],
[
"0.11755",
"0.857",
"0.4878",
"0.857",
"0.4878",
"0.8706",
"0.11755",
"0.8706"
],
[
"0.1142",
"0.8726",
"0.4902",
"0.8726",
"0.4902",
"0.8857",
"0.1142",
"0.8857"
],
[
"0.509",
"0.8706",
"0.8833",
"0.8716",
"0.8833",
"0.8857",
"0.509",
"0.8853"
],
[
"0.1142",
"0.8896",
"0.4878",
"0.8896",
"0.4878",
"0.903",
"0.1142",
"0.903"
],
[
"0.509",
"0.887",
"0.882",
"0.888",
"0.882",
"0.903",
"0.509",
"0.902"
],
[
"0.11584",
"0.905",
"0.3938",
"0.905",
"0.3938",
"0.9185",
"0.11584",
"0.9185"
],
[
"0.51",
"0.904",
"0.8833",
"0.903",
"0.8833",
"0.9175",
"0.51",
"0.9185"
]
] | [
"5.0",
"4.2Implementation Details",
"Number of sentences in the story",
"Length of image sequences",
"5.0",
"Our models are built on top of BART-base (Lewis",
"Number of tokens per story",
"52.3",
"et al., 2020) and finetuned with a learning rate",
" Number of QA pairs per story",
"11.1",
"Number of tokens per QA pair",
"10.3",
"of 3e-5, batch size of 64, and warm-up ratio of 0.05.",
" Number of tokens per story plus QA pair",
"166.2",
"We select the best checkpoint on the validation set",
"using a QA-based metric which quantifies the ex-",
"Table 1: VIST dataset Statistics (average values).",
"tent to which the output story follows its blueprint",
"(see Section 4.4). During inference, we employ",
"beam search (size 5). For our visual prefix, we",
"Si represents the i-th sentence in the story, and b",
"employed the Clarifai Concept Detector which was",
"its associated blueprint. Each b; consists of answer-",
"trained on a dataset containing 9,098 concepts and",
"question pairs, denoted as (ai, qi, ..., ai(), qi()],",
"20 million images (multiple concepts are assigned",
"where l() is the number of pairs in the i-th",
"to each image), and is integrated with the Incep",
"blueprint. The training objective for the iterative",
"tionV2 architecture (Szegedy et al., 2016) .",
"model is defined as follows:",
"4.3 Comparison Systems",
"Nk",
"maxgpe, (b+, +11:, 1:,5) (4)",
"We compared our models against several base-",
"0.0",
"lines and state-of-the-art visual storytelling mod",
"j-1 i=1",
"els. These included a vanilla BART-base mode",
"where bi:i and si: refer to the blueprint and sen-",
"with the same encoder and visual prefix as ours",
"but no planning (VP-BART; it generates the story",
"tences generated so far, from step 1 to step i.",
"At each time step i, the encoder takes image",
"directlyin an autoregressive manner withoutthe",
"blueprint). KG-Story (Hsu et al., 2020) predicts a",
"sequence I as input, and the decoder takes the",
"set of words representative of the image sequence",
"context (i.e., blueprint and sentences generated so",
"enriches them using external knowledge graphs",
"far [bi,b2,.., bi; $i, 82,., si]) as a prompt to",
"and generates stories based on the enriched word",
"predict the next blueprint bi+1 and sentence si+1.",
"set. PR-VIST (Hsu et al., 2021) is a state-of-the-art",
"Therefore, the iterative model is trained on samples",
"model which constructs a graph representing the",
"[I, (bi:, S1:), bi+1, Si+1]. We prefix (b1:, $1:),",
"relations between elements in the image sequence,",
"bi+1, and si+1 with Context:, Plan., and Next Sen-",
"identifies the best storyline captured therein, and",
"tence, respectively. To handle the first time step,",
"proceeds to generate a story based on it. The pro-",
"we introduce special token (START) as context to",
"cess of constructing the story graph can be viewed",
"predict bi and si. We also use (END) to indicate",
"as a form of planning. Along similar lines, Chen",
"the completion of an iteration (see Figure 3 for an",
"et al. (2021) build a common sense knowledge",
"illustration). It is important to note that (b1:, S1:)",
"graph capturing concepts in the image sequence,",
"aremaskedoutwhencomputingthelossbecause",
"and useMCSM,a Maximal Clique Selection Mod-",
"they serve as prompts to the decoder. We want to",
"ule to identify which ones to write a story about",
"avoid the model repeatedly predicting and overly",
"They use BART-large to generate the story based",
"optimizing the blueprints and sentences that appear",
"on selected concepts (and image features).",
"at the beginning of the output.",
"We also compared against an LLM that gener-",
"ates stories via prompting. We provide GPT-3.5",
"4",
"Experimental Setting",
"with a visual prefix, namely the concepts identified",
"in the image sequence, and a prompt which ex-",
"4.1 Dataset",
"plains how to create the blueprint and generate the",
"We performed experiments on the widely used",
"story together with examples (in-context learning)",
"VIST dataset (Huang et al., 2016), which contains",
"Details on the prompt can be found in Appendix A.",
"10,117 Flickr albums and 210,819 unique photos",
"4.4Automatic Evaluation",
"Each training sample consists of k = 5 images and",
"a corresponding story of k -- 5 sentences. As de-",
"We evaluated our stories using BLEU, ROUGE,",
" scribed in Section 3.1, we augment each story with",
"METEOR, and CIDER, mainly to compare to pre-",
"an automatically generated blueprint.",
"vious work. Several studies (Hsu et al., 2022, 2021,"
] | {
"arxiv_id": "2310.05295",
"language": "en",
"page_idx": 4
} |
||
2211.06444-en-0 | 2211.06444-en-0.jpg | ## Probabilistic Debiasing of Scene Graphs
Bashirul Azam Biswas and Qiang Ji
Rensselaer Polytechnic Institute, Troy, NY- 12180
$\{$biswab, jiq$\}$erpi.edu
## $\mathbf{Abstract}$
The quality of scene graphs generated by the state-ofthe-art (SOTA) models is compromised due to the long-tail nature of the relationships and their parent object pairs. Training of the scene graphs is dominated by the majority relationships of the majority pairs and, therefore, the $object- conditional\textit{ distributions of relationship in the mi- }$ nority pairs are not preserved after the training is con-
$\begin{array}{ccc}{}relationships~such~as~'on'and~'wearing'\end{array}$
dence incorporated within-triplet Bayesian Network (BN) to preserve the object-conditional distribution of the relationship label and to eradicate the bias created by the marginal probability of the relationships. The insufficient number of relationships in the minority classes poses a significant problem in learning the within- triplet Bayesian network. We address this insufficiency by embedding-based augmentation of triplets where we borrow samples of the minority triplet classes from its neighborhood triplets in the semantic space. We perform experiments on two different datasets and achieve a significant improvement in the mean recall of the relationships. We also achieve better balance between recall and mean recall performance compared to the SOTA de-biasing techniques of scene graph models. Code is publicly available at https: / github. com/ bashirulazam/within-triplet-debias.
## $1. \textbf{Introduction}$
Any visual relationship can be expressed as a triplet subject-relationship-object and all triplets in an image can be represented as a concise graph called Scene Graph (SG) [19] where the nodes represent the objects and the edges represent relationships. This representation has been proven useful for many downstream tasks such as image captioning [37], visual reasoning [24], and image generation[10]. Scene Graph Generation (SGG) has become one of the ma-
jor computer vision research arenas after the introduction of Visual Genome (VG) dataset[11]. The distribution of triplets in VG images has two distinct characteristics: (1) the presence of strong within-triplet prior, and (2) the longtail distribution of the relationship. As shown in Figure l (a), the within-triplet prior dictates that 'window' will most likely be‘on’ the‘building’ rather than‘eating’it. Zeller $etal.[43]$ has utilized this within-triplet prior as the conditional probability of relationships given subject and object long-tailed issue by considering a causal interpretation of the biased prediction. We argue that these two seemingly different characteristics of the relationship distribution are interrelated. The abundance of the head classes of the relationship distribution in Figure l (c), such as ‘on’and‘wearing', arises from the abundance of their parent subject and object lying in the head region of Figure l (b).
As indicated by[5], the long-tailed distribution exists both in relationship and object label. Since relationship labels are dependent on their object pair because of the within-triplet prior, the long-tail distribution of the relationship worsens due to the long-tail nature of the object pairs. Crowd-collection of VG images creates selection bias and crowd- annotation of these images create $label-bias$ [29] and $co-occurring-bias[25].$ To analyze such biases, we investigate the distribution of the object pair of the triplets in VG database. As shown in Figure l (b), 'window-building' and `man-shirt' are the most frequently annotated pairs and top $1\%$ object pair covers $33\%$ of all triplets. As a result, the dominant relationships in these head pairs, such as ‘on’and “wearing’, dominate the marginal distribution of Figure 1 (c).
In training a deep-learning-based SGG model, samplers will sample more relationships from the head pairs. As a result, the Maximum Likelihood Estimation (MLE) of the parameters is biased to predict the relationship classes in the head pairs [27] and the object-conditional representation of the relationship in the tail pairs will be lost in the training process. Therefore, various deep learning-based models, | [
[
"0.2852",
"0.1332",
"0.6855",
"0.1332",
"0.6855",
"0.1509",
"0.2852",
"0.1509"
],
[
"0.343",
"0.1831",
"0.6284",
"0.1831",
"0.6284",
"0.1976",
"0.343",
"0.1976"
],
[
"0.2893",
"0.1989",
"0.681",
"0.1976",
"0.681",
"0.2153",
"0.2893",
"0.2166"
],
[
"0.392",
"0.2184",
"0.5786",
"0.2197",
"0.5786",
"0.2343",
"0.392",
"0.2329"
],
[
"0.02287",
"0.2583",
"0.06046",
"0.2583",
"0.06046",
"0.3738",
"0.02287",
"0.3738"
],
[
"0.2361",
"0.2676",
"0.3145",
"0.2676",
"0.3145",
"0.2854",
"0.2361",
"0.2854"
],
[
"0.5024",
"0.2708",
"0.8906",
"0.2708",
"0.8906",
"0.2842",
"0.5024",
"0.2842"
],
[
"0.5024",
"0.286",
"0.892",
"0.286",
"0.892",
"0.2998",
"0.5024",
"0.2998"
],
[
"0.10046",
"0.3018",
"0.4673",
"0.3018",
"0.4673",
"0.315",
"0.10046",
"0.315"
],
[
"0.5024",
"0.3018",
"0.8906",
"0.3018",
"0.8906",
"0.315",
"0.5024",
"0.315"
],
[
"0.0809",
"0.3176",
"0.468",
"0.3176",
"0.468",
"0.3308",
"0.0809",
"0.3308"
],
[
"0.501",
"0.315",
"0.8896",
"0.3164",
"0.8896",
"0.3308",
"0.501",
"0.33"
],
[
"0.0809",
"0.3333",
"0.468",
"0.3333",
"0.468",
"0.3467",
"0.0809",
"0.3467"
],
[
"0.5034",
"0.332",
"0.8896",
"0.332",
"0.8896",
"0.3452",
"0.5034",
"0.3452"
],
[
"0.0809",
"0.348",
"0.468",
"0.348",
"0.468",
"0.361",
"0.0809",
"0.361"
],
[
"0.5034",
"0.3467",
"0.8936",
"0.3467",
"0.8936",
"0.361",
"0.5034",
"0.361"
],
[
"0.0809",
"0.3623",
"0.468",
"0.3623",
"0.468",
"0.3757",
"0.0809",
"0.3757"
],
[
"0.501",
"0.361",
"0.8906",
"0.361",
"0.8906",
"0.375",
"0.501",
"0.375"
],
[
"0.5024",
"0.377",
"0.8906",
"0.377",
"0.8906",
"0.3901",
"0.5024",
"0.3901"
],
[
"0.02696",
"0.363",
"0.05554",
"0.363",
"0.05554",
"0.3914",
"0.02696",
"0.3914"
],
[
"0.0809",
"0.3774",
"0.468",
"0.3774",
"0.468",
"0.3914",
"0.0809",
"0.3914"
],
[
"0.0801",
"0.3933",
"0.468",
"0.3933",
"0.468",
"0.4072",
"0.0801",
"0.4072"
],
[
"0.5024",
"0.3926",
"0.8906",
"0.3926",
"0.8906",
"0.4072",
"0.5024",
"0.4072"
],
[
"0.02533",
"0.406",
"0.05554",
"0.406",
"0.05554",
"0.4507",
"0.02533",
"0.4507"
],
[
"0.0809",
"0.4077",
"0.4714",
"0.4077",
"0.4714",
"0.4216",
"0.0809",
"0.4216"
],
[
"0.5024",
"0.4077",
"0.8906",
"0.4077",
"0.8906",
"0.4216",
"0.5024",
"0.4216"
],
[
"0.0801",
"0.4224",
"0.4714",
"0.4224",
"0.4714",
"0.4363",
"0.0801",
"0.4363"
],
[
"0.5024",
"0.4224",
"0.892",
"0.4224",
"0.892",
"0.4363",
"0.5024",
"0.4363"
],
[
"0.03595",
"0.4382",
"0.0474",
"0.4363",
"0.04984",
"0.4414",
"0.03757",
"0.4443"
],
[
"0.0809",
"0.4382",
"0.4697",
"0.4382",
"0.4697",
"0.452",
"0.0809",
"0.452"
],
[
"0.5024",
"0.4375",
"0.892",
"0.4375",
"0.892",
"0.4507",
"0.5024",
"0.4507"
],
[
"0.03105",
"0.4495",
"0.058",
"0.4495",
"0.058",
"0.481",
"0.03105",
"0.481"
],
[
"0.0784",
"0.4526",
"0.468",
"0.452",
"0.468",
"0.4666",
"0.0784",
"0.4673"
],
[
"0.5034",
"0.4539",
"0.8906",
"0.4539",
"0.8906",
"0.464",
"0.5034",
"0.464"
],
[
"0.0801",
"0.4685",
"0.468",
"0.4685",
"0.468",
"0.4817",
"0.0801",
"0.4817"
],
[
"0.501",
"0.4666",
"0.8896",
"0.4673",
"0.8896",
"0.4817",
"0.501",
"0.481"
],
[
"0.0809",
"0.4849",
"0.4673",
"0.4849",
"0.4673",
"0.4956",
"0.0809",
"0.4956"
],
[
"0.505",
"0.4841",
"0.8896",
"0.4841",
"0.8896",
"0.4944",
"0.505",
"0.4944"
],
[
"0.02942",
"0.4937",
"0.05554",
"0.4976",
"0.05148",
"0.515",
"0.02533",
"0.511"
],
[
"0.0801",
"0.4988",
"0.4656",
"0.4988",
"0.4656",
"0.512",
"0.0801",
"0.512"
],
[
"0.5024",
"0.4976",
"0.8906",
"0.4976",
"0.8906",
"0.511",
"0.5024",
"0.511"
],
[
"0.03513",
"0.509",
"0.04984",
"0.509",
"0.04984",
"0.5264",
"0.03513",
"0.5264"
],
[
"0.0801",
"0.513",
"0.4697",
"0.513",
"0.4697",
"0.5264",
"0.0801",
"0.5264"
],
[
"0.501",
"0.513",
"0.8896",
"0.513",
"0.8896",
"0.5264",
"0.501",
"0.5264"
],
[
"0.03595",
"0.5264",
"0.04575",
"0.5264",
"0.04575",
"0.5386",
"0.03595",
"0.5386"
],
[
"0.0801",
"0.529",
"0.4697",
"0.529",
"0.4697",
"0.5425",
"0.0801",
"0.5425"
],
[
"0.5034",
"0.53",
"0.8906",
"0.53",
"0.8906",
"0.5405",
"0.5034",
"0.5405"
],
[
"0.0784",
"0.5435",
"0.468",
"0.5444",
"0.468",
"0.558",
"0.0784",
"0.5566"
],
[
"0.5024",
"0.5435",
"0.8096",
"0.5435",
"0.8096",
"0.5566",
"0.5024",
"0.5566"
],
[
"0.02942",
"0.558",
"0.05148",
"0.558",
"0.05148",
"0.5728",
"0.02942",
"0.5728"
],
[
"0.0801",
"0.5596",
"0.468",
"0.5596",
"0.468",
"0.5728",
"0.0801",
"0.5728"
],
[
"0.5215",
"0.563",
"0.892",
"0.563",
"0.892",
"0.5776",
"0.5215",
"0.5776"
],
[
"0.03186",
"0.5737",
"0.03757",
"0.5737",
"0.03757",
"0.5835",
"0.03186",
"0.5835"
],
[
"0.0809",
"0.5757",
"0.468",
"0.5757",
"0.468",
"0.586",
"0.0809",
"0.586"
],
[
"0.5024",
"0.579",
"0.892",
"0.579",
"0.892",
"0.5933",
"0.5024",
"0.5933"
],
[
"0.0801",
"0.589",
"0.4697",
"0.589",
"0.4697",
"0.603",
"0.0801",
"0.603"
],
[
"0.5024",
"0.5947",
"0.892",
"0.5947",
"0.892",
"0.608",
"0.5024",
"0.608"
],
[
"0.0801",
"0.603",
"0.4673",
"0.603",
"0.4673",
"0.616",
"0.0801",
"0.616"
],
[
"0.02777",
"0.613",
"0.05392",
"0.613",
"0.05392",
"0.632",
"0.02777",
"0.632"
],
[
"0.5024",
"0.61",
"0.8906",
"0.61",
"0.8906",
"0.6235",
"0.5024",
"0.6235"
],
[
"0.0809",
"0.6206",
"0.468",
"0.6206",
"0.468",
"0.631",
"0.0809",
"0.631"
],
[
"0.5024",
"0.6235",
"0.8906",
"0.625",
"0.8906",
"0.638",
"0.5024",
"0.637"
],
[
"0.0784",
"0.6333",
"0.4697",
"0.634",
"0.4697",
"0.6484",
"0.0784",
"0.648"
],
[
"0.02942",
"0.625",
"0.058",
"0.63",
"0.05228",
"0.648",
"0.0237",
"0.6426"
],
[
"0.03513",
"0.6426",
"0.05148",
"0.6426",
"0.05148",
"0.664",
"0.03513",
"0.664"
],
[
"0.5024",
"0.6396",
"0.892",
"0.6396",
"0.892",
"0.653",
"0.5024",
"0.653"
],
[
"0.0801",
"0.6494",
"0.468",
"0.6494",
"0.468",
"0.6636",
"0.0801",
"0.6636"
],
[
"0.5034",
"0.656",
"0.8906",
"0.656",
"0.8906",
"0.6665",
"0.5034",
"0.6665"
],
[
"0.0433",
"0.6665",
"0.05392",
"0.672",
"0.04575",
"0.6816",
"0.03513",
"0.677"
],
[
"0.0801",
"0.6655",
"0.4697",
"0.6655",
"0.4697",
"0.6787",
"0.0801",
"0.6787"
],
[
"0.02777",
"0.675",
"0.05392",
"0.674",
"0.05637",
"0.7085",
"0.03105",
"0.709"
],
[
"0.5024",
"0.6694",
"0.8906",
"0.6694",
"0.8906",
"0.683",
"0.5024",
"0.683"
],
[
"0.0801",
"0.6797",
"0.468",
"0.6797",
"0.468",
"0.6934",
"0.0801",
"0.6934"
],
[
"0.501",
"0.6846",
"0.8896",
"0.6846",
"0.8896",
"0.6978",
"0.501",
"0.6978"
],
[
"0.0825",
"0.6963",
"0.4673",
"0.6963",
"0.4673",
"0.707",
"0.0825",
"0.707"
],
[
"0.5034",
"0.6997",
"0.892",
"0.6997",
"0.892",
"0.7134",
"0.5034",
"0.7134"
],
[
"0.0809",
"0.709",
"0.4192",
"0.709",
"0.4192",
"0.7227",
"0.0809",
"0.7227"
],
[
"0.505",
"0.7153",
"0.8906",
"0.7153",
"0.8906",
"0.729",
"0.505",
"0.729"
],
[
"0.5034",
"0.73",
"0.8906",
"0.73",
"0.8906",
"0.744",
"0.5034",
"0.744"
],
[
"0.5034",
"0.7456",
"0.892",
"0.7456",
"0.892",
"0.759",
"0.5034",
"0.759"
],
[
"0.0809",
"0.755",
"0.2075",
"0.755",
"0.2075",
"0.7695",
"0.0809",
"0.7695"
],
[
"0.505",
"0.761",
"0.8906",
"0.761",
"0.8906",
"0.7744",
"0.505",
"0.7744"
],
[
"0.0972",
"0.78",
"0.4697",
"0.781",
"0.4697",
"0.7954",
"0.0972",
"0.7944"
],
[
"0.5034",
"0.7764",
"0.5293",
"0.7764",
"0.5293",
"0.789",
"0.5034",
"0.789"
],
[
"0.0809",
"0.797",
"0.4697",
"0.797",
"0.4697",
"0.81",
"0.0809",
"0.81"
],
[
"0.5215",
"0.7954",
"0.892",
"0.7954",
"0.892",
"0.81",
"0.5215",
"0.81"
],
[
"0.0801",
"0.811",
"0.468",
"0.811",
"0.468",
"0.8247",
"0.0801",
"0.8247"
],
[
"0.5034",
"0.811",
"0.892",
"0.811",
"0.892",
"0.8247",
"0.5034",
"0.8247"
],
[
"0.0809",
"0.827",
"0.4697",
"0.827",
"0.4697",
"0.8403",
"0.0809",
"0.8403"
],
[
"0.5024",
"0.827",
"0.892",
"0.827",
"0.892",
"0.8403",
"0.5024",
"0.8403"
],
[
"0.0809",
"0.8413",
"0.468",
"0.8413",
"0.468",
"0.855",
"0.0809",
"0.855"
],
[
"0.5034",
"0.8413",
"0.892",
"0.8413",
"0.892",
"0.855",
"0.5034",
"0.855"
],
[
"0.0801",
"0.857",
"0.4673",
"0.857",
"0.4673",
"0.8706",
"0.0801",
"0.8706"
],
[
"0.5024",
"0.857",
"0.8936",
"0.857",
"0.8936",
"0.8706",
"0.5024",
"0.8706"
],
[
"0.0801",
"0.872",
"0.468",
"0.872",
"0.468",
"0.886",
"0.0801",
"0.886"
],
[
"0.501",
"0.8706",
"0.892",
"0.872",
"0.892",
"0.886",
"0.501",
"0.8853"
],
[
"0.0801",
"0.887",
"0.4697",
"0.887",
"0.4697",
"0.901",
"0.0801",
"0.901"
],
[
"0.501",
"0.887",
"0.8906",
"0.887",
"0.8906",
"0.901",
"0.501",
"0.901"
]
] | [
"Probabilistic Debiasing of Scene Graphs",
"Bashirul Azam Biswas and Qiang Ji",
"Rensselaer Polytechnic Institute, Troy, NY-12180",
"[biswab,jigl@rpi.edu",
"",
"Abstract",
"jor computer vision research arenas after the introduction",
"of Visual Genome (VG) dataset [11]. The distribution of",
"The quality of scene graphs generated by the state-of-",
"triplets in VG images has two distinct characteristics: (1)",
"the-art (SOTA) models is compromised due to the long-tail",
"the presence of strong within-triplet prior, and (2) the long-",
"nature of the relationships and their parent object pairs.",
"tail distribution of the relationship. As shown in Figure 1",
"Training of the scene graphs is dominated by the major-",
"(a), the within-triplet prior dictates that window’ will most",
"ity relationships of the majority pairs and, therefore, the",
"likely be ‘on' the “building' rather than 'eating' it. Zeller",
"et al. [43] has utilized this within-triplet prior as the condi-",
"4",
"object-conditional distributions of relationship in the mi-",
"nority pairs are not preserved after the training is con-",
"tional probability of relationships given subject and object",
"LAD",
"verged. Consequently, the biased model performs well on",
"by proposing a frequency baseline in the SGG task. On",
"more frequent relationships in the marginal distribution of",
"the other hand, the distribution of relationship labels suffers",
"",
"relationships such as ‘on' and 'wearing', and performs",
"from a long-tailed nature and Tang et al. [27] addressed this",
"S",
"poorly on the less frequent relationships such as 'eating'",
"long-tailed issueby considering a causal interpretation of",
"or ‘hanging from'. In this work, we propose virtual evi-",
"the biased prediction. We argue that these two seemingly",
"dence incorporated within-triplet Bayesian Network(BN)",
"different characteristics of the relationship distribution are",
"2",
"to preserve the object-conditional distribution of the re-",
"interrelated. The abundance of the head classes of the rela-",
"A",
"lationship label and to eradicate the bias created by the",
"tionship distribution in Figure 1 (c), such as 'on' and 'wear-",
"7",
"marginal probability of the relationships. The insufficient",
"ing',arisesfrom the abundance of their parent subject and",
"number of relationships in the minority classes poses a sig-",
"object lying in the head region of Figure 1 (b).",
"6",
"nificant problem in learning the within-triplet Bayesian net-",
"As indicated by [5], the long-tailed distribution exists",
"",
"work.We address this insufficiency by embedding-based",
"both in relationship and object label. Since relationship",
"augmentation of triplets where we borrow samples of the",
"labels are dependent on their object pair because of the",
"minority triplet classes from its neighborhood triplets in the",
"2",
"within-triplet prior, the long-tail distribution of the relation-",
"semantic space.We perform experiments on two different",
"ship worsens due to the long-tail nature of the object pairs.",
"datasets and achieve a significant improvement in the mean",
"2",
".A",
"Crowd-collection of VG images creates selection bias and",
"recall of the relationships. We also achieve better balance",
"crowd-annotationoftheseimagescreatelabel-bias[29and",
"",
"between recall and mean recall performance compared to",
"XI",
"co-occurring-bias [25]. To analyze such biases, we investi-",
"the SOTA de-biasing techniques of scene graph models.",
"gate the distribution of the object pair of the triplets in VG",
"Codeispubliclyavailableathttps://github.com/",
"database. As shown in Figure 1 (b), window-building' and",
"bashirulazam/within-triplet-debias.",
"'man-shirt' are the most frequently annotated pairs and top",
"1% object pair covers 33% of all triplets. As a result, the",
"dominant relationships in these head pairs, such as ‘on' and",
"1. Introduction",
"wearing', dominate the marginal distribution of Figure 1",
"Any visual relationship can be expressed as a triplet",
"(c).",
"subject-relationship-object and all triplets in an image can",
"In training a deep-learning-based SGG model, samplers",
"be represented as a concise graph called Scene Graph (SG)",
"will sample more relationships from the head pairs. As a",
"[19] where the nodes represent the objects and the edges",
"result, the Maximum Likelihood Estimation (MLE) of the",
"represent relationships. This representation has been proven",
"parameters is biased to predict the relationship classes in the",
"useful for many downstream tasks such as image caption-",
"head pairs [27] and the object-conditional representation of",
"ing [37], visual reasoning [24], and image generation [10].",
"the relationship in the tail pairs will be lost in the training",
"Scene Graph Generation (SGG) has become one of the ma-",
"process. Therefore, various deep learning-based models,"
] | {
"arxiv_id": "2211.06444",
"language": "en",
"page_idx": 0
} |
||
2211.06444-en-3 | 2211.06444-en-3.jpg | $B_o.$ We consider the semantic components as random categorical variables and we aim to model their joint distribution $P(S,R,O).$ We assume there exists no latent confounder between the subject and object node whereas the relationship node depends on its parent subject and object. Formally, we assume the following statements hold true for any triplet -
1. Relationship label of a triplet depends on its subject
and object $\Longrightarrow S\to R\leftarrow O;$
2. Subject and object are independent, not given the rela-
tionship $\Longrightarrow S\perp O\not|R;$
3. Subject and object becomes dependent, given the rela-
tionship $\Longrightarrow S\not\bot O\mid R$
Based on these assumptions, we can build a Bayesian network of triplet, as shown in Figure 3, which encodes the joint distribution using the chain rule as follows
$$P(S,R,O)=P(S)P(O)P(R|S,O)$$
(2)
where $P(S)$ and $P(O)$ represent the marginal distribution of the parents $S$ and $O$, and $P(R|S,O)$ represents the conditional distribution of the relationship $R$ given its parent subject and object. This Bayesian network encodes the prior joint distribution which resides within a triplet and hence we term it as within-triplet Bayesian network. In the next subsection, we discuss how we can debias the measurement probability of a relationship by incorporating them as uncertain evidence into this Bayesian network to perform posterior inference.
<FigureHere>
Figure 3. Within-triplet Bayesian network where prior probabilities of subject and object are combined with conditional probability of relationship.
### 4.2. Uncertain evidence
We denote any trained SGG model with parameter $\theta_m$ which generates measurement probabilities of subject, object, and relationship of every triplet for an image $I$ as $P_{\theta _{m}, I}( S) $, $P_{\theta _{m}, I}( O) $,and$P_{\theta _{m}, I}( R) .$ We consider these measurements as uncertain evidence of the nodes in the within-triplet BN in Figure 3. Since this evidence is uncertain, we incorporate it as virtual evidence nodes into the BN as shown in Figure 4. Following the virtual evidence method proposed by Judea Pearl in [20], we introduce binary virtual evidence nodes $Z_s,Z_o$, and $Z_r$ as children of
<FigureHere>
Figure 4. Uncertain evidence of each entity of triplet is incorporated into the Bayesian network as the conditional probability of virtual evidence node denoted as $Z_s,Z_o$, and $Z_r.$
their respective parent evidence nodes $S,R$, and $O$ and instantiate them as True. According to Theorem 5 in [1], the conditional distributions of these virtual nodes maintain the following likelihood ratios
$$P(Z_{s}=1|s_{1}):...:P(Z_{s}=1|s_{n})=\frac{P_{I,\theta_{m}}(s_{1})}{P(s_{1})}:...:\frac{P_{I,\theta_{m}}(s_{n})}{P(s_{n})}\\P(Z_{o}=1|o_{1}):...:P(Z_{o}=1|o_{n})=\frac{P_{I,\theta_{m}}(o_{1})}{P(o_{1})}:...:\frac{P_{I,\theta_{m}}(o_{n})}{P(o_{n})}\\P(Z_{r}=1|r_{1}):...:P(Z_{r}=1|r_{n})=\frac{P_{I,\theta_{m}}(r_{1})}{P(r_{1})}:...:\frac{P_{I,\theta_{m}}(r_{n})}{P(r_{n})}$$
where $P( S$ = $s) , P( O$ = $o) $, and $P( R$ = $r)$ are the marginal probabilities of subject, object, and relationship ${\mathrm{node~and~}P_{\theta_{m},I}(S=s),P_{\theta_{m},I}(O=o),{\mathrm{and~}}P_{\theta_{m},I}(R=r)}$ are their observed measurement probabilities from image $I$ with model $\theta_m.$ Now, we have a complete Bayesian network in Figure 4 with well-defined marginal and conditional probabilities. A brief discussion on uncertain evidence and its incorporation into the Bayesian network is discussed in Appendix A of the supplementary material.
4.3. Within-Triplet Inference (WTI) of triplets
### After the evidence incorporation as virtual evidence nodes, the posterior joint distribution of triplet nodes becomes
$$\begin{aligned}&P(S,R,O|Z_{s}=1,Z_{o}=1,Z_{r}=1)\\&\propto P(Z_{s}=1|S)P(S)P(Z_{o}=1|O)P(O)P(Z_{r}=1|R)P(R|S,O)\\&\propto P_{I,\theta_{m}}(S)P_{I,\theta_{m}}(O)\frac{P_{I,\theta_{m}}(R)}{P(R)}P(R|S,O)\end{aligned}$$
The Maximum a-Posterior (MAP) of this posterior joint
distribution becomes
$$\begin{aligned}&S^{*},R^{*},O^{*}=\arg\max_{S,R,O}P(S,R,O|Z_{s}=1,Z_{o}=1,Z_{r}=1)\\&=\arg\max_{S,R,O}P_{I,\theta_{m}}(S)P_{I,\theta_{m}}(O)\frac{P_{I,\theta_{m}}(R)}{P(R)}P(R|S,O)\end{aligned}$$ | [
[
"0.597",
"0.0884",
"0.626",
"0.0884",
"0.626",
"0.101",
"0.597",
"0.101"
],
[
"0.767",
"0.089",
"0.7974",
"0.089",
"0.7974",
"0.101",
"0.767",
"0.101"
],
[
"0.0801",
"0.108",
"0.468",
"0.1067",
"0.468",
"0.1206",
"0.0801",
"0.1218"
],
[
"0.0825",
"0.09216",
"0.468",
"0.09216",
"0.468",
"0.1061",
"0.0825",
"0.1061"
],
[
"0.7695",
"0.11554",
"0.777",
"0.11554",
"0.777",
"0.1218",
"0.7695",
"0.1218"
],
[
"0.0801",
"0.1225",
"0.468",
"0.1225",
"0.468",
"0.1364",
"0.0801",
"0.1364"
],
[
"0.6733",
"0.1225",
"0.7295",
"0.1225",
"0.7295",
"0.137",
"0.6733",
"0.137"
],
[
"0.0801",
"0.1383",
"0.4697",
"0.1383",
"0.4697",
"0.1515",
"0.0801",
"0.1515"
],
[
"0.5596",
"0.1445",
"0.609",
"0.1445",
"0.609",
"0.1604",
"0.5596",
"0.1604"
],
[
"0.7837",
"0.1445",
"0.835",
"0.1445",
"0.835",
"0.159",
"0.7837",
"0.159"
],
[
"0.0801",
"0.154",
"0.4673",
"0.154",
"0.4673",
"0.1674",
"0.0801",
"0.1674"
],
[
"0.0784",
"0.1666",
"0.4697",
"0.1674",
"0.4697",
"0.1818",
"0.0784",
"0.1812"
],
[
"0.0809",
"0.1843",
"0.1602",
"0.1843",
"0.1602",
"0.1957",
"0.0809",
"0.1957"
],
[
"0.699",
"0.1787",
"0.7466",
"0.1787",
"0.7466",
"0.1901",
"0.699",
"0.1901"
],
[
"0.6006",
"0.1881",
"0.615",
"0.1881",
"0.615",
"0.1963",
"0.6006",
"0.1963"
],
[
"0.0907",
"0.2058",
"0.468",
"0.207",
"0.468",
"0.2216",
"0.0907",
"0.2203"
],
[
"0.1111",
"0.2229",
"0.3186",
"0.2229",
"0.3186",
"0.2361",
"0.1111",
"0.2361"
],
[
"0.0907",
"0.2474",
"0.4673",
"0.2474",
"0.4673",
"0.2615",
"0.0907",
"0.2615"
],
[
"0.5034",
"0.2456",
"0.888",
"0.2456",
"0.888",
"0.2556",
"0.5034",
"0.2556"
],
[
"0.1111",
"0.2632",
"0.2932",
"0.2632",
"0.2932",
"0.2766",
"0.1111",
"0.2766"
],
[
"0.5034",
"0.2583",
"0.892",
"0.2583",
"0.892",
"0.2715",
"0.5034",
"0.2715"
],
[
"0.5024",
"0.2727",
"0.7935",
"0.2727",
"0.7935",
"0.286",
"0.5024",
"0.286"
],
[
"0.0907",
"0.286",
"0.4673",
"0.2874",
"0.4673",
"0.3018",
"0.0907",
"0.3005"
],
[
"0.1095",
"0.3018",
"0.2893",
"0.3018",
"0.2893",
"0.315",
"0.1095",
"0.315"
],
[
"0.5024",
"0.31",
"0.8906",
"0.31",
"0.8906",
"0.324",
"0.5024",
"0.324"
],
[
"0.0801",
"0.327",
"0.4673",
"0.327",
"0.4673",
"0.3403",
"0.0801",
"0.3403"
],
[
"0.5024",
"0.3257",
"0.892",
"0.3257",
"0.892",
"0.339",
"0.5024",
"0.339"
],
[
"0.0801",
"0.3423",
"0.4697",
"0.3423",
"0.4697",
"0.356",
"0.0801",
"0.356"
],
[
"0.5024",
"0.3403",
"0.892",
"0.3403",
"0.892",
"0.3542",
"0.5024",
"0.3542"
],
[
"0.0784",
"0.358",
"0.3945",
"0.358",
"0.3945",
"0.3718",
"0.0784",
"0.3718"
],
[
"0.5024",
"0.356",
"0.6807",
"0.356",
"0.6807",
"0.3694",
"0.5024",
"0.3694"
],
[
"0.152",
"0.3838",
"0.3962",
"0.3838",
"0.3962",
"0.3984",
"0.152",
"0.3984"
],
[
"0.4468",
"0.385",
"0.4697",
"0.385",
"0.4697",
"0.3984",
"0.4468",
"0.3984"
],
[
"0.7295",
"0.3882",
"0.8037",
"0.3882",
"0.8037",
"0.4028",
"0.7295",
"0.4028"
],
[
"0.8267",
"0.3882",
"0.898",
"0.3882",
"0.898",
"0.4028",
"0.8267",
"0.4028"
],
[
"0.7417",
"0.404",
"0.786",
"0.404",
"0.786",
"0.4185",
"0.7417",
"0.4185"
],
[
"0.5024",
"0.3926",
"0.74",
"0.3882",
"0.74",
"0.4092",
"0.5024",
"0.4136"
],
[
"0.8423",
"0.404",
"0.888",
"0.404",
"0.888",
"0.4185",
"0.8423",
"0.4185"
],
[
"0.0809",
"0.4124",
"0.468",
"0.4124",
"0.468",
"0.4255",
"0.0809",
"0.4255"
],
[
"0.4983",
"0.4204",
"0.9043",
"0.4136",
"0.9043",
"0.4382",
"0.4983",
"0.4458"
],
[
"0.0801",
"0.4268",
"0.468",
"0.4268",
"0.468",
"0.44",
"0.0801",
"0.44"
],
[
"0.7427",
"0.4363",
"0.785",
"0.4363",
"0.785",
"0.4475",
"0.7427",
"0.4475"
],
[
"0.839",
"0.435",
"0.886",
"0.435",
"0.886",
"0.4495",
"0.839",
"0.4495"
],
[
"0.0801",
"0.4426",
"0.468",
"0.4426",
"0.468",
"0.4558",
"0.0801",
"0.4558"
],
[
"0.7305",
"0.4507",
"0.8013",
"0.4507",
"0.8013",
"0.4653",
"0.7305",
"0.4653"
],
[
"0.8267",
"0.4507",
"0.898",
"0.4507",
"0.898",
"0.4653",
"0.8267",
"0.4653"
],
[
"0.0809",
"0.459",
"0.468",
"0.459",
"0.468",
"0.4697",
"0.0809",
"0.4697"
],
[
"0.5034",
"0.4583",
"0.728",
"0.4583",
"0.728",
"0.4717",
"0.5034",
"0.4717"
],
[
"0.7417",
"0.4666",
"0.786",
"0.4666",
"0.786",
"0.481",
"0.7417",
"0.481"
],
[
"0.8423",
"0.4666",
"0.8857",
"0.4666",
"0.8857",
"0.481",
"0.8423",
"0.481"
],
[
"0.0801",
"0.473",
"0.4697",
"0.473",
"0.4697",
"0.486",
"0.0801",
"0.486"
],
[
"0.871",
"0.4797",
"0.8936",
"0.4797",
"0.8936",
"0.4937",
"0.871",
"0.4937"
],
[
"0.0801",
"0.4873",
"0.4697",
"0.4873",
"0.4697",
"0.5005",
"0.0801",
"0.5005"
],
[
"0.0801",
"0.5024",
"0.4697",
"0.5024",
"0.4697",
"0.5166",
"0.0801",
"0.5166"
],
[
"0.5034",
"0.5005",
"0.892",
"0.5005",
"0.892",
"0.515",
"0.5034",
"0.515"
],
[
"0.0801",
"0.5176",
"0.468",
"0.5176",
"0.468",
"0.531",
"0.0801",
"0.531"
],
[
"0.5024",
"0.5166",
"0.8906",
"0.5166",
"0.8906",
"0.53",
"0.5024",
"0.53"
],
[
"0.0801",
"0.5327",
"0.468",
"0.5327",
"0.468",
"0.547",
"0.0801",
"0.547"
],
[
"0.5034",
"0.532",
"0.8906",
"0.532",
"0.8906",
"0.5454",
"0.5034",
"0.5454"
],
[
"0.0809",
"0.55",
"0.1847",
"0.55",
"0.1847",
"0.56",
"0.0809",
"0.56"
],
[
"0.5024",
"0.5454",
"0.892",
"0.547",
"0.892",
"0.561",
"0.5024",
"0.56"
],
[
"0.5034",
"0.561",
"0.8906",
"0.561",
"0.8906",
"0.5747",
"0.5034",
"0.5747"
],
[
"0.1814",
"0.5713",
"0.2091",
"0.5713",
"0.2091",
"0.5835",
"0.1814",
"0.5835"
],
[
"0.339",
"0.5693",
"0.3718",
"0.5693",
"0.3718",
"0.5854",
"0.339",
"0.5854"
],
[
"0.5034",
"0.579",
"0.8906",
"0.579",
"0.8906",
"0.589",
"0.5034",
"0.589"
],
[
"0.3464",
"0.5967",
"0.353",
"0.5967",
"0.353",
"0.604",
"0.3464",
"0.604"
],
[
"0.5034",
"0.5923",
"0.892",
"0.5923",
"0.892",
"0.606",
"0.5034",
"0.606"
],
[
"0.1871",
"0.6016",
"0.1929",
"0.5967",
"0.1969",
"0.5986",
"0.1912",
"0.604"
],
[
"0.2532",
"0.603",
"0.304",
"0.603",
"0.304",
"0.617",
"0.2532",
"0.617"
],
[
"0.505",
"0.6094",
"0.8906",
"0.6094",
"0.8906",
"0.619",
"0.505",
"0.619"
],
[
"0.5024",
"0.622",
"0.7876",
"0.622",
"0.7876",
"0.636",
"0.5024",
"0.636"
],
[
"0.267",
"0.629",
"0.2737",
"0.627",
"0.2778",
"0.634",
"0.2712",
"0.636"
],
[
"0.505",
"0.646",
"0.8555",
"0.646",
"0.8555",
"0.659",
"0.505",
"0.659"
],
[
"0.0801",
"0.657",
"0.468",
"0.657",
"0.468",
"0.6704",
"0.0801",
"0.6704"
],
[
"0.0801",
"0.672",
"0.468",
"0.672",
"0.468",
"0.685",
"0.0801",
"0.685"
],
[
"0.523",
"0.6685",
"0.892",
"0.6685",
"0.892",
"0.6816",
"0.523",
"0.6816"
],
[
"0.0801",
"0.685",
"0.1896",
"0.685",
"0.1896",
"0.6987",
"0.0801",
"0.6987"
],
[
"0.5024",
"0.6846",
"0.8906",
"0.6846",
"0.8906",
"0.6978",
"0.5024",
"0.6978"
],
[
"0.5024",
"0.7007",
"0.5474",
"0.7007",
"0.5474",
"0.712",
"0.5024",
"0.712"
],
[
"0.5034",
"0.7173",
"0.726",
"0.7173",
"0.726",
"0.731",
"0.5034",
"0.731"
],
[
"0.0801",
"0.726",
"0.2607",
"0.7266",
"0.2607",
"0.741",
"0.0801",
"0.7407"
],
[
"0.5063",
"0.735",
"0.91",
"0.735",
"0.91",
"0.7485",
"0.5063",
"0.7485"
],
[
"0.10046",
"0.7505",
"0.4673",
"0.7505",
"0.4673",
"0.7646",
"0.10046",
"0.7646"
],
[
"0.501",
"0.75",
"0.7812",
"0.75",
"0.7812",
"0.779",
"0.501",
"0.779"
],
[
"0.0809",
"0.7666",
"0.4673",
"0.7666",
"0.4673",
"0.78",
"0.0809",
"0.78"
],
[
"0.653",
"0.7695",
"0.6978",
"0.7695",
"0.6978",
"0.781",
"0.653",
"0.781"
],
[
"0.0784",
"0.782",
"0.4697",
"0.781",
"0.4697",
"0.7944",
"0.0784",
"0.7954"
],
[
"0.871",
"0.782",
"0.8936",
"0.782",
"0.8936",
"0.7954",
"0.871",
"0.7954"
],
[
"0.0809",
"0.797",
"0.4697",
"0.797",
"0.4697",
"0.811",
"0.0809",
"0.811"
],
[
"0.5215",
"0.802",
"0.892",
"0.803",
"0.892",
"0.8174",
"0.5215",
"0.8164"
],
[
"0.0809",
"0.8125",
"0.4673",
"0.8125",
"0.4673",
"0.8228",
"0.0809",
"0.8228"
],
[
"0.5024",
"0.8184",
"0.6436",
"0.8184",
"0.6436",
"0.832",
"0.5024",
"0.832"
],
[
"0.0784",
"0.8257",
"0.4673",
"0.827",
"0.4673",
"0.8403",
"0.0784",
"0.839"
],
[
"0.5073",
"0.834",
"0.8857",
"0.834",
"0.8857",
"0.8594",
"0.5073",
"0.8594"
],
[
"0.0801",
"0.8413",
"0.468",
"0.8413",
"0.468",
"0.855",
"0.0801",
"0.855"
],
[
"0.0801",
"0.856",
"0.4697",
"0.856",
"0.4697",
"0.869",
"0.0801",
"0.869"
],
[
"0.7036",
"0.856",
"0.846",
"0.863",
"0.8447",
"0.884",
"0.7017",
"0.877"
],
[
"0.0801",
"0.872",
"0.468",
"0.872",
"0.468",
"0.8853",
"0.0801",
"0.8853"
],
[
"0.517",
"0.867",
"0.716",
"0.867",
"0.716",
"0.882",
"0.517",
"0.882"
],
[
"0.5557",
"0.8774",
"0.5923",
"0.8774",
"0.5923",
"0.888",
"0.5557",
"0.888"
],
[
"0.722",
"0.8755",
"0.7646",
"0.8755",
"0.7646",
"0.89",
"0.722",
"0.89"
],
[
"0.0801",
"0.886",
"0.4714",
"0.886",
"0.4714",
"0.901",
"0.0801",
"0.901"
],
[
"0.871",
"0.8896",
"0.8936",
"0.8896",
"0.8936",
"0.903",
"0.871",
"0.903"
]
] | [
"P(S)",
"P(O)",
"egorical variables and we aim to model their joint distri-",
"Bo. We consider the semantic components as random cat-",
"7",
"bution P(S, R, O). We assume there exists no latent con-",
"P(R|S,O)",
"founder between the subject and object node whereas the",
"P(ZsIS)",
"P(Zol0)",
"relationship node depends on its parent subject and object.",
"Formally,we assume the following statements hold true for",
"any triplet -",
"P(Zr|R)",
"Z.",
"1. Relationship label of a triplet depends on its subject",
"andobject → S→RO;",
"2. Subject and object are independent, not given the rela-",
"Figure4.Uncertain evidenceof eachentity of triplet is incorpo-",
"tionship → S I O XR;",
"rated into the Bayesian network as the conditional probability of",
"virtual evidence node denoted as Zs, Zo, and Zr.",
"3. Subject and object becomes dependent, given the rela-",
"tionship → S OR",
"their respective parent evidence nodes S, R, and O and in-",
"Based on these assumptions, we can build a Bayesian net-",
"stantiate them as True. According to Theorem 5 in [Il, the",
"work of triplet, as shown in Figure 3, which encodes the",
"conditional distributions of these virtual nodes maintain the",
"joint distribution using the chain rule as follows",
"following likelihood ratios",
"P(S, R,O) = P(S)P(O)P(RS,O)",
"(2)",
"Pr,em(s1)",
"Pr,om(sn)",
"P(s1)",
"P(Ze= 1s1) :.: P(Z. = 1sn) = I",
"P(sn)",
"where P(S) and P(O) represent the marginal distribution",
"P(Zo=1o1) .: (Z。=1lon) = Pror (o1) ..Prom(on)",
"of the parents S and O, and P(R|S, O) represents the con-",
"P(oi)",
"P(on)",
"ditional distribution of the relationship R given its parent",
"Pr.0m(ri)",
"Pr,om(rn)",
"subject and object.This Bayesian network encodestheprior",
"P(Z, = 1|r1) :.. : P(Zr = 1|rn) =",
"P(ri)",
"P(rn)",
"joint distribution which resides within a triplet and hence",
"(3)",
"we term it as within-triplet Bayesian network. In the next",
"subsection, we discuss how we can debias the measurement",
"where P(S = s), P(O = o), and P(R = r) are the",
"probability of a relationship by incorporating them as un-",
"marginal probabilities of subject, object, and relationship",
"certain evidence into this Bayesian network to perform pos-",
"node and Pem1(S = s), Pem.r(O= o), and Pem.1(R = r)",
"teriorinference.",
"are their observed measurement probabilities from image",
"I with model m. Now, we have a complete Bayesian net-",
"P(S)",
"P(0)",
"workinFigure4withwell-definedmarginalandconditional",
"0",
"probabilities. A brief discussion on uncertain evidence and",
"r",
"P(RIS, O)",
"itsincorporationintotheBayesiannetworkisdiscussedin",
"Appendix A of the supplementary material.",
"R",
"4.3. Within-Triplet Inference (WTI) of triplets",
"Figure 3. Within-triplet Bayesian network where prior probabili-",
"ties of subject and object are combined with conditional probabil-",
"After the evidence incorporation as virtual evidence",
"ity of relationship.",
"nodes, the posterior joint distribution of triplet nodes be-",
"comes",
"P(S, R, OZ.= 1, Z= 1, Z, = 1)",
"4.2. Uncertain evidence",
"α P(Z= 1|S)P(S)P(Z。= 1|O)P(O)P(Z=1|R)P(R|S, O)",
"We denote any trained SGG model with parameter Om",
"x Prom(S)Pi0m(0)P(R P(RIS,O)",
"which generates measurement probabilities of subject, ob-",
"P(R)",
"ject, and relationship of every triplet for an image I as",
"(4)",
"Pem,I(S), Pom,I(O), and Pem,1(R). We consider these",
"The Maximum a-Posterior (MAP) of this posterior joint",
"measurements as uncertain evidence of the nodes in the",
"distribution becomes",
"within-triplet BN in Figure 3. Since this evidence is un-",
"S\",R,0\"= arg gma P(S, R,O|Z.= 1,Z。 = 1,Z, =1)",
"certain, we incorporate it as virtual evidence nodes into the",
"BN as shown in Figure 4. Following the virtual evidence",
") Pr.0m (R) P(RIS, O)",
"method proposed by Judea Pearl in [20], we introduce bi-",
"= arg max Pr,om (S)Pr,0m (O)",
"S,R.O",
"P(R)",
"nary virtual evidence nodes Zs, Zo, and Z, as children of",
"(5)"
] | {
"arxiv_id": "2211.06444",
"language": "en",
"page_idx": 3
} |
||
2306.16678-en-0 | 2306.16678-en-0.jpg | BinaryViT: Pushing Binary Vision Transformers Towards Convolutional Models
Phuoc-Hoan Charles Le $^{*}$
le.charles55@gmail.com
Xinlin Li
Huawei Noah's Ark Lab xinlin.lil@huawei.com
## $\mathbf{Abstract}$
With the increasing popularity and the increasing size $ofvision\textit{ transformers ( ViTs) , there has been an increasing}$ interest in making them more efficient and less computationally costly for deployment on edge devices with limited computing resources. Binarization can be used to help reduce the size of ViT models and their computational cost significantly, using popcount operations when the weights and the activations are in binary. However, ViTs suffer a larger performance drop when directly applying convolutional neural network (CNN) binarization methods or existing binarization methods to binarize ViTs compared to CNNs on datasets with a large number of classes such as $ImageNet-Ik.With$ extensive analysis, we find that binary
vamilla ViTs such as DeiT miss out on a lot of key architec- tions can facilitate the use of popcount operations to reduce to porations to reduce to pons contions to reduce to pons contions to reduce o to pons con fons to reduce o to s to to s to to
vanilla ViT. Therefore, we propose BinaryViT, in which inspired by the CNN architecture, we include operations from the CNN architecture into a pure ViT architecture to enrich the representational capability of a binary ViT without introducing convolutions. These include an average pooling layer instead of a token pooling layer, a block that contains multiple average pooling branches, an affine transformation right before the addition of each main residual connection, and a pyramid structure. Experimental results on the ImageNet- lk dataset show the effectiveness of these operations that allow a binary pure ViT model to be competitive with previous state-of-the-art (SOTA) binary CNN models.
of data with proper data augmentation and regularization.
With the rising popularity and the increasing size of vision transformers (ViTs), there has been a rising interest in making them more efficient and less computationally costly to deploy them onto edge devices with limited resources such as smartphones, smart watches, etc. Therefore, model compression techniques such as quantization [20,21,26,30,47,50], distillation [18,30,50], pruning [12,34,48], etc. have been actively studied to reduce the model size and computational cost of transformers. Among these compression methods, quantization can not only reduce the memory requirement of the model but can also replace expensive floating-point operations with simpler fixed-point operations. An extreme form of quantization is binarization. Binarization of weights and activamatching the performance of a binary transformer with binary weights and activations with its full-precision counterpart is still a challenging task.
Previous works have shown that the drop in the performance can be mitigated with distillation methods from [30,50] to encourage the binary transformers to mimic the full-precision model. Also, a scaling factor can be applied to reduce the drop in the performance of DNNs due to binarization to minimize the quantization error [33]. The scaling factor can also be determined by setting it as a learnable parameter and training it to minimize the task loss [24]. For CNNs,[27] proposes Bi-RealNet a binary CNN that has extra residual connections to preserve more information to increase its representational capability to make it more accurate on tasks that have lots of classes like ImageNet-1k [7].[25] proposes ReActNet which further improves upon the work Bi-RealNet [27] by adding a learnable threshold before the sign function and by adding the RPReLU activation function after each residual connection to help reshape the output distribution at near zero extra cost. For MLPs, BiMLP [43] tries to solve the limited representational capability of fully-connected layers by proposing the multibranch block where they allow patch mixing and channel mixing to happen simultaneously.
## $1. \textbf{Introduction}$
Transformers[38] have attracted a lot of attention in natural language processing tasks such as BERT [8] and GPT [5]. Also, they are gaining a lot of attention in computer vision tasks[10,36], since they have been able to outperform most CNNs when being pre-trained on large amounts
sThis work was done when Phuoc-Hoan Charles Le was an intern at
Huawei Noah's Ark Lab Montreal Research Center.
1 | [
[
"0.0866",
"0.1338",
"0.8857",
"0.1338",
"0.8857",
"0.1483",
"0.0866",
"0.1483"
],
[
"0.2607",
"0.1831",
"0.4592",
"0.1831",
"0.4592",
"0.1976",
"0.2607",
"0.1976"
],
[
"0.5767",
"0.1812",
"0.6567",
"0.1812",
"0.6567",
"0.1989",
"0.5767",
"0.1989"
],
[
"0.259",
"0.202",
"0.4558",
"0.202",
"0.4558",
"0.2153",
"0.259",
"0.2153"
],
[
"0.5195",
"0.2008",
"0.7134",
"0.2008",
"0.7134",
"0.214",
"0.5195",
"0.214"
],
[
"0.5195",
"0.2197",
"0.709",
"0.2197",
"0.709",
"0.2329",
"0.5195",
"0.2329"
],
[
"0.02287",
"0.2664",
"0.06046",
"0.2664",
"0.06046",
"0.3738",
"0.02287",
"0.3738"
],
[
"0.2361",
"0.2676",
"0.3145",
"0.2676",
"0.3145",
"0.2854",
"0.2361",
"0.2854"
],
[
"0.5024",
"0.2695",
"0.8784",
"0.2708",
"0.8784",
"0.2854",
"0.5024",
"0.2842"
],
[
"0.523",
"0.2917",
"0.892",
"0.2917",
"0.892",
"0.3062",
"0.523",
"0.3062"
],
[
"0.0989",
"0.2998",
"0.468",
"0.3005",
"0.468",
"0.315",
"0.0989",
"0.3145"
],
[
"0.5024",
"0.307",
"0.892",
"0.307",
"0.892",
"0.3208",
"0.5024",
"0.3208"
],
[
"0.0809",
"0.3176",
"0.468",
"0.3176",
"0.468",
"0.3308",
"0.0809",
"0.3308"
],
[
"0.5024",
"0.3225",
"0.8936",
"0.3225",
"0.8936",
"0.336",
"0.5024",
"0.336"
],
[
"0.0801",
"0.332",
"0.468",
"0.332",
"0.468",
"0.3452",
"0.0801",
"0.3452"
],
[
"0.5034",
"0.3384",
"0.8936",
"0.3384",
"0.8936",
"0.3516",
"0.5034",
"0.3516"
],
[
"0.0801",
"0.3467",
"0.4714",
"0.3467",
"0.4714",
"0.3599",
"0.0801",
"0.3599"
],
[
"0.5034",
"0.3542",
"0.8906",
"0.3542",
"0.8906",
"0.3643",
"0.5034",
"0.3643"
],
[
"0.02124",
"0.363",
"0.05392",
"0.3599",
"0.05963",
"0.3945",
"0.02696",
"0.3977"
],
[
"0.0809",
"0.3623",
"0.468",
"0.3623",
"0.468",
"0.3757",
"0.0809",
"0.3757"
],
[
"0.5024",
"0.3662",
"0.8906",
"0.3674",
"0.8906",
"0.3818",
"0.5024",
"0.3806"
],
[
"0.0809",
"0.377",
"0.4697",
"0.377",
"0.4697",
"0.3901",
"0.0809",
"0.3901"
],
[
"0.5034",
"0.3833",
"0.8906",
"0.3833",
"0.8906",
"0.3965",
"0.5034",
"0.3965"
],
[
"0.0809",
"0.3926",
"0.468",
"0.3926",
"0.468",
"0.4072",
"0.0809",
"0.4072"
],
[
"0.5024",
"0.3977",
"0.8906",
"0.3977",
"0.8906",
"0.411",
"0.5024",
"0.411"
],
[
"0.02533",
"0.406",
"0.05554",
"0.406",
"0.05554",
"0.4507",
"0.02533",
"0.4507"
],
[
"0.0801",
"0.4077",
"0.4697",
"0.4077",
"0.4697",
"0.4216",
"0.0801",
"0.4216"
],
[
"0.5024",
"0.4136",
"0.892",
"0.4136",
"0.892",
"0.4268",
"0.5024",
"0.4268"
],
[
"0.0801",
"0.4224",
"0.468",
"0.4224",
"0.468",
"0.4375",
"0.0801",
"0.4375"
],
[
"0.5024",
"0.4287",
"0.892",
"0.4287",
"0.892",
"0.4426",
"0.5024",
"0.4426"
],
[
"0.0809",
"0.4375",
"0.468",
"0.4375",
"0.468",
"0.4507",
"0.0809",
"0.4507"
],
[
"0.03186",
"0.449",
"0.058",
"0.449",
"0.058",
"0.481",
"0.03186",
"0.481"
],
[
"0.5024",
"0.443",
"0.892",
"0.443",
"0.892",
"0.457",
"0.5024",
"0.457"
],
[
"0.0809",
"0.4539",
"0.468",
"0.4539",
"0.468",
"0.464",
"0.0809",
"0.464"
],
[
"0.5024",
"0.4602",
"0.8896",
"0.4602",
"0.8896",
"0.4702",
"0.5024",
"0.4702"
],
[
"0.0809",
"0.4673",
"0.4697",
"0.4673",
"0.4697",
"0.481",
"0.0809",
"0.481"
],
[
"0.5024",
"0.4734",
"0.892",
"0.4734",
"0.892",
"0.4873",
"0.5024",
"0.4873"
],
[
"0.0801",
"0.483",
"0.468",
"0.483",
"0.468",
"0.4968",
"0.0801",
"0.4968"
],
[
"0.5024",
"0.488",
"0.892",
"0.488",
"0.892",
"0.502",
"0.5024",
"0.502"
],
[
"0.0237",
"0.5005",
"0.05963",
"0.5005",
"0.05963",
"0.565",
"0.0237",
"0.565"
],
[
"0.0801",
"0.4976",
"0.468",
"0.4976",
"0.468",
"0.511",
"0.0801",
"0.511"
],
[
"0.5034",
"0.504",
"0.8936",
"0.504",
"0.8936",
"0.5176",
"0.5034",
"0.5176"
],
[
"0.0801",
"0.513",
"0.4697",
"0.513",
"0.4697",
"0.5264",
"0.0801",
"0.5264"
],
[
"0.5024",
"0.518",
"0.892",
"0.518",
"0.892",
"0.532",
"0.5024",
"0.532"
],
[
"0.0801",
"0.528",
"0.468",
"0.528",
"0.468",
"0.5415",
"0.0801",
"0.5415"
],
[
"0.5024",
"0.534",
"0.8906",
"0.534",
"0.8906",
"0.5474",
"0.5024",
"0.5474"
],
[
"0.0801",
"0.5435",
"0.468",
"0.5435",
"0.468",
"0.5566",
"0.0801",
"0.5566"
],
[
"0.5024",
"0.549",
"0.8906",
"0.549",
"0.8906",
"0.5625",
"0.5024",
"0.5625"
],
[
"0.02777",
"0.5566",
"0.05392",
"0.5566",
"0.05392",
"0.577",
"0.02777",
"0.577"
],
[
"0.0809",
"0.558",
"0.468",
"0.558",
"0.468",
"0.5728",
"0.0809",
"0.5728"
],
[
"0.5024",
"0.5645",
"0.699",
"0.5645",
"0.699",
"0.5776",
"0.5024",
"0.5776"
],
[
"0.0801",
"0.5737",
"0.4697",
"0.5737",
"0.4697",
"0.587",
"0.0801",
"0.587"
],
[
"0.02533",
"0.587",
"0.05637",
"0.587",
"0.05637",
"0.634",
"0.02533",
"0.634"
],
[
"0.0801",
"0.589",
"0.468",
"0.589",
"0.468",
"0.603",
"0.0801",
"0.603"
],
[
"0.5195",
"0.5835",
"0.8896",
"0.584",
"0.8896",
"0.5986",
"0.5195",
"0.5977"
],
[
"0.0784",
"0.603",
"0.4697",
"0.604",
"0.4697",
"0.617",
"0.0784",
"0.616"
],
[
"0.5024",
"0.5996",
"0.8906",
"0.5996",
"0.8906",
"0.613",
"0.5024",
"0.613"
],
[
"0.0801",
"0.6187",
"0.468",
"0.6187",
"0.468",
"0.632",
"0.0801",
"0.632"
],
[
"0.5034",
"0.6143",
"0.892",
"0.6143",
"0.892",
"0.6274",
"0.5034",
"0.6274"
],
[
"0.02124",
"0.6235",
"0.05637",
"0.6226",
"0.05963",
"0.709",
"0.0237",
"0.7104"
],
[
"0.0801",
"0.634",
"0.468",
"0.634",
"0.468",
"0.648",
"0.0801",
"0.648"
],
[
"0.5024",
"0.629",
"0.892",
"0.63",
"0.892",
"0.6445",
"0.5024",
"0.643"
],
[
"0.5024",
"0.6445",
"0.8906",
"0.6445",
"0.8906",
"0.6577",
"0.5024",
"0.6577"
],
[
"0.0809",
"0.651",
"0.4673",
"0.651",
"0.4673",
"0.661",
"0.0809",
"0.661"
],
[
"0.0801",
"0.664",
"0.4697",
"0.664",
"0.4697",
"0.678",
"0.0801",
"0.678"
],
[
"0.5024",
"0.66",
"0.8906",
"0.66",
"0.8906",
"0.674",
"0.5024",
"0.674"
],
[
"0.0801",
"0.6797",
"0.4697",
"0.6797",
"0.4697",
"0.6934",
"0.0801",
"0.6934"
],
[
"0.501",
"0.675",
"0.8906",
"0.674",
"0.8906",
"0.688",
"0.501",
"0.6895"
],
[
"0.0801",
"0.6943",
"0.4697",
"0.6943",
"0.4697",
"0.7085",
"0.0801",
"0.7085"
],
[
"0.5024",
"0.69",
"0.8906",
"0.69",
"0.8906",
"0.704",
"0.5024",
"0.704"
],
[
"0.0809",
"0.711",
"0.4673",
"0.711",
"0.4673",
"0.7217",
"0.0809",
"0.7217"
],
[
"0.5024",
"0.7056",
"0.8936",
"0.7056",
"0.8936",
"0.7197",
"0.5024",
"0.7197"
],
[
"0.5034",
"0.72",
"0.8936",
"0.72",
"0.8936",
"0.7344",
"0.5034",
"0.7344"
],
[
"0.5034",
"0.7373",
"0.8896",
"0.7373",
"0.8896",
"0.7476",
"0.5034",
"0.7476"
],
[
"0.505",
"0.7524",
"0.8896",
"0.7524",
"0.8896",
"0.763",
"0.505",
"0.763"
],
[
"0.0809",
"0.7666",
"0.2059",
"0.7666",
"0.2059",
"0.78",
"0.0809",
"0.78"
],
[
"0.505",
"0.7666",
"0.892",
"0.7666",
"0.892",
"0.78",
"0.505",
"0.78"
],
[
"0.5034",
"0.781",
"0.892",
"0.781",
"0.892",
"0.7944",
"0.5034",
"0.7944"
],
[
"0.0989",
"0.7925",
"0.468",
"0.7925",
"0.468",
"0.806",
"0.0989",
"0.806"
],
[
"0.5024",
"0.797",
"0.892",
"0.797",
"0.892",
"0.81",
"0.5024",
"0.81"
],
[
"0.0801",
"0.8066",
"0.468",
"0.806",
"0.468",
"0.821",
"0.0801",
"0.8213"
],
[
"0.5034",
"0.811",
"0.892",
"0.811",
"0.892",
"0.8247",
"0.5034",
"0.8247"
],
[
"0.0801",
"0.8237",
"0.4673",
"0.8237",
"0.4673",
"0.837",
"0.0801",
"0.837"
],
[
"0.5034",
"0.827",
"0.892",
"0.827",
"0.892",
"0.8403",
"0.5034",
"0.8403"
],
[
"0.0809",
"0.8384",
"0.468",
"0.8384",
"0.468",
"0.8516",
"0.0809",
"0.8516"
],
[
"0.5024",
"0.839",
"0.892",
"0.8403",
"0.892",
"0.855",
"0.5024",
"0.854"
],
[
"0.0801",
"0.854",
"0.4697",
"0.854",
"0.4697",
"0.8687",
"0.0801",
"0.8687"
],
[
"0.5034",
"0.857",
"0.8906",
"0.857",
"0.8906",
"0.8706",
"0.5034",
"0.8706"
],
[
"0.1013",
"0.8735",
"0.4714",
"0.875",
"0.4714",
"0.8896",
"0.1013",
"0.888"
],
[
"0.5024",
"0.872",
"0.892",
"0.872",
"0.892",
"0.886",
"0.5024",
"0.886"
],
[
"0.0801",
"0.887",
"0.353",
"0.888",
"0.353",
"0.8994",
"0.0801",
"0.8984"
],
[
"0.501",
"0.887",
"0.7256",
"0.887",
"0.7256",
"0.901",
"0.501",
"0.901"
],
[
"0.4836",
"0.928",
"0.4885",
"0.928",
"0.4885",
"0.933",
"0.4836",
"0.933"
]
] | [
"BinaryViT: Pushing Binary Vision Transformers Towards Convolutional Models",
"Phuoc-Hoan Charles Le *",
"Xinlin Li",
"le.charles55@gmail.com",
"Huawei Noah's Ark Lab",
"xinlin.lil@huawei.com",
" ",
"Abstract",
"of data with proper data augmentation and regularization.",
"With the rising popularity and the increasing size of",
"With the increasing popularity and the increasing size",
"vision transformers (ViTs), there has been a rising inter-",
"of vision transformers (ViTs), there has been an increasing",
"est in making them more efficient and less computation-",
"interest in making them more efficient and less computa-",
"ally costly to deploy them onto edge devices with lim-",
"tionally costly for deployment on edge devices with limited",
"ited resources such as smartphones.smart watches,etc.",
"2",
"computing resources. Binarization can be used to help re-",
"Therefore, model compression techniques such as quanti-",
"duce the size of ViT models and their computational cost",
"zation [20,21,26,30, 47, 50], distillation [18, 30, 50], prun-",
"significantly, using popcount operations when the weights",
"ing [12, 34, 48], etc. have been actively studied to re-",
"LAD",
"and the activations are in binary. However, ViTs suffer a",
"duce the model size and computational cost of transform-",
"larger performance drop when directly applying convolu-",
"ers. Among these compression methods, quantization can",
"tional neural network (CNN) binarization methods or ex-",
"SO",
"not only reduce the memory requirement of the model but",
"isting binarizationmethodsto binarizeViTscompared to",
"can also replace expensive floating-point operations with",
"CNNs on datasets with a large number of classes such as",
"simpler fixed-point operations. An extreme form of quan-",
"ImageNet-lk. With extensive analysis, we find that binary",
"tization is binarization. Binarization of weights and activa-",
"4829",
"vanilla ViTs such as DeiT miss out on a lot of key architec-",
"tions can facilitate the use of popcount operations to reduce",
"tural properties that CNNs have that allow binary CNNs to",
"the model size and reduce computational cost. However,",
"have much higher representational capability than binary",
"matching the performance of a binary transformer with bi-",
"vanilla ViT. Therefore, we propose BinaryViT, in which in-",
"nary weights and activations with its full-precision counter-",
"6",
"spired by the CNN architecture, we include operations from",
"part is still a challenging task.",
"the CNN architecture into a pure ViT architecture to enrich",
"906",
"the representational capability of a binary ViT without in-",
"Previous works have shown that the drop in the per-",
"troducing convolutions. These include an average pooling",
"formance can be mitigated with distillation methods from",
"layer instead of a token pooling layer, a block that contains",
"[30, 50] to encourage the binary transformers to mimic the",
"Z:AIXI",
"multiple average pooling branches, an affine transforma-",
"full-precision model. Also, a scaling factor can be applied",
"to reduce the drop in the performance of DNNs due to bina-",
"tion right before the addition ofeach main residual connec-",
"tion, and a pyramid structure. Experimental results on the",
"rization to minimize the quantization error [33]. The scal-",
"ImageNet-lk dataset show the effectiveness of these opera-",
"ing factor can also be determined by setting it as a learn-",
"tions that allow a binary pure ViT model to be competitive",
"able parameter and training it to minimize the task loss [24]",
"withpreviousstate-of-the-art(SOTA)binaryCNNmodels.",
"For CNNs, [27] proposes Bi-RealNet a binary CNN that has",
"extra residual connections to preserve more information to",
"increase its representational capability to make it more ac-",
"curate on tasks that have lots of classes like ImageNet-lk",
"1. Introduction",
"[7]. [25] proposes ReActNet which further improves upon",
"the work Bi-RealNet [27] by adding a learnable threshold",
"Transformers [38] have attracted a lot of attention in nat-",
"before the sign function and by adding the RPReLU activa-",
"ural language processing tasks such as BERT [8] and GPT",
"tion function after each residual connection to help reshape",
"[5]. Also, they are gaining a lot of attention in computer",
"the output distribution at near zero extra cost. For MLPs,",
"vision tasks [10, 36], since they have been able to outper-",
"BiMLP [43] tries to solve the limited representational ca-",
"form most CNNs when being pre-trained on large amounts",
"pability of fully-connected layers by proposing the multi-",
"This work was done when Phuoc-Hoan Charles Le was an intern at",
"branch block where they allow patch mixing and channel",
"Huawei Noah's Ark Lab Montreal Research Center.",
"mixing to happen simultaneously.",
"1"
] | {
"arxiv_id": "2306.16678",
"language": "en",
"page_idx": 0
} |
||
2306.16678-en-2 | 2306.16678-en-2.jpg | <table>
<tbody>
<tr>
<th>Method</th>
<th>Top- 1$\left ( \% \right )$</th>
</tr>
<tr>
<td>$\operatorname{BinaryBERT}[2]$</td>
<td>1.4</td>
</tr>
<tr>
<td>$\operatorname{BiBERT}[30]$</td>
<td>33.5</td>
</tr>
<tr>
<td>BiT[24]</td>
<td>45.7</td>
</tr>
<tr>
<td>+Remove FFN-Distill</td>
<td>46.5</td>
</tr>
<tr>
<td>+SGBERT $\left[1\right]$ </td>
<td>47.4</td>
</tr>
<tr>
<td>+ReActNet [25]</td>
<td>48.5</td>
</tr>
</tbody>
</table>
Table 1. Top 1 accuracy on ImageNet with existing transformer-
based binarization methods applied on the binary DeiT-S.
been explored in previous works relating to full-precision ViTs, but their effects on the accuracy of binary neural networks or on binary ViTs have not yet been explored. Also, to the best of our knowledge, with these operations, we are the first to be competitive with prior binary CNN models on the ImageNet-1k dataset in terms of accuracy and the number of operations (OPs), using a pure ViT architecture. The source code of this work is available on GitHub $^{\mathrm{l}}.$
## $2. \textbf{Designing a binary ViT baseline}$
First, we design a baseline binary pure ViT model using the DeiT [10,36] backbone, since, to the best of our knowledge, there has never been works on a pure ViT model with binary weights and activations. In this section, we use existing binarization techniques applied on CNN, and on BERT[8] to design the baseline binary pure ViT model as shown in Figure la and 2a.
## $2. 1. \textbf{Binary fully- connected layer}$
For the forward pass, for each matrix multiplication, a sign function is used for each input activation matrix, $\mathbf{X}\in\mathbb{R}^{N\times D_{in}}$, and weight matrix, $\hat{\mathbf{W}}\in\mathbb{R}^{D_{\mathrm{in}}\times D_{\mathrm{out}}}.$ A threshold vector, $\beta_{\mathbf{X}}\in\mathbb{R}^{D_{\mathrm{in}}}$, can be applied to the real value inputs right before applying the sign function to allow these inputs to have some distributional shift. For the weights, the threshold, $\mu ( \mathbf{W} ) \in \mathbb{R} ^{D_{\mathrm{out}}}$, can be determined by computing the mean value of all elements inside the matrix as in[30-32]. For the activations, the threshold parameter can be determined using backpropagation to minimize the task loss as in [25,44]. After applying the sign function, a scaling factor, $\alpha_\mathbf{W}\in\mathbb{R}$, where $\alpha_\mathbf{W}=\frac1n\|\mathbf{W}\|_1$, is applied as in [30]. Then, the matrix multiplication output, Y$( \mathbf{X} ) \in \mathbb{R} ^{N\times D_{\mathrm{out}}}$,can be calculated using popcount, $\otimes$,as
(1)
$$\mathrm{Y}(\mathbf{X})=\alpha_\mathbf{W}\mathrm{Rsign}(\mathbf{X})\otimes\mathrm{sign}(\mathbf{W}-\mu(\mathbf{W}))$$
where Rsign=sign$(\mathbf{X}+\beta_{\mathbf{X}})$as in[25].
For each binary fully-connected layer, BiFC, in a binary transformer, we connect a residual connection, R, from the input, $\mathbf{X}$, to the output of the linear layer, BN(Y(X)), as:
$$\mathrm{BiFC}(\mathbf{X})=\mathrm{RPReLU}(\mathrm{BN}(\mathbf{Y})+\mathrm{R}(\mathbf{X}))$$
(2)
to preserve the information from the previous layer as in [25,27]. RPReLU activation function proposed by[25] is used after each residual connection. Also, we replace all layer normalization in the ViT model with batch normalization [17], BN, since all linear layers have a normalization layer after it, as in [45], to enable faster inference and also faster training compared to layer normalization. The residual function, R, from [25] can be an identity function if the
$^{1}$https://github.com/Phuoc-Hoan-Le/BinaryViT
input and output dimensions are the same. If the input dimension is smaller than the output channel by $n$ times, the residual function will be the input concatenation with itself $n$ times. If the input dimension is larger than the output channel by $n$ times, the residual function will be the average pooling function with a stride of $n$ and a window size of $n.$ Therefore, R$(\mathbf{X})$ can be defined as
$$\mathrm{R}(\mathbf{X})=\begin{cases}\mathbf{X},&c_{in}=c_{out}\\\mathrm{Cat}([\mathbf{X},\text{for i in range(n)}],\text{dim=2}),&nc_{in}=c_{out}\\\frac{1}{n}\sum_{i=1}^{n}\mathbf{X}(:,:,\frac{i-1}{n}d_{in}:\frac{i}{n}d_{in}))&c_{in}=nc_{out}\\\end{cases}$$
where Cat is the concatenation function.
Using a straight-through estimator [3], we approximate
the derivative of sign with respect to an input as
$$\frac{\partial\operatorname{sign}(x)}{\partial x}\approx\begin{cases}1,&|x|\le1\\0,&\text{otherwis}\end{cases}$$
(4)
## 2.2. Binary vision transformer
A ViT uses N number of transformer encoder blocks and each transformer block contains one multi-head attention (MHA) module and one feed-forward network (FFN) module. Initially, in the embedding layer, the image, x $\in\mathbb{R}^{H\times W\times C}$gets split up into fixed-size patches, $\mathbf{x}_p\in$ $\mathbb{R}^{N\times(P^{2}\cdot C)}$, where $(H,W)$ is the image dimensions, C is the number of channels of the input image, $(P,P)$ is the dimension of each image patch, and $N=HW/P^2$ is the number of patches. These patches are then linearly projected by $\mathbf{E}\in\mathbb{R}^{(P^{2}\cdot C)\times D}$ before being appended with a cls-token embedding, $\mathbf{x}_{\mathrm{class}}\in\mathbb{R}^D$, and summed with the position embeddings, $\mathbf{E}_{\mathrm{pos}}\in\mathbb{R}^{(N+1)\times D}$, as in Eq. (5)
$$\mathbf H_1=[\mathbf x_{\mathrm{class}};\:\mathbf x_p^1\mathbf E;\:\mathbf x_p^2\mathbf E;...;\:\mathbf x_p^N\mathbf E]+\mathbf E_{\mathrm{pos}}$$
(5)
Like in previous works on binarization, the operations in the
first layer are always kept in full precision.
The output of the embedding layer $\mathbf{H}_1\in\mathbb{R}^{(N+1)\times D}$ then becomes the input for the first transformer block. In each transformer block, the input H$\in\mathbb{R}^{(N+1)\times D}$ first gets
3 | [
[
"0.6323",
"0.0884",
"0.687",
"0.0884",
"0.687",
"0.1029",
"0.6323",
"0.1029"
],
[
"0.74",
"0.089",
"0.8096",
"0.089",
"0.8096",
"0.1042",
"0.74",
"0.1042"
],
[
"0.763",
"0.1042",
"0.786",
"0.1042",
"0.786",
"0.11743",
"0.763",
"0.11743"
],
[
"0.0801",
"0.09216",
"0.468",
"0.09216",
"0.468",
"0.1061",
"0.0801",
"0.1061"
],
[
"0.0809",
"0.1067",
"0.468",
"0.1067",
"0.468",
"0.1206",
"0.0809",
"0.1206"
],
[
"0.6094",
"0.1048",
"0.709",
"0.1048",
"0.709",
"0.11615",
"0.6094",
"0.11615"
],
[
"0.619",
"0.1193",
"0.701",
"0.1193",
"0.701",
"0.1307",
"0.619",
"0.1307"
],
[
"0.7573",
"0.1187",
"0.789",
"0.1187",
"0.789",
"0.1307",
"0.7573",
"0.1307"
],
[
"0.0809",
"0.125",
"0.4673",
"0.125",
"0.4673",
"0.1351",
"0.0809",
"0.1351"
],
[
"0.6323",
"0.1338",
"0.687",
"0.1338",
"0.687",
"0.1459",
"0.6323",
"0.1459"
],
[
"0.7573",
"0.132",
"0.79",
"0.132",
"0.79",
"0.1477",
"0.7573",
"0.1477"
],
[
"0.0801",
"0.1383",
"0.4697",
"0.1383",
"0.4697",
"0.1515",
"0.0801",
"0.1515"
],
[
"0.5923",
"0.1465",
"0.726",
"0.1465",
"0.726",
"0.1604",
"0.5923",
"0.1604"
],
[
"0.759",
"0.1465",
"0.79",
"0.1465",
"0.79",
"0.159",
"0.759",
"0.159"
],
[
"0.0801",
"0.1528",
"0.4697",
"0.1528",
"0.4697",
"0.1666",
"0.0801",
"0.1666"
],
[
"0.614",
"0.1578",
"0.705",
"0.1604",
"0.705",
"0.1749",
"0.614",
"0.1724"
],
[
"0.7573",
"0.1604",
"0.7915",
"0.1604",
"0.7915",
"0.1755",
"0.7573",
"0.1755"
],
[
"0.0784",
"0.1666",
"0.468",
"0.1674",
"0.468",
"0.1818",
"0.0784",
"0.1812"
],
[
"0.6113",
"0.1755",
"0.7075",
"0.1755",
"0.7075",
"0.1875",
"0.6113",
"0.1875"
],
[
"0.759",
"0.1749",
"0.789",
"0.1749",
"0.789",
"0.1875",
"0.759",
"0.1875"
],
[
"0.0801",
"0.1831",
"0.468",
"0.1831",
"0.468",
"0.1963",
"0.0801",
"0.1963"
],
[
"0.0784",
"0.1976",
"0.4028",
"0.1976",
"0.4028",
"0.2108",
"0.0784",
"0.2108"
],
[
"0.5024",
"0.2008",
"0.8896",
"0.2008",
"0.8896",
"0.214",
"0.5024",
"0.214"
],
[
"0.5",
"0.2134",
"0.8433",
"0.214",
"0.8433",
"0.2291",
"0.5",
"0.2279"
],
[
"0.0809",
"0.2279",
"0.3677",
"0.2279",
"0.3677",
"0.2424",
"0.0809",
"0.2424"
],
[
"0.0972",
"0.252",
"0.468",
"0.2507",
"0.468",
"0.2651",
"0.0972",
"0.2664"
],
[
"0.0801",
"0.2683",
"0.468",
"0.2683",
"0.468",
"0.2822",
"0.0801",
"0.2822"
],
[
"0.5024",
"0.2695",
"0.8896",
"0.2695",
"0.8896",
"0.2827",
"0.5024",
"0.2827"
],
[
"0.0784",
"0.2827",
"0.468",
"0.2822",
"0.468",
"0.2966",
"0.0784",
"0.2974"
],
[
"0.5024",
"0.2854",
"0.892",
"0.2854",
"0.892",
"0.2986",
"0.5024",
"0.2986"
],
[
"0.0801",
"0.2974",
"0.4697",
"0.2974",
"0.4697",
"0.3113",
"0.0801",
"0.3113"
],
[
"0.5024",
"0.2998",
"0.892",
"0.2998",
"0.892",
"0.3132",
"0.5024",
"0.3132"
],
[
"0.0801",
"0.3132",
"0.4697",
"0.3132",
"0.4697",
"0.327",
"0.0801",
"0.327"
],
[
"0.5024",
"0.315",
"0.892",
"0.315",
"0.892",
"0.3289",
"0.5024",
"0.3289"
],
[
"0.0809",
"0.3276",
"0.4697",
"0.3276",
"0.4697",
"0.3423",
"0.0809",
"0.3423"
],
[
"0.5024",
"0.33",
"0.892",
"0.33",
"0.892",
"0.3435",
"0.5024",
"0.3435"
],
[
"0.0801",
"0.3435",
"0.259",
"0.3435",
"0.259",
"0.3567",
"0.0801",
"0.3567"
],
[
"0.5024",
"0.3452",
"0.892",
"0.3452",
"0.892",
"0.3591",
"0.5024",
"0.3591"
],
[
"0.5024",
"0.3599",
"0.7686",
"0.3599",
"0.7686",
"0.3738",
"0.5024",
"0.3738"
],
[
"0.0801",
"0.3687",
"0.3308",
"0.3687",
"0.3308",
"0.3818",
"0.0801",
"0.3818"
],
[
"0.0972",
"0.3901",
"0.468",
"0.3914",
"0.468",
"0.406",
"0.0972",
"0.4048"
],
[
"0.5703",
"0.4016",
"0.5933",
"0.4016",
"0.5933",
"0.4153",
"0.5703",
"0.4153"
],
[
"0.814",
"0.4016",
"0.8896",
"0.404",
"0.8896",
"0.4185",
"0.814",
"0.4167"
],
[
"0.0784",
"0.4185",
"0.4714",
"0.4185",
"0.4714",
"0.4363",
"0.0784",
"0.4363"
],
[
"0.0801",
"0.4072",
"0.468",
"0.4072",
"0.468",
"0.4204",
"0.0801",
"0.4204"
],
[
"0.5024",
"0.4192",
"0.5522",
"0.4192",
"0.5522",
"0.4343",
"0.5024",
"0.4343"
],
[
"0.576",
"0.4204",
"0.8057",
"0.4204",
"0.8057",
"0.4343",
"0.576",
"0.4343"
],
[
"0.8022",
"0.4204",
"0.899",
"0.4224",
"0.899",
"0.4363",
"0.8022",
"0.4343"
],
[
"0.0768",
"0.433",
"0.4697",
"0.4343",
"0.4697",
"0.452",
"0.0768",
"0.4507"
],
[
"0.574",
"0.4375",
"0.7837",
"0.4375",
"0.7837",
"0.455",
"0.574",
"0.455"
],
[
"0.8145",
"0.44",
"0.899",
"0.44",
"0.899",
"0.455",
"0.8145",
"0.455"
],
[
"0.0801",
"0.452",
"0.468",
"0.452",
"0.468",
"0.4653",
"0.0801",
"0.4653"
],
[
"0.869",
"0.455",
"0.8936",
"0.455",
"0.8936",
"0.4685",
"0.869",
"0.4685"
],
[
"0.0809",
"0.4685",
"0.4673",
"0.4685",
"0.4673",
"0.4792",
"0.0809",
"0.4792"
],
[
"0.505",
"0.4702",
"0.7686",
"0.4702",
"0.7686",
"0.481",
"0.505",
"0.481"
],
[
"0.0801",
"0.481",
"0.4697",
"0.481",
"0.4697",
"0.4956",
"0.0801",
"0.4956"
],
[
"0.5215",
"0.483",
"0.892",
"0.4841",
"0.892",
"0.4988",
"0.5215",
"0.4976"
],
[
"0.0801",
"0.4968",
"0.4673",
"0.4968",
"0.4673",
"0.511",
"0.0801",
"0.511"
],
[
"0.501",
"0.5",
"0.8164",
"0.5",
"0.8164",
"0.513",
"0.501",
"0.513"
],
[
"0.0809",
"0.5146",
"0.4673",
"0.5146",
"0.4673",
"0.5244",
"0.0809",
"0.5244"
],
[
"0.0801",
"0.528",
"0.4697",
"0.528",
"0.4697",
"0.5415",
"0.0801",
"0.5415"
],
[
"0.596",
"0.529",
"0.6685",
"0.529",
"0.6685",
"0.547",
"0.596",
"0.547"
],
[
"0.697",
"0.529",
"0.777",
"0.529",
"0.777",
"0.5435",
"0.697",
"0.5435"
],
[
"0.8677",
"0.537",
"0.8945",
"0.537",
"0.8945",
"0.5537",
"0.8677",
"0.5537"
],
[
"0.0784",
"0.5435",
"0.4673",
"0.5435",
"0.4673",
"0.5566",
"0.0784",
"0.5566"
],
[
"0.619",
"0.5474",
"0.6436",
"0.5474",
"0.6436",
"0.561",
"0.619",
"0.561"
],
[
"0.694",
"0.5474",
"0.7974",
"0.547",
"0.7974",
"0.561",
"0.694",
"0.5625"
],
[
"0.0801",
"0.5566",
"0.468",
"0.5566",
"0.468",
"0.5713",
"0.0801",
"0.5713"
],
[
"0.0801",
"0.5728",
"0.4673",
"0.5728",
"0.4673",
"0.587",
"0.0801",
"0.587"
],
[
"0.5034",
"0.5776",
"0.738",
"0.5776",
"0.738",
"0.5913",
"0.5034",
"0.5913"
],
[
"0.0801",
"0.5854",
"0.4697",
"0.5854",
"0.4697",
"0.603",
"0.0801",
"0.603"
],
[
"0.5215",
"0.601",
"0.892",
"0.601",
"0.892",
"0.6143",
"0.5215",
"0.6143"
],
[
"0.5034",
"0.616",
"0.8906",
"0.616",
"0.8906",
"0.63",
"0.5034",
"0.63"
],
[
"0.1316",
"0.631",
"0.4192",
"0.631",
"0.4192",
"0.6445",
"0.1316",
"0.6445"
],
[
"0.4478",
"0.6333",
"0.4656",
"0.6333",
"0.4656",
"0.6426",
"0.4478",
"0.6426"
],
[
"0.5034",
"0.6333",
"0.8906",
"0.6333",
"0.8906",
"0.643",
"0.5034",
"0.643"
],
[
"0.501",
"0.646",
"0.892",
"0.648",
"0.892",
"0.662",
"0.501",
"0.66"
],
[
"0.0801",
"0.6577",
"0.3464",
"0.657",
"0.3464",
"0.672",
"0.0801",
"0.6724"
],
[
"0.5",
"0.656",
"0.8945",
"0.66",
"0.8945",
"0.681",
"0.5",
"0.677"
],
[
"0.0989",
"0.6724",
"0.468",
"0.6724",
"0.468",
"0.6875",
"0.0989",
"0.6875"
],
[
"0.501",
"0.674",
"0.8936",
"0.678",
"0.8936",
"0.6963",
"0.501",
"0.6924"
],
[
"0.0801",
"0.6895",
"0.4697",
"0.6895",
"0.4697",
"0.7026",
"0.0801",
"0.7026"
],
[
"0.5034",
"0.6934",
"0.892",
"0.6934",
"0.892",
"0.707",
"0.5034",
"0.707"
],
[
"0.0801",
"0.705",
"0.452",
"0.705",
"0.452",
"0.7183",
"0.0801",
"0.7183"
],
[
"0.5",
"0.707",
"0.8936",
"0.7056",
"0.8936",
"0.7236",
"0.5",
"0.7246"
],
[
"0.5024",
"0.7236",
"0.8896",
"0.7236",
"0.8896",
"0.7373",
"0.5024",
"0.7373"
],
[
"0.1396",
"0.7344",
"0.407",
"0.7344",
"0.407",
"0.7476",
"0.1396",
"0.7476"
],
[
"0.4468",
"0.735",
"0.468",
"0.735",
"0.468",
"0.7476",
"0.4468",
"0.7476"
],
[
"0.5",
"0.735",
"0.8965",
"0.7363",
"0.8965",
"0.757",
"0.5",
"0.756"
],
[
"0.0801",
"0.7583",
"0.468",
"0.7583",
"0.468",
"0.7715",
"0.0801",
"0.7715"
],
[
"0.5",
"0.754",
"0.8936",
"0.7524",
"0.8936",
"0.771",
"0.5",
"0.7715"
],
[
"0.0809",
"0.773",
"0.468",
"0.773",
"0.468",
"0.787",
"0.0809",
"0.787"
],
[
"0.4983",
"0.7676",
"0.864",
"0.765",
"0.864",
"0.786",
"0.4983",
"0.7886"
],
[
"0.0801",
"0.7886",
"0.468",
"0.7886",
"0.468",
"0.802",
"0.0801",
"0.802"
],
[
"0.556",
"0.8003",
"0.8433",
"0.8003",
"0.8433",
"0.8184",
"0.556",
"0.8184"
],
[
"0.8677",
"0.803",
"0.8906",
"0.803",
"0.8906",
"0.815",
"0.8677",
"0.815"
],
[
"0.0801",
"0.8184",
"0.468",
"0.8184",
"0.468",
"0.832",
"0.0801",
"0.832"
],
[
"0.0809",
"0.8047",
"0.4673",
"0.8047",
"0.4673",
"0.815",
"0.0809",
"0.815"
],
[
"0.5024",
"0.8257",
"0.892",
"0.8257",
"0.892",
"0.839",
"0.5024",
"0.839"
],
[
"0.0809",
"0.8354",
"0.4673",
"0.8354",
"0.4673",
"0.845",
"0.0809",
"0.845"
],
[
"0.5024",
"0.8413",
"0.785",
"0.8413",
"0.785",
"0.855",
"0.5024",
"0.855"
],
[
"0.0809",
"0.8496",
"0.4656",
"0.8496",
"0.4656",
"0.86",
"0.0809",
"0.86"
],
[
"0.517",
"0.853",
"0.8936",
"0.8496",
"0.8936",
"0.8706",
"0.517",
"0.8735"
],
[
"0.0801",
"0.863",
"0.4697",
"0.863",
"0.4697",
"0.877",
"0.0801",
"0.877"
],
[
"0.5024",
"0.872",
"0.892",
"0.872",
"0.892",
"0.8853",
"0.5024",
"0.8853"
],
[
"0.0972",
"0.887",
"0.3472",
"0.887",
"0.3472",
"0.901",
"0.0972",
"0.901"
],
[
"0.5",
"0.882",
"0.8965",
"0.882",
"0.8965",
"0.903",
"0.5",
"0.903"
],
[
"0.4795",
"0.924",
"0.492",
"0.924",
"0.492",
"0.9375",
"0.4795",
"0.9375"
]
] | [
"Method",
"Top-1 (%)",
"1.4",
"been explored in previous works relating to full-precision",
"ViTs, but their effects on the accuracy of binary neural net-",
"BinaryBERT[2]",
"BiBERT[30]",
"33.5",
"works oron binaryViTshavenotyetbeenexplored.Also.",
"BiT [24]",
"45.7",
"to the best of our knowledge, with these operations, we are",
"+Remove FFN-Distill",
"46.5",
"the first to be competitive with prior binary CNN models on",
"+SGBERT [1]",
"47.4",
"the ImageNet-1k dataset in terms of accuracy and the num-",
"+ReActNet [25]",
"48.5",
"ber of operations (OPs), using a pure ViT architecture. The",
"source code of this work is available on GitHub1",
"Table 1. Top 1 accuracy on ImageNet with existing transformer-",
"based binarization methods applied on the binary DeiT-S.",
"2. Designing a binary ViT baseline",
"First, we design a baseline binary pure ViT model us-",
"ing the DeiT [10, 36] backbone, since, to the best of our",
"input and output dimensions are the same. If the input di-",
"knowledge, there has never been works on a pure ViT model",
"mension is smaller than the output channel by n times, the",
"with binary weights and activations. In this section, we use",
"residual function will be the input concatenation with itself",
"existing binarization techniques applied on CNN, and on",
"n times.If the input dimension is larger than the output",
"BERT [8] to design the baseline binary pure ViT model as",
"channel by n times, the residual function will be the aver-",
"shown in Figure la and 2a.",
"age pooling function with a stride of n and a window size",
"of n. Therefore, R(X) can be defined as",
"2.1. Binary fully-connected layer",
"For the forward pass, for each matrix multiplication,",
"X,",
"Cin = Cout",
"X e RNxDan, and weight matrix, w e RDinxDout, A",
"a sign function is used for each input activation matrix,",
"R(X) =",
"Cat([X, for i in range(n)l, dim=2),",
"nCin=Cout",
"threshold vector, βx e RDin, can be applied to the real",
"n-X(,:, -ldin:din)",
"Cin = nCout",
"value inputs right before applying the sign function to al-",
"(3)",
"lowthese inputsto havesomedistributional shift. Forthe",
"whereCatistheconcatenationfunction",
"weights, the threshold, μ(W) RDout, can be determined",
"Using a straight-through estimator [3], we approximate",
"by computing the mean value of all elements inside the ma-",
"the derivative of sign with respect to an input as",
"trix as in30-321.Forthe activations, the threshold param-",
"eter can be determined using backpropagation to minimize",
" sign(r)",
"1,≤1",
"(4)",
"the task loss as in [25, 44]. After applying the sign func-",
"ar",
"0,otherwise",
"tion, a scaling factor, aw e R, where aw = Wi, is",
"applied as in [30]. Then, the matrix multiplication output,",
"2.2. Binary vision transformer",
"Y(X) e IRNxDout, can be calculated using popcount, , as",
"A ViT uses N number of transformer encoder blocks",
"and each transformer block contains one multi-head atten-",
"Y(X) = Qw Rsign(X) sign(W - μ(W))",
"(1)",
"tion(MHA)moduleandonefeed-forwardnetwork(FFN)",
"module.Initially, in the embedding layer, the image, x",
"where Rsign = sign(X + βx) as in [25]",
"E IRHxWxC gets split up into fixed-size patches, Xp E",
"For each binary fully-connected layer, BiFC, in a binary",
"RNx(P2.C), where (H, W) is the image dimensions, C is",
"transformer, we connect a residual connection, R, from the",
"the number of channels of the input image, (P, P) is the",
"input, X, to the output of the linear layer, BN(Y(X)), as:",
"dimension of each image patch, and N = HW/p2 is the",
"number of patches. These patches are then linearly pro-",
"BiFC(X) =RPReLU(BN(Y) + R(X))",
"(2)",
"jected by E e IR(P2.C)xD before being appended with a",
"to preserve the information from the previous layer as in",
"cls-token embedding, xelass E RD, and summed with the",
"[25, 27]. RPReLU activation function proposed by [25] is",
"position embeddings, Epos E IR(N+1)xD, as in Eq. (5)",
"used after each residual connection. Also, we replace all",
"HI =[xelass; x,E; x,E; ., x,E] +Epos",
"(5)",
"tion [17], BN, since all linear layers have a normalization",
"layernormalizationintheViTmodelwithbatchnormaliza-",
"Like in previous works on binarization, the operations in the",
"layerafterit,asin[45l,toenablefasterinferenceandalso",
"first layer are always kept in full precision.",
"fastertrainingcomparedtolayernormalization.Theresid-",
"The output of the embedding layer HI E IR(N+1)xD",
"ual function, R, from [25] can be an identity function if the",
"then becomes the input for the first transformer block. In",
"'https://github.com/Phuoc-Hoan-Le/Binary ViT",
"each transformer block, the input He IR(N+1)xD first gets",
"3"
] | {
"arxiv_id": "2306.16678",
"language": "en",
"page_idx": 2
} |
||
2306.16678-en-3 | 2306.16678-en-3.jpg | Model
$\mathbb{R}(...)$
$\overline{\text{DeiT-S}}$
$\overline{153,216}$
ResNet- 34 71,193,472
Table 2. Element-wise representational capability $\mathbb{R}(...)$ for the
binary DeiT-S and for the binary ResNet-34.
normalized by a pre-batch-normalization layer as $\hat{\mathbf{H}}=$ $\mathrm{BN}_{1}( \mathbf{H} ) .$ Then, in the MHA module of the binary transformer with $N_H$ attention heads, the output of that batch- $\begin{array}{l}\text{normalization is then used to calculate the query,}&\mathbf{Q}_{h}&\in\\\mathbb{R}^{(N+1)\times D_{h}};&\mathrm{key,~K}_{h}&\in\mathbb{R}^{(N+1)\times D_{h}};&\mathrm{and~value,~}&\mathbf{V}_{h}&\in\\\mathbb{R}^{(N+1)\times D_{h}},&\mathrm{for~each~head,~}h,\mathrm{as:}\end{array}$
$$\mathbf{Q}_{h}=\mathrm{BiFC}_{\mathrm{Q}_{h}}(\hat{\mathbf{H}}_{h})\\\mathbf{K}_{h}=\mathrm{BiFC}_{\mathrm{K}_{h}}(\hat{\mathbf{H}}_{h})\\\mathbf{V}_{h}=\mathrm{BiFC}_{\mathrm{V}_{h}}(\hat{\mathbf{H}}_{h})$$
(6)
using Eq.(1) and Eq.(2), where $D_h=D/N_H.$ We then calculate the attention score as, $\mathbf{A}_{h}=$
$\mathop{\mathrm{Rsign}}(\mathbf{Q}_{h})$Rsign$(\mathbf{K}_{h}^{\top}).$
From [24], to get the binary attention probability matrix $\begin{array}{l}{\mathrm{of~each~head,~P_{h}\in~\mathbb{R}^{(N+1)\times(N+1)}~we~apply~the~softmax}}\\{\mathrm{function~on~the~attention~score,~A_{h}\in~\mathbb{R}^{(N+1)\times(N+1)},~and}}\end{array}$ then apply the round-to-nearest-integer function, $\lceil\cdot\rfloor$,
(7)
$$\mathbf{P}_h=\alpha_\mathbf{P}\Big[\sigma\Big(\frac{1}{\alpha_\mathbf{P}}\:\mathrm{Softmax}\Big(\frac{\mathbf{A}_h}{\sqrt{D_h}}\Big),0,1\Big)\Big]$$
where $\alpha_\mathbf{p}\in\mathbb{R}$ is a learnable scaling factor trained using the method from [11] and the $\sigma(x,r_1,r_2)$ function keeps the output to be between $r_1$ and $r_2.$ A boolean function from [30] can also be applied on the attention score, but in Table l, we find that it performs 12% worse than applying the softmax function and then rounding the output.
To get the output of each head, head$_h\in\mathbb{R}^{(N+1)\times D_h}$,the binary attention probability matrix, $\mathbf{P}_h$, will be multiplied by the binary value matrix, $\bar{\mathbf{V}}_{h}=$ Rsign$(\mathbf{V}_{h}).$ Also, to preserve the information of the query, the key, and the value, we also add the query, the key, and the value to the head output as shown below
(8)
$$\mathrm{head}_h=\mathrm{RPReLU}(\mathrm{BN}_\mathrm{at}(\mathbf{P}_h\bar{\mathbf{V}}_h)+\mathbf{Q}_h+\mathbf{K}_h+\mathbf{V}_h)$$
and as shown in Figure 2a.
The outputs from all heads are then concatenated with each other and used by the fully-connected layer, BiFC$_{\mathrm{O}}$, to get the multi-head attention output. A main residual connection is then applied to the output of MHA as
$$\mathbf{F}=\mathrm{BiFC}_{\mathrm{O}}(\mathrm{Cat}(\mathrm{head}_{1},...,\mathrm{head}_{N_{H}}))+\mathbf{H}$$
(9)
Then the residual output, F $\in\mathbb{R}^{(N+1)\times D}$, gets normalized by a 2nd pre-batch-normalization layer, BN$_{2}$, as $\hat{\mathbf{F} } = \mathrm{BN}_{2}( \mathbf{F} )$ and goes through the binary feed-forward network layer, BiFFN, which has two binary fully-connected linear (BiFC) layers. Finally, a second main residual connection is then applied to the FFN output to get R
$$\mathbf{R}=\mathrm{BiFFN}(\hat{\mathbf{F}})+\mathbf{F}$$
(10)
which gets inputted to the next transformer block.
Like in previous works in binarization, the parameter biases, the parameters at the classifier, operations in the softmax, and normalization layers are kept in full precision.
To improve the performance of the binary ViT, we distill the knowledge of a full precision model to a model with binary weights and activations by minimizing the soft crossentropy loss between the student's logit and the teacher's logit as in [25]. We do not use distillation loss for the attention scores and output as the performance will significantly degrade just as shown in [24,30]. Also, we do not distill the FFN outputs from [2,24,50] as it causes a slight loss of accuracy by 1.2%. Using the partially-random-initialization method from [l] where we initialize the first patch embedding layer from the full precision model, improves the performance by $0.9\%.$
# 3. What else do binary CNNs have that binary transformers do not have?
In Table 1, using all of the aforementioned techniques in Section 2 will only get us 48.5% top-1 accuracy on ImageNet-1k which is far below the accuracy for most SOTA binary CNNs. Therefore, in this section, we further analyze the details/properties of the CNN architecture that have not been explored in the context of binary neural networks and that can help pure ViTs with binary weights and binary activations improve their representational capability to improve their accuracy on a dataset with a large number of classes without introducing convolutions and without significantly increasing the number of operations and the number of parameters. From Table 4,even increasing the number of parameters by either increasing the width or depth of the baseline vanilla binary ViT, it has trouble surpassing SOTA binary CNN models.
Analysis of representational capability. First, let's analyze and compare the element-wise representational capability of a binary CNN such as a binary ResNet34, $\mathbb{R}($ResNet-34) versus a binary ViT such as a binary DeiTS, $\mathbb{R}($DeiT-S) with both of these models using the ReActNet[25] design. As in [27,43], we quantify the elementwise representational capability as the number of possible absolute values that each element in a matrix/tensor can have. We calculate the element-wise representational capability of these models by calculating the element-wise representational capability of the tensor that will be the input for the classifier layer, following the steps from [27,43].
4 | [
[
"0.634",
"0.0884",
"0.6807",
"0.0884",
"0.6807",
"0.1029",
"0.634",
"0.1029"
],
[
"0.721",
"0.089",
"0.759",
"0.089",
"0.759",
"0.1029",
"0.721",
"0.1029"
],
[
"0.4495",
"0.09344",
"0.4534",
"0.0985",
"0.451",
"0.101",
"0.4453",
"0.09534"
],
[
"0.0784",
"0.09155",
"0.4468",
"0.0903",
"0.4468",
"0.1048",
"0.0784",
"0.1061"
],
[
"0.6323",
"0.1048",
"0.6807",
"0.1048",
"0.6807",
"0.11615",
"0.6323",
"0.11615"
],
[
"0.7173",
"0.1048",
"0.7686",
"0.1048",
"0.7686",
"0.11615",
"0.7173",
"0.11615"
],
[
"0.6216",
"0.1187",
"0.693",
"0.1187",
"0.693",
"0.13",
"0.6216",
"0.13"
],
[
"0.0809",
"0.1092",
"0.4673",
"0.1092",
"0.4673",
"0.1193",
"0.0809",
"0.1193"
],
[
"0.701",
"0.1187",
"0.7754",
"0.1187",
"0.7754",
"0.13",
"0.701",
"0.13"
],
[
"0.0809",
"0.125",
"0.4673",
"0.125",
"0.4673",
"0.1351",
"0.0809",
"0.1351"
],
[
"0.0809",
"0.1383",
"0.4697",
"0.1383",
"0.4697",
"0.1515",
"0.0809",
"0.1515"
],
[
"0.0768",
"0.1465",
"0.4714",
"0.1465",
"0.4714",
"0.1674",
"0.0768",
"0.1674"
],
[
"0.501",
"0.1427",
"0.892",
"0.1433",
"0.892",
"0.1578",
"0.501",
"0.1572"
],
[
"0.076",
"0.1622",
"0.3022",
"0.1642",
"0.3022",
"0.1831",
"0.076",
"0.1812"
],
[
"0.501",
"0.1578",
"0.7646",
"0.1578",
"0.7646",
"0.1718",
"0.501",
"0.1718"
],
[
"0.2051",
"0.1913",
"0.3423",
"0.1895",
"0.3423",
"0.207",
"0.2051",
"0.209"
],
[
"0.2051",
"0.2122",
"0.3406",
"0.2122",
"0.3406",
"0.2299",
"0.2051",
"0.2299"
],
[
"0.4453",
"0.214",
"0.4697",
"0.214",
"0.4697",
"0.2279",
"0.4453",
"0.2279"
],
[
"0.5215",
"0.209",
"0.8906",
"0.2102",
"0.8906",
"0.2247",
"0.5215",
"0.2235"
],
[
"0.2051",
"0.2317",
"0.343",
"0.2349",
"0.343",
"0.2524",
"0.2051",
"0.25"
],
[
"0.5034",
"0.2267",
"0.8906",
"0.2267",
"0.8906",
"0.2405",
"0.5034",
"0.2405"
],
[
"0.5024",
"0.2412",
"0.869",
"0.2412",
"0.869",
"0.2551",
"0.5024",
"0.2551"
],
[
"0.0784",
"0.26",
"0.468",
"0.26",
"0.468",
"0.2778",
"0.0784",
"0.2778"
],
[
"0.5215",
"0.2556",
"0.892",
"0.2556",
"0.892",
"0.2695",
"0.5215",
"0.2695"
],
[
"0.0801",
"0.2766",
"0.4346",
"0.2766",
"0.4346",
"0.2903",
"0.0801",
"0.2903"
],
[
"0.5024",
"0.2715",
"0.892",
"0.2715",
"0.892",
"0.2854",
"0.5024",
"0.2854"
],
[
"0.455",
"0.2822",
"0.464",
"0.2822",
"0.464",
"0.286",
"0.455",
"0.286"
],
[
"0.0801",
"0.2922",
"0.2378",
"0.2922",
"0.2378",
"0.307",
"0.0801",
"0.307"
],
[
"0.5034",
"0.2886",
"0.8906",
"0.2886",
"0.8906",
"0.2986",
"0.5034",
"0.2986"
],
[
"0.0989",
"0.307",
"0.4673",
"0.307",
"0.4673",
"0.3208",
"0.0989",
"0.3208"
],
[
"0.5024",
"0.3018",
"0.892",
"0.3018",
"0.892",
"0.315",
"0.5024",
"0.315"
],
[
"0.0784",
"0.3193",
"0.4714",
"0.3193",
"0.4714",
"0.3372",
"0.0784",
"0.3372"
],
[
"0.5024",
"0.3176",
"0.8906",
"0.3176",
"0.8906",
"0.3308",
"0.5024",
"0.3308"
],
[
"0.0784",
"0.334",
"0.4697",
"0.334",
"0.4697",
"0.3516",
"0.0784",
"0.3516"
],
[
"0.5034",
"0.334",
"0.8896",
"0.334",
"0.8896",
"0.3447",
"0.5034",
"0.3447"
],
[
"0.0784",
"0.351",
"0.4248",
"0.3528",
"0.4248",
"0.3674",
"0.0784",
"0.3655"
],
[
"0.5034",
"0.348",
"0.892",
"0.348",
"0.892",
"0.361",
"0.5034",
"0.361"
],
[
"0.505",
"0.3643",
"0.8906",
"0.3643",
"0.8906",
"0.375",
"0.505",
"0.375"
],
[
"0.5034",
"0.377",
"0.892",
"0.377",
"0.892",
"0.3901",
"0.5034",
"0.3901"
],
[
"0.134",
"0.3833",
"0.1978",
"0.385",
"0.1978",
"0.3997",
"0.134",
"0.3977"
],
[
"0.2279",
"0.3833",
"0.3096",
"0.3833",
"0.3096",
"0.401",
"0.2279",
"0.401"
],
[
"0.3245",
"0.385",
"0.3489",
"0.385",
"0.3489",
"0.3894",
"0.3245",
"0.3894"
],
[
"0.3545",
"0.3838",
"0.4004",
"0.3838",
"0.4004",
"0.3997",
"0.3545",
"0.3997"
],
[
"0.4478",
"0.3894",
"0.4656",
"0.3894",
"0.4656",
"0.3977",
"0.4478",
"0.3977"
],
[
"0.5034",
"0.3926",
"0.8906",
"0.3926",
"0.8906",
"0.406",
"0.5034",
"0.406"
],
[
"0.2214",
"0.3997",
"0.2345",
"0.3997",
"0.2345",
"0.404",
"0.2214",
"0.404"
],
[
"0.3218",
"0.3945",
"0.3555",
"0.3945",
"0.3555",
"0.406",
"0.3218",
"0.406"
],
[
"0.0801",
"0.4136",
"0.468",
"0.414",
"0.468",
"0.4287",
"0.0801",
"0.428"
],
[
"0.5034",
"0.4077",
"0.8906",
"0.4077",
"0.8906",
"0.4216",
"0.5034",
"0.4216"
],
[
"0.5024",
"0.4224",
"0.63",
"0.4224",
"0.63",
"0.4363",
"0.5024",
"0.4363"
],
[
"0.0801",
"0.43",
"0.4697",
"0.43",
"0.4697",
"0.4443",
"0.0801",
"0.4443"
],
[
"0.0809",
"0.4443",
"0.468",
"0.4443",
"0.468",
"0.4583",
"0.0809",
"0.4583"
],
[
"0.5024",
"0.4495",
"0.888",
"0.4507",
"0.888",
"0.4653",
"0.5024",
"0.464"
],
[
"0.0801",
"0.4602",
"0.468",
"0.4602",
"0.468",
"0.4734",
"0.0801",
"0.4734"
],
[
"0.0801",
"0.4734",
"0.468",
"0.4749",
"0.468",
"0.4893",
"0.0801",
"0.488"
],
[
"0.5317",
"0.4685",
"0.7534",
"0.4673",
"0.7534",
"0.4817",
"0.5317",
"0.483"
],
[
"0.0809",
"0.4912",
"0.4102",
"0.4912",
"0.4102",
"0.502",
"0.0809",
"0.502"
],
[
"0.5215",
"0.4937",
"0.8906",
"0.4937",
"0.8906",
"0.507",
"0.5215",
"0.507"
],
[
"0.09644",
"0.5024",
"0.4697",
"0.5005",
"0.4697",
"0.518",
"0.09644",
"0.521"
],
[
"0.5024",
"0.509",
"0.892",
"0.509",
"0.892",
"0.5234",
"0.5024",
"0.5234"
],
[
"0.0801",
"0.5195",
"0.468",
"0.5195",
"0.468",
"0.5327",
"0.0801",
"0.5327"
],
[
"0.5034",
"0.526",
"0.8906",
"0.526",
"0.8906",
"0.536",
"0.5034",
"0.536"
],
[
"0.0801",
"0.535",
"0.4697",
"0.535",
"0.4697",
"0.549",
"0.0801",
"0.549"
],
[
"0.5024",
"0.539",
"0.892",
"0.539",
"0.892",
"0.553",
"0.5024",
"0.553"
],
[
"0.0784",
"0.55",
"0.468",
"0.549",
"0.468",
"0.563",
"0.0784",
"0.5645"
],
[
"0.5024",
"0.5547",
"0.892",
"0.5547",
"0.892",
"0.5684",
"0.5024",
"0.5684"
],
[
"0.0801",
"0.565",
"0.4697",
"0.565",
"0.4697",
"0.579",
"0.0801",
"0.579"
],
[
"0.5034",
"0.571",
"0.8896",
"0.571",
"0.8896",
"0.5806",
"0.5034",
"0.5806"
],
[
"0.0801",
"0.58",
"0.2303",
"0.58",
"0.2303",
"0.5933",
"0.0801",
"0.5933"
],
[
"0.5034",
"0.584",
"0.892",
"0.584",
"0.892",
"0.5986",
"0.5034",
"0.5986"
],
[
"0.5034",
"0.601",
"0.8896",
"0.601",
"0.8896",
"0.6113",
"0.5034",
"0.6113"
],
[
"0.08826",
"0.619",
"0.4673",
"0.619",
"0.4673",
"0.634",
"0.08826",
"0.634"
],
[
"0.5034",
"0.616",
"0.8896",
"0.616",
"0.8896",
"0.627",
"0.5034",
"0.627"
],
[
"0.5024",
"0.63",
"0.8906",
"0.63",
"0.8906",
"0.643",
"0.5024",
"0.643"
],
[
"0.0768",
"0.648",
"0.255",
"0.646",
"0.255",
"0.66",
"0.0768",
"0.662"
],
[
"0.5034",
"0.646",
"0.892",
"0.646",
"0.892",
"0.659",
"0.5034",
"0.659"
],
[
"0.0989",
"0.662",
"0.468",
"0.662",
"0.468",
"0.6753",
"0.0989",
"0.6753"
],
[
"0.5024",
"0.66",
"0.8906",
"0.66",
"0.8906",
"0.674",
"0.5024",
"0.674"
],
[
"0.0801",
"0.678",
"0.4697",
"0.678",
"0.4697",
"0.6914",
"0.0801",
"0.6914"
],
[
"0.501",
"0.675",
"0.892",
"0.674",
"0.892",
"0.688",
"0.501",
"0.6895"
],
[
"0.0809",
"0.6943",
"0.4673",
"0.6943",
"0.4673",
"0.705",
"0.0809",
"0.705"
],
[
"0.5024",
"0.69",
"0.892",
"0.69",
"0.892",
"0.704",
"0.5024",
"0.704"
],
[
"0.0784",
"0.707",
"0.3945",
"0.7056",
"0.3945",
"0.72",
"0.0784",
"0.7217"
],
[
"0.5024",
"0.7056",
"0.738",
"0.7056",
"0.738",
"0.7197",
"0.5024",
"0.7197"
],
[
"0.5034",
"0.72",
"0.892",
"0.72",
"0.892",
"0.7344",
"0.5034",
"0.7344"
],
[
"0.1299",
"0.733",
"0.4207",
"0.733",
"0.4207",
"0.7476",
"0.1299",
"0.7476"
],
[
"0.4453",
"0.7344",
"0.468",
"0.7344",
"0.468",
"0.747",
"0.4453",
"0.747"
],
[
"0.505",
"0.7373",
"0.8896",
"0.7373",
"0.8896",
"0.7476",
"0.505",
"0.7476"
],
[
"0.09644",
"0.754",
"0.4714",
"0.754",
"0.4714",
"0.7744",
"0.09644",
"0.7744"
],
[
"0.5034",
"0.7505",
"0.8906",
"0.7505",
"0.8906",
"0.7646",
"0.5034",
"0.7646"
],
[
"0.5024",
"0.765",
"0.8896",
"0.765",
"0.8896",
"0.779",
"0.5024",
"0.779"
],
[
"0.0801",
"0.7734",
"0.4697",
"0.7734",
"0.4697",
"0.7886",
"0.0801",
"0.7886"
],
[
"0.5024",
"0.781",
"0.8906",
"0.781",
"0.8906",
"0.7944",
"0.5024",
"0.7944"
],
[
"0.0801",
"0.789",
"0.468",
"0.789",
"0.468",
"0.803",
"0.0801",
"0.803"
],
[
"0.5024",
"0.797",
"0.8906",
"0.797",
"0.8906",
"0.81",
"0.5024",
"0.81"
],
[
"0.0801",
"0.8047",
"0.468",
"0.8047",
"0.468",
"0.8184",
"0.0801",
"0.8184"
],
[
"0.5024",
"0.811",
"0.8906",
"0.811",
"0.8906",
"0.8247",
"0.5024",
"0.8247"
],
[
"0.0809",
"0.8193",
"0.468",
"0.8193",
"0.468",
"0.8335",
"0.0809",
"0.8335"
],
[
"0.5034",
"0.827",
"0.892",
"0.827",
"0.892",
"0.8403",
"0.5034",
"0.8403"
],
[
"0.0801",
"0.8354",
"0.407",
"0.8354",
"0.407",
"0.8486",
"0.0801",
"0.8486"
],
[
"0.5034",
"0.8413",
"0.892",
"0.8413",
"0.892",
"0.855",
"0.5034",
"0.855"
],
[
"0.2051",
"0.86",
"0.343",
"0.86",
"0.343",
"0.8735",
"0.2051",
"0.8735"
],
[
"0.501",
"0.855",
"0.8906",
"0.857",
"0.8906",
"0.8706",
"0.501",
"0.8687"
],
[
"0.4387",
"0.8613",
"0.4697",
"0.8613",
"0.4697",
"0.8735",
"0.4387",
"0.8735"
],
[
"0.5024",
"0.872",
"0.892",
"0.872",
"0.892",
"0.8853",
"0.5024",
"0.8853"
],
[
"0.0801",
"0.8853",
"0.4043",
"0.886",
"0.4043",
"0.901",
"0.0801",
"0.8994"
],
[
"0.501",
"0.8853",
"0.871",
"0.886",
"0.871",
"0.901",
"0.501",
"0.8994"
],
[
"0.4805",
"0.925",
"0.492",
"0.925",
"0.492",
"0.935",
"0.4805",
"0.935"
]
] | [
"Model ",
"R(...)",
"",
"normalized by a pre-batch-normalization layer as H",
"DeiT-S",
"153,216",
"ResNet-34",
"BNi(H).Then.inthe MHA moduleof the binary trans-",
"71,193.472",
"former with NH attention heads,the output of that batch-",
"normalization is then used to calculate the query, Qh E",
"IR(N+1)xDn; key, Kh E IR(N+1)xDr; and value, Vh ",
"Table 2. Element-wise representational capability R(...) for the",
"IR(N+1)xDn, for each head, h, as:",
"binary DeiT-S and for the binary ResNet-34.",
"Qn = BiFCQ, (Hh)",
"Kh = BiFCkn (Hh)",
"(6)",
"Like in previous works in binarization, the parameter bi-",
"Vh = BiFCv, (Hn)",
"ases, the parameters at the classifier, operations in the soft-",
"max, and normalization layers are kept in full precision.",
"using Eq.(1) and Eq.(2), where Dh = D/NH.",
"To improve the performance of the binary ViT, we distill",
"We then calculate the attention score as, Ah",
"the knowledge of a full precision model to a model with bi-",
"",
"Rsign(Qh)Rsign(Kh ).",
"nary weights and activations by minimizing the soft cross-",
"From [24], to get the binary attention probability matrix",
"entropy loss between the student's logit and the teacher's",
"of each head, Ph e IR(N+1)x(N+1) we apply the softmax",
"logit as in [25]. We do not use distillation loss for the atten-",
"function on the attention score, Ah E IR(N+i)x(N+1), and",
"tion scores and output as the performance will significantly",
"then apply the round-to-nearest-integer function, [-],",
"degrade just as shown in [24, 30]. Also, we do not distill",
"theFFNoutputsfrom[2.24,501asitcausesaslightlossof",
"accuracy by 1.2%. Using the partially-random-initialization",
"Ph=op",
" Softmax",
"h",
"),0,1",
"7",
"method from [l] where we initialize the first patch embed-",
"",
"/D.",
"where Qp E IR is a learnable scaling factor trained using",
"ding layer from the full precision model, improves the per-",
"formance by 0.9%.",
"the method from [11] and the o(α, ri, r2) function keeps",
"the output to be between ri and r2. A boolean function",
"3. What else do binary CNNs have that binary",
"from [30] can also be applied on the attention score, but in",
"Table 1, we find that it performs 12% worse than applying",
"transformers do not have?",
"thesoftmaxfunctionandthenroundingtheoutput",
"In Table 1, using all of the aforementioned techniques",
"To get the output of each head, headh e R(N+1)xDn, the",
"in Section 2 will only get us 48.5% top-1 accuracy on",
"binary attention probability matrix, Ph, will be multiplied",
"ImageNet-lkwhichis far below the accuracyfor most",
"by the binary value matrix, Vh = Rsign(Vh). Also, to",
"SOTA binary CNNs. Therefore, in this section, we further",
"preserve the information of the query, the key, and the value,",
"analyze the details/properties of the CNN architecture that",
"we also add the query, the key, and the value to the head",
"havenot been explored in thecontextof binaryneural net-",
"output as shown below",
"works and that can help pure ViTs with binary weights and",
"binaryactivationsimprovetheirrepresentational capability",
"headh = RPReLU(BNat(P,Vh) + Qh +Kh + Vh) (8)",
"to improve their accuracy on a dataset with a large num-",
"ber of classes without introducing convolutions and with-",
"and as shown in Figure 2a.",
"out significantly increasing the number of operations and",
"The outputs from all heads are then concatenated with",
"the number of parameters. From Table 4, even increasing",
"each other and used by the fully-connected layer, BiFCo,",
"the number of parameters by either increasing the width or",
"toget themulti-head attention output.A mainresidualcon-",
"depth of the baseline vanilla binary ViT, it has trouble sur-",
"nection is then applied to the output of MHA as",
"passing SOTA binary CNN models.",
"Analysis of representational capability.First, let's an-",
"F = BiFCo(Cat(headi,.., headN)) +H",
"(9)",
"alyze and comparethe element-wise representational ca-",
"Then the residual output, F e R(N+1)xD, gets nor-",
"pability of a binary CNN such as a binary ResNet34,",
"R(ResNet-34) versus a binary ViT such as a binary DeiT-",
"malized by a 2nd pre-batch-normalization layer, BN2, as",
"S, R(DeiT-S) with both of these models using the ReAct-",
"F = BN2(F) and goes through the binary feed-forward net-",
"Net [25] design. As in [27, 43], we quantify the element-",
"work layer, BiFFN, which has two binary fully-connected",
"wise representational capability as the number of possible",
"linear (BiFC) layers. Finally, a second main residual con-",
"absolute values that each element in a matrix/tensor can",
"nection is then applied to the FFN output to get R",
"have. We calculate the element-wise representational capa-",
"R =BiFFN(F) +F",
"bility of these models by calculating the element-wise rep-",
"(10)",
"resentational capability of the tensor that will be the input",
"which gets inputted to the next transformer block.",
"for the classifier layer, following the steps from [27,43].",
"4"
] | {
"arxiv_id": "2306.16678",
"language": "en",
"page_idx": 3
} |
||
1908.08039-en-1 | 1908.08039-en-1.jpg | with a different sentiment, we do not have to create a new sentence from scratch. Instead, we simply replace the sentimental tokens of the sentence with other expressions indicative of a different sentiment. For example, given a negative review“terrible scenery and poor service", a straightforward positive review is “beautiful scenery and good service”. Such a process is very similar to the task of text infilling [Zhu et $al.,2019]$ or Cloze[Taylor, 1953]. Given the template“
scenery and\_ service", we human feel easy to fill the right words because we have enough prior linguistic knowledge to predict the missing words from their contexts.
We transform the problem of sentiment transfer to the task of text infilling by a pre-trained Masked Language Model (MLM). We utilize a pre-trained deep bidirectional language model (corresponding to human linguistic knowledge) to infill the masked positions. Our approach comprises two stages: Mask and Infill (Figure 1). In the mask stage, we explicitly identify and mask the sentiment tokens in a given sentence. In the infill stage, we fill the masked positions with new expressions conditioned on their context and the target sentiment. There exist two feature-based methods for identifying sentiment attribute markers (i.e. words or phrases that are indicative of a particular sentiment attribute). Attention-based method [Xu $et$ al., 2018] trains a self-attention[Bahdanau $et$ $al.,2015]$ sentiment classifer, where the learned attention weights can be used as signals to identify sentiment markers, and the tokens with weights higher than the average are regarded as sentiment attribute markers. The attention-based method prefers isolated single words to n-gram phrases, which restricts MLM to produce sentences of diverse expressions. Frequency-ratio method [Li et al., 2018] constructs an attribute marker dictionary of n-gram phrases for each sentiment attribute and looks up the dictionary to identify sentiment phrases in sentences. Frequency-ratio method is inclined to mask longer phrases, and its performance heavily relies on the quality of attribute marker dictionaries. Thus, we explore a simple fused method to utilize the merits of both methods. Specifically, we combine the attention-based classifer with the frequency-ratio method, filtering out false attribute markers and discovering new single sentiment words. When the frequency-ratio method fails to identify any attribute marker or recognize too many ones (with insufficient content left), we utilize the attention-based method directly.
MLM predicts the masked tokens only conditioned on their context, considering no attribute information. To fill the masked positions in a sentence with expressions compatible to a particular sentiment, we retrofit MLM to Attribute Conditional Masked Language Model (AC-MLM) by integrating attribute embeddings with the original input for MLM. Furthermore, we introduce a pre-trained sentiment classifier to constrain AC-MLM, which ensures the generated sentences compatible with the target sentiment. To deal with the discrete nature of language generation, we utilize softsampling [Hu et al., 2017] to guide the optimization of ACMLM, where soft-sampling is used to back-propagate gradients through the sampling process by using an approximation of the sampled word vector.
We evaluate our models on two review datasets Yelp and Amazon by quantitative, qualitative, and human evaluations.
Experimental results show that our method achieves state-ofthe-art results on accuracy, BLEU[Papineni et al., 2002] and human evaluations.
Our contributions are summarized as follows:
$\bullet$ We propose a two-stage “Mask and Infill” approach to
sentiment transfer task, capable of identifying both simple and complex sentiment markers and producing highquality sentences.
$\bullet$Experimental results show that our approach outper-
forms most state-of-the-art models in terms of both
BLEU and accuracy scores.
$\bullet$ We retrofit MLM to AC-MLM for labeled sentence gen-
eration. To the best of our knowledge, it is the first work to apply a pre-trained masked language model to labeled sentence generation task.
## 2 $\textbf{Related Work}$
2.1 Non-parallel Style Transfer
[Shen $etal.,2017;$ Fu $etal.,2018;$ Xu et $al.,2018;$ Li $et$ $al.,2018;$ Yang et $al.,2018]$ are most relevant to our work. [Shen $etal.,2\tilde{0}$17] assumed a shared latent content distribution across different text corpora, and leverages refined alignment of latent representations. [Fu $et$ al.,2018] learned a representation that only contains the content information by multi-task learning and adversarial training.[Xu et al., 2018] proposed a cycled reinforcement learning method by collaboration between a neutralization module and an emotionalization module.[Li et al., 2018] separated attribute and content by deleting attribute markers, and attempted to reconstruct the sentence using the content and the retrieved target attribute markers with an RNN.[Yang et al., 2018] used a target domain language model as the discriminator to guide the generator to produce sentences. Unlike these models, which adopt RNN as encoder and decoder, we utilize pre-trained MLM to capture longer linguistic structure and better language representation.
2.2 Fine-tuning on Pre-trained Language Model
Language model pre-training has attracted widespread attention, and fine-tuning on pre-trained language models has proven to be effective in improving many downstream natural $\hat{\text{language processing tasks.}}[$Dai and Le, 2015] improved Sequence Learning with recurrent networks by pre-training on unlabeled data. [Radford et al., 2018] improved the performance largely on many sentence-level tasks from the GLUE benchmarks[Wang et al., 2018]. [Howard and Ruder, 2018] proposed Universal Language Model Fine-tuning (ULMFiT), $\hat{\text{which was a general transfer learning method for fine-tuning}}$ a language model. [Radford et al., 2018] demonstrated that by generative pre-training language model on unlabeled text from diverse corpora, large gains could be achieved on a diverse range of tasks. By introducing a deep bidirectional masked language model, BERT [Devlin et al., 2018] obtained new state-of-the-art results on a broad range of tasks. Unlike previous works fine-tuning pre-trained language models to perform discriminative tasks, we aim to apply a pre-trained masked language model on generative task. | [
[
"0.0866",
"0.0682",
"0.4844",
"0.06946",
"0.4844",
"0.0833",
"0.0866",
"0.0821"
],
[
"0.513",
"0.06946",
"0.9087",
"0.06946",
"0.9087",
"0.0833",
"0.513",
"0.0833"
],
[
"0.0866",
"0.084",
"0.486",
"0.084",
"0.486",
"0.09784",
"0.0866",
"0.09784"
],
[
"0.513",
"0.084",
"0.9126",
"0.084",
"0.9126",
"0.09784",
"0.513",
"0.09784"
],
[
"0.0866",
"0.09784",
"0.4885",
"0.09784",
"0.4885",
"0.1111",
"0.0866",
"0.1111"
],
[
"0.5107",
"0.09784",
"0.638",
"0.09784",
"0.638",
"0.1111",
"0.5107",
"0.1111"
],
[
"0.08496",
"0.11237",
"0.486",
"0.11237",
"0.486",
"0.1256",
"0.08496",
"0.1256"
],
[
"0.5293",
"0.1111",
"0.83",
"0.1111",
"0.83",
"0.125",
"0.5293",
"0.125"
],
[
"0.08496",
"0.1256",
"0.486",
"0.1256",
"0.486",
"0.1395",
"0.08496",
"0.1395"
],
[
"0.5293",
"0.13",
"0.914",
"0.13",
"0.914",
"0.1445",
"0.5293",
"0.1445"
],
[
"0.0866",
"0.1395",
"0.486",
"0.1395",
"0.486",
"0.1528",
"0.0866",
"0.1528"
],
[
"0.544",
"0.1445",
"0.91",
"0.1445",
"0.91",
"0.1578",
"0.544",
"0.1578"
],
[
"0.08496",
"0.154",
"0.486",
"0.1528",
"0.486",
"0.1642",
"0.08496",
"0.1654"
],
[
"0.546",
"0.159",
"0.911",
"0.159",
"0.911",
"0.1724",
"0.546",
"0.1724"
],
[
"0.08496",
"0.1654",
"0.452",
"0.1654",
"0.452",
"0.1787",
"0.08496",
"0.1787"
],
[
"0.546",
"0.1749",
"0.661",
"0.1749",
"0.661",
"0.1849",
"0.546",
"0.1849"
],
[
"0.0866",
"0.1818",
"0.1708",
"0.1818",
"0.1708",
"0.1932",
"0.0866",
"0.1932"
],
[
"0.1854",
"0.1812",
"0.486",
"0.1787",
"0.486",
"0.1925",
"0.1854",
"0.1945"
],
[
"0.0866",
"0.1945",
"0.486",
"0.1945",
"0.486",
"0.2084",
"0.0866",
"0.2084"
],
[
"0.5293",
"0.1895",
"0.911",
"0.1901",
"0.911",
"0.2052",
"0.5293",
"0.2039"
],
[
"0.08496",
"0.2084",
"0.3882",
"0.2084",
"0.3882",
"0.2216",
"0.08496",
"0.2216"
],
[
"0.544",
"0.2052",
"0.914",
"0.2052",
"0.914",
"0.2184",
"0.544",
"0.2184"
],
[
"0.10297",
"0.2229",
"0.4836",
"0.2229",
"0.4836",
"0.2361",
"0.10297",
"0.2361"
],
[
"0.544",
"0.2172",
"0.7295",
"0.2197",
"0.7295",
"0.2329",
"0.544",
"0.2311"
],
[
"0.08496",
"0.2349",
"0.486",
"0.2361",
"0.486",
"0.25",
"0.08496",
"0.2488"
],
[
"0.5293",
"0.2349",
"0.91",
"0.2374",
"0.91",
"0.252",
"0.5293",
"0.25"
],
[
"0.08496",
"0.2507",
"0.4844",
"0.2507",
"0.4844",
"0.2644",
"0.08496",
"0.2644"
],
[
"0.5435",
"0.252",
"0.9126",
"0.252",
"0.9126",
"0.2651",
"0.5435",
"0.2651"
],
[
"0.08826",
"0.2644",
"0.4844",
"0.2644",
"0.4844",
"0.2778",
"0.08826",
"0.2778"
],
[
"0.544",
"0.2664",
"0.9126",
"0.2664",
"0.9126",
"0.2798",
"0.544",
"0.2798"
],
[
"0.08905",
"0.2798",
"0.4844",
"0.2798",
"0.4844",
"0.2903",
"0.08905",
"0.2903"
],
[
"0.544",
"0.2798",
"0.7134",
"0.2798",
"0.7134",
"0.2935",
"0.544",
"0.2935"
],
[
"0.0866",
"0.2922",
"0.4844",
"0.2922",
"0.4844",
"0.3062",
"0.0866",
"0.3062"
],
[
"0.08826",
"0.3062",
"0.4878",
"0.3062",
"0.4878",
"0.3193",
"0.08826",
"0.3193"
],
[
"0.5415",
"0.308",
"0.661",
"0.308",
"0.661",
"0.3225",
"0.5415",
"0.3225"
],
[
"0.08905",
"0.3213",
"0.482",
"0.3213",
"0.482",
"0.332",
"0.08905",
"0.332"
],
[
"0.08496",
"0.3333",
"0.4714",
"0.3333",
"0.4714",
"0.3467",
"0.08496",
"0.3467"
],
[
"0.513",
"0.3289",
"0.763",
"0.3289",
"0.763",
"0.3435",
"0.513",
"0.3435"
],
[
"0.10297",
"0.348",
"0.4836",
"0.348",
"0.4836",
"0.361",
"0.10297",
"0.361"
],
[
"0.5146",
"0.3467",
"0.914",
"0.3467",
"0.914",
"0.361",
"0.5146",
"0.361"
],
[
"0.08826",
"0.3623",
"0.4836",
"0.3623",
"0.4836",
"0.3726",
"0.08826",
"0.3726"
],
[
"0.513",
"0.361",
"0.9126",
"0.361",
"0.9126",
"0.375",
"0.513",
"0.375"
],
[
"0.0866",
"0.375",
"0.486",
"0.375",
"0.486",
"0.3882",
"0.0866",
"0.3882"
],
[
"0.5146",
"0.375",
"0.9126",
"0.375",
"0.9126",
"0.3894",
"0.5146",
"0.3894"
],
[
"0.0866",
"0.3894",
"0.4885",
"0.3894",
"0.4885",
"0.4028",
"0.0866",
"0.4028"
],
[
"0.513",
"0.3894",
"0.911",
"0.3894",
"0.911",
"0.4028",
"0.513",
"0.4028"
],
[
"0.0866",
"0.4028",
"0.4878",
"0.4028",
"0.4878",
"0.4167",
"0.0866",
"0.4167"
],
[
"0.5146",
"0.4048",
"0.911",
"0.4048",
"0.911",
"0.4153",
"0.5146",
"0.4153"
],
[
"0.0866",
"0.4172",
"0.4844",
"0.4172",
"0.4844",
"0.4312",
"0.0866",
"0.4312"
],
[
"0.513",
"0.4172",
"0.9126",
"0.4172",
"0.9126",
"0.4312",
"0.513",
"0.4312"
],
[
"0.0866",
"0.4312",
"0.486",
"0.4312",
"0.486",
"0.4443",
"0.0866",
"0.4443"
],
[
"0.513",
"0.4312",
"0.9126",
"0.4312",
"0.9126",
"0.4443",
"0.513",
"0.4443"
],
[
"0.08826",
"0.4463",
"0.486",
"0.4463",
"0.486",
"0.457",
"0.08826",
"0.457"
],
[
"0.511",
"0.4458",
"0.911",
"0.4443",
"0.911",
"0.4583",
"0.511",
"0.459"
],
[
"0.0866",
"0.459",
"0.4878",
"0.459",
"0.4878",
"0.473",
"0.0866",
"0.473"
],
[
"0.5146",
"0.4602",
"0.91",
"0.4602",
"0.91",
"0.4702",
"0.5146",
"0.4702"
],
[
"0.08826",
"0.473",
"0.4878",
"0.473",
"0.4878",
"0.486",
"0.08826",
"0.486"
],
[
"0.513",
"0.473",
"0.9126",
"0.473",
"0.9126",
"0.486",
"0.513",
"0.486"
],
[
"0.08826",
"0.488",
"0.486",
"0.488",
"0.486",
"0.4988",
"0.08826",
"0.4988"
],
[
"0.5146",
"0.488",
"0.9126",
"0.488",
"0.9126",
"0.4988",
"0.5146",
"0.4988"
],
[
"0.0866",
"0.5005",
"0.4844",
"0.5005",
"0.4844",
"0.5146",
"0.0866",
"0.5146"
],
[
"0.513",
"0.5005",
"0.9126",
"0.5005",
"0.9126",
"0.5146",
"0.513",
"0.5146"
],
[
"0.0866",
"0.5146",
"0.4844",
"0.5146",
"0.4844",
"0.528",
"0.0866",
"0.528"
],
[
"0.513",
"0.5146",
"0.9126",
"0.5146",
"0.9126",
"0.528",
"0.513",
"0.528"
],
[
"0.0866",
"0.529",
"0.4844",
"0.529",
"0.4844",
"0.5425",
"0.0866",
"0.5425"
],
[
"0.513",
"0.528",
"0.911",
"0.528",
"0.911",
"0.5415",
"0.513",
"0.5415"
],
[
"0.0866",
"0.5425",
"0.4844",
"0.5425",
"0.4844",
"0.556",
"0.0866",
"0.556"
],
[
"0.513",
"0.5415",
"0.9126",
"0.5415",
"0.9126",
"0.5547",
"0.513",
"0.5547"
],
[
"0.0866",
"0.556",
"0.486",
"0.556",
"0.486",
"0.5693",
"0.0866",
"0.5693"
],
[
"0.511",
"0.5537",
"0.914",
"0.5547",
"0.914",
"0.5693",
"0.511",
"0.5684"
],
[
"0.08905",
"0.571",
"0.4844",
"0.571",
"0.4844",
"0.5806",
"0.08905",
"0.5806"
],
[
"0.511",
"0.5693",
"0.91",
"0.5693",
"0.91",
"0.5835",
"0.511",
"0.5835"
],
[
"0.0866",
"0.5854",
"0.4836",
"0.5854",
"0.4836",
"0.595",
"0.0866",
"0.595"
],
[
"0.513",
"0.5854",
"0.5786",
"0.5854",
"0.5786",
"0.5967",
"0.513",
"0.5967"
],
[
"0.0866",
"0.5977",
"0.4844",
"0.5977",
"0.4844",
"0.6113",
"0.0866",
"0.6113"
],
[
"0.5146",
"0.608",
"0.5513",
"0.608",
"0.5513",
"0.619",
"0.5146",
"0.619"
],
[
"0.5483",
"0.606",
"0.892",
"0.607",
"0.892",
"0.622",
"0.5483",
"0.6206"
],
[
"0.0866",
"0.6123",
"0.4836",
"0.6123",
"0.4836",
"0.6226",
"0.0866",
"0.6226"
],
[
"0.5483",
"0.6113",
"0.5537",
"0.6113",
"0.5537",
"0.617",
"0.5483",
"0.617"
],
[
"0.08496",
"0.625",
"0.486",
"0.625",
"0.486",
"0.638",
"0.08496",
"0.638"
],
[
"0.511",
"0.6255",
"0.91",
"0.625",
"0.91",
"0.6396",
"0.511",
"0.64"
],
[
"0.0866",
"0.6396",
"0.4878",
"0.6396",
"0.4878",
"0.653",
"0.0866",
"0.653"
],
[
"0.513",
"0.64",
"0.911",
"0.64",
"0.911",
"0.654",
"0.513",
"0.654"
],
[
"0.0866",
"0.654",
"0.4739",
"0.654",
"0.4739",
"0.664",
"0.0866",
"0.664"
],
[
"0.513",
"0.653",
"0.911",
"0.653",
"0.911",
"0.6665",
"0.513",
"0.6665"
],
[
"0.10455",
"0.6685",
"0.4836",
"0.6685",
"0.4836",
"0.6787",
"0.10455",
"0.6787"
],
[
"0.511",
"0.6675",
"0.91",
"0.6665",
"0.91",
"0.681",
"0.511",
"0.6816"
],
[
"0.08496",
"0.681",
"0.486",
"0.681",
"0.486",
"0.6943",
"0.08496",
"0.6943"
],
[
"0.511",
"0.6816",
"0.9126",
"0.6816",
"0.9126",
"0.696",
"0.511",
"0.696"
],
[
"0.0866",
"0.6963",
"0.4844",
"0.6963",
"0.4844",
"0.707",
"0.0866",
"0.707"
],
[
"0.513",
"0.6963",
"0.911",
"0.6963",
"0.911",
"0.707",
"0.513",
"0.707"
],
[
"0.08826",
"0.709",
"0.4836",
"0.709",
"0.4836",
"0.7197",
"0.08826",
"0.7197"
],
[
"0.5146",
"0.711",
"0.911",
"0.711",
"0.911",
"0.7217",
"0.5146",
"0.7217"
],
[
"0.0866",
"0.7227",
"0.486",
"0.7227",
"0.486",
"0.7363",
"0.0866",
"0.7363"
],
[
"0.511",
"0.7236",
"0.9126",
"0.7236",
"0.9126",
"0.7373",
"0.511",
"0.7373"
],
[
"0.08905",
"0.7373",
"0.4836",
"0.7373",
"0.4836",
"0.7476",
"0.08905",
"0.7476"
],
[
"0.513",
"0.7373",
"0.911",
"0.7373",
"0.911",
"0.7505",
"0.513",
"0.7505"
],
[
"0.0866",
"0.7505",
"0.4844",
"0.7505",
"0.4844",
"0.7646",
"0.0866",
"0.7646"
],
[
"0.511",
"0.75",
"0.911",
"0.75",
"0.911",
"0.7646",
"0.511",
"0.7646"
],
[
"0.0866",
"0.7646",
"0.4844",
"0.7646",
"0.4844",
"0.778",
"0.0866",
"0.778"
],
[
"0.513",
"0.765",
"0.9126",
"0.765",
"0.9126",
"0.779",
"0.513",
"0.779"
],
[
"0.08496",
"0.778",
"0.486",
"0.778",
"0.486",
"0.7915",
"0.08496",
"0.7915"
],
[
"0.513",
"0.779",
"0.911",
"0.778",
"0.911",
"0.7915",
"0.513",
"0.7925"
],
[
"0.08496",
"0.7925",
"0.486",
"0.7925",
"0.486",
"0.806",
"0.08496",
"0.806"
],
[
"0.511",
"0.7925",
"0.91",
"0.7925",
"0.91",
"0.806",
"0.511",
"0.806"
],
[
"0.0866",
"0.8047",
"0.486",
"0.8047",
"0.486",
"0.8193",
"0.0866",
"0.8193"
],
[
"0.511",
"0.806",
"0.911",
"0.806",
"0.911",
"0.8193",
"0.511",
"0.8193"
],
[
"0.08496",
"0.8193",
"0.4844",
"0.8193",
"0.4844",
"0.8335",
"0.08496",
"0.8335"
],
[
"0.5107",
"0.8193",
"0.911",
"0.8193",
"0.911",
"0.8335",
"0.5107",
"0.8335"
],
[
"0.08417",
"0.832",
"0.486",
"0.8335",
"0.486",
"0.8477",
"0.08417",
"0.8467"
],
[
"0.513",
"0.834",
"0.91",
"0.834",
"0.91",
"0.8477",
"0.513",
"0.8477"
],
[
"0.08496",
"0.8477",
"0.2695",
"0.8477",
"0.2695",
"0.8613",
"0.08496",
"0.8613"
],
[
"0.513",
"0.8477",
"0.9126",
"0.8477",
"0.9126",
"0.8623",
"0.513",
"0.8623"
],
[
"0.10297",
"0.86",
"0.4878",
"0.8613",
"0.4878",
"0.875",
"0.10297",
"0.8735"
],
[
"0.513",
"0.8623",
"0.914",
"0.8623",
"0.914",
"0.8755",
"0.513",
"0.8755"
],
[
"0.0866",
"0.8755",
"0.486",
"0.8755",
"0.486",
"0.8896",
"0.0866",
"0.8896"
],
[
"0.511",
"0.8755",
"0.7974",
"0.8755",
"0.7974",
"0.8896",
"0.511",
"0.8896"
]
] | [
"with a different sentiment, we do not have to create a new sen-",
"Experimental results show that our method achieves state-of.",
"tence from scratch. Instead, we simply replace the sentimen-",
"the-art results on accuracy, BLEU[Papineni et al., 2002]l and",
"tal tokens of the sentence with other expressions indicative",
"human evaluations",
"of a different sentiment. For example, given a negative re-",
"Our contributions are summarized as follows:",
"view “terrible scenery and poor service\", a straightforward",
". We propose a two-stage “Mask and Infill\" approach to",
"positive review is “beautiful scenery and good service\". Such",
"sentiment transfer task, capable of identifying both sim-",
"a process is very similartothetask oftextinfilling[Zhu et",
"ple and complex sentiment markers and producing high-",
"al., 2019] or Cloze [Taylor, 1953]. Given the template",
"qualitysentences",
"scenery and",
"service\",wehumanfeel easytofill theright",
"words because we have enough prior linguistic knowledge to",
".Experimental results show that our approach outper-",
"predict the missing words from their contexts.",
"forms most state-of-the-art models in terms of both",
"We transform the problem of sentiment transfer to the task",
"BLEU and accuracy scores.",
"of text infilling by a pre-trained Masked Language Model",
". We retrofit MLM to AC-MLM for labeled sentence gen-",
"(MLM). We utilize a pre-trained deep bidirectional language",
"eration. To the best of our knowledge, it is the first work",
"model (corresponding to human linguistic knowledge) to in-",
"to apply a pre-trained masked language model to labeled",
"fillthemaskedpositions.Ourapproachcomprisestwostages:",
"sentence generation task.",
"Mask and Infill (Figure 1). In the mask stage, we explicitly",
"identify and mask the sentiment tokens in a given sentence. In",
"Related Work",
"theinfill stage,wefillthemaskedpositionswithnewexpres",
"sions conditioned on their context and the target sentiment.",
"2.1 Non-parallel Style Transfer",
"There exist two feature-based methods for identifying sen-",
"[Shen et al., 2017; Fu et al., 2018; Xu et al., 2018; Li et",
"timent attribute markers (i.e.words or phrases that are in-",
"al., 2018; Yang et al., 20181 are most relevant to our work.",
"dicative of a particular sentiment attribute). Attention-based",
"[Shen et al., 2017] assumed a shared latent content distribu-",
"method [Xu et al., 2018] trains a self-attention[Bahdanau et",
"tion across different text corpora, and leverages refined align-",
"al., 20151 sentiment classifier, where the learned attention",
"ment oflatent representations.[Fu et al.,2018]learned a",
"weights can be used as signals to identify sentiment mark-",
"representation that only contains the content information by",
"ers, and the tokens with weights higher than the average are",
"multi-task learning and adversarial training.[Xu et al., 2018]",
"regarded as sentiment attributemarkers.The attention-based",
"proposed a cycled reinforcement learningmethodbycollabo-",
"method prefers isolated single words to n-gram phrases,",
"ration betweenaneutralizationmoduleand anemotionaliza-",
"which restricts MLM to produce sentences of diverse expres-",
"tion module. [Li et al., 2018] separated attribute and content",
"sions.Frequency-ratio method[Li et al.,2018] constructsan",
"bydeletingattributemarkers,andattemptedtoreconstructthe",
"attribute marker dictionary of n-gram phrases for each sen-",
"sentence using the content and the retrieved target attribute",
"timent attribute and looks up the dictionary to identify sen-",
"markers with an RNN. [Yang et al., 2018] used a target do-",
"timent phrases in sentences. Frequency-ratio method is in-",
"main language model as the discriminator to guide the gener-",
"clined to mask longer phrases, and its performance heavily",
"ator to produce sentences. Unlike these models, which adopt",
"relies on the quality of attribute marker dictionaries. Thus,",
"RNN as encoder and decoder, we utilize pre-trained MLM to",
"weexploreasimplefusedmethodtoutilizethemeritsofboth",
"capture longer linguistic structure and better language repre-",
"methods.Specifically,we combinetheattention-based clas-",
"sentation.",
"sifier with the frequency-ratio method, filtering out false at-",
"2.2",
"Fine-tuning on Pre-trained Language Model",
"tributemarkersanddiscoveringnewsinglesentimentwords",
"",
"When the frequency-ratio method fails to identify any at-",
"Language model pre-training has attracted widespread at",
"tribute marker or recognize too many ones (with insufficient",
"tention, and fine-tuning on pre-trained language models has",
"contentleft).weutilizetheattention-basedmethoddirectly",
"proven to be effective in improving many downstream natural",
"MLMpredictsthemaskedtokensonlyconditionedontheir",
"language processing tasks. [Dai and Le, 2015] improved Se",
"context, considering no attribute information. To fill the",
"quence Learning with recurrent networks by pre-training on",
"masked positions in a sentence withexpressions compatible",
"unlabeled data.[Radford et al.,2018] improved the perfor-",
"to a particular sentiment.weretrofitMLM toAttributeCon-",
"mancelargelyonmanysentence-leveltasksfromtheGLUE",
"ditional Masked Language Model (AC-MLM) by integrat-",
"benchmarks [Wang et al., 2018]. [Howard and Ruder, 2018]",
"ing attribute embeddingswiththeoriginal input forMLM",
"proposed Universal Language Model Fine-tuning (ULMFiT),",
"Furthermore, we introduce a pre-trained sentiment classi-",
"which was a general transfer learning method for fine-tuning",
"fier to constrain AC-MLM, which ensures the generated sen-",
"a language model. [Radford et al., 2018] demonstrated that",
"tences compatible with the target sentiment. To deal with",
"by generative pre-training language model on unlabeled text",
"the discrete nature of language generation, we utilize soft-",
"from diverse corpora, large gains could be achieved on a di-",
"sampling [Hu et al., 2017] to guide the optimization of AC-",
"verse range of tasks. By introducing a deep bidirectional",
"MLM, where soft-sampling is used to back-propagate gradi-",
"masked language model, BERT [Devlin et al., 2018] obtained",
"ents through the sampling process by using an approximation",
"new state-of-the-art results on a broad range of tasks. Un-",
"of the sampled word vector.",
"like previous works fine-tuning pre-trained language models",
"We evaluate our models on two review datasets Yelp and",
"to perform discriminative tasks, we aim to apply a pre-trained",
"Amazon by quantitative, qualitative, and human evaluations.",
"masked language model on generative task."
] | {
"arxiv_id": "1908.08039",
"language": "en",
"page_idx": 1
} |
||
1908.08039-en-2 | 1908.08039-en-2.jpg | <FigureHere>
Figure 2: The overall model architecture consists of two modules. The Mask module is the fusion-method introduced in Section 3.2. The Infill module consists of two parts, reconstruction and discrimination, corresponding to Equation 10.
# 3 Approach
In this section, we introduce our proposed approach. An overview is presented in Section 3.1. The mask process is introduced in Section 3.2. The details of the infill process are shown in Section 3.3.
## 3.1 0verview
Given a corpus of labeled sentences:
$D$ =
$\{(S_{1},a_{1}),...,(\bar{S_{m}},a_{m})\}$,where $S_i$ is a sentence and $a_i$ is the corresponding attribute. We define $\mathcal{A}$ as the set of all possible attributes, $\mathcal{D} _a$ = $\{ S$ : $( S, a)$ $\in$ $\mathcal{D} \}$ as the set of sentences with attribute $a.$
$\begin{array}{c}\mathrm{Our~goal~is~to~learn~a~model~that~takes~as~input~}(S,\hat{a}),\\\mathrm{where~}S\text{ is a sentence with original sentiment}\mathrm{~attribute~and}\end{array}$ $\hat{a}$ is target sentiment attribute, outputs a sentence $\hat{S}$ that retains the semantic content of $S$while exhibiting $\hat{a} .$
The proposed approach consists of two parts: a mask module and an infill module, as shown in Figure 2. The mask module combines the strength of two methods, utilizing a candidate phrase vocabulary to find the potential attribute markers in sentences, then performs mask operation. Then the masked sentences are sent into the infill module, which infills masked positions via AC-MLM. During the second stage, for better attribute compatibility of generated sentences, we introduce a pre-trained classifier as the discriminator. Due to the discrete nature of language generation, we draw support from soft-sampling to guide the optimization of our model. The implement details are shown in Algorithm 1.
3.2 Identify and Mask Attribute Marker
We first introduce the frequency-ratio method, then the attention-base method. At last, we propose a fusion-method by combing the strength of these two methods.
Frequency-ratio Method
We define for any $a$ in $\mathcal{A}$,let $count(u,\mathcal{D}_a)$ denotes the number of times an n-gram $u$ appear in $D_a$,we score the salience of $u$ with respect to $a$ by its smoothed frequency-ratio:
(1)
$$s_{c}(u,a)=\frac{count(u,\mathcal{D}_{a})+\lambda}{\left(\sum_{a'\in\mathcal{A},a'\neq a}count(u,\mathcal{D}_{a'})\right)+\lambda}$$
where $\lambda$ is the smoothing parameter. We declare $u$ to be a candidate attribute marker for $a$ if $s_c(u,a)$ is above a specified threshold $\gamma_c$, these candidate attribute markers constitute a candidate attribute marker vocabulary $\mathcal{V}_a.$
## Attention-based Method
Each word in a sentence contributes differently to the label or attribute of the whole sentence, we train an attention-based classifier to extract the attribute relative extent. Given a sequence $S$ of N tokens, < $t_1,t_2,...,t_N>$, we adopt a bidirectional LSTM to encode the sentence and concatenate the forward hidden state and the backward hidden state for each word, and obtain:
(2)
$$\mathbf{H}=(h_1,h_2,\cdots,h_N)$$
where $N$ is the length of given sentence. The attention mechanism produces a vector of attention weights a and a vector of weighted hidden states c. Finally a softmax layer is followed to transform c into probability distribution y:
(3) (4)
$$\begin{aligned}&\mathbf{a}=softmax(\mathbf{w}\cdot tanh(\mathbf{W}\mathbf{H}^{T}))\\&\mathbf{c}=\mathbf{a}\cdot\mathbf{H}\\&\mathbf{y}=\mathbf{softmax}(\mathbf{W}^{\prime}\cdot\mathbf{c})\end{aligned}$$
(5)
where w, W, W' are projection parameters. After being well-trained, the attention-based classifier can be used to extract attention weights, which capture the sentiment information of each word. For simplicity, following [Xu et al., 2018], we set the averaged attention value in a sentence as the threshold, words with attention weights higher than the threshold are viewed as attribute markers.
## Fusion-Method However, there are fake attribute markers in the candidate
vocabulary$^{2}.$ So we use a pre-trained classifier, i.e. the attention-based classifier, to calculate the probability $p$ of being real for each candidate attribute marker with Equation 5 and update the salience score:
(6)
$$s(u,a)=s_c(u,a)*p$$
We filter out fake attribute markers with $s(u,a)$ below threshold $\gamma$, obtaining a high-quality attribute marker vocabulary $\mathcal{V}$ for attribute $a.$
After the vocabulary constructed, we identify and mask as many attribute markers as possible by matching the sentences, because within a sentence style words are usually less than content words. Lastly, if frequency-ratio method identifies no attribute marker or too many attribute markers (with insufficient content left, especially for short sentences), we re-process it with the attention-based method3 by identifying the positions with high weights.
3.3 $\textbf{Text Infilling with MLM}$
Attribute Conditional Masked Language Model
Unlike traditional language models which predict tokens
based on previously generated tokens, Masked Language Model (MLM) predicts the masked tokens according to their
$^{2}$In the original positive vocabulary of Yelp dataset,“he did a”
scores as high as 26.0, yet it is not a positive attribute marker.
$^{3}$Note that we use the attention-based classifier twice: the first time, it is used as discriminator to improve the quality of candidate attribute vocabulary, the second time, it is used to extract attention weights for a sentence. | [
[
"0.3096",
"0.0726",
"0.3962",
"0.0726",
"0.3962",
"0.0833",
"0.3096",
"0.0833"
],
[
"0.5146",
"0.06946",
"0.6855",
"0.06946",
"0.6855",
"0.0833",
"0.5146",
"0.0833"
],
[
"0.2769",
"0.084",
"0.308",
"0.0884",
"0.3057",
"0.101",
"0.2737",
"0.0966"
],
[
"0.411",
"0.0833",
"0.464",
"0.0833",
"0.464",
"0.0947",
"0.411",
"0.0947"
],
[
"0.513",
"0.0852",
"0.9126",
"0.0852",
"0.9126",
"0.0985",
"0.513",
"0.0985"
],
[
"0.3423",
"0.0966",
"0.3757",
"0.0966",
"0.3757",
"0.108",
"0.3423",
"0.108"
],
[
"0.4224",
"0.09784",
"0.46",
"0.09784",
"0.46",
"0.1111",
"0.4224",
"0.1111"
],
[
"0.513",
"0.101",
"0.911",
"0.101",
"0.911",
"0.1111",
"0.513",
"0.1111"
],
[
"0.1969",
"0.1061",
"0.2239",
"0.1061",
"0.2239",
"0.11426",
"0.1969",
"0.11426"
],
[
"0.1054",
"0.1016",
"0.1602",
"0.1042",
"0.1602",
"0.11554",
"0.1054",
"0.113"
],
[
"0.2566",
"0.10986",
"0.308",
"0.10986",
"0.308",
"0.1218",
"0.2566",
"0.1218"
],
[
"0.4224",
"0.113",
"0.433",
"0.113",
"0.433",
"0.11743",
"0.4224",
"0.11743"
],
[
"0.4436",
"0.113",
"0.4534",
"0.113",
"0.4534",
"0.11743",
"0.4436",
"0.11743"
],
[
"0.513",
"0.11426",
"0.911",
"0.11426",
"0.911",
"0.125",
"0.513",
"0.125"
],
[
"0.1871",
"0.11554",
"0.2303",
"0.11554",
"0.2303",
"0.127",
"0.1871",
"0.127"
],
[
"0.1209",
"0.11743",
"0.1544",
"0.1193",
"0.1528",
"0.132",
"0.11926",
"0.13"
],
[
"0.343",
"0.1256",
"0.3757",
"0.1256",
"0.3757",
"0.137",
"0.343",
"0.137"
],
[
"0.4143",
"0.1237",
"0.464",
"0.1237",
"0.464",
"0.1351",
"0.4143",
"0.1351"
],
[
"0.511",
"0.127",
"0.9126",
"0.127",
"0.9126",
"0.1401",
"0.511",
"0.1401"
],
[
"0.4224",
"0.1383",
"0.46",
"0.1383",
"0.46",
"0.1515",
"0.4224",
"0.1515"
],
[
"0.3105",
"0.1459",
"0.392",
"0.1459",
"0.392",
"0.1572",
"0.3105",
"0.1572"
],
[
"0.513",
"0.1427",
"0.9126",
"0.1427",
"0.9126",
"0.1528",
"0.513",
"0.1528"
],
[
"0.511",
"0.1547",
"0.914",
"0.1547",
"0.914",
"0.1686",
"0.511",
"0.1686"
],
[
"0.327",
"0.1635",
"0.3987",
"0.1635",
"0.3987",
"0.1749",
"0.327",
"0.1749"
],
[
"0.513",
"0.1686",
"0.6313",
"0.1686",
"0.6313",
"0.1818",
"0.513",
"0.1818"
],
[
"0.6357",
"0.1895",
"0.7876",
"0.1913",
"0.7876",
"0.2058",
"0.6357",
"0.2039"
],
[
"0.8906",
"0.1913",
"0.914",
"0.1913",
"0.914",
"0.2052",
"0.8906",
"0.2052"
],
[
"0.0866",
"0.2084",
"0.4844",
"0.2084",
"0.4844",
"0.2216",
"0.0866",
"0.2216"
],
[
"0.08496",
"0.1945",
"0.4844",
"0.1957",
"0.4844",
"0.209",
"0.08496",
"0.2084"
],
[
"0.513",
"0.2134",
"0.91",
"0.2134",
"0.91",
"0.2279",
"0.513",
"0.2279"
],
[
"0.08826",
"0.2229",
"0.4836",
"0.2229",
"0.4836",
"0.2329",
"0.08826",
"0.2329"
],
[
"0.0866",
"0.2349",
"0.3137",
"0.2349",
"0.3137",
"0.2456",
"0.0866",
"0.2456"
],
[
"0.5146",
"0.2291",
"0.911",
"0.2291",
"0.911",
"0.2393",
"0.5146",
"0.2393"
],
[
"0.5107",
"0.2412",
"0.9126",
"0.2412",
"0.9126",
"0.2551",
"0.5107",
"0.2551"
],
[
"0.5107",
"0.254",
"0.8105",
"0.2551",
"0.8105",
"0.2695",
"0.5107",
"0.2683"
],
[
"0.08417",
"0.2632",
"0.2018",
"0.2632",
"0.2018",
"0.281",
"0.08417",
"0.281"
],
[
"0.596",
"0.2766",
"0.826",
"0.2747",
"0.826",
"0.2922",
"0.596",
"0.2942"
],
[
"0.888",
"0.2766",
"0.915",
"0.2766",
"0.915",
"0.2942",
"0.888",
"0.2942"
],
[
"0.5986",
"0.2942",
"0.671",
"0.2942",
"0.671",
"0.3093",
"0.5986",
"0.3093"
],
[
"0.08417",
"0.2822",
"0.486",
"0.2827",
"0.486",
"0.2974",
"0.08417",
"0.2966"
],
[
"0.08496",
"0.2974",
"0.4878",
"0.2974",
"0.4878",
"0.3113",
"0.08496",
"0.3113"
],
[
"0.888",
"0.2954",
"0.915",
"0.2954",
"0.915",
"0.3125",
"0.888",
"0.3125"
],
[
"0.08496",
"0.3113",
"0.4878",
"0.3113",
"0.4878",
"0.3245",
"0.08496",
"0.3245"
],
[
"0.596",
"0.3145",
"0.7563",
"0.3125",
"0.7563",
"0.33",
"0.596",
"0.332"
],
[
"0.888",
"0.315",
"0.915",
"0.315",
"0.915",
"0.332",
"0.888",
"0.332"
],
[
"0.08496",
"0.327",
"0.2296",
"0.327",
"0.2296",
"0.3372",
"0.08496",
"0.3372"
],
[
"0.509",
"0.3384",
"0.914",
"0.339",
"0.914",
"0.3567",
"0.509",
"0.356"
],
[
"0.08496",
"0.348",
"0.2018",
"0.348",
"0.2018",
"0.361",
"0.08496",
"0.361"
],
[
"0.511",
"0.356",
"0.9126",
"0.356",
"0.9126",
"0.3694",
"0.511",
"0.3694"
],
[
"0.08417",
"0.363",
"0.4019",
"0.363",
"0.4019",
"0.377",
"0.08417",
"0.377"
],
[
"0.423",
"0.3643",
"0.4468",
"0.3643",
"0.4468",
"0.3757",
"0.423",
"0.3757"
],
[
"0.513",
"0.3718",
"0.91",
"0.3718",
"0.91",
"0.3818",
"0.513",
"0.3818"
],
[
"0.0866",
"0.379",
"0.486",
"0.377",
"0.486",
"0.3914",
"0.0866",
"0.3933"
],
[
"0.5107",
"0.3833",
"0.9126",
"0.3833",
"0.9126",
"0.3965",
"0.5107",
"0.3965"
],
[
"0.08496",
"0.3926",
"0.486",
"0.3926",
"0.486",
"0.406",
"0.08496",
"0.406"
],
[
"0.5107",
"0.3977",
"0.911",
"0.3965",
"0.911",
"0.4104",
"0.5107",
"0.411"
],
[
"0.08417",
"0.406",
"0.4878",
"0.4048",
"0.4878",
"0.4192",
"0.08417",
"0.4204"
],
[
"0.5107",
"0.411",
"0.9126",
"0.4104",
"0.9126",
"0.4248",
"0.5107",
"0.4255"
],
[
"0.0866",
"0.4216",
"0.2573",
"0.4216",
"0.2573",
"0.432",
"0.0866",
"0.432"
],
[
"0.511",
"0.4255",
"0.7236",
"0.4255",
"0.7236",
"0.4395",
"0.511",
"0.4395"
],
[
"0.10297",
"0.433",
"0.4844",
"0.433",
"0.4844",
"0.4463",
"0.10297",
"0.4463"
],
[
"0.08496",
"0.4475",
"0.4878",
"0.4475",
"0.4878",
"0.461",
"0.08496",
"0.461"
],
[
"0.511",
"0.4443",
"0.6235",
"0.4458",
"0.6235",
"0.4602",
"0.511",
"0.459"
],
[
"0.08496",
"0.4634",
"0.4844",
"0.4634",
"0.4844",
"0.4778",
"0.08496",
"0.4778"
],
[
"0.5146",
"0.4634",
"0.911",
"0.4634",
"0.911",
"0.4734",
"0.5146",
"0.4734"
],
[
"0.0866",
"0.4792",
"0.415",
"0.4792",
"0.415",
"0.4893",
"0.0866",
"0.4893"
],
[
"0.513",
"0.476",
"0.9126",
"0.476",
"0.9126",
"0.4893",
"0.513",
"0.4893"
],
[
"0.10297",
"0.4912",
"0.4844",
"0.4912",
"0.4844",
"0.505",
"0.10297",
"0.505"
],
[
"0.511",
"0.4905",
"0.911",
"0.4905",
"0.911",
"0.504",
"0.511",
"0.504"
],
[
"0.08496",
"0.505",
"0.486",
"0.505",
"0.486",
"0.518",
"0.08496",
"0.518"
],
[
"0.511",
"0.504",
"0.9126",
"0.504",
"0.9126",
"0.5176",
"0.511",
"0.5176"
],
[
"0.0866",
"0.521",
"0.4878",
"0.521",
"0.4878",
"0.531",
"0.0866",
"0.531"
],
[
"0.511",
"0.5176",
"0.7134",
"0.5176",
"0.7134",
"0.531",
"0.511",
"0.531"
],
[
"0.08496",
"0.5327",
"0.486",
"0.5327",
"0.486",
"0.547",
"0.08496",
"0.547"
],
[
"0.6396",
"0.5405",
"0.786",
"0.5405",
"0.786",
"0.5547",
"0.6396",
"0.5547"
],
[
"0.8906",
"0.5405",
"0.914",
"0.5405",
"0.914",
"0.5537",
"0.8906",
"0.5537"
],
[
"0.0866",
"0.547",
"0.4878",
"0.547",
"0.4878",
"0.56",
"0.0866",
"0.56"
],
[
"0.0866",
"0.561",
"0.482",
"0.561",
"0.482",
"0.5713",
"0.0866",
"0.5713"
],
[
"0.528",
"0.561",
"0.9126",
"0.5625",
"0.9126",
"0.577",
"0.528",
"0.5757"
],
[
"0.0866",
"0.5747",
"0.4844",
"0.5747",
"0.4844",
"0.5884",
"0.0866",
"0.5884"
],
[
"0.513",
"0.577",
"0.911",
"0.577",
"0.911",
"0.5903",
"0.513",
"0.5903"
],
[
"0.08496",
"0.587",
"0.4878",
"0.587",
"0.4878",
"0.601",
"0.08496",
"0.601"
],
[
"0.511",
"0.5903",
"0.6665",
"0.5903",
"0.6665",
"0.604",
"0.511",
"0.604"
],
[
"0.08826",
"0.604",
"0.486",
"0.604",
"0.486",
"0.6143",
"0.08826",
"0.6143"
],
[
"0.5293",
"0.606",
"0.9126",
"0.606",
"0.9126",
"0.616",
"0.5293",
"0.616"
],
[
"0.0866",
"0.616",
"0.4878",
"0.616",
"0.4878",
"0.63",
"0.0866",
"0.63"
],
[
"0.511",
"0.6187",
"0.911",
"0.6187",
"0.911",
"0.632",
"0.511",
"0.632"
],
[
"0.0866",
"0.63",
"0.486",
"0.63",
"0.486",
"0.643",
"0.0866",
"0.643"
],
[
"0.511",
"0.632",
"0.9126",
"0.632",
"0.9126",
"0.646",
"0.511",
"0.646"
],
[
"0.0866",
"0.643",
"0.4102",
"0.643",
"0.4102",
"0.657",
"0.0866",
"0.657"
],
[
"0.5146",
"0.648",
"0.91",
"0.648",
"0.91",
"0.6577",
"0.5146",
"0.6577"
],
[
"0.08496",
"0.664",
"0.4084",
"0.6655",
"0.4084",
"0.6797",
"0.08496",
"0.6787"
],
[
"0.513",
"0.66",
"0.9126",
"0.66",
"0.9126",
"0.674",
"0.513",
"0.674"
],
[
"0.511",
"0.674",
"0.914",
"0.674",
"0.914",
"0.6875",
"0.511",
"0.6875"
],
[
"0.0866",
"0.683",
"0.4878",
"0.683",
"0.4878",
"0.6963",
"0.0866",
"0.6963"
],
[
"0.5107",
"0.686",
"0.9126",
"0.6875",
"0.9126",
"0.702",
"0.5107",
"0.7007"
],
[
"0.0866",
"0.6978",
"0.4878",
"0.6978",
"0.4878",
"0.711",
"0.0866",
"0.711"
],
[
"0.511",
"0.702",
"0.722",
"0.702",
"0.722",
"0.7153",
"0.511",
"0.7153"
],
[
"0.08496",
"0.711",
"0.3938",
"0.711",
"0.3938",
"0.7246",
"0.08496",
"0.7246"
],
[
"0.513",
"0.7246",
"0.7417",
"0.7246",
"0.7417",
"0.7393",
"0.513",
"0.7393"
],
[
"0.0866",
"0.731",
"0.2607",
"0.731",
"0.2607",
"0.744",
"0.0866",
"0.744"
],
[
"0.0866",
"0.7456",
"0.486",
"0.7456",
"0.486",
"0.759",
"0.0866",
"0.759"
],
[
"0.513",
"0.7437",
"0.847",
"0.7437",
"0.847",
"0.7583",
"0.513",
"0.7583"
],
[
"0.08417",
"0.7583",
"0.4885",
"0.759",
"0.4885",
"0.7734",
"0.08417",
"0.773"
],
[
"0.513",
"0.759",
"0.9126",
"0.759",
"0.9126",
"0.773",
"0.513",
"0.773"
],
[
"0.08496",
"0.7734",
"0.4495",
"0.7734",
"0.4495",
"0.787",
"0.08496",
"0.787"
],
[
"0.511",
"0.7734",
"0.9126",
"0.7734",
"0.9126",
"0.787",
"0.511",
"0.787"
],
[
"0.2607",
"0.7905",
"0.384",
"0.7905",
"0.384",
"0.8047",
"0.2607",
"0.8047"
],
[
"0.513",
"0.787",
"0.9126",
"0.787",
"0.9126",
"0.8003",
"0.513",
"0.8003"
],
[
"0.1323",
"0.802",
"0.1936",
"0.802",
"0.1936",
"0.813",
"0.1323",
"0.813"
],
[
"0.464",
"0.7974",
"0.4878",
"0.8003",
"0.4836",
"0.8145",
"0.46",
"0.811"
],
[
"0.2303",
"0.813",
"0.4084",
"0.8066",
"0.4102",
"0.8257",
"0.232",
"0.832"
],
[
"0.5312",
"0.81",
"0.911",
"0.811",
"0.911",
"0.8257",
"0.5312",
"0.8247"
],
[
"0.513",
"0.8257",
"0.875",
"0.8257",
"0.875",
"0.836",
"0.513",
"0.836"
],
[
"0.0866",
"0.8335",
"0.4878",
"0.8335",
"0.4878",
"0.8477",
"0.0866",
"0.8477"
],
[
"0.5317",
"0.836",
"0.9126",
"0.837",
"0.9126",
"0.851",
"0.5317",
"0.8496"
],
[
"0.0866",
"0.8477",
"0.4885",
"0.8477",
"0.4885",
"0.8613",
"0.0866",
"0.8613"
],
[
"0.511",
"0.851",
"0.9126",
"0.851",
"0.9126",
"0.8643",
"0.511",
"0.8643"
],
[
"0.08496",
"0.8613",
"0.4885",
"0.8623",
"0.4885",
"0.8755",
"0.08496",
"0.875"
],
[
"0.511",
"0.863",
"0.914",
"0.863",
"0.914",
"0.877",
"0.511",
"0.877"
],
[
"0.08417",
"0.875",
"0.3694",
"0.8755",
"0.3694",
"0.89",
"0.08417",
"0.8896"
],
[
"0.513",
"0.8774",
"0.6504",
"0.8774",
"0.6504",
"0.888",
"0.513",
"0.888"
]
] | [
"Reconstruction",
"Attention-based Method",
"a1",
"content",
"Each word in a sentence contributes differently to the label",
"MLM",
"stylei",
"orattribute ofthewhole sentence,wetrain an attention-based",
"Mask",
"content",
"content",
"-",
"-",
"classifierto extract the attribute relative extent. Given a se-",
"Module",
"stylet",
"MLM",
"content",
"quence S of N tokens, < ti,t2, ...,tn >, we adopt a bidi-",
"style2",
"Discrimination",
"rectionalLSTM to encodethe sentence and concatenate the",
"forward hidden state and the backward hidden state for each",
"Discriminator",
"word, and obtain:",
"H= (hi,h2,*.., hn)",
"(2)",
"The Mask module is the fusion-method introduced in Section 3.2.",
"Figure 2:The overall model architecture consists of two modules.",
"where N is the length of given sentence. The attention mech-",
"TheInfill moduleconsists of two parts,reconstruction anddiscrim",
"ination,correspondingtoEquation10.",
"anismproducesavectorofattentionweightsaandavectorof",
"weighted hidden states c. Finally a softmax layer is followed",
"to transform c into probability distribution y:",
"3Approach",
"a = softmar(w -tanh(WHT))",
"(3)",
"c=a.H",
"In this section, we introduce our proposed approach.An",
"overview is presented in Section 3.1. The mask process is",
"(4)",
"introduced in Section 3.2. The details of the infill process are",
"y = softmax(w .c)",
"(5)",
"shown in Section 3.3.",
"where w, W, W\" are projection parameters. After being",
"3.1 Overview",
"well-trained, the attention-based classifier can be used to ex-",
"Given a corpus of labeled sentences:",
"D",
"tractattentionweights,whichcapturethe sentimentinforma-",
"[(Si,ai),..,(Sm,am)), where S, is a sentence and ai",
"tion of each word. For simplicity, following [Xu et al., 2018],",
"is the corresponding attribute. We define A as the set of all",
"we set the averaged attention value in a sentence as the thresh-",
"possible attributes, Da = (S : (S,a) E D) as the set of",
"old, words with attention weights higher than the threshold",
"sentences withattributea.",
"are viewed as attribute markers.",
"Our goal is to learn a model that takes as input (S,a),",
"where S is a sentence with original sentiment attribute and",
"Fusion-Method",
"a is target sentiment attribute, outputs a sentence S that re-",
"Howeverthere are fake attribute markers inthe candidate",
"tainsthe semantic content of Swhile exhibiting a.",
"vocabulary?.So we use a pre-trained classifier, i.e. the",
"The proposed approach consists of two parts: a mask mod-",
"attention-based classifier, to calculate the probability p of be-",
"ule and an infill module, as shown in Figure 2. The mask",
"ing real for each candidate attribute marker with Equation 5",
"module combines the strength of two methods,utilizing a",
"and update the salience score:",
"candidate phrase vocabulary to find the potential attribute",
"s(u,a) = sc(u,a) *p",
"(6)",
"markers in sentences, then performs mask operation. Then",
"themaskedsentencesare sent intotheinfill module.whichin",
"We filter out fake attribute markers with s(u,a) below",
"fills masked positions via AC-MLM. During the second stage,",
"threshold , obtaining a high-quality attribute marker vocab-",
"for better attribute compatibility of generated sentences, we",
"ulary V for attribute a.",
"introduceapre-trained classifierasthediscriminator.Dueto",
"Afterthe vocabulary constructed.weidentify and mask",
"the discrete nature of language generation, we draw support",
"as many attribute markers as possible by matching the sen-",
"from soft-sampling to guide the optimization of our model.",
"tences, because within a sentence style words are usually less",
"The implement details are shown in Algorithm 1.",
"thancontentwords.Lastly,iffrequency-ratiomethod identi-",
"3.2 Identify and Mask Attribute Marker",
"fies no attribute marker or too many attribute markers (with",
"insufficient content left, especially for short sentences), we",
"We first introduce the frequency-ratio method, then the",
"re-process it with the attention-based method3 by identifying",
"attention-base method. At last, we propose a fusion-method",
"the positions with high weights.",
"by combing the strength of these two methods.",
"3.3Text Infilling with MLM",
"Frequency-ratio Method",
"We define for any a in A, let count(u, D.) denotes the num-",
"Attribute Conditional Masked Language Model",
"ber of times an n-gram u appear in Da, we score the salience",
"Unlike traditional language models which predict tokens",
"of u with respect to a by its smoothed frequency-ratio :",
"based on previously generated tokens, Masked Language",
"count(u, Da) +^",
"Model (MLM) predicts the masked tokens according to their",
"sc(u,a)",
"(1)",
"Ca'eA,a'+a count(u, Da))",
"2In the original positive vocabulary of Yelp dataset, “he did a\"",
"scores as high as 26.0,yet it is not a positive attributemarker.",
"where is the smoothing parameter. We declare u to be a",
"3Note that we use the attention-based classifier twice: the first",
"candidate attribute marker for a if se(u, a) is above a speci-",
"time, it is used as discriminator to improve the quality of candidate",
"fied threshold e,these candidate attribute markers constitute",
"attribute vocabulary, the second time, it is used to extract attention",
"a candidate attribute marker vocabulary Va",
"weights for a sentence."
] | {
"arxiv_id": "1908.08039",
"language": "en",
"page_idx": 2
} |
||
1908.08039-en-3 | 1908.08039-en-3.jpg | context, rather than reconstructing the entire input. For the task of sentiment transfer, infilled words or phrases should be consistent to a target attribute. Therefore, we propose Attribute Conditional Masked Language Model (AC-MLM) to infill masked positions conditioned on both their context and a target label, by adding attribute embeddings into original input embeddings. Let $\hat{a}\in\mathcal{A}$ be a target attribute, $S=$ $\langle t_{1},t_{2},...,t_{N}\rangle$ a sentence of N tokens, $M=\{t_i_{1},...,t_{i_{m}}\}$ the set of masked tokens, $\overline{S}=S\setminus M$ the set of content tokens.
Reconstruction
Lacking parallel sentiment sentence pairs, so we train the AC-MLM to reconstruct the original sentence $S$ conditioned on the content $\overline{S}$ and original attribute $a.$ The AC-MLM is optimized to minimize the reconstruction error of original masked words:
$$\mathcal{L}_{rec}=-\sum_{a\in\mathcal{A},t_{i}\in M}logp(t_{i}|\overline{S},a)$$
(7)
Well-trained AC-MLM model can take the sentence content $\overline{S}$ and a target attribute $\hat{a}$ as input, and output $p(\cdot|\hat{a},\overline{S})$ for each masked position. After all masked positions are infilled, we get the transferred sentence:
$$\hat{S}=\text{AC-MLM}(\overline{S},\hat{a})$$
(8)
Specifically, we use the pre-trained BERT as our MLM, and substitute segmentation embeddings (which is useless when training single sentence) with attribute embeddings.
### Pre-trained Classifier Constraint
To further improve the transfer accuracy, we introduce a pre-
trained sentiment classifier to the AC-MLM, i.e. AC-MLMSS,“-SS" indicates with soft-sampling which will be introduced later.
Discrimination Given the content $\overline{S}$ and a target attribute $\hat{a}$, when a sentence is successfully transferred, the label predicted by classifier should be consistent to the target attribute. So the AC-MLM-SS is further optimized to minimize the discrimination discrepancy:
$$\mathcal{L}_{acc}=-logp(\hat{a}|\hat{S})$$
(9)
where $\hat{S}$ indicates transferred sentence generated by AC-
$MLM.$
By combining Equation 7 and Equation 9 we obtain the
objective function of AC-MLM-SS:
$$min_{\theta}\mathcal{L}=\mathcal{L}_{rec}+\eta\mathcal{L}_{acc}$$
(10)
where $\theta$ is AC-MLM-SS's parameters and $\eta$ is a balancing
parameter.
To propagate gradients from the discriminator through the
discrete samples, we adopt soft-sampling.
Soft-sampling Soft-sampling back-propagates gradients by using an approximation of the sampled word vector. The approximation replaces the sampled token $t_i$ (represented as a one-hot vector) at each step with the probability vector
$$\hat{t}_{i}\sim softmax(\mathbf{o}_{t}/\tau)$$
(11)
which is differentiable w.r.t the AC-MLM's parameters. The resulting soft generated sentence is fed into the discriminator to measure the fitness to the target attribute.
<table>
<tbody>
<tr>
<th>$\overline{\mathbf{Alg}}$ -</th>
<th>$0\mathbf{rithm}$ and Infill 1 1 lmple $\because\mathbf{Mas}$ abbroacn</th>
</tr>
<tr>
<td>1 · 1 1 ·</td>
<td>TC (Eq. 2- 5)</td>
</tr>
<tr>
<td>2: 11 一</td>
<td>( Fa16</td>
</tr>
<tr>
<td>3: 一 1 K 1 I</td>
<td>77, $d_{0}$ 11n'</td>
</tr>
<tr>
<td>4: 1 一 1</td>
<td> </td>
</tr>
<tr>
<td>5: $r$</td>
<td>ifSis $Or$ 人 1 15</td>
</tr>
<tr>
<td>6:</td>
<td>$(\mathrm{Fo}^{3}$</td>
</tr>
<tr>
<td>7: $w$</td>
<td>end if</td>
</tr>
<tr>
<td>8:</td>
<td>end $1for$</td>
</tr>
<tr>
<td>9: 1. 一 1</td>
<td>$2, . . . . . . , \textbf{M do}$</td>
</tr>
<tr>
<td>10:</td>
<td>C with samnl aftrihite</td>
</tr>
<tr>
<td>1 1</td>
<td>amr</td>
</tr>
<tr>
<td>12: </td>
<td>$\hat{a}=$the starget attribute</td>
</tr>
<tr>
<td>13:</td>
<td>Construct $\ddot{S}\left(\mathrm{Eq.8}\right)$</td>
</tr>
<tr>
<td>1. 4: 1 1 1</td>
<td>calculatino $(\mathbf{Eq.9})$</td>
</tr>
<tr>
<td>15: </td>
<td>IIndate oarameters $\theta$ amode</td>
</tr>
<tr>
<td>16:</td>
<td>end for</td>
</tr>
</tbody>
</table>
## 4 $\textbf{Experiment}$
In this section, we evaluate our method on two review datasets Yelp and Amazon. Firstly, we introduce datasets, baseline models, training details, and evaluation metrics. Secondly, we compare our approach to several state-of-the-art systems. Finally, we present our experimental results and analyze each component in detail.
## 4.1 Datasets
We experiment our methods on two review datasets from [Li $etal.,2018]:$ Yelp and Amazon [He and McAuley, 2016], each of which is randomly split into training, validation and testing sets. Examples in YELP are sentences from business reviews on Yelp, and examples in AMAZON are sentences from product reviews on Amazon. Each example is labeled as having either positive or negative sentiment. The statistics of datasets are shown in Table 1.
## 4.2 Baselines
We compared our methods to existing relevant models: CrossAligned[Shen $etal.,2017]$, StyleEmbedding and MultiDecoder [Fu et al., 2018], CycledReinforce[Xu $et$ $al.,2018]^4$, TemplateBased, RetrievalOnly, DeleteOnly and DeleteAndRetrieval [Li et al., 2018], LM+Classifer [Yang $et$ $al.,2018].$
## 4.3 Experiment Details
Mask step
For the frequency-ratio method, we consider n-gram up to 4-gram and set the smoothing parameter $\lambda$ to 1. Other hyperparameters are following [Li $et$ al., 2018]. For the attentionbased method, we train the attention-based classifier for 10 epochs with accuracy 97% for Yelp and 82% for Amazon. For the fusion-method, we consider a sentence with content less than 5 tokens as insufficient content, and re-process it with the attention-based method.
$^{4}$https://github.com/lancopku/Unpaired-Sentiment-Translation | [
[
"0.08496",
"0.06946",
"0.4878",
"0.06946",
"0.4878",
"0.0833",
"0.08496",
"0.0833"
],
[
"0.513",
"0.06946",
"0.91",
"0.06946",
"0.91",
"0.0833",
"0.513",
"0.0833"
],
[
"0.08496",
"0.084",
"0.4878",
"0.084",
"0.4878",
"0.09784",
"0.08496",
"0.09784"
],
[
"0.5186",
"0.0859",
"0.8164",
"0.0871",
"0.8164",
"0.101",
"0.5186",
"0.09973"
],
[
"0.08417",
"0.0966",
"0.486",
"0.09784",
"0.486",
"0.1111",
"0.08417",
"0.10986"
],
[
"0.517",
"0.0985",
"0.8286",
"0.09973",
"0.8286",
"0.11426",
"0.517",
"0.113"
],
[
"0.08496",
"0.11237",
"0.4878",
"0.11237",
"0.4878",
"0.1256",
"0.08496",
"0.1256"
],
[
"0.517",
"0.113",
"0.7134",
"0.113",
"0.7134",
"0.127",
"0.517",
"0.127"
],
[
"0.0866",
"0.127",
"0.4844",
"0.127",
"0.4844",
"0.137",
"0.0866",
"0.137"
],
[
"0.5195",
"0.1276",
"0.5337",
"0.1276",
"0.5337",
"0.1364",
"0.5195",
"0.1364"
],
[
"0.5522",
"0.1256",
"0.9067",
"0.1256",
"0.9067",
"0.1395",
"0.5522",
"0.1395"
],
[
"0.08496",
"0.1395",
"0.486",
"0.1395",
"0.486",
"0.1528",
"0.08496",
"0.1528"
],
[
"0.5186",
"0.1401",
"0.5337",
"0.1401",
"0.5337",
"0.1497",
"0.5186",
"0.1497"
],
[
"0.5513",
"0.1383",
"0.8105",
"0.1383",
"0.8105",
"0.1515",
"0.5513",
"0.1515"
],
[
"0.08496",
"0.1528",
"0.4656",
"0.1528",
"0.4656",
"0.1666",
"0.08496",
"0.1666"
],
[
"0.5186",
"0.1528",
"0.5337",
"0.1528",
"0.5337",
"0.1622",
"0.5186",
"0.1622"
],
[
"0.569",
"0.1509",
"0.9043",
"0.1509",
"0.9043",
"0.1642",
"0.569",
"0.1642"
],
[
"0.5195",
"0.1642",
"0.5337",
"0.1642",
"0.5337",
"0.1736",
"0.5195",
"0.1736"
],
[
"0.5522",
"0.1642",
"0.5947",
"0.1642",
"0.5947",
"0.1755",
"0.5522",
"0.1755"
],
[
"0.5186",
"0.1768",
"0.585",
"0.1768",
"0.585",
"0.1881",
"0.5186",
"0.1881"
],
[
"0.0866",
"0.1666",
"0.486",
"0.1666",
"0.486",
"0.1812",
"0.0866",
"0.1812"
],
[
"0.08417",
"0.1818",
"0.4739",
"0.1818",
"0.4739",
"0.1957",
"0.08417",
"0.1957"
],
[
"0.5156",
"0.1875",
"0.726",
"0.1875",
"0.726",
"0.2008",
"0.5156",
"0.2008"
],
[
"0.513",
"0.202",
"0.535",
"0.202",
"0.535",
"0.214",
"0.513",
"0.214"
],
[
"0.5537",
"0.202",
"0.8228",
"0.202",
"0.8228",
"0.2153",
"0.5537",
"0.2153"
],
[
"0.513",
"0.214",
"0.535",
"0.214",
"0.535",
"0.2267",
"0.513",
"0.2267"
],
[
"0.08826",
"0.2039",
"0.1936",
"0.2039",
"0.1936",
"0.214",
"0.08826",
"0.214"
],
[
"0.08496",
"0.2172",
"0.486",
"0.2172",
"0.486",
"0.2311",
"0.08496",
"0.2311"
],
[
"0.5537",
"0.214",
"0.903",
"0.214",
"0.903",
"0.2279",
"0.5537",
"0.2279"
],
[
"0.513",
"0.2267",
"0.535",
"0.2267",
"0.535",
"0.2393",
"0.513",
"0.2393"
],
[
"0.5537",
"0.2267",
"0.6895",
"0.2267",
"0.6895",
"0.2405",
"0.5537",
"0.2405"
],
[
"0.5537",
"0.2405",
"0.671",
"0.2405",
"0.671",
"0.254",
"0.5537",
"0.254"
],
[
"0.0866",
"0.2311",
"0.4878",
"0.2311",
"0.4878",
"0.2443",
"0.0866",
"0.2443"
],
[
"0.08496",
"0.2443",
"0.4878",
"0.2443",
"0.4878",
"0.2583",
"0.08496",
"0.2583"
],
[
"0.513",
"0.2412",
"0.536",
"0.2412",
"0.536",
"0.2524",
"0.513",
"0.2524"
],
[
"0.513",
"0.254",
"0.536",
"0.254",
"0.536",
"0.2651",
"0.513",
"0.2651"
],
[
"0.5522",
"0.254",
"0.697",
"0.254",
"0.697",
"0.2676",
"0.5522",
"0.2676"
],
[
"0.5537",
"0.2651",
"0.721",
"0.2651",
"0.721",
"0.279",
"0.5537",
"0.279"
],
[
"0.08496",
"0.2588",
"0.486",
"0.2588",
"0.486",
"0.2727",
"0.08496",
"0.2727"
],
[
"0.0866",
"0.2734",
"0.1871",
"0.2734",
"0.1871",
"0.2854",
"0.0866",
"0.2854"
],
[
"0.5156",
"0.2676",
"0.5337",
"0.2676",
"0.5337",
"0.2766",
"0.5156",
"0.2766"
],
[
"0.5146",
"0.2798",
"0.5864",
"0.2798",
"0.5864",
"0.2917",
"0.5146",
"0.2917"
],
[
"0.1765",
"0.289",
"0.2279",
"0.2917",
"0.2263",
"0.3062",
"0.1748",
"0.3037"
],
[
"0.3015",
"0.2886",
"0.3962",
"0.2886",
"0.3962",
"0.303",
"0.3015",
"0.303"
],
[
"0.464",
"0.2903",
"0.4878",
"0.2903",
"0.4878",
"0.3037",
"0.464",
"0.3037"
],
[
"0.2427",
"0.307",
"0.3096",
"0.307",
"0.3096",
"0.318",
"0.2427",
"0.318"
],
[
"0.5215",
"0.3193",
"0.528",
"0.3245",
"0.5215",
"0.33",
"0.5146",
"0.3245"
],
[
"0.54",
"0.3176",
"0.6445",
"0.3176",
"0.6445",
"0.332",
"0.54",
"0.332"
],
[
"0.08496",
"0.336",
"0.4844",
"0.336",
"0.4844",
"0.3499",
"0.08496",
"0.3499"
],
[
"0.1013",
"0.3208",
"0.4844",
"0.3213",
"0.4844",
"0.336",
"0.1013",
"0.3352"
],
[
"0.513",
"0.3423",
"0.9126",
"0.3423",
"0.9126",
"0.356",
"0.513",
"0.356"
],
[
"0.08496",
"0.3499",
"0.4878",
"0.3499",
"0.4878",
"0.363",
"0.08496",
"0.363"
],
[
"0.08496",
"0.3623",
"0.3381",
"0.363",
"0.3381",
"0.3774",
"0.08496",
"0.377"
],
[
"0.513",
"0.3567",
"0.9126",
"0.3567",
"0.9126",
"0.3706",
"0.513",
"0.3706"
],
[
"0.513",
"0.3706",
"0.9126",
"0.3706",
"0.9126",
"0.3838",
"0.513",
"0.3838"
],
[
"0.2157",
"0.3806",
"0.357",
"0.3818",
"0.357",
"0.3965",
"0.2157",
"0.3958"
],
[
"0.464",
"0.3838",
"0.4878",
"0.3838",
"0.4878",
"0.3977",
"0.464",
"0.3977"
],
[
"0.513",
"0.385",
"0.911",
"0.385",
"0.911",
"0.3984",
"0.513",
"0.3984"
],
[
"0.10297",
"0.4016",
"0.4844",
"0.4016",
"0.4844",
"0.4153",
"0.10297",
"0.4153"
],
[
"0.511",
"0.3984",
"0.914",
"0.3984",
"0.914",
"0.4124",
"0.511",
"0.4124"
],
[
"0.08496",
"0.4167",
"0.4878",
"0.4167",
"0.4878",
"0.43",
"0.08496",
"0.43"
],
[
"0.511",
"0.4124",
"0.649",
"0.4124",
"0.649",
"0.4255",
"0.511",
"0.4255"
],
[
"0.08496",
"0.428",
"0.4631",
"0.43",
"0.4631",
"0.4443",
"0.08496",
"0.4426"
],
[
"0.5146",
"0.4414",
"0.539",
"0.4414",
"0.539",
"0.4539",
"0.5146",
"0.4539"
],
[
"0.5513",
"0.4426",
"0.619",
"0.4426",
"0.619",
"0.4539",
"0.5513",
"0.4539"
],
[
"0.0866",
"0.4495",
"0.3186",
"0.4495",
"0.3186",
"0.4634",
"0.0866",
"0.4634"
],
[
"0.0866",
"0.464",
"0.4844",
"0.464",
"0.4844",
"0.4778",
"0.0866",
"0.4778"
],
[
"0.513",
"0.4634",
"0.914",
"0.4634",
"0.914",
"0.4778",
"0.513",
"0.4778"
],
[
"0.08496",
"0.4778",
"0.486",
"0.4778",
"0.486",
"0.4912",
"0.08496",
"0.4912"
],
[
"0.511",
"0.4778",
"0.911",
"0.4778",
"0.911",
"0.4912",
"0.511",
"0.4912"
],
[
"0.08496",
"0.4912",
"0.4878",
"0.4912",
"0.4878",
"0.505",
"0.08496",
"0.505"
],
[
"0.511",
"0.4912",
"0.914",
"0.4912",
"0.914",
"0.505",
"0.511",
"0.505"
],
[
"0.0866",
"0.507",
"0.1642",
"0.507",
"0.1642",
"0.518",
"0.0866",
"0.518"
],
[
"0.511",
"0.505",
"0.914",
"0.505",
"0.914",
"0.518",
"0.511",
"0.518"
],
[
"0.0866",
"0.5234",
"0.4878",
"0.5234",
"0.4878",
"0.537",
"0.0866",
"0.537"
],
[
"0.511",
"0.518",
"0.9126",
"0.518",
"0.9126",
"0.532",
"0.511",
"0.532"
],
[
"0.08496",
"0.537",
"0.486",
"0.537",
"0.486",
"0.5503",
"0.08496",
"0.5503"
],
[
"0.511",
"0.5327",
"0.9126",
"0.5327",
"0.9126",
"0.547",
"0.511",
"0.547"
],
[
"0.0866",
"0.55",
"0.4844",
"0.55",
"0.4844",
"0.563",
"0.0866",
"0.563"
],
[
"0.511",
"0.547",
"0.914",
"0.547",
"0.914",
"0.56",
"0.511",
"0.56"
],
[
"0.511",
"0.56",
"0.728",
"0.56",
"0.728",
"0.5737",
"0.511",
"0.5737"
],
[
"0.0866",
"0.5664",
"0.4844",
"0.5664",
"0.4844",
"0.577",
"0.0866",
"0.577"
],
[
"0.08417",
"0.5776",
"0.2524",
"0.58",
"0.2524",
"0.5933",
"0.08417",
"0.5913"
],
[
"0.5146",
"0.5903",
"0.5522",
"0.5903",
"0.5522",
"0.604",
"0.5146",
"0.604"
],
[
"0.546",
"0.587",
"0.6274",
"0.589",
"0.6274",
"0.604",
"0.546",
"0.6016"
],
[
"0.2173",
"0.5986",
"0.353",
"0.595",
"0.353",
"0.613",
"0.2173",
"0.616"
],
[
"0.464",
"0.5986",
"0.4878",
"0.5986",
"0.4878",
"0.613",
"0.464",
"0.613"
],
[
"0.511",
"0.61",
"0.911",
"0.6113",
"0.911",
"0.6255",
"0.511",
"0.625"
],
[
"0.0866",
"0.6187",
"0.4844",
"0.6187",
"0.4844",
"0.6333",
"0.0866",
"0.6333"
],
[
"0.513",
"0.6255",
"0.914",
"0.6255",
"0.914",
"0.6396",
"0.513",
"0.6396"
],
[
"0.0866",
"0.6333",
"0.1299",
"0.6333",
"0.1299",
"0.6445",
"0.0866",
"0.6445"
],
[
"0.513",
"0.64",
"0.914",
"0.64",
"0.914",
"0.654",
"0.513",
"0.654"
],
[
"0.511",
"0.654",
"0.914",
"0.654",
"0.914",
"0.6675",
"0.511",
"0.6675"
],
[
"0.10297",
"0.66",
"0.4885",
"0.66",
"0.4885",
"0.674",
"0.10297",
"0.674"
],
[
"0.511",
"0.6665",
"0.914",
"0.6675",
"0.914",
"0.6816",
"0.511",
"0.681"
],
[
"0.08496",
"0.675",
"0.3284",
"0.675",
"0.3284",
"0.688",
"0.08496",
"0.688"
],
[
"0.5107",
"0.681",
"0.5825",
"0.6787",
"0.5825",
"0.6934",
"0.5107",
"0.696"
],
[
"0.2018",
"0.6914",
"0.3652",
"0.6914",
"0.3652",
"0.7056",
"0.2018",
"0.7056"
],
[
"0.4575",
"0.6934",
"0.4878",
"0.6934",
"0.4878",
"0.7056",
"0.4575",
"0.7056"
],
[
"0.08496",
"0.709",
"0.4878",
"0.7104",
"0.4878",
"0.7246",
"0.08496",
"0.7236"
],
[
"0.513",
"0.7104",
"0.7036",
"0.7104",
"0.7036",
"0.7246",
"0.513",
"0.7246"
],
[
"0.0866",
"0.726",
"0.156",
"0.726",
"0.156",
"0.7373",
"0.0866",
"0.7373"
],
[
"0.511",
"0.732",
"0.5884",
"0.7344",
"0.5884",
"0.7485",
"0.511",
"0.747"
],
[
"0.1013",
"0.738",
"0.4885",
"0.7373",
"0.4885",
"0.752",
"0.1013",
"0.7524"
],
[
"0.0866",
"0.7524",
"0.361",
"0.7524",
"0.361",
"0.7666",
"0.0866",
"0.7666"
],
[
"0.511",
"0.75",
"0.914",
"0.7505",
"0.914",
"0.765",
"0.511",
"0.7646"
],
[
"0.08496",
"0.768",
"0.4878",
"0.7695",
"0.4878",
"0.784",
"0.08496",
"0.7827"
],
[
"0.513",
"0.765",
"0.911",
"0.765",
"0.911",
"0.779",
"0.513",
"0.779"
],
[
"0.08496",
"0.784",
"0.4878",
"0.784",
"0.4878",
"0.7974",
"0.08496",
"0.7974"
],
[
"0.511",
"0.779",
"0.911",
"0.778",
"0.911",
"0.7915",
"0.511",
"0.7925"
],
[
"0.08496",
"0.799",
"0.4878",
"0.799",
"0.4878",
"0.8125",
"0.08496",
"0.8125"
],
[
"0.5146",
"0.7944",
"0.911",
"0.7944",
"0.911",
"0.8047",
"0.5146",
"0.8047"
],
[
"0.08496",
"0.8125",
"0.4592",
"0.8125",
"0.4592",
"0.8257",
"0.08496",
"0.8257"
],
[
"0.513",
"0.8066",
"0.911",
"0.8066",
"0.911",
"0.821",
"0.513",
"0.821"
],
[
"0.513",
"0.821",
"0.914",
"0.821",
"0.914",
"0.834",
"0.513",
"0.834"
],
[
"0.2133",
"0.829",
"0.3586",
"0.83",
"0.3586",
"0.8447",
"0.2133",
"0.8433"
],
[
"0.4558",
"0.8306",
"0.486",
"0.8306",
"0.486",
"0.8433",
"0.4558",
"0.8433"
],
[
"0.513",
"0.8354",
"0.914",
"0.8354",
"0.914",
"0.8486",
"0.513",
"0.8486"
],
[
"0.0866",
"0.8477",
"0.486",
"0.8477",
"0.486",
"0.8613",
"0.0866",
"0.8613"
],
[
"0.511",
"0.8486",
"0.728",
"0.8486",
"0.728",
"0.8623",
"0.511",
"0.8623"
],
[
"0.08496",
"0.8623",
"0.4878",
"0.8623",
"0.4878",
"0.8755",
"0.08496",
"0.8755"
],
[
"0.08496",
"0.8755",
"0.3733",
"0.8755",
"0.3733",
"0.8896",
"0.08496",
"0.8896"
],
[
"0.5337",
"0.8755",
"0.903",
"0.8755",
"0.903",
"0.8896",
"0.5337",
"0.8896"
]
] | [
"context, rather than reconstructing the entire input. For the",
"Algorithm 1 Implementation of “Mask and Infill\" approach.",
"task of sentiment transfer, infilled words or phrases should",
"1: Pre-train attention-based classifier Cls (Eq.2-5)",
"be consistent to a target attribute. Therefore, we propose At-",
"2: Construct attribute marker vocabulary V (Eq.1,6)",
"tribute Conditional Masked Language Model (AC-MLM) to",
"3: for every sentence S in D do",
"infillmaskedpositionsconditionedonboththeircontextand",
"4:",
"Mask attribute markers within S by looking up V, getting S",
"a target label, by adding attribute embeddings into original",
"5:",
"if S is too short or S is the same as S then",
"input embeddings. Let a E A be a target attribute, S",
"6:",
"Re-mask with attention weights calculated by Cls (Eq.3)",
"7:",
"endif",
"8: end for",
"(ti,t2, ..,tn) a sentence ofNtokens, M = {tii,.., tim} the",
"set of masked tokens, $ = S M the set of content tokens.",
"9: for each iteration i=1,2....M do",
"10:",
"Sample a masked sentence S with attribute a",
"11:",
"Reconstruction",
"Lacking parallel sentiment sentence pairs, so we train the",
"Reconstruct S with S and a, calculating Cree based (Eq.7)",
"12:",
"a = the target attribute",
"Construct S (Eq.8)",
"AC-MLM to reconstruct the original sentence S conditioned",
"on the content S and original attribute a. The AC-MLM is",
"13:",
"14:",
"calculating Cace (Eq.9)",
"Update model parameters ",
"optimized to minimize the reconstruction error of original",
"masked words:",
"15:",
"16:end for",
"Crec=",
"logp(ti/s, a)",
"(7)",
"aeA,tiEM",
"M",
"Experiment",
"tent S and a target attribute a as input, and output p(Ja, $)",
"Well-trained AC-MLM model can take the sentence con-",
"In this section, we evaluate our method on two review datasets",
"for each masked position. After all masked positions are in-",
"filled, we get the transferred sentence:",
"Yelp and Amazon. Firstly, we introduce datasets, baseline",
"models, training details, and evaluation metrics. Secondly,",
"S = AC-MLM(S,a)",
"(8)",
"we compare our approach to several state-of-the-art systems.",
"Specifically, we use the pre-trained BERT as our MLM,",
"Finally, we present our experimental results and analyze each",
"and substitute segmentation embeddings (which is useless",
"component in detail.",
"when training single sentence) with attribute embeddings.",
"4.1",
"Datasets",
"Pre-trained Classifier Constraint",
"To further improve the transfer accuracy, we introduce a pre-",
"We experiment our methods on two review datasets from [Li",
"trained sentiment classifier to the AC-MLM, i.e. AC-MLM-",
"et al., 2018]: Yelp and Amazon [He and McAuley, 2016],",
"SS, \"-Ss\" indicates with soft-sampling which will be intro-",
"each of which is randomly split into training, validation and",
"duced later.",
"testing sets. Examples in YELP are sentences from business",
"Discrimination Given the content S and a target attribute",
"reviews on Yelp, and examples in AMAZON are sentences",
"a, when a sentence is successfully transferred, the label pre-",
"from product reviews on Amazon. Each example is labeled",
"dicted by classifier should be consistent to the target attribute.",
"as having either positive or negative sentiment. The statistics",
"of datasets are shown in Table 1.",
"SotheAC-MLM-SSisfurther optimizedtominimizethedis-",
"crimination discrepancy:",
"4.2",
"Baselines",
"Lace = -logp(alS)",
"(9)",
"We compared our methods to existing relevant models:",
"where S indicates transferred sentence generated by AC-",
"CrossAligned [Shen et al., 2017], StyleEmbedding and",
"MLM.",
"MultiDecoder [Fu et al., 2018], CycledReinforce [Xu et",
"al., 2018J4, TemplateBased, RetrievalOnly, DeleteOnly and",
"By combining Equation 7 and Equation 9 we obtain the",
"DeleteAndRetrieval [Li et al., 2018], LM+Classifer [Yang et",
"objective function of AC-MLM-SS :",
"al., 2018].",
"mineC=Lrec+nLaco",
"(10)",
"where is AC-MLM-SS's parameters and n is a balancing",
"4.3Experiment Details",
"parameter.",
"Mask step",
"To propagate gradients from the discriminator through the",
"discrete samples, we adopt soft-sampling.",
"For the frequency-ratio method, we consider n-gram up to",
"Soft-sampling Soft-sampling back-propagates gradients",
"4-gram and set the smoothing parameter ^ to 1. Other hyper-",
"by using an approximation of the sampled word vector. The",
"parameters are following [Li et al.,2018].For the attention-",
"approximation replaces the sampled token t; (represented as",
"based method,wetraintheattention-based classifierfor 10",
"a one-hot vector) at each step with the probability vector",
"epochs with accuracy 97% for Yelp and 82% for Amazon.",
"For the fusion-method, we consider a sentence with content",
"t ~ softmar(ot/1)",
"(11)",
"less than 5 tokens as insufficient content, and re-process it",
"which is differentiable w.r.t the AC-MLM's parameters. The",
"with the attention-based method.",
"resulting soft generated sentence is fed into the discriminator",
"to measure the fitness to the target attribute.",
"\"https://github.com/lancopku/Unpaired-Sentiment-Translation"
] | {
"arxiv_id": "1908.08039",
"language": "en",
"page_idx": 3
} |
||
1908.08039-en-4 | 1908.08039-en-4.jpg | <table>
<tbody>
<tr>
<th>Dataset</th>
<th>Attributes</th>
<th>Train</th>
<th>Dev</th>
<th>Test</th>
</tr>
<tr>
<td rowspan="2">$\mathbf{YELP}$</td>
<td>Positive</td>
<td>$270K$</td>
<td>2000</td>
<td>500</td>
</tr>
<tr>
<td>Negative</td>
<td>$180K$</td>
<td>2000</td>
<td>500</td>
</tr>
<tr>
<td rowspan="2">AMAZON</td>
<td>Positive</td>
<td>$277K$</td>
<td>985</td>
<td>500</td>
</tr>
<tr>
<td>Negative</td>
<td>$278K$</td>
<td>1015</td>
<td>500</td>
</tr>
</tbody>
</table>
Table 1: Dataset statistics.
# Infill step We use pre-trained BERT$_base$ as MLM and substitute seg-
ment embedding layer with attribute embedding layer. The input size is kept compatible with original BERT and rele- $\hat{\text{vant hyperparameters can be found in[Devlin et al., 2018]}}.$ The pre-trained discriminator is a CNN-based classifier[Kim, $201\bar{4}]$ with convolutional filters of size 3,4,5 and use WordPiece embeddings. The hyperparameters in Equation 10 and 11 are selected by a grid-search method using the validation set. We fine-tune BERT to AC-MLM for 10 epochs, and further train 6 epochs to apply discriminator constraint.
## 4.4 Evaluation
Automatic Evaluation
We compute automatic evaluation metrics by employing au-
tomatic evaluation tools from [Li et al., 2018] and[Yang $et$ $al.,2018].$ Accuracy score assesses whether outputs have the desired attribute. BLEU score is computed between the outputs and the human references. A high BLEU score primar-
${^{5}\text{https:}//\mathrm{github.com/lijuncen/Sentiment-and-Style-Transfer}}$
$^{6}[$Yang $etal.,2018]$ only evaluated their models on Yelp. For fair
comparison, we only evaluate our models on Yelp too.
<table>
<tbody>
<tr>
<th> </th>
<th colspan="2">$\mathbf{YELP}$</th>
<th colspan="2">AMAZON</th>
</tr>
<tr>
<th> </th>
<th>$\operatorname{ACC}\left(\%\right)$</th>
<th>BLEU</th>
<th>$\operatorname{ACC}(\%)$</th>
<th>BLEU</th>
</tr>
<tr>
<td>CrossAligned</td>
<td>73.7</td>
<td>3.1</td>
<td>74.1</td>
<td>0.4</td>
</tr>
<tr>
<td>StyleEmbedding</td>
<td>8.7</td>
<td>11.8</td>
<td>43.3</td>
<td>10.0</td>
</tr>
<tr>
<td>MultiDecoder</td>
<td>47.6</td>
<td>7.1</td>
<td>68.3</td>
<td>5.0</td>
</tr>
<tr>
<td>CycledReinforce</td>
<td>85.2</td>
<td>9.9</td>
<td>77.3</td>
<td>0.1</td>
</tr>
<tr>
<td>TemplateBased</td>
<td>81.7</td>
<td>11.8</td>
<td>68.7</td>
<td>27.1</td>
</tr>
<tr>
<td>RetrievalOnly</td>
<td>95.4</td>
<td>0.4</td>
<td>70.3</td>
<td>0.9</td>
</tr>
<tr>
<td>DeleteOnly</td>
<td>85.7</td>
<td>7.5</td>
<td>45.6</td>
<td>24.6</td>
</tr>
<tr>
<td>DeleteAndRetrieval</td>
<td>88.7</td>
<td>8.4</td>
<td>48.0</td>
<td>22.8</td>
</tr>
<tr>
<td>w/frequency-ratio</td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
</tr>
<tr>
<td>$\mathbf{AC-MLM}$</td>
<td>55.0</td>
<td>12.7</td>
<td>28.7</td>
<td>31.0</td>
</tr>
<tr>
<td>AC- MLM- SS</td>
<td>90.5</td>
<td>11.6</td>
<td>75.7</td>
<td>26.0</td>
</tr>
<tr>
<td>w/attention-based</td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
</tr>
<tr>
<td>$\mathbf{AC-MLM}$</td>
<td>41.5</td>
<td>15.9</td>
<td>31.2</td>
<td>32.1</td>
</tr>
<tr>
<td>AC- MLM- SS</td>
<td>97.3</td>
<td>14.1</td>
<td>75.9</td>
<td>28.5</td>
</tr>
<tr>
<td>w/fusion-method</td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
</tr>
<tr>
<td>$\mathbf{AC-MLM}$</td>
<td>43.5</td>
<td>15.3</td>
<td>42.9</td>
<td>30.7</td>
</tr>
<tr>
<td>AC- MLM- SS</td>
<td>97.3</td>
<td>14.4</td>
<td>84.5</td>
<td>28.5</td>
</tr>
</tbody>
</table>
Table 2: Automatic evaluation performed by tools from [Li et al., $2018]^5.“$ACC” indicates accuracy,“BLEU”measures content similarity between the outputs and the human references.“AC-MLM”, represents attribute conditional masked language model.“w" represents“with”:“-SS” represents with soft-sampling.
<table>
<tbody>
<tr>
<th> </th>
<th>$\operatorname{ACC}\left(\%\right)$</th>
<th>$\mathbf{BLEU}$</th>
</tr>
<tr>
<td>Yang's results</td>
<td> </td>
<td> </td>
</tr>
<tr>
<td>$LM$</td>
<td>85.4</td>
<td>13.4</td>
</tr>
<tr>
<td>LM+Classifier</td>
<td>90.0</td>
<td>22.3</td>
</tr>
<tr>
<td>w/frequency-ratio</td>
<td> </td>
<td> </td>
</tr>
<tr>
<td>$\mathbf{AC-MLM}$</td>
<td>58.0</td>
<td>18.7</td>
</tr>
<tr>
<td>AC- MLM- SS</td>
<td>93.7</td>
<td>17.5</td>
</tr>
<tr>
<td>w/attention-based</td>
<td> </td>
<td> </td>
</tr>
<tr>
<td>$\mathbf{AC-MLM}$</td>
<td>40.0</td>
<td>21.8</td>
</tr>
<tr>
<td>AC- MLM- SS</td>
<td>98.5</td>
<td>20.5</td>
</tr>
<tr>
<td>w/fusion-method</td>
<td> </td>
<td> </td>
</tr>
<tr>
<td>$\mathbf{AC-MLM}$</td>
<td>44.1</td>
<td>21.3</td>
</tr>
<tr>
<td>AC- MLM- SS</td>
<td>97.3</td>
<td>20.7</td>
</tr>
</tbody>
</table>
Table 3: Automatic evaluation on Yelp dataset performed by tools
from [Yang et al., 2018]$^6.$
ily indicates that the system can correctly preserve content by retaining the same words from the source sentence as the reference.
Table 2 shows the performance obtained with [Li et al., 2018]'s tools. Our base model AC-MLM achieves all the best BLEU, but performs poorly in accuracy. After applying the discriminator constraint to apply constraint (i.e. ACMLM-SS), the accuracy improves significantly, though with the slight decline of BLEU. Previous models are hard to perform well on accuracy and BLEU simultaneously. Among our models, AC-MLM-SS using fusion-methods in mask step achieves the most satisfactory performance considering on both accuracy and BLEU. Table 3 shows the performance on [Yang et al., 2018]'s tools. Compared to Yang's best model LM+Classifer, we perform better on accuracy, but slightly lower on BLEU.
## Human Evaluation We hired three annotators (not authors) to rate outputs for
our models and[Li et al., 2018]'s best model(DeleteAndRetrieval). We adopt three criteria range from 1 to 5 (1 is very bad and 5 is very good): grammaticality, similarity to the target attribute, and preservation of the source content. For each dataset, we randomly sample 200 examples for each target attribute. Table 5 shows the human evaluation results. Among them, our AC-MLM-SS with the fusion-method achieves best results. Pre-trained MLM infills the masked positions considering the context from both directions, being able to meet grammatical requirements better. Benefiting from explicitly separating content and style, our sentences' structures are kept and we perform better on the preservation of the source content. Then by introducing the pre-trained discriminator, MLM infills more accurate attribute words.
## 5 $\mathbf{Analysis}$
Trade-off between Content and Attribute Our loss function in Equation 10 consists of two parts: reconstruction loss and discriminative loss. Reconstruction loss guides MLM to infill contextual-compatible words. Compared to weak labelconstraint introduced by attribute embeddings, discriminative | [
[
"0.2239",
"0.0682",
"0.2893",
"0.0707",
"0.2893",
"0.0852",
"0.2239",
"0.0833"
],
[
"0.789",
"0.06946",
"0.835",
"0.06946",
"0.835",
"0.084",
"0.789",
"0.084"
],
[
"0.4019",
"0.0707",
"0.4355",
"0.0707",
"0.4355",
"0.0859",
"0.4019",
"0.0859"
],
[
"0.1503",
"0.0707",
"0.2009",
"0.0707",
"0.2009",
"0.0852",
"0.1503",
"0.0852"
],
[
"0.3022",
"0.07135",
"0.3406",
"0.07135",
"0.3406",
"0.0852",
"0.3022",
"0.0852"
],
[
"0.357",
"0.07135",
"0.3865",
"0.07135",
"0.3865",
"0.084",
"0.357",
"0.084"
],
[
"0.7173",
"0.0707",
"0.7793",
"0.0707",
"0.7793",
"0.0852",
"0.7173",
"0.0852"
],
[
"0.3015",
"0.0884",
"0.3423",
"0.0884",
"0.3423",
"0.1042",
"0.3015",
"0.1042"
],
[
"0.353",
"0.0884",
"0.3906",
"0.0884",
"0.3906",
"0.1042",
"0.353",
"0.1042"
],
[
"0.4028",
"0.0884",
"0.4346",
"0.0884",
"0.4346",
"0.1042",
"0.4028",
"0.1042"
],
[
"0.2279",
"0.089",
"0.282",
"0.089",
"0.282",
"0.1042",
"0.2279",
"0.1042"
],
[
"0.1528",
"0.089",
"0.1953",
"0.0871",
"0.1969",
"0.1029",
"0.1544",
"0.1048"
],
[
"0.5923",
"0.089",
"0.6787",
"0.089",
"0.6787",
"0.1042",
"0.5923",
"0.1042"
],
[
"0.353",
"0.09973",
"0.3906",
"0.1016",
"0.3896",
"0.1187",
"0.3513",
"0.11615"
],
[
"0.8022",
"0.101",
"0.835",
"0.101",
"0.835",
"0.11743",
"0.8022",
"0.11743"
],
[
"0.3022",
"0.1016",
"0.3406",
"0.1016",
"0.3406",
"0.11743",
"0.3022",
"0.11743"
],
[
"0.2263",
"0.1016",
"0.2852",
"0.1016",
"0.2852",
"0.11743",
"0.2263",
"0.11743"
],
[
"0.4043",
"0.1029",
"0.433",
"0.1029",
"0.433",
"0.11743",
"0.4043",
"0.11743"
],
[
"0.591",
"0.1029",
"0.621",
"0.1029",
"0.621",
"0.11554",
"0.591",
"0.11554"
],
[
"0.744",
"0.1029",
"0.777",
"0.1029",
"0.777",
"0.11554",
"0.744",
"0.11554"
],
[
"0.744",
"0.11426",
"0.778",
"0.11426",
"0.778",
"0.13",
"0.744",
"0.13"
],
[
"0.8013",
"0.113",
"0.835",
"0.113",
"0.835",
"0.13",
"0.8013",
"0.13"
],
[
"0.3015",
"0.1193",
"0.3406",
"0.1193",
"0.3406",
"0.1351",
"0.3015",
"0.1351"
],
[
"0.5933",
"0.11615",
"0.6807",
"0.11615",
"0.6807",
"0.1276",
"0.5933",
"0.1276"
],
[
"0.2296",
"0.1206",
"0.282",
"0.1206",
"0.282",
"0.1351",
"0.2296",
"0.1351"
],
[
"0.1405",
"0.1206",
"0.21",
"0.1206",
"0.21",
"0.1351",
"0.1405",
"0.1351"
],
[
"0.357",
"0.1218",
"0.3857",
"0.1218",
"0.3857",
"0.1338",
"0.357",
"0.1338"
],
[
"0.4028",
"0.1206",
"0.4346",
"0.1206",
"0.4346",
"0.1364",
"0.4028",
"0.1364"
],
[
"0.2263",
"0.1332",
"0.2852",
"0.1332",
"0.2852",
"0.1477",
"0.2263",
"0.1477"
],
[
"0.3545",
"0.1332",
"0.3896",
"0.1332",
"0.3896",
"0.1483",
"0.3545",
"0.1483"
],
[
"0.4028",
"0.1332",
"0.4346",
"0.1332",
"0.4346",
"0.1483",
"0.4028",
"0.1483"
],
[
"0.591",
"0.1338",
"0.7017",
"0.1338",
"0.7017",
"0.1477",
"0.591",
"0.1477"
],
[
"0.744",
"0.1459",
"0.778",
"0.1459",
"0.778",
"0.161",
"0.744",
"0.161"
],
[
"0.3022",
"0.1338",
"0.3406",
"0.1338",
"0.3406",
"0.1477",
"0.3022",
"0.1477"
],
[
"0.5933",
"0.1477",
"0.6543",
"0.1477",
"0.6543",
"0.159",
"0.5933",
"0.159"
],
[
"0.8022",
"0.1459",
"0.835",
"0.1459",
"0.835",
"0.161",
"0.8022",
"0.161"
],
[
"0.7427",
"0.1572",
"0.7793",
"0.1604",
"0.777",
"0.1749",
"0.7417",
"0.1724"
],
[
"0.8022",
"0.1572",
"0.839",
"0.1604",
"0.8364",
"0.1749",
"0.8013",
"0.1724"
],
[
"0.5933",
"0.1604",
"0.6763",
"0.1604",
"0.6763",
"0.1718",
"0.5933",
"0.1718"
],
[
"0.2075",
"0.1622",
"0.3635",
"0.1622",
"0.3635",
"0.1755",
"0.2075",
"0.1755"
],
[
"0.5933",
"0.1799",
"0.699",
"0.1799",
"0.699",
"0.1901",
"0.5933",
"0.1901"
],
[
"0.746",
"0.1913",
"0.777",
"0.1913",
"0.777",
"0.2039",
"0.746",
"0.2039"
],
[
"0.8013",
"0.1895",
"0.835",
"0.1895",
"0.835",
"0.2052",
"0.8013",
"0.2052"
],
[
"0.08417",
"0.1963",
"0.156",
"0.1989",
"0.156",
"0.2134",
"0.08417",
"0.2108"
],
[
"0.5933",
"0.1925",
"0.6562",
"0.1925",
"0.6562",
"0.2039",
"0.5933",
"0.2039"
],
[
"0.7427",
"0.2008",
"0.7793",
"0.2026",
"0.777",
"0.2184",
"0.7417",
"0.2153"
],
[
"0.8013",
"0.202",
"0.835",
"0.202",
"0.835",
"0.2172",
"0.8013",
"0.2172"
],
[
"0.0825",
"0.2108",
"0.4878",
"0.2134",
"0.4878",
"0.2311",
"0.0825",
"0.2291"
],
[
"0.5933",
"0.2039",
"0.6763",
"0.2039",
"0.6763",
"0.2153",
"0.5933",
"0.2153"
],
[
"0.08496",
"0.2279",
"0.4878",
"0.2279",
"0.4878",
"0.2412",
"0.08496",
"0.2412"
],
[
"0.5923",
"0.2229",
"0.697",
"0.2229",
"0.697",
"0.2349",
"0.5923",
"0.2349"
],
[
"0.7427",
"0.2317",
"0.7793",
"0.2349",
"0.777",
"0.25",
"0.7417",
"0.2474"
],
[
"0.8013",
"0.2329",
"0.8364",
"0.2329",
"0.8364",
"0.2488",
"0.8013",
"0.2488"
],
[
"0.08496",
"0.2424",
"0.486",
"0.2424",
"0.486",
"0.2556",
"0.08496",
"0.2556"
],
[
"0.5933",
"0.2349",
"0.6543",
"0.2349",
"0.6543",
"0.2468",
"0.5933",
"0.2468"
],
[
"0.7427",
"0.2456",
"0.778",
"0.2456",
"0.778",
"0.2615",
"0.7427",
"0.2615"
],
[
"0.8013",
"0.2443",
"0.8364",
"0.2474",
"0.834",
"0.2632",
"0.8",
"0.26"
],
[
"0.08417",
"0.2556",
"0.4844",
"0.2551",
"0.4844",
"0.2683",
"0.08417",
"0.2695"
],
[
"0.5933",
"0.2474",
"0.6772",
"0.2474",
"0.6772",
"0.2588",
"0.5933",
"0.2588"
],
[
"0.0866",
"0.2695",
"0.486",
"0.2695",
"0.486",
"0.2827",
"0.0866",
"0.2827"
],
[
"0.513",
"0.276",
"0.9126",
"0.276",
"0.9126",
"0.289",
"0.513",
"0.289"
],
[
"0.08826",
"0.2854",
"0.4844",
"0.2854",
"0.4844",
"0.2954",
"0.08826",
"0.2954"
],
[
"0.511",
"0.2886",
"0.6665",
"0.2886",
"0.6665",
"0.3018",
"0.511",
"0.3018"
],
[
"0.0866",
"0.2974",
"0.4878",
"0.2974",
"0.4878",
"0.3113",
"0.0866",
"0.3113"
],
[
"0.0866",
"0.3113",
"0.486",
"0.3113",
"0.486",
"0.3245",
"0.0866",
"0.3245"
],
[
"0.08496",
"0.3245",
"0.4878",
"0.3245",
"0.4878",
"0.3384",
"0.08496",
"0.3384"
],
[
"0.5107",
"0.3208",
"0.9126",
"0.3213",
"0.9126",
"0.336",
"0.5107",
"0.3352"
],
[
"0.08417",
"0.3384",
"0.4272",
"0.3384",
"0.4272",
"0.3516",
"0.08417",
"0.3516"
],
[
"0.513",
"0.336",
"0.9126",
"0.336",
"0.9126",
"0.3499",
"0.513",
"0.3499"
],
[
"0.513",
"0.351",
"0.5786",
"0.351",
"0.5786",
"0.3623",
"0.513",
"0.3623"
],
[
"0.08496",
"0.3623",
"0.2115",
"0.3623",
"0.2115",
"0.377",
"0.08496",
"0.377"
],
[
"0.5293",
"0.363",
"0.9126",
"0.363",
"0.9126",
"0.377",
"0.5293",
"0.377"
],
[
"0.513",
"0.3774",
"0.6177",
"0.3774",
"0.6177",
"0.3901",
"0.513",
"0.3901"
],
[
"0.6094",
"0.3774",
"0.9126",
"0.3774",
"0.9126",
"0.3914",
"0.6094",
"0.3914"
],
[
"0.08826",
"0.3838",
"0.2418",
"0.3838",
"0.2418",
"0.3945",
"0.08826",
"0.3945"
],
[
"0.513",
"0.3914",
"0.9126",
"0.3914",
"0.9126",
"0.4048",
"0.513",
"0.4048"
],
[
"0.08496",
"0.3977",
"0.486",
"0.3977",
"0.486",
"0.411",
"0.08496",
"0.411"
],
[
"0.08417",
"0.4104",
"0.4878",
"0.411",
"0.4878",
"0.4255",
"0.08417",
"0.4248"
],
[
"0.5107",
"0.406",
"0.91",
"0.406",
"0.91",
"0.4192",
"0.5107",
"0.4192"
],
[
"0.511",
"0.4192",
"0.9126",
"0.4192",
"0.9126",
"0.433",
"0.511",
"0.433"
],
[
"0.08496",
"0.4255",
"0.4878",
"0.4255",
"0.4878",
"0.4395",
"0.08496",
"0.4395"
],
[
"0.513",
"0.433",
"0.9126",
"0.433",
"0.9126",
"0.4463",
"0.513",
"0.4463"
],
[
"0.08496",
"0.4395",
"0.486",
"0.4395",
"0.486",
"0.4526",
"0.08496",
"0.4526"
],
[
"0.08496",
"0.4526",
"0.486",
"0.4526",
"0.486",
"0.4673",
"0.08496",
"0.4673"
],
[
"0.513",
"0.4475",
"0.9126",
"0.4475",
"0.9126",
"0.461",
"0.513",
"0.461"
],
[
"0.513",
"0.461",
"0.9126",
"0.461",
"0.9126",
"0.4749",
"0.513",
"0.4749"
],
[
"0.10706",
"0.4778",
"0.452",
"0.4778",
"0.452",
"0.4912",
"0.10706",
"0.4912"
],
[
"0.511",
"0.4749",
"0.914",
"0.4749",
"0.914",
"0.488",
"0.511",
"0.488"
],
[
"0.10706",
"0.4912",
"0.4878",
"0.4912",
"0.4878",
"0.505",
"0.10706",
"0.505"
],
[
"0.511",
"0.488",
"0.914",
"0.488",
"0.914",
"0.502",
"0.511",
"0.502"
],
[
"0.511",
"0.502",
"0.738",
"0.502",
"0.738",
"0.515",
"0.511",
"0.515"
],
[
"0.7344",
"0.502",
"0.914",
"0.502",
"0.914",
"0.515",
"0.7344",
"0.515"
],
[
"0.511",
"0.5166",
"0.914",
"0.5166",
"0.914",
"0.53",
"0.511",
"0.53"
],
[
"0.0866",
"0.506",
"0.4043",
"0.506",
"0.4043",
"0.5166",
"0.0866",
"0.5166"
],
[
"0.513",
"0.53",
"0.6772",
"0.53",
"0.6772",
"0.5435",
"0.513",
"0.5435"
],
[
"0.2688",
"0.5405",
"0.312",
"0.5405",
"0.312",
"0.556",
"0.2688",
"0.556"
],
[
"0.3882",
"0.5425",
"0.4534",
"0.5425",
"0.4534",
"0.5537",
"0.3882",
"0.5537"
],
[
"0.232",
"0.5537",
"0.2932",
"0.5537",
"0.2932",
"0.5684",
"0.232",
"0.5684"
],
[
"0.3057",
"0.5537",
"0.3489",
"0.5537",
"0.3489",
"0.5684",
"0.3057",
"0.5684"
],
[
"0.3635",
"0.5537",
"0.4224",
"0.5537",
"0.4224",
"0.5684",
"0.3635",
"0.5684"
],
[
"0.433",
"0.5537",
"0.478",
"0.5537",
"0.478",
"0.5684",
"0.433",
"0.5684"
],
[
"0.513",
"0.5503",
"0.648",
"0.5503",
"0.648",
"0.5645",
"0.513",
"0.5645"
],
[
"0.513",
"0.565",
"0.914",
"0.565",
"0.914",
"0.579",
"0.513",
"0.579"
],
[
"0.259",
"0.5713",
"0.2932",
"0.5713",
"0.2932",
"0.5884",
"0.259",
"0.5884"
],
[
"0.326",
"0.5728",
"0.3489",
"0.5728",
"0.3489",
"0.587",
"0.326",
"0.587"
],
[
"0.3882",
"0.5713",
"0.4207",
"0.5713",
"0.4207",
"0.5884",
"0.3882",
"0.5884"
],
[
"0.455",
"0.5728",
"0.4795",
"0.5728",
"0.4795",
"0.586",
"0.455",
"0.586"
],
[
"0.511",
"0.5776",
"0.91",
"0.5776",
"0.91",
"0.5913",
"0.511",
"0.5913"
],
[
"0.09314",
"0.5835",
"0.1969",
"0.586",
"0.1969",
"0.601",
"0.09314",
"0.5977"
],
[
"0.3179",
"0.5854",
"0.3506",
"0.5854",
"0.3506",
"0.601",
"0.3179",
"0.601"
],
[
"0.3882",
"0.5854",
"0.4207",
"0.5854",
"0.4207",
"0.601",
"0.3882",
"0.601"
],
[
"0.2646",
"0.586",
"0.2932",
"0.586",
"0.2932",
"0.601",
"0.2646",
"0.601"
],
[
"0.0972",
"0.5747",
"0.1805",
"0.5747",
"0.1805",
"0.586",
"0.0972",
"0.586"
],
[
"0.4478",
"0.586",
"0.478",
"0.586",
"0.478",
"0.5986",
"0.4478",
"0.5986"
],
[
"0.5107",
"0.5913",
"0.911",
"0.5923",
"0.911",
"0.607",
"0.5107",
"0.606"
],
[
"0.0948",
"0.5977",
"0.1805",
"0.5977",
"0.1805",
"0.6094",
"0.0948",
"0.6094"
],
[
"0.2607",
"0.5986",
"0.2917",
"0.5986",
"0.2917",
"0.6113",
"0.2607",
"0.6113"
],
[
"0.3228",
"0.5986",
"0.3489",
"0.5986",
"0.3489",
"0.6123",
"0.3228",
"0.6123"
],
[
"0.3896",
"0.5986",
"0.4192",
"0.5986",
"0.4192",
"0.6113",
"0.3896",
"0.6113"
],
[
"0.452",
"0.5986",
"0.4795",
"0.5986",
"0.4795",
"0.6113",
"0.452",
"0.6113"
],
[
"0.0948",
"0.6094",
"0.1978",
"0.6113",
"0.1978",
"0.625",
"0.0948",
"0.6226"
],
[
"0.259",
"0.61",
"0.2932",
"0.61",
"0.2932",
"0.6255",
"0.259",
"0.6255"
],
[
"0.3245",
"0.61",
"0.3489",
"0.61",
"0.3489",
"0.6235",
"0.3245",
"0.6235"
],
[
"0.3882",
"0.61",
"0.4207",
"0.61",
"0.4207",
"0.6255",
"0.3882",
"0.6255"
],
[
"0.4534",
"0.6113",
"0.478",
"0.6113",
"0.478",
"0.6235",
"0.4534",
"0.6235"
],
[
"0.511",
"0.607",
"0.9126",
"0.607",
"0.9126",
"0.6206",
"0.511",
"0.6206"
],
[
"0.3179",
"0.6226",
"0.3506",
"0.6226",
"0.3506",
"0.638",
"0.3179",
"0.638"
],
[
"0.3882",
"0.6226",
"0.4207",
"0.6226",
"0.4207",
"0.638",
"0.3882",
"0.638"
],
[
"0.5107",
"0.619",
"0.91",
"0.619",
"0.91",
"0.6333",
"0.5107",
"0.6333"
],
[
"0.4478",
"0.6235",
"0.4763",
"0.6235",
"0.4763",
"0.636",
"0.4478",
"0.636"
],
[
"0.09644",
"0.625",
"0.1896",
"0.625",
"0.1896",
"0.636",
"0.09644",
"0.636"
],
[
"0.2607",
"0.6235",
"0.29",
"0.6235",
"0.29",
"0.636",
"0.2607",
"0.636"
],
[
"0.259",
"0.6353",
"0.2932",
"0.6353",
"0.2932",
"0.651",
"0.259",
"0.651"
],
[
"0.3882",
"0.6353",
"0.4207",
"0.6353",
"0.4207",
"0.651",
"0.3882",
"0.651"
],
[
"0.513",
"0.634",
"0.914",
"0.634",
"0.914",
"0.6484",
"0.513",
"0.6484"
],
[
"0.452",
"0.636",
"0.478",
"0.636",
"0.478",
"0.6494",
"0.452",
"0.6494"
],
[
"0.09644",
"0.637",
"0.179",
"0.637",
"0.179",
"0.6484",
"0.09644",
"0.6484"
],
[
"0.3245",
"0.636",
"0.3489",
"0.636",
"0.3489",
"0.6494",
"0.3245",
"0.6494"
],
[
"0.259",
"0.648",
"0.2917",
"0.648",
"0.2917",
"0.6636",
"0.259",
"0.6636"
],
[
"0.3882",
"0.648",
"0.4207",
"0.648",
"0.4207",
"0.6636",
"0.3882",
"0.6636"
],
[
"0.4468",
"0.648",
"0.4795",
"0.648",
"0.4795",
"0.6636",
"0.4468",
"0.6636"
],
[
"0.511",
"0.6484",
"0.911",
"0.6484",
"0.911",
"0.662",
"0.511",
"0.662"
],
[
"0.326",
"0.6484",
"0.3489",
"0.6484",
"0.3489",
"0.6636",
"0.326",
"0.6636"
],
[
"0.09644",
"0.6494",
"0.1666",
"0.6494",
"0.1666",
"0.661",
"0.09644",
"0.661"
],
[
"0.259",
"0.66",
"0.2917",
"0.66",
"0.2917",
"0.6753",
"0.259",
"0.6753"
],
[
"0.3228",
"0.66",
"0.3513",
"0.66",
"0.3513",
"0.677",
"0.3228",
"0.677"
],
[
"0.3882",
"0.66",
"0.4207",
"0.66",
"0.4207",
"0.6753",
"0.3882",
"0.6753"
],
[
"0.4468",
"0.659",
"0.482",
"0.662",
"0.4795",
"0.678",
"0.4453",
"0.675"
],
[
"0.511",
"0.662",
"0.914",
"0.662",
"0.914",
"0.6753",
"0.511",
"0.6753"
],
[
"0.513",
"0.6753",
"0.9126",
"0.6753",
"0.9126",
"0.6895",
"0.513",
"0.6895"
],
[
"0.0948",
"0.661",
"0.2157",
"0.661",
"0.2157",
"0.675",
"0.0948",
"0.675"
],
[
"0.09314",
"0.6797",
"0.2051",
"0.6797",
"0.2051",
"0.6934",
"0.09314",
"0.6934"
],
[
"0.259",
"0.6914",
"0.2932",
"0.6914",
"0.2932",
"0.707",
"0.259",
"0.707"
],
[
"0.3179",
"0.6914",
"0.3506",
"0.6914",
"0.3506",
"0.707",
"0.3179",
"0.707"
],
[
"0.3882",
"0.6914",
"0.4207",
"0.6914",
"0.4207",
"0.707",
"0.3882",
"0.707"
],
[
"0.4468",
"0.6914",
"0.4805",
"0.6914",
"0.4805",
"0.707",
"0.4468",
"0.707"
],
[
"0.513",
"0.69",
"0.914",
"0.69",
"0.914",
"0.704",
"0.513",
"0.704"
],
[
"0.511",
"0.7026",
"0.9126",
"0.704",
"0.9126",
"0.7173",
"0.511",
"0.7163"
],
[
"0.0972",
"0.6934",
"0.1602",
"0.6934",
"0.1602",
"0.705",
"0.0972",
"0.705"
],
[
"0.259",
"0.704",
"0.2932",
"0.704",
"0.2932",
"0.7197",
"0.259",
"0.7197"
],
[
"0.3882",
"0.704",
"0.4207",
"0.704",
"0.4207",
"0.7197",
"0.3882",
"0.7197"
],
[
"0.4453",
"0.704",
"0.4795",
"0.704",
"0.4795",
"0.7197",
"0.4453",
"0.7197"
],
[
"0.3179",
"0.705",
"0.3513",
"0.705",
"0.3513",
"0.72",
"0.3179",
"0.72"
],
[
"0.0972",
"0.7056",
"0.1805",
"0.7056",
"0.1805",
"0.7173",
"0.0972",
"0.7173"
],
[
"0.513",
"0.7173",
"0.9126",
"0.7173",
"0.9126",
"0.731",
"0.513",
"0.731"
],
[
"0.0948",
"0.7246",
"0.2051",
"0.7246",
"0.2051",
"0.738",
"0.0948",
"0.738"
],
[
"0.3186",
"0.735",
"0.3506",
"0.735",
"0.3506",
"0.7505",
"0.3186",
"0.7505"
],
[
"0.4468",
"0.7363",
"0.4795",
"0.7363",
"0.4795",
"0.752",
"0.4468",
"0.752"
],
[
"0.513",
"0.733",
"0.9126",
"0.733",
"0.9126",
"0.7437",
"0.513",
"0.7437"
],
[
"0.3896",
"0.7373",
"0.4192",
"0.7373",
"0.4192",
"0.75",
"0.3896",
"0.75"
],
[
"0.0972",
"0.738",
"0.1602",
"0.738",
"0.1602",
"0.75",
"0.0972",
"0.75"
],
[
"0.2607",
"0.7373",
"0.2917",
"0.7373",
"0.2917",
"0.75",
"0.2607",
"0.75"
],
[
"0.259",
"0.7485",
"0.2917",
"0.7485",
"0.2917",
"0.7646",
"0.259",
"0.7646"
],
[
"0.3186",
"0.7485",
"0.3506",
"0.7485",
"0.3506",
"0.7646",
"0.3186",
"0.7646"
],
[
"0.3882",
"0.7485",
"0.4207",
"0.7485",
"0.4207",
"0.7646",
"0.3882",
"0.7646"
],
[
"0.4468",
"0.7485",
"0.4805",
"0.7485",
"0.4805",
"0.7646",
"0.4468",
"0.7646"
],
[
"0.513",
"0.7456",
"0.914",
"0.7456",
"0.914",
"0.759",
"0.513",
"0.759"
],
[
"0.511",
"0.759",
"0.9126",
"0.759",
"0.9126",
"0.773",
"0.511",
"0.773"
],
[
"0.09644",
"0.7505",
"0.179",
"0.7505",
"0.179",
"0.762",
"0.09644",
"0.762"
],
[
"0.09314",
"0.768",
"0.2009",
"0.768",
"0.2009",
"0.782",
"0.09314",
"0.782"
],
[
"0.513",
"0.7734",
"0.6055",
"0.7734",
"0.6055",
"0.785",
"0.513",
"0.785"
],
[
"0.259",
"0.78",
"0.2932",
"0.78",
"0.2932",
"0.7954",
"0.259",
"0.7954"
],
[
"0.3179",
"0.78",
"0.3506",
"0.78",
"0.3506",
"0.7954",
"0.3179",
"0.7954"
],
[
"0.3882",
"0.78",
"0.4207",
"0.78",
"0.4207",
"0.7954",
"0.3882",
"0.7954"
],
[
"0.4468",
"0.78",
"0.4795",
"0.78",
"0.4795",
"0.7954",
"0.4468",
"0.7954"
],
[
"0.2573",
"0.7905",
"0.2942",
"0.7935",
"0.2917",
"0.8096",
"0.2566",
"0.8066"
],
[
"0.09644",
"0.782",
"0.1602",
"0.782",
"0.1602",
"0.7935",
"0.09644",
"0.7935"
],
[
"0.3186",
"0.7925",
"0.3506",
"0.7925",
"0.3506",
"0.808",
"0.3186",
"0.808"
],
[
"0.3882",
"0.7925",
"0.4207",
"0.7925",
"0.4207",
"0.808",
"0.3882",
"0.808"
],
[
"0.4453",
"0.7925",
"0.4795",
"0.7925",
"0.4795",
"0.8096",
"0.4453",
"0.8096"
],
[
"0.539",
"0.8003",
"0.6167",
"0.8003",
"0.6167",
"0.815",
"0.539",
"0.815"
],
[
"0.09314",
"0.7935",
"0.1805",
"0.7915",
"0.1805",
"0.806",
"0.09314",
"0.808"
],
[
"0.5146",
"0.802",
"0.527",
"0.802",
"0.527",
"0.8125",
"0.5146",
"0.8125"
],
[
"0.08496",
"0.8213",
"0.486",
"0.8213",
"0.486",
"0.8354",
"0.08496",
"0.8354"
],
[
"0.5146",
"0.8213",
"0.8057",
"0.8213",
"0.8057",
"0.832",
"0.5146",
"0.832"
],
[
"0.8013",
"0.8228",
"0.8164",
"0.8228",
"0.8164",
"0.83",
"0.8013",
"0.83"
],
[
"0.8105",
"0.8213",
"0.9067",
"0.8213",
"0.9067",
"0.832",
"0.8105",
"0.832"
],
[
"0.08496",
"0.834",
"0.486",
"0.834",
"0.486",
"0.8477",
"0.08496",
"0.8477"
],
[
"0.513",
"0.834",
"0.914",
"0.834",
"0.914",
"0.8477",
"0.513",
"0.8477"
],
[
"0.08496",
"0.8467",
"0.4885",
"0.8467",
"0.4885",
"0.86",
"0.08496",
"0.86"
],
[
"0.511",
"0.8477",
"0.9126",
"0.8477",
"0.9126",
"0.8613",
"0.511",
"0.8613"
],
[
"0.08417",
"0.8594",
"0.4878",
"0.86",
"0.4878",
"0.8735",
"0.08417",
"0.8726"
],
[
"0.513",
"0.8623",
"0.911",
"0.8623",
"0.911",
"0.8755",
"0.513",
"0.8755"
],
[
"0.08496",
"0.8726",
"0.3896",
"0.8726",
"0.3896",
"0.886",
"0.08496",
"0.886"
],
[
"0.511",
"0.8755",
"0.914",
"0.8755",
"0.914",
"0.8896",
"0.511",
"0.8896"
]
] | [
"Attributes",
"BLEU",
"Test",
"Dataset",
"Train",
"Dev",
"ACC (%)",
"270K",
"2000",
"500",
"Positive",
"YELP",
"Yang's results",
"2000",
"13.4",
"180K",
"Negative",
"500",
"LM",
"85.4",
"90.0",
"22.3",
"277K",
"LM+Classifier",
"Positive",
"AMAZON",
"985",
"500",
"Negative",
"1015",
"500",
"w/frequency-ratio",
"58.0",
"278K",
"AC-MLM",
"18.7",
"93.7",
"17.5",
"AC-MLM-SS",
"Table 1: Dataset statistics.",
"w/attention-based",
"40.0",
"21.8",
"Infill step",
"AC-MLM",
"98.5",
"20.5",
"We use pre-trained BERTbase as MLM and substitute seg-",
"AC-MLM-SS",
"ment embedding layer with attribute embedding layer. The",
"w/fusion-method",
"44.1",
"21.3",
"input size is kept compatible with original BERT and rele-",
"AC-MLM",
"97.3",
"20.7",
"vant hyperparameters can be found in [Devlin et al., 2018].",
"AC-MLM-SS",
"The pre-trained discriminator is a CNN-based classifier [Kim,",
"Table 3: Automatic evaluation on Yelp dataset performed by tools",
"2014]withconvolutionalfiltersofsize3,4,5anduseWord-",
"from [Yang et al., 2018]°.",
"Piece embeddings. The hyperparameters in Equation 10 and",
"11 are selected by a grid-search method using the validation",
"set. We fine-tune BERT to AC-MLM for 10 epochs, and fur-",
"ily indicates that the system can correctly preserve content",
"ther train 6 epochs to apply discriminator constraint.",
"by retaining the same words from the source sentence as the",
"reference.",
"4.4 Evaluation",
"Table 2 shows the performance obtained with [Li et al.,",
"2018l's tools.",
"Our base model AC-MLM achieves all the",
"AutomaticEvaluation",
"best BLEU, but performs poorly in accuracy. After apply-",
"We compute automatic evaluation metrics by employing au-",
"tomatic evaluation tools from [Li et al., 20i8] and [Yang et",
"ing the discriminator constraint to apply constraint (i.e. AC-",
"MLM-SS), the accuracy improves significantly, though with",
"al., 20181. Accuracy score assesses whether outputs have the",
"the slight decline of BLEU. Previous models are hard to per-",
"desired attribute. BLEU score is computed between the out-",
"puts and the human references. A high BLEU score primar-",
"form well on accuracy and BLEU simultaneously. Among",
"our models, AC-MLM-SS using fusion-methods in mask step",
"https://github.com/lijuncen/Sentiment-and-Style-Transfer",
"achieves the most satisfactory performance considering on",
"6[Yang et al., 2018] only evaluated their models on Yelp. For fair",
"both accuracy and BLEU. Table 3 shows the performance",
"on [Yang et al., 2018]'s tools.",
"Compared to Yang's best",
"model LM+Classifier, we perform better on accuracy, but",
"comparison, we only evaluate our models on Yelptoo",
"slightly lower on BLEU.",
"YELP",
"AMAZON",
"ACC (%)",
"BLEU",
"ACC(%)",
"BLEU",
"Human Evaluation",
"We hired three annotators (not authors) to rate outputs for",
"73.7",
"3.1",
"74.1",
"0.4",
"our models and [Li et al., 2018]'s best model (DeleteAn-",
"StyleEmbedding",
"11.8",
"43.3",
"8.7",
"CrossAligned",
"10.0",
"dRetrieval).We adopt three criteria range from 1 to 5 (1",
"MultiDecoder",
"47.6",
"7.1",
"68.3",
"5.0",
"CycledReinforce",
"85.2",
"9.9",
"77.3",
"0.1",
"is very bad and 5 is very good): grammaticality, similarity",
"11.8",
"68.7",
"to the target attribute, and preservation of the source con-",
"27.1",
"TemplateBased",
"81.7",
"95.4",
"70.3",
"tent. For each dataset, we randomly sample 200 examples",
"0.9",
"RetrievalOnly",
"0.4",
"85.7",
"45.6",
"24.6",
"for each target attribute.Table 5 shows the human eval-",
"7.5",
"DeleteOnly",
"88.7",
"8.4",
"48.0",
"22.8",
"uation results.Among them, our AC-MLM-SS with the",
"fusion-method achieves best results. Pre-trained MLM in-",
"DeleteAndRetrieval",
"w/frequency-ratio",
"55.0",
"12.7",
"28.7",
"31.0",
"fills the masked positions considering the context from both",
"directions,being able to meetgrammatical requirements bet-",
"AC-MLM",
"90.5",
"75.7",
"26.0",
"11.6",
"AC-MLM-SS",
"ter. Benefiting from explicitly separating content and style,",
"w/attention-based",
"15.9",
"32.1",
"our sentences'structures arekeptand we perform betteron",
"31.2",
"AC-MLM",
"41.5",
"97.3",
"14.1",
"75.9",
"28.5",
"the preservation of the source content. Then by introducing",
"the pre-trained discriminator, MLM infills more accurate at-",
"AC-MLM-SS",
"w/fusion-method",
"tribute words.",
"43.5",
"15.3",
"42.9",
"30.7",
"97.3",
"AC-MLM",
"14.4",
"84.5",
"28.5",
"Analysis",
"AC-MLM-SS",
"5",
"Table 2: Automatic evaluation performed by tools from [Li et al.,",
"Trade-off between Content and Attribute",
"",
"Ourlossfunc",
"2018]. \"ACC\" indicates accuracy,\"BLEU\"measures content sim-",
"tion in Equation 10 consists of two parts: reconstruction loss",
"ilarity between the outputs and the human references. “AC-MLM\",",
"and discriminative loss. Reconstruction loss guides MLM to",
"represents attribute conditional masked languagemodel.\"w/\"rep-",
"infill contextual-compatible words. Compared to weak label-",
"resents “\"with\"\"-ss\" represents with soft-sampling.",
"constraint introduced by attribute embeddings, discriminative"
] | {
"arxiv_id": "1908.08039",
"language": "en",
"page_idx": 4
} |
||
1812.03324-en-0 | 1812.03324-en-0.jpg | Tight Bounds on the Rényi Entropy via Majorization
with Applications to Guessing and Compression
Igal Sason
Department of Electrical Engineering Technion-Israel Institute of Technology
Haifa 3200003, Israel
E-mail: sason@ee.technion.ac.il.
### Abstract
This paper provides tight bounds on the Rényi entropy of a function of a discrete random variable with a finite number of possible values, where the considered function is not one-to-one. To that end, a tight lower bound on the Rényi entropy of a discrete random variable with a finite support is derived as a function of the size of the support, and the ratio of the maximal to minimal probability masses. This work was inspired by the recently published paper by Cicalese $et$ al., which is focused on the Shannon entropy, and it strengthens and generalizes the results of that paper to Rényi entropies of arbitrary positive orders. In view of these generalized bounds and the works by Arikan and Campbell, non-asymptotic bounds are derived for guessing moments and lossless data compression of discrete memoryless sources.
Keywords: Majorization, Rényi entropy, Rényi divergence, cumulant generating functions, guessing moments,
lossless source coding, fixed-to-variable source codes, Huffman algorithm, Tunstall codes.
The paper was published in the Entropy journal (special issue on Probabilistic Methods in Information Theory, Hypothesis Testing,
and Coding), vol. 20, no. 12, paper no. 896, November 22,2018. Online available at https://www.mdpi.com/1099-4300/20/12/896
DRAFT
March 3, 2022 | [
[
"0.1478",
"0.0852",
"0.8555",
"0.0852",
"0.8555",
"0.1061",
"0.1478",
"0.1061"
],
[
"0.179",
"0.1307",
"0.822",
"0.1307",
"0.822",
"0.1515",
"0.179",
"0.1515"
],
[
"0.455",
"0.1831",
"0.5435",
"0.1831",
"0.5435",
"0.2008",
"0.455",
"0.2008"
],
[
"0.3472",
"0.2052",
"0.6504",
"0.2058",
"0.6504",
"0.2235",
"0.3472",
"0.2229"
],
[
"0.3381",
"0.2267",
"0.6587",
"0.2291",
"0.6587",
"0.2468",
"0.3381",
"0.2443"
],
[
"0.411",
"0.2524",
"0.5864",
"0.2524",
"0.5864",
"0.2676",
"0.411",
"0.2676"
],
[
"0.02287",
"0.276",
"0.06046",
"0.276",
"0.06046",
"0.406",
"0.02287",
"0.406"
],
[
"0.3677",
"0.2747",
"0.6274",
"0.2747",
"0.6274",
"0.289",
"0.3677",
"0.289"
],
[
"0.468",
"0.3499",
"0.5337",
"0.3499",
"0.5337",
"0.3655",
"0.468",
"0.3655"
],
[
"0.11926",
"0.385",
"0.9062",
"0.385",
"0.9062",
"0.3997",
"0.11926",
"0.3997"
],
[
"0.0196",
"0.4072",
"0.05963",
"0.406",
"0.0621",
"0.486",
"0.02287",
"0.4873"
],
[
"0.09235",
"0.4092",
"0.903",
"0.4092",
"0.903",
"0.4236",
"0.09235",
"0.4236"
],
[
"0.09235",
"0.432",
"0.9087",
"0.432",
"0.9087",
"0.4463",
"0.09235",
"0.4463"
],
[
"0.09235",
"0.455",
"0.9067",
"0.455",
"0.9067",
"0.4685",
"0.09235",
"0.4685"
],
[
"0.08905",
"0.4749",
"0.9087",
"0.476",
"0.9087",
"0.4937",
"0.08905",
"0.4924"
],
[
"0.02287",
"0.486",
"0.05963",
"0.486",
"0.05963",
"0.705",
"0.02287",
"0.705"
],
[
"0.09235",
"0.5",
"0.9087",
"0.5",
"0.9087",
"0.5146",
"0.09235",
"0.5146"
],
[
"0.09235",
"0.5215",
"0.794",
"0.5215",
"0.794",
"0.536",
"0.09235",
"0.536"
],
[
"0.10864",
"0.5537",
"0.9087",
"0.556",
"0.9087",
"0.5737",
"0.10864",
"0.5713"
],
[
"0.09235",
"0.582",
"0.7563",
"0.582",
"0.7563",
"0.595",
"0.09235",
"0.595"
],
[
"0.10455",
"0.6226",
"0.9067",
"0.6226",
"0.9067",
"0.637",
"0.10455",
"0.637"
],
[
"0.09235",
"0.643",
"0.8813",
"0.643",
"0.8813",
"0.657",
"0.09235",
"0.657"
],
[
"0.08905",
"0.94",
"0.1748",
"0.94",
"0.1748",
"0.955",
"0.08905",
"0.955"
],
[
"0.8613",
"0.94",
"0.91",
"0.94",
"0.91",
"0.955",
"0.8613",
"0.955"
]
] | [
"Tight Bounds on the Rényi Entropy via Majorization",
"with Applications to Guessing and Compression",
"Igal Sason",
"Department of Electrical Engineering",
"Technion-Israel Institute of Technology",
"Haifa 3200003, Israel",
"81208",
"E-mail: sason@ee.technion.ac.il.",
"Abstract",
"This paper provides tight bounds on the Rényi entropy of a function of a discrete random variable with a finite number",
"[LI's]",
"of possible values, where the considered function is not one-to-one. To that end, a tight lower bound on the Renyi entropy",
"of a discrete random variable with a finite support is derived as a function of the size of the support, and the ratio of",
"the maximal to minimal probability masses.This work was inspired by the recently published paper by Cicalese et al.,",
"which is focused on the Shannon entropy, and it strengthens and generalizes the results of that paper to Renyi entropies",
"1700210",
"of arbitrary positive orders. In view of these generalized bounds and the works by Arikan and Campbell, non-asymptotic",
"bounds are derived for guessing moments and lossless data compression of discrete memoryless sources.",
"Keywords: Majorization, Rényi entropy, Renyi divergence, cumulant generating functions, guessing moments,",
"lossless source coding,fixed-to-variable source codes,Huffman algorithm,Tunstall codes.",
"The paper was published in the Entropy journal (special issue on Probabilistic Methods in Information Theory, Hypothesis Testing,",
"and Coding), vol. 20, no.12, paper no.896, November 22, 2018.Online available at https:/www.mdpi.com/1099-4300/20/12/896.",
"March 3,2022",
"DRAFT"
] | {
"arxiv_id": "1812.03324",
"language": "en",
"page_idx": 0
} |
||
1812.03324-en-1 | 1812.03324-en-1.jpg | 2
I. INTRODUCTION
Majorization theory is a simple and productive concept in the theory of inequalities, which also unifies a variety of familiar bounds [33],[53]. The concept of majorization finds various applications in diverse fields (see, e.g.,[6]) such as economics [7],[19],[53], combinatorial analysis [53],[73], geometric inequalities [53] matrix theory [9],[40],[53],[73],Shannon theory [8],[18],[19],[20],[21],[22],[34],[37],[38],[39],[44], [45],[47],[61],[67],[79],[83],[84], and wireless communications [42],[46],[58],[63],[69],[80],[81],[82]
This work, which relies on the majorization theory, has been greatly inspired by the recent insightful paper by Cicalese $etal.[21].^1$ The work in [21] provides tight bounds on the Shannon entropy of a function of a discrete random variable with a finite number of possible values, where the considered function is not one-toone. For that purpose, and while being of interest by its own right (see [21,Section 6]),a tight lower bound on the Shannon entropy of a discrete random variable with a finite support was derived in [21] as a function of the size of the support, and the ratio of the maximal to minimal probability masses. The present paper aims to extend the bounds in [21] to Rényi entropies of arbitrary positive orders (note that the Shannon entropy is equal to the Rényi entropy of order l), and to study the information-theoretic applications of these (non-trivial) generalizations in the context of non-asymptotic analysis of guessing moments and lossless data compression.
The motivation for this work is rooted in the diverse information-theoretic applications of Rényi measures [62]. These include (but are not limited to) asymptotically tight bounds on guessing moments[1], informationtheoretic applications such as guessing subject to distortion [2], joint source-channel coding and guessing with application to sequential decoding [3], guessing with a prior access to a malicious oracle [14], guessing while allowing the guesser to give up and declare an error [50], guessing in secrecy problems [56],[75], guessing with limited memory [64], and guessing under source uncertainty [74]; encoding tasks [12],[13]; Bayesian hypothesis testing [8],[67],[79],and composite hypothesis testing $[ 71] , [ 77] ; \text{R\'{e} nyi generalizations of the rejection sampling}$ problem in [35], motivated by the communication complexity in distributed channel simulation, where these generalizations distinguish between causal and non-causal sampler scenarios [52]; Wyner's common information in distributed source simulation under Rényi divergence measures [87]; various other source coding theorems [15],[23],[24],[36],[49],[50],[68],[76],[78],[79],channel coding theorems [4],[5],[26],[60],[66],[78], [79],[86], including coding theorems in quantum information theory [27],[51],[57].
The presentation in this paper is structured as follows: Section II provides notation and essential preliminaries for the analysis in this paper. Sections III and IV strengthen and generalize, in a non-trivial way, the bounds on the Shannon entropy in [21] to Rényi entropies of arbitrary positive orders (see Theorems 1 and 2). Section V relies on the generalized bound from Section IV and the work by Arikan [1] to derive non-asymptotic bounds
$^{1}$The research work in the present paper has been initialized while the author handled[21] as an associate editor.
DRAFT
March 3, 2022 | [
[
"0.4207",
"0.089",
"0.576",
"0.09155",
"0.576",
"0.1061",
"0.4207",
"0.1042"
],
[
"0.1095",
"0.1206",
"0.9067",
"0.1206",
"0.9067",
"0.1351",
"0.1095",
"0.1351"
],
[
"0.09235",
"0.1465",
"0.9067",
"0.1465",
"0.9067",
"0.161",
"0.09235",
"0.161"
],
[
"0.09235",
"0.1718",
"0.9067",
"0.1718",
"0.9067",
"0.1863",
"0.09235",
"0.1863"
],
[
"0.09235",
"0.1963",
"0.9067",
"0.1963",
"0.9067",
"0.2108",
"0.09235",
"0.2108"
],
[
"0.09235",
"0.2216",
"0.9067",
"0.2216",
"0.9067",
"0.2361",
"0.09235",
"0.2361"
],
[
"0.10864",
"0.2474",
"0.9067",
"0.2474",
"0.9067",
"0.262",
"0.10864",
"0.262"
],
[
"0.0907",
"0.2727",
"0.9087",
"0.2727",
"0.9087",
"0.2874",
"0.0907",
"0.2874"
],
[
"0.0907",
"0.2966",
"0.9062",
"0.2966",
"0.9062",
"0.3113",
"0.0907",
"0.3113"
],
[
"0.09235",
"0.324",
"0.9087",
"0.324",
"0.9087",
"0.3372",
"0.09235",
"0.3372"
],
[
"0.09235",
"0.3484",
"0.9087",
"0.3484",
"0.9087",
"0.363",
"0.09235",
"0.363"
],
[
"0.09235",
"0.3738",
"0.9067",
"0.3738",
"0.9067",
"0.3882",
"0.09235",
"0.3882"
],
[
"0.0907",
"0.3997",
"0.9087",
"0.3997",
"0.9087",
"0.4136",
"0.0907",
"0.4136"
],
[
"0.09235",
"0.4248",
"0.9067",
"0.4248",
"0.9067",
"0.4382",
"0.09235",
"0.4382"
],
[
"0.09235",
"0.4495",
"0.9043",
"0.4495",
"0.9043",
"0.464",
"0.09235",
"0.464"
],
[
"0.1095",
"0.4749",
"0.9087",
"0.4749",
"0.9087",
"0.488",
"0.1095",
"0.488"
],
[
"0.0907",
"0.4988",
"0.9067",
"0.5",
"0.9067",
"0.5146",
"0.0907",
"0.513"
],
[
"0.09235",
"0.5244",
"0.9067",
"0.5244",
"0.9067",
"0.539",
"0.09235",
"0.539"
],
[
"0.09235",
"0.5503",
"0.9067",
"0.5503",
"0.9067",
"0.565",
"0.09235",
"0.565"
],
[
"0.09235",
"0.5757",
"0.9067",
"0.5757",
"0.9067",
"0.5903",
"0.09235",
"0.5903"
],
[
"0.0907",
"0.601",
"0.9067",
"0.601",
"0.9067",
"0.6157",
"0.0907",
"0.6157"
],
[
"0.09235",
"0.6255",
"0.9067",
"0.6255",
"0.9067",
"0.64",
"0.09235",
"0.64"
],
[
"0.0907",
"0.6514",
"0.9087",
"0.651",
"0.9087",
"0.6655",
"0.0907",
"0.6665"
],
[
"0.0907",
"0.677",
"0.9067",
"0.6753",
"0.9067",
"0.69",
"0.0907",
"0.6914"
],
[
"0.0907",
"0.702",
"0.9067",
"0.702",
"0.9067",
"0.7163",
"0.0907",
"0.7163"
],
[
"0.08905",
"0.7246",
"0.9087",
"0.726",
"0.9087",
"0.7437",
"0.08905",
"0.742"
],
[
"0.09235",
"0.7524",
"0.7236",
"0.7524",
"0.7236",
"0.7676",
"0.09235",
"0.7676"
],
[
"0.1095",
"0.778",
"0.9067",
"0.778",
"0.9067",
"0.7925",
"0.1095",
"0.7925"
],
[
"0.09235",
"0.803",
"0.9067",
"0.803",
"0.9067",
"0.8174",
"0.09235",
"0.8174"
],
[
"0.0907",
"0.8276",
"0.9067",
"0.8276",
"0.9067",
"0.8423",
"0.0907",
"0.8423"
],
[
"0.09235",
"0.854",
"0.9087",
"0.854",
"0.9087",
"0.8687",
"0.09235",
"0.8687"
],
[
"0.10706",
"0.8945",
"0.7915",
"0.896",
"0.7915",
"0.91",
"0.10706",
"0.909"
],
[
"0.0907",
"0.9414",
"0.1732",
"0.9414",
"0.1732",
"0.9526",
"0.0907",
"0.9526"
],
[
"0.8613",
"0.9395",
"0.911",
"0.9395",
"0.911",
"0.955",
"0.8613",
"0.955"
]
] | [
"I.INTRODUCTION",
"Majorization theory is a simple and productive concept in the theory of inequalities, which also unifies a",
"variety of familiar bounds [33], [53]. The concept of majorization finds various applications in diverse fields",
"(see, e.g., [6]) such as economics [7], [19], [53], combinatorial analysis [53], [73], geometric inequalities [53],",
"matrix theory [9], [40], [53], [73], Shannon theory [8], [18], [19], [20], [21], [22], [34], [37], [38], [39], [44],",
"[45], [47], [61], [67], [79], [83], [84], and wireless communications [42], [46], [58], [63], [69], [80], [81], [82].",
"This work, which relies on the majorization theory, has been greatly inspired by the recent insightful paper",
"by Cicalese et al. [21].] The work in [21] provides tight bounds on the Shannon entropy of a function of a",
"discrete random variable with a finite number of possible values, where the considered function is not one-to-",
"one. For that purpose, and while being of interest by its own right (see [2l, Section 6]), a tight lower bound",
"on the Shannon entropy of a discrete random variable with a finite support was derived in [21] as a function",
"of the size of the support, and the ratio of the maximal to minimal probability masses. The present paper aims",
"to extend the bounds in [21] to Rényi entropies of arbitrary positive orders (note that the Shannon entropy is",
"equal to the Rényi entropy of order 1), and to study the information-theoretic applications of these (non-trivial)",
"generalizations in the context of non-asymptotic analysis of guessing moments and lossless data compression.",
"The motivation for this work is rooted in the diverse information-theoretic applications of Renyi measures",
"[62]. These include (but are not limited to) asymptotically tight bounds on guessing moments [1], information-",
"theoretic applications such as guessing subject to distortion [2], joint source-channel coding and guessing with",
"application to sequential decoding [3], guessing with a prior access to a malicious oracle [14], guessing while",
"allowing the guesser to give up and declare an error [50], guessing in secrecy problems [56], [75], guessing with",
"limited memory [64], and guessing under source uncertainty [74]; encoding tasks [12], [13]; Bayesian hypothesis",
"testing [8], [67], [79], and composite hypothesis testing [71], [77]; Rényi generalizations of the rejection sampling",
"problem in [35], motivated by the communication complexity in distributed channel simulation, where these",
"generalizations distinguish between causal and non-causal sampler scenarios [52]; Wyner's common information",
"in distributed source simulation under Renyi divergence measures [87]; various other source coding theorems",
"[15], [23], [24], [36], [49], [50], [68], [76], [78], [79], channel coding theorems [4], [5], [26], [60], [66], [78],",
"[79], [86], including coding theorems in quantum information theory [27], [51], [57].",
"The presentation in this paper is structured as follows: Section II provides notation and essential preliminaries",
"for the analysis in this paper. Sections IlI and IV strengthen and generalize, in a non-trivial way, the bounds on",
"the Shannon entropy in [21] to Rényi entropies of arbitrary positive orders (see Theorems 1 and 2). Section V",
"relies on the generalized bound from Section IV and the work by Arikan [1] to derive non-asymptotic bounds",
"'The research work in the present paper has been initialized while the author handled [21] as an associate editor",
"March 3,2022",
"DRAFT"
] | {
"arxiv_id": "1812.03324",
"language": "en",
"page_idx": 1
} |
||
1812.03324-en-2 | 1812.03324-en-2.jpg | 3
for guessing moments (see Theorem 3); Section V also relies on the generalized bound in Section IV and the source coding theorem by Campbell [15] (see Theorem 4) for the derivation of non-asymptotic bounds for lossless compression of discrete memoryless sources (see Theorem 5).
II. NOTATION AND PRELIMINARIES
Let
$\bullet$ $P$ be a probability mass function defined on a finite set $\mathcal{X} ;$ $\bullet$ $p_{\max }$ and $p_{\min }$ be, respectively, the maximal and minimal positive masses of $P;$
$\bullet$ $G_P( k)$ be the sum of the $k$ largest masses of $P$ for $k\in\{1,\ldots,|\mathcal{X}|\}$ (note that $G_P(1)=p_{\max}$ and
$G_P(|\mathcal{X}|)=1);$
$\bullet$ $\mathcal{P} _n$, for an integer $n\geq2$, be the set of all probability mass functions defined on $\mathcal{X}$ with $|\mathcal{X}|=n;$ without
any loss of generality, let $\mathcal{X}=\{1,\ldots,n\};$
$\bullet$ $\mathcal{P} _n( \rho ) $, for $\rho\geq1$ and an integer $n\geq2$, be the subset of all probability measures $P\in\mathcal{P}_n$ such that
(1)
$$\frac{p_{\max}}{p_{\min}}\leq\rho.$$
$Definition~I~($majorization): Consider discrete probability mass functions $P$ and $Q$ defined on the same (finite or countably infinite) set $\mathcal{X}.$ It is said that $P$ is majorize$d$ by $Q$ (or $Q$ majorizes $P)$, and it is denoted by $P\prec Q$, if$G_P(k)\leq G_Q(k)$ for all $k\in\{1,\ldots,|\mathcal{X}|-1\}$ (recall that $G_P(|\mathcal{X}|)=G_Q(|\mathcal{X}|)=1).$ If $P$ and $Q$ are defined on finite sets of different cardinalities, then the probability mass function which is defined over the smaller set is first padded by zeros for making the cardinalities of these sets be equal.
$\textbf{By Definition 1, a unit mass majorizes any other distribution; on the other hand, the equiprobable distribution}$
on a finite set is majorized by any other distribution defined on the same set.
$Definition2($Schur-convexity/concavity): A function $f:\mathcal{P}_n\to\mathbb{R}$ is said to be Schur-convex if for every $P,Q\in\mathcal{P}_n$ such that $P\prec Q$,we have $f(P)\leq f(Q).$ Likewise, $f$ is said to be Schur-concave if $-f$ is Schur- convex, i. e. , $P, Q\in \mathcal{P} _{n}$ and $P\prec Q$ imply that $f(P)\geq f(Q).$
Definition 3 (Rényi entropy [62]): Let $X$ be a random variable taking values on a finite or countably infinite set $\mathcal{X}$,and let $P_X$ be its probability mass function. The Rényi entropy of order $\alpha\in(0,1)\cup(1,\infty)$ is given by$^2$
(2)
$$H_\alpha(X)=H_\alpha(P_X)=\frac{1}{1-\alpha}\:\log\left(\sum_{x\in\mathcal{X}}P_X^\alpha(x)\right).$$
$^{2}$Unless explicitly stated, the logarithm base can be chosen by the reader, with exp indicating the inverse function of log.
DRAFT
$\mathbf{March}3,2022$ | [
[
"0.09235",
"0.09155",
"0.9087",
"0.09155",
"0.9087",
"0.1061",
"0.09235",
"0.1061"
],
[
"0.09235",
"0.11615",
"0.9087",
"0.11615",
"0.9087",
"0.1307",
"0.09235",
"0.1307"
],
[
"0.0907",
"0.1414",
"0.6094",
"0.1414",
"0.6094",
"0.1559",
"0.0907",
"0.1559"
],
[
"0.3545",
"0.1843",
"0.642",
"0.1849",
"0.642",
"0.1995",
"0.3545",
"0.1989"
],
[
"0.1054",
"0.2134",
"0.1381",
"0.2134",
"0.1381",
"0.2299",
"0.1054",
"0.2299"
],
[
"0.1111",
"0.2424",
"0.574",
"0.2424",
"0.574",
"0.2568",
"0.1111",
"0.2568"
],
[
"0.10706",
"0.2676",
"0.7295",
"0.2651",
"0.7295",
"0.2827",
"0.10706",
"0.2854"
],
[
"0.1095",
"0.2922",
"0.91",
"0.2922",
"0.91",
"0.31",
"0.1095",
"0.31"
],
[
"0.1356",
"0.318",
"0.2467",
"0.318",
"0.2467",
"0.3333",
"0.1356",
"0.3333"
],
[
"0.1111",
"0.3435",
"0.9087",
"0.3435",
"0.9087",
"0.358",
"0.1111",
"0.358"
],
[
"0.134",
"0.3687",
"0.4495",
"0.3687",
"0.4495",
"0.3833",
"0.134",
"0.3833"
],
[
"0.1111",
"0.3945",
"0.869",
"0.3945",
"0.869",
"0.4092",
"0.1111",
"0.4092"
],
[
"0.4836",
"0.4185",
"0.5664",
"0.4312",
"0.5605",
"0.4495",
"0.478",
"0.4375"
],
[
"0.8823",
"0.428",
"0.911",
"0.428",
"0.911",
"0.4443",
"0.8823",
"0.4443"
],
[
"0.4836",
"0.4414",
"0.524",
"0.4414",
"0.524",
"0.4558",
"0.4836",
"0.4558"
],
[
"0.1095",
"0.4634",
"0.9067",
"0.4634",
"0.9067",
"0.4778",
"0.1095",
"0.4778"
],
[
"0.09235",
"0.4893",
"0.9043",
"0.4893",
"0.9043",
"0.504",
"0.09235",
"0.504"
],
[
"0.0907",
"0.5146",
"0.9087",
"0.5146",
"0.9087",
"0.529",
"0.0907",
"0.529"
],
[
"0.0907",
"0.5386",
"0.9087",
"0.539",
"0.9087",
"0.5537",
"0.0907",
"0.553"
],
[
"0.0907",
"0.565",
"0.6445",
"0.565",
"0.6445",
"0.58",
"0.0907",
"0.58"
],
[
"0.1095",
"0.5903",
"0.9067",
"0.5903",
"0.9067",
"0.605",
"0.1095",
"0.605"
],
[
"0.09235",
"0.616",
"0.664",
"0.616",
"0.664",
"0.63",
"0.09235",
"0.63"
],
[
"0.10706",
"0.638",
"0.9087",
"0.6396",
"0.9087",
"0.657",
"0.10706",
"0.656"
],
[
"0.09314",
"0.6665",
"0.9067",
"0.6665",
"0.9067",
"0.681",
"0.09314",
"0.681"
],
[
"0.09235",
"0.6914",
"0.5933",
"0.6914",
"0.5933",
"0.7056",
"0.09235",
"0.7056"
],
[
"0.1095",
"0.7163",
"0.9087",
"0.7163",
"0.9087",
"0.731",
"0.1095",
"0.731"
],
[
"0.08905",
"0.7407",
"0.9067",
"0.738",
"0.9067",
"0.756",
"0.08905",
"0.7583"
],
[
"0.3162",
"0.7744",
"0.4763",
"0.778",
"0.4763",
"0.7954",
"0.3162",
"0.7925"
],
[
"0.5034",
"0.7744",
"0.5107",
"0.7744",
"0.5107",
"0.779",
"0.5034",
"0.779"
],
[
"0.58",
"0.7764",
"0.653",
"0.7764",
"0.653",
"0.7944",
"0.58",
"0.7944"
],
[
"0.8823",
"0.7764",
"0.911",
"0.7764",
"0.911",
"0.7944",
"0.8823",
"0.7944"
],
[
"0.4902",
"0.789",
"0.527",
"0.789",
"0.527",
"0.8",
"0.4902",
"0.8"
],
[
"0.528",
"0.78",
"0.564",
"0.78",
"0.564",
"0.7954",
"0.528",
"0.7954"
],
[
"0.58",
"0.8003",
"0.601",
"0.8003",
"0.601",
"0.8047",
"0.58",
"0.8047"
],
[
"0.10046",
"0.8335",
"0.8433",
"0.8354",
"0.8433",
"0.853",
"0.10046",
"0.851"
],
[
"0.0907",
"0.9414",
"0.1732",
"0.9414",
"0.1732",
"0.9526",
"0.0907",
"0.9526"
],
[
"0.8613",
"0.9395",
"0.911",
"0.9395",
"0.911",
"0.955",
"0.8613",
"0.955"
]
] | [
"for guessing moments (see Theorem 3); Section V also relies on the generalized bound in Section IV and the",
"source coding theorem by Campbell [15] (see Theorem 4) for the derivation of non-asymptotic bounds for",
"lossless compression of discrete memoryless sources (see Theorem 5)",
"II.NOTATIONANDPRELIMINARIES",
"Let",
"·P be a probability mass function defined on a finite set X;",
". Pmax and pmin be, respectively, the maximal and minimal positive masses of P;",
". Gp(k) be the sum of the k largest masses of P for k e [1,...,|Xl) (note that Gp(1) = Pmax and",
"Gp(IX) = 1);",
" Pn, for an integer n ≥ 2, be the set of all probability mass functions defined on X with X= n; without",
"any loss of generality, let X = {1, ..., n]:",
". Pn(p), for p ≥ 1 and an integer n ≥ 2, be the subset of all probability measures P e Pn such that",
"Pmax ≤ p.",
"(1)",
"Pmin",
"Definition 1 (majorization): Consider discrete probability mass functions P and Q defined on the same (finite",
"or countably infinite) set X. It is said that P is majorized by Q (or Q majorizes P), and it is denoted by P Q",
"if Gp(k) ≤ GQ(k) for all k E {1, :::, /X|- 1) (recall that Gp(/XD) = G(IXD) = 1). If P and Q are defined",
"on finite sets of different cardinalities, then the probability mass function which is defined over the smaller set",
"is first padded by zeros for making the cardinalities of these sets be equal.",
"By Definition 1, a unit mass majorizes any other distribution; on the other hand, the equiprobable distribution",
"on a finite set is majorized by any other distribution defined on the same set.",
"Definition 2 (Schur-convexity/concavity): A function f: Pn → R is said to be Schur-convex if for every",
"P,Q E Pn such that P < Q, we have f(P) ≤ f(Q). Likewise, f is said to be Schur-concave if -f is",
"Schur-convex, i.e., P, Q E Pn and P < Q imply that f(P) ≥ f(Q).",
"Definition 3 (Renyi entropy [62]): Let X be a random variable taking values on a finite or countably infinite",
"set X, and let Px be its probability mass function. The Rényi entropy of order α E (O, 1) U (1, oo) is given by?",
"Hα(X) = Ha(Px) =",
"E",
" P(α",
"(2)",
"1-0",
"log",
"YC",
"2Unless explicitly stated, the logarithm base can be chosen by the reader, with exp indicating the inverse function of log.",
"March 3.2022",
"DRAFT"
] | {
"arxiv_id": "1812.03324",
"language": "en",
"page_idx": 2
} |
||
1909.13155-en-0 | 1909.13155-en-0.jpg | Weakly Supervised Energy-Based Learning for Action Segmentation
Jun Li
Oregon State University liju2@oregonstate.edu
Peng Lei$^*$
Amazon.com Services, Inc.
leipeng@amazon.com
Sinisa Todorovic
Oregon State University sinisagoregonstate.edu
$U2$
$സ്പെ$
$ଶ୍ନି$
## Abstract
This paper is about labeling video frames with action classes under weak supervision in training, where we have
$\textcircled {\mathrm{D} }$ $classesunderweaksuperviston in traming, where we haveaccess to a temporal ordering of actions, but their start$
and end frames in training videos are unknown. Following $priorwork,weuseanHMMgroundedonaGatedRecur-$ rent Unit (GRU) for frame labeling. Our key contribution is a new constrained discriminative forward loss (CDFL)
$\sum_{\begin{array}{cc}&thatweusefortraining\:the\:HMM\:and\:GRU\:under\:weak\:su-\\&pervision.\:While\:prior\:work\:typically\:estimates\:the\:loss\:on\\&a\:single,\:inferred\:video\:segmentaion,\:our\:CDFL\:discrimi-\end{array}}$ $\begin{array}{lll}{&a\:single,\:inferred\:video\:segmentation,\:our\:CDFL\:discrimi-\\\dot{o}_{2}&nates\:between\:the\:energy\:of\:all\:valid\:and\:invalid\:frame\:la-\\\dot{O}&belings\:of\:a\:training\:video.\:A\:valid\:frame\:labeling\:satisfes}\end{array}$
the ground-truth temporal ordering of actions, whereas an invalid one violates the ground truth. We specify an efficient recursive algorithm for computing the CDFL in terms of the $logaddfunction of the segmentation energy. Our evaluation$ on action segmentation and alignment gives superior results to those of the state of the art on the benchmark Breakfast $Action$, Hollywood $Extended$, and $50Salads$ $datasets. ^{\dagger }$
## $1. \textbf{Introduction}$
This paper presents an approach to weakly supervised action segmentation by labeling video frames with action classes. Weak supervision means that in training our approach has access only to the temporal ordering of actions, but their ground-truth start and end frames are not provided. This is an important problem with a wide range of applications, since the more common fully supervised action segmentation typically requires expensive manual annotations of action occurrences in every video frame.
Our fundamental challenge is that the set of all possible segmentations of a training video may consist of multiple distinct valid segmentations that satisfy the provided ground-truth ordering of actions, along with invalid segmentations that violate the ground truth. It is not clear how
*The work was done at the Oregon State University before Peng Lei joined Amazon. $^{\dagger}$The code is available at https://github.com/JunLiGalios/CDFL.
to estimate loss (and subsequently train the segmenter) over
multiple valid segmentations.
$\textbf{Motivation: Prior work[ 8, 12, 20, 7, 22] typically uses}$ a temporal model (e.g., deep neural network, or HMM) to infer a $single$, valid, optimal video segmentation, and takes this inference result as a pseudo ground truth for estimating the incurred loss. However, a particular training video may exhibit a significant variation (not yet captured by the model along the course of training), which may negatively affect estimation of the pseudo ground truth, such that the inferred action segmentation is significantly different from the true one. In turn, the loss estimated on the incorrect pseudo ground truth may corrupt training by reducing, instead of maximizing, the discriminative margin between the ground truth and other valid segmentations. In this paper, we seek to alleviate these issues.
Contributions: Prior work shows that a statistical language model is useful for weakly supervised learning and modeling of video sequences [17,9,19,22,3]. Following [22], we also adopt a Hidden Markov Model (HMM) grounded on a Gated Recurrent Unit (GRU)[4] for labeling video frames. The major difference is that we do not generate a unique pseudo ground truth for training. Instead, we efficiently account for all candidate segmentations of a training video when estimating the loss. To this end, we formulate a new Constrained Discriminative Forward Loss (CDFL) as a difference between the energy of valid and invalid candidate video segmentations. In comparison with prior work, the CDFL improves robustness of our training, because minimizing the CDFL amounts to maximizing the discrimination margin between candidate segmentations that satisfy and violate ground truth, whereas prior work solely optimizes a score of the inferred single valid segmentation. Robustness of training is further improved when the CDFL takes into account only hard invalid segmentations whose edge energy is lower than that of valid ones. Along with the new CDFL formulation, our key contribution is a new recursive algorithm for efficiently estimating the CDFL in terms of the logadd function of the segmentation energy.
Our Approach: Fig. 1 shows an overview of our weakly supervised training of the HMM with GRU that consists of | [
[
"0.1438",
"0.1332",
"0.8286",
"0.1332",
"0.8286",
"0.1509",
"0.1438",
"0.1509"
],
[
"0.2059",
"0.1818",
"0.2615",
"0.1818",
"0.2615",
"0.1976",
"0.2059",
"0.1976"
],
[
"0.4453",
"0.1812",
"0.528",
"0.1812",
"0.528",
"0.1989",
"0.4453",
"0.1989"
],
[
"0.6685",
"0.1831",
"0.8022",
"0.1831",
"0.8022",
"0.1976",
"0.6685",
"0.1976"
],
[
"0.1396",
"0.2008",
"0.3284",
"0.2008",
"0.3284",
"0.2153",
"0.1396",
"0.2153"
],
[
"0.3743",
"0.2008",
"0.5923",
"0.2008",
"0.5923",
"0.2153",
"0.3743",
"0.2153"
],
[
"0.6406",
"0.2008",
"0.8286",
"0.2008",
"0.8286",
"0.2153",
"0.6406",
"0.2153"
],
[
"0.1405",
"0.2197",
"0.327",
"0.2197",
"0.327",
"0.2329",
"0.1405",
"0.2329"
],
[
"0.4028",
"0.2203",
"0.564",
"0.2203",
"0.564",
"0.2343",
"0.4028",
"0.2343"
],
[
"0.638",
"0.2216",
"0.831",
"0.2216",
"0.831",
"0.2317",
"0.638",
"0.2317"
],
[
"0.02124",
"0.26",
"0.0645",
"0.26",
"0.0645",
"0.3687",
"0.02124",
"0.3687"
],
[
"0.2361",
"0.2676",
"0.3145",
"0.2676",
"0.3145",
"0.2854",
"0.2361",
"0.2854"
],
[
"0.501",
"0.2695",
"0.892",
"0.2708",
"0.892",
"0.2854",
"0.501",
"0.2842"
],
[
"0.5034",
"0.2886",
"0.6978",
"0.2886",
"0.6978",
"0.2986",
"0.5034",
"0.2986"
],
[
"0.10046",
"0.3018",
"0.4697",
"0.3018",
"0.4697",
"0.315",
"0.10046",
"0.315"
],
[
"0.5215",
"0.305",
"0.892",
"0.305",
"0.892",
"0.318",
"0.5215",
"0.318"
],
[
"0.0809",
"0.3176",
"0.468",
"0.3176",
"0.468",
"0.3308",
"0.0809",
"0.3308"
],
[
"0.501",
"0.3193",
"0.8906",
"0.3193",
"0.8906",
"0.3333",
"0.501",
"0.3333"
],
[
"0.0801",
"0.332",
"0.4697",
"0.332",
"0.4697",
"0.3452",
"0.0801",
"0.3452"
],
[
"0.5034",
"0.3372",
"0.8906",
"0.3372",
"0.8906",
"0.348",
"0.5034",
"0.348"
],
[
"0.0801",
"0.348",
"0.4697",
"0.348",
"0.4697",
"0.361",
"0.0801",
"0.361"
],
[
"0.5024",
"0.351",
"0.8906",
"0.351",
"0.8906",
"0.3643",
"0.5024",
"0.3643"
],
[
"0.03513",
"0.3599",
"0.04984",
"0.3599",
"0.04984",
"0.3674",
"0.03513",
"0.3674"
],
[
"0.0801",
"0.361",
"0.4673",
"0.361",
"0.4673",
"0.375",
"0.0801",
"0.375"
],
[
"0.501",
"0.3655",
"0.8896",
"0.3655",
"0.8896",
"0.379",
"0.501",
"0.379"
],
[
"0.0825",
"0.379",
"0.468",
"0.379",
"0.468",
"0.3894",
"0.0825",
"0.3894"
],
[
"0.02533",
"0.3623",
"0.058",
"0.3623",
"0.058",
"0.3965",
"0.02533",
"0.3965"
],
[
"0.5024",
"0.3806",
"0.892",
"0.3806",
"0.892",
"0.3945",
"0.5024",
"0.3945"
],
[
"0.0809",
"0.3926",
"0.468",
"0.3926",
"0.468",
"0.406",
"0.0809",
"0.406"
],
[
"0.5024",
"0.3965",
"0.8906",
"0.3965",
"0.8906",
"0.4104",
"0.5024",
"0.4104"
],
[
"0.02533",
"0.406",
"0.05637",
"0.406",
"0.05637",
"0.4507",
"0.02533",
"0.4507"
],
[
"0.0801",
"0.4072",
"0.4673",
"0.4072",
"0.4673",
"0.4204",
"0.0801",
"0.4204"
],
[
"0.5034",
"0.411",
"0.892",
"0.411",
"0.892",
"0.4248",
"0.5034",
"0.4248"
],
[
"0.0801",
"0.4224",
"0.4697",
"0.4224",
"0.4697",
"0.4375",
"0.0801",
"0.4375"
],
[
"0.5024",
"0.4268",
"0.892",
"0.4268",
"0.892",
"0.44",
"0.5024",
"0.44"
],
[
"0.0801",
"0.4382",
"0.4697",
"0.4382",
"0.4697",
"0.452",
"0.0801",
"0.452"
],
[
"0.5024",
"0.4414",
"0.892",
"0.4414",
"0.892",
"0.455",
"0.5024",
"0.455"
],
[
"0.03186",
"0.449",
"0.058",
"0.449",
"0.058",
"0.481",
"0.03186",
"0.481"
],
[
"0.0801",
"0.4526",
"0.468",
"0.4526",
"0.468",
"0.4666",
"0.0801",
"0.4666"
],
[
"0.5024",
"0.457",
"0.892",
"0.457",
"0.892",
"0.4702",
"0.5024",
"0.4702"
],
[
"0.0801",
"0.4685",
"0.468",
"0.4685",
"0.468",
"0.4817",
"0.0801",
"0.4817"
],
[
"0.5024",
"0.4717",
"0.8936",
"0.4717",
"0.8936",
"0.4849",
"0.5024",
"0.4849"
],
[
"0.0801",
"0.483",
"0.468",
"0.483",
"0.468",
"0.4968",
"0.0801",
"0.4968"
],
[
"0.5034",
"0.486",
"0.8936",
"0.486",
"0.8936",
"0.5005",
"0.5034",
"0.5005"
],
[
"0.0237",
"0.502",
"0.058",
"0.502",
"0.058",
"0.539",
"0.0237",
"0.539"
],
[
"0.0809",
"0.5",
"0.468",
"0.5",
"0.468",
"0.5103",
"0.0809",
"0.5103"
],
[
"0.5024",
"0.502",
"0.7183",
"0.502",
"0.7183",
"0.515",
"0.5024",
"0.515"
],
[
"0.0801",
"0.513",
"0.4697",
"0.513",
"0.4697",
"0.5264",
"0.0801",
"0.5264"
],
[
"0.524",
"0.5215",
"0.8896",
"0.5215",
"0.8896",
"0.532",
"0.524",
"0.532"
],
[
"0.03105",
"0.528",
"0.06046",
"0.5386",
"0.04822",
"0.5566",
"0.0196",
"0.547"
],
[
"0.0801",
"0.529",
"0.468",
"0.529",
"0.468",
"0.5425",
"0.0801",
"0.5425"
],
[
"0.5024",
"0.535",
"0.892",
"0.534",
"0.892",
"0.549",
"0.5024",
"0.55"
],
[
"0.0801",
"0.5435",
"0.4697",
"0.5435",
"0.4697",
"0.5566",
"0.0801",
"0.5566"
],
[
"0.5034",
"0.5503",
"0.8906",
"0.5503",
"0.8906",
"0.5645",
"0.5034",
"0.5645"
],
[
"0.02777",
"0.5566",
"0.05554",
"0.5566",
"0.05554",
"0.577",
"0.02777",
"0.577"
],
[
"0.0801",
"0.5596",
"0.468",
"0.5596",
"0.468",
"0.5728",
"0.0801",
"0.5728"
],
[
"0.5024",
"0.565",
"0.8896",
"0.565",
"0.8896",
"0.579",
"0.5024",
"0.579"
],
[
"0.0809",
"0.5728",
"0.4307",
"0.5728",
"0.4307",
"0.586",
"0.0809",
"0.586"
],
[
"0.0196",
"0.5835",
"0.058",
"0.5835",
"0.06046",
"0.712",
"0.02287",
"0.712"
],
[
"0.5024",
"0.5806",
"0.8906",
"0.58",
"0.8906",
"0.5947",
"0.5024",
"0.595"
],
[
"0.03595",
"0.5884",
"0.0474",
"0.5884",
"0.0474",
"0.5913",
"0.03595",
"0.5913"
],
[
"0.5034",
"0.5967",
"0.892",
"0.5967",
"0.892",
"0.61",
"0.5034",
"0.61"
],
[
"0.0809",
"0.616",
"0.2059",
"0.616",
"0.2059",
"0.63",
"0.0809",
"0.63"
],
[
"0.5034",
"0.6123",
"0.8906",
"0.6123",
"0.8906",
"0.6255",
"0.5034",
"0.6255"
],
[
"0.501",
"0.625",
"0.892",
"0.6255",
"0.892",
"0.64",
"0.501",
"0.6396"
],
[
"0.09644",
"0.64",
"0.4714",
"0.6416",
"0.4714",
"0.659",
"0.09644",
"0.6577"
],
[
"0.5034",
"0.6416",
"0.892",
"0.6416",
"0.892",
"0.655",
"0.5034",
"0.655"
],
[
"0.0801",
"0.6577",
"0.4697",
"0.6577",
"0.4697",
"0.6724",
"0.0801",
"0.6724"
],
[
"0.5024",
"0.656",
"0.892",
"0.656",
"0.892",
"0.6694",
"0.5024",
"0.6694"
],
[
"0.0801",
"0.674",
"0.468",
"0.674",
"0.468",
"0.6875",
"0.0801",
"0.6875"
],
[
"0.5024",
"0.672",
"0.8906",
"0.672",
"0.8906",
"0.685",
"0.5024",
"0.685"
],
[
"0.5034",
"0.688",
"0.8906",
"0.688",
"0.8906",
"0.6987",
"0.5034",
"0.6987"
],
[
"0.03513",
"0.677",
"0.04004",
"0.677",
"0.04004",
"0.6846",
"0.03513",
"0.6846"
],
[
"0.0801",
"0.6895",
"0.468",
"0.6895",
"0.468",
"0.7026",
"0.0801",
"0.7026"
],
[
"0.0809",
"0.704",
"0.468",
"0.704",
"0.468",
"0.7173",
"0.0809",
"0.7173"
],
[
"0.5034",
"0.702",
"0.8906",
"0.702",
"0.8906",
"0.7153",
"0.5034",
"0.7153"
],
[
"0.0825",
"0.7183",
"0.468",
"0.7183",
"0.468",
"0.732",
"0.0825",
"0.732"
],
[
"0.5024",
"0.7173",
"0.892",
"0.7173",
"0.892",
"0.731",
"0.5024",
"0.731"
],
[
"0.0801",
"0.7344",
"0.468",
"0.7344",
"0.468",
"0.7476",
"0.0801",
"0.7476"
],
[
"0.5034",
"0.7344",
"0.8896",
"0.7344",
"0.8896",
"0.744",
"0.5034",
"0.744"
],
[
"0.0801",
"0.75",
"0.4697",
"0.75",
"0.4697",
"0.763",
"0.0801",
"0.763"
],
[
"0.5024",
"0.7476",
"0.8936",
"0.7476",
"0.8936",
"0.761",
"0.5024",
"0.761"
],
[
"0.0801",
"0.7646",
"0.3628",
"0.7646",
"0.3628",
"0.778",
"0.0801",
"0.778"
],
[
"0.5034",
"0.762",
"0.892",
"0.762",
"0.892",
"0.776",
"0.5034",
"0.776"
],
[
"0.0972",
"0.779",
"0.468",
"0.778",
"0.468",
"0.7925",
"0.0972",
"0.7935"
],
[
"0.5034",
"0.779",
"0.8906",
"0.779",
"0.8906",
"0.789",
"0.5034",
"0.789"
],
[
"0.0801",
"0.7944",
"0.468",
"0.7944",
"0.468",
"0.808",
"0.0801",
"0.808"
],
[
"0.5024",
"0.7925",
"0.892",
"0.7925",
"0.892",
"0.806",
"0.5024",
"0.806"
],
[
"0.0801",
"0.81",
"0.4697",
"0.81",
"0.4697",
"0.8237",
"0.0801",
"0.8237"
],
[
"0.5024",
"0.806",
"0.892",
"0.8066",
"0.892",
"0.8213",
"0.5024",
"0.821"
],
[
"0.0801",
"0.8237",
"0.468",
"0.8247",
"0.468",
"0.839",
"0.0801",
"0.8384"
],
[
"0.5024",
"0.8228",
"0.892",
"0.8228",
"0.892",
"0.836",
"0.5024",
"0.836"
],
[
"0.0801",
"0.8403",
"0.468",
"0.8403",
"0.468",
"0.854",
"0.0801",
"0.854"
],
[
"0.5034",
"0.839",
"0.8906",
"0.839",
"0.8906",
"0.8496",
"0.5034",
"0.8496"
],
[
"0.5024",
"0.853",
"0.8896",
"0.853",
"0.8896",
"0.866",
"0.5024",
"0.866"
],
[
"0.0972",
"0.863",
"0.468",
"0.863",
"0.468",
"0.877",
"0.0972",
"0.877"
],
[
"0.0801",
"0.8755",
"0.468",
"0.8755",
"0.468",
"0.8896",
"0.0801",
"0.8896"
],
[
"0.5215",
"0.872",
"0.8906",
"0.872",
"0.8906",
"0.8853",
"0.5215",
"0.8853"
],
[
"0.0809",
"0.888",
"0.1569",
"0.888",
"0.1569",
"0.8994",
"0.0809",
"0.8994"
],
[
"0.501",
"0.886",
"0.8936",
"0.8853",
"0.8936",
"0.8994",
"0.501",
"0.901"
]
] | [
"Weakly Supervised Energy-Based Learning for Action Segmentation",
"Jun Li",
"Peng Lei*",
"Sinisa Todorovic",
"Oregon State University",
"Amazon.com Services, Inc.",
"Oregon State University",
"liju2@oregonstate.edu",
"leipeng@amazon.com",
"sinisa@oregonstate.edu",
"610z des",
"Abstract",
"to estimate loss (and subsequently train the segmenter) over",
"multiplevalid segmentations.",
"This paper is about labeling video frames with action",
"Motivation: Prior work [8, 12, 20, 7, 22] typically uses",
"classes under weak supervision in training, where we have",
"a temporal model (e.g., deep neural network, or HMM) to",
"access to a temporal ordering of actions, but their start",
"infera single,valid,optimal video segmentation,and takes",
"and end frames in training videos are unknown. Following",
"this inference result as a pseudo ground truth for estimat-",
"",
"prior work, we use an HMM grounded on a Gated Recur-",
"ing the incurred loss. However, a particular training video",
"rentUnit(GRU)for framelabeling.Ourkeycontribution",
"8",
"may exhibit a significant variation (not yet captured by the",
"is a new constrained discriminative forward loss (CDFL)",
"model along the course of training), which may negatively",
"LAD",
"that we use for training the HMM and GRU under weak su-",
"affect estimation of the pseudo ground truth, such that the",
"pervision. While prior work typically estimates the loss on",
"inferred action segmentation is significantly different from",
"a single, inferred video segmentation, our CDFL discrimi-",
"the true one. In turn, the loss estimated on the incorrect",
"SO",
"nates between the energy of all valid and invalid frame la-",
"pseudo ground truth may corrupt training by reducing, in-",
"belings of a training video. A valid frame labeling satisfies",
"stead of maximizing, the discriminative margin between the",
"the ground-truth temporal ordering of actions, whereas an",
"ground truth and other valid segmentations. In this paper,",
"5",
"invalid oneviolatesthegroundtruth.Wespecifyan efficient",
"we seek to alleviate these issues.",
"recursive algorithm for computing the CDFL in terms of the",
"Contributions: Prior work shows that a statistical lan-",
"5",
"logadd function of the segmentation energyOur evaluation",
"guage model is useful for weakly supervised learning and",
"on action segmentation and alignment gives superior results",
"modeling of video sequences [17, 9, 19, 22, 3]. Follow-",
"3",
"to those of the state of the art on the benchmark Breakfast",
"ing [22], we also adopt a Hidden Markov Model (HMM)",
"Action, Hollywood Extended, and 5oSalads datasets.t",
"6061:AIXe",
"grounded on a Gated Recurrent Unit (GRU) [4] for label-",
"",
"ing video frames. The major difference is that we do not",
"1.Introduction",
"generate a unique pseudo ground truth for training. Instead,",
"we efficiently account for all candidate segmentations of a",
"This paper presents an approach to weakly supervised",
"training video when estimating the loss. To this end, we",
"action segmentation by labeling video frames with action",
"formulate a new Constrained Discriminative Forward Loss",
"classes. Weak supervision means that in training our ap-",
"(CDFL) as a difference between the energy of valid and in-",
"valid candidate video segmentations.In comparison with",
"",
"proach has access only to the temporal ordering of actions,",
"but their ground-truth start and end frames are not provided.",
"prior work, the CDFL improves robustness of our train-",
"This is an important problem with a wide range of applica-",
"ing, because minimizing the CDFL amounts to maximizing",
"tions, since the more common fully supervised action seg-",
"thediscriminationmarginbetweencandidatesegmentations",
"mentation typically requires expensive manual annotations",
"that satisfy and violate ground truth, whereas prior work",
"of action occurrences in every video frame.",
"solely optimizes a score of the inferred single valid segmen-",
"Our fundamental challenge is that the set of all possi-",
"tation.Robustnessoftrainingisfurtherimprovedwhenthe",
"ble segmentations of a training video may consist of mul-",
"CDFL takes into account only hard invalid segmentations",
"tiple distinct valid segmentations that satisfy the provided",
"whose edge energy is lower than that of valid ones. Along",
"ground-truth ordering of actions, along with invalid seg-",
"with the new CDFL formulation, our key contribution is a",
"mentations that violate the ground truth. It is not clear how",
"newrecursivealgorithmforefficientlyestimatingtheCDFL",
"in terms of the logadd function of the segmentation energy.",
"*The work was done at the Oregon State University before Peng",
"Lei joined Amazon. tThe code is available at https://github.com/JunLi-",
"Our Approach: Fig. 1 shows an overview of our weakly",
"Galios/CDFL.",
"supervised training of the HMM with GRU that consists of"
] | {
"arxiv_id": "1909.13155",
"language": "en",
"page_idx": 0
} |
||
1909.13155-en-2 | 1909.13155-en-2.jpg | ual frames. In contrast, we use an energy-based framework with the following differences. First, in training, we minimize the total energy of valid paths in the segmentation graph rather than optimize labeling probabilities of each frame. Second, instead of considering a single optimal valid path in the segmentation graph, we specify a loss function in terms of all valid paths. Hence, the Viterbi-initialized training on pseudo-labels of frames [22] represents a special case of our training done only for one valid path. In addition, our loss enforces discriminative training by accounting for invalid paths in the segmentation graph. Unlike [3] that randomly selects invalid paths, we efficiently account for all hard invalid paths in training. Finally, our training is not iterative as in [12, 20], and does not require iterative refinement of action boundaries as in [7].
Our CDFL extends the loss used for training of the Graph Transformer Network (GTN) [15,13,2,5]. To the best of our knowledge, the GTN has been used only for text parsing, and never for action segmentation. In comparison with the GTN training, we significantly reduce complexity by building the video's segmentation graph. Also, while the loss used for training the GTN accounts for both valid and invalid text parses, it cannot handle the special case when valid parses have lower scores than invalid ones. In contrast, our CDFL effectively accounts for the energy of valid and invalid paths, even when valid paths have significantly lower energy than invalid paths in the segmentation graph.
## $3. \textbf{Our Model for Action Segmentation}$
Problem Setup: For each training video of length $T$, we are given unsupervised frame-level features, $x_{1:T}=$ $[x_1,x_2,...,x_T]$, and the ground-truth ordering of action classes $\boldsymbol{a}_1:N=[a_1,a_2,...,a_N]$, also referred to as the transcript. $N$ is the length of the annotation sequence, and $a_n$ is $n$th action class in $a_1:N$ that belongs to the set of $K$ action classes, $a_n\in\mathcal{A}=\{1,2,...,K\}.$ Note that $T$ and $N$ may vary across the training set, and that there may be more than one occurrences of the same action class spread out in $a_{1:N}$ (but of course $a_n\neq a_n+1).$
In inference, given frame features $x_{1:T}$ of a video, our goal is to find an optimal segmentation $(\hat{a}_{1:\hat{N}},\hat{\boldsymbol{l}}_{1:\hat{N}})$,where $\hat{N}$ is the predicted length of the action sequence, and $\hat{\boldsymbol{l}} _{1: \hat{N} }$ = $[ \hat{l} _{1}, \hat{l} _{2}, \cdots , \hat{l} _{\hat{N} }]$ includes the predicted number of video frames $\hat{l}_n$ occupied by the predicted action $\hat{a}_n.$
The Model: We use an HMM to model the posterior distribution of a video segmentation $(\boldsymbol a_{1:N},\boldsymbol l_{1:N})$ given $x_1:T$ as
$$\begin{aligned}p(\boldsymbol{a}_{1:N},\boldsymbol{l}_{1:N}|\boldsymbol{x}_{1:T})\\&\propto p(\boldsymbol{x}_{1:T}|\boldsymbol{a}_{1:N},\boldsymbol{l}_{1:N})p(\boldsymbol{l}_{1:N}|\boldsymbol{a}_{1:N})p(\boldsymbol{a}_{1:N}),\\&=\left(\prod_{t=1}^{T}p(x_{t}|a_{n(t)})\right)\left(\prod_{n=1}^{N}p(l_{n}|a_{n})\right)p(\boldsymbol{a}_{1:N}).\end{aligned}$$
(1)
In (1),the likelihood $p(x_t|a)$ is estimated as
$$p(x_t|a)\propto\frac{p(a|x_t)}{p(a)},$$
(2)
where $p(a|x_t)$ is the GRU's softmax score for action $a\in\mathcal{A}$ at frame $t$,and the prior distribution of action classes $p(a)$ is an normalized frame frequency of action occurrences in the training dataset. The likelihood of action length is modeled as a class-dependent Poisson distribution
$$p(l|a)=\frac{\lambda_a^l}{l!}e^{-\lambda_a},$$
(3)
where $\lambda_a$ is the mean length for class $a\in\mathcal{A}.$ Finally, the joint prior $p(\boldsymbol{a}_{1:N})$ is a constant if the transcript $a_1:N$ exists in the training set; otherwise, $p(\boldsymbol{a}_1:N)=0.$ The same modeling formulation was well-motivated and used in state of the art [22].
Constrained Viterbi Algorithm: Given a training video, we frst find an optimal valid action segmentation $(\hat{\boldsymbol{a}}_{1:\hat{N}},\hat{\boldsymbol{l}}_{1:\hat{N}})$ by maximizing (1) with a constrained Viterbi algorithm, which ensures that $\hat{a}_{1:\hat{N}}$ is equal to the annotated transcript, $\hat{\boldsymbol{a}}_{1:\hat{N}}=\boldsymbol{a}_{1:N}.$ Similarly, for inference on a test video, we first perform the constrained Viterbi algorithm against all transcripts $\{a_{1:N}\}$ seen in training, i.e., ensure that the predicted $\hat{a}_{1:\hat{N}}$ has been seen at least once in training. Thus, the initial step of our inference on a training or test video is the same as in [22].
Our key difference from [22], is that we use the initial $(\hat{\boldsymbol{a}}_{1:\hat{N}},\hat{\boldsymbol{l}}_{1:\hat{N}})$ to efficiently build a fully connected segmentation graph of the video, as explained in Sec. 4. Importantly, in training, the segmentation graph is not constructed to find a more optimal video segmentation that improves upon the initial prediction. Instead, the graph is used to efficiently account for all valid and invalid segmentations.
Given a video $x_1:T$ and a transcript $a_{1:N}$, the constrained Viterbi algorithm recursively maximizes the posterior in (1) such that the first $n$ action labels of the transcript $a_1:n=$ $[a_1,...,a_n]\subseteq a_{1:N}$ are respected at time $t:$
$$p(\boldsymbol{a}_{1:n},\hat{\boldsymbol{l}}_{1:n}|\boldsymbol{x}_{1:t})=\max_{t^{\prime},\:t^{\prime}<t}\bigg\{\:p(\boldsymbol{a}_{1:n-1},\hat{\boldsymbol{l}}_{1:n-1}|\boldsymbol{x}_{1:t^{\prime}})\\\cdot\left(\prod_{s=t^{\prime}}^{t}p(x_{s}|a_{n(s)})\right)\cdot p(l_{n}|a_{n})\cdot p(\boldsymbol{a}_{1:n})\bigg\},$$
(4)
where $l_n=t-t^{\prime}.$ We set $p(\cdot|\boldsymbol x_{1:0})=1$, and $p(\boldsymbol a_{1:n})=\kappa$,
where $\kappa>0$ is a constant.
# 4. Constructing the Segmentation Graph
Given a video $x_{1:T}$, we first run the constrained Viterbi algorithm to obtain an initial video segmentation $(\hat{\boldsymbol{a}}_{1:\hat{N}},\hat{\boldsymbol{l}}_{1:\hat{N}}).$For simplicity, in the following, we ignore the symbol[INDEX] This initial segmentation is characterized by | [
[
"0.0801",
"0.09216",
"0.468",
"0.09216",
"0.468",
"0.1061",
"0.0801",
"0.1061"
],
[
"0.501",
"0.0903",
"0.7935",
"0.09155",
"0.7935",
"0.1061",
"0.501",
"0.1048"
],
[
"0.0801",
"0.108",
"0.468",
"0.108",
"0.468",
"0.1218",
"0.0801",
"0.1218"
],
[
"0.0801",
"0.1225",
"0.4697",
"0.1225",
"0.4697",
"0.1364",
"0.0801",
"0.1364"
],
[
"0.625",
"0.125",
"0.755",
"0.1111",
"0.759",
"0.1332",
"0.6284",
"0.1465"
],
[
"0.8677",
"0.125",
"0.8945",
"0.125",
"0.8945",
"0.1414",
"0.8677",
"0.1414"
],
[
"0.0801",
"0.137",
"0.4673",
"0.137",
"0.4673",
"0.1509",
"0.0801",
"0.1509"
],
[
"0.7134",
"0.1364",
"0.7485",
"0.1364",
"0.7485",
"0.1483",
"0.7134",
"0.1483"
],
[
"0.0801",
"0.1528",
"0.468",
"0.1528",
"0.468",
"0.1666",
"0.0801",
"0.1666"
],
[
"0.5024",
"0.1578",
"0.8906",
"0.1578",
"0.8906",
"0.1724",
"0.5024",
"0.1724"
],
[
"0.0801",
"0.1674",
"0.4697",
"0.1666",
"0.4697",
"0.1812",
"0.0801",
"0.1818"
],
[
"0.5024",
"0.1749",
"0.892",
"0.1749",
"0.892",
"0.1881",
"0.5024",
"0.1881"
],
[
"0.0809",
"0.1843",
"0.468",
"0.1843",
"0.468",
"0.1945",
"0.0809",
"0.1945"
],
[
"0.5024",
"0.1901",
"0.892",
"0.1901",
"0.892",
"0.2039",
"0.5024",
"0.2039"
],
[
"0.0784",
"0.1963",
"0.468",
"0.1976",
"0.468",
"0.2122",
"0.0784",
"0.2108"
],
[
"0.5024",
"0.2058",
"0.892",
"0.2058",
"0.892",
"0.2197",
"0.5024",
"0.2197"
],
[
"0.0801",
"0.2134",
"0.468",
"0.2134",
"0.468",
"0.2267",
"0.0801",
"0.2267"
],
[
"0.501",
"0.2203",
"0.773",
"0.2203",
"0.773",
"0.2343",
"0.501",
"0.2343"
],
[
"0.0809",
"0.2279",
"0.468",
"0.2279",
"0.468",
"0.2412",
"0.0809",
"0.2412"
],
[
"0.0801",
"0.2437",
"0.468",
"0.2437",
"0.468",
"0.2568",
"0.0801",
"0.2568"
],
[
"0.6978",
"0.2437",
"0.7563",
"0.2507",
"0.7524",
"0.2695",
"0.694",
"0.262"
],
[
"0.6357",
"0.2551",
"0.694",
"0.2551",
"0.694",
"0.2695",
"0.6357",
"0.2695"
],
[
"0.8677",
"0.254",
"0.8945",
"0.254",
"0.8945",
"0.2708",
"0.8677",
"0.2708"
],
[
"0.0809",
"0.26",
"0.468",
"0.26",
"0.468",
"0.2708",
"0.0809",
"0.2708"
],
[
"0.7017",
"0.2651",
"0.7183",
"0.2651",
"0.7183",
"0.2734",
"0.7017",
"0.2734"
],
[
"0.0801",
"0.2727",
"0.468",
"0.2727",
"0.468",
"0.286",
"0.0801",
"0.286"
],
[
"0.0784",
"0.289",
"0.4714",
"0.289",
"0.4714",
"0.303",
"0.0784",
"0.303"
],
[
"0.5034",
"0.2854",
"0.892",
"0.2854",
"0.892",
"0.2986",
"0.5034",
"0.2986"
],
[
"0.0801",
"0.3037",
"0.3513",
"0.3037",
"0.3513",
"0.3176",
"0.0801",
"0.3176"
],
[
"0.501",
"0.3005",
"0.8896",
"0.2998",
"0.8896",
"0.3145",
"0.501",
"0.315"
],
[
"0.0989",
"0.318",
"0.468",
"0.318",
"0.468",
"0.332",
"0.0989",
"0.332"
],
[
"0.5024",
"0.3164",
"0.892",
"0.3164",
"0.892",
"0.33",
"0.5024",
"0.33"
],
[
"0.0809",
"0.3352",
"0.468",
"0.3352",
"0.468",
"0.3452",
"0.0809",
"0.3452"
],
[
"0.5024",
"0.3308",
"0.892",
"0.3308",
"0.892",
"0.3447",
"0.5024",
"0.3447"
],
[
"0.0801",
"0.3484",
"0.468",
"0.3484",
"0.468",
"0.3623",
"0.0801",
"0.3623"
],
[
"0.5034",
"0.3467",
"0.597",
"0.3467",
"0.597",
"0.358",
"0.5034",
"0.358"
],
[
"0.0801",
"0.363",
"0.468",
"0.363",
"0.468",
"0.377",
"0.0801",
"0.377"
],
[
"0.523",
"0.3599",
"0.892",
"0.3599",
"0.892",
"0.3738",
"0.523",
"0.3738"
],
[
"0.0801",
"0.379",
"0.468",
"0.379",
"0.468",
"0.3926",
"0.0801",
"0.3926"
],
[
"0.5024",
"0.3757",
"0.892",
"0.3757",
"0.892",
"0.3894",
"0.5024",
"0.3894"
],
[
"0.0801",
"0.3945",
"0.4697",
"0.3945",
"0.4697",
"0.4077",
"0.0801",
"0.4077"
],
[
"0.5024",
"0.3914",
"0.892",
"0.3894",
"0.892",
"0.404",
"0.5024",
"0.406"
],
[
"0.0801",
"0.4092",
"0.468",
"0.4092",
"0.468",
"0.4224",
"0.0801",
"0.4224"
],
[
"0.501",
"0.4048",
"0.8936",
"0.406",
"0.8936",
"0.4204",
"0.501",
"0.4192"
],
[
"0.0809",
"0.4255",
"0.468",
"0.4255",
"0.468",
"0.4363",
"0.0809",
"0.4363"
],
[
"0.5024",
"0.4216",
"0.8936",
"0.4216",
"0.8936",
"0.4363",
"0.5024",
"0.4363"
],
[
"0.0801",
"0.4395",
"0.4673",
"0.4395",
"0.4673",
"0.4526",
"0.0801",
"0.4526"
],
[
"0.501",
"0.4363",
"0.892",
"0.435",
"0.892",
"0.4495",
"0.501",
"0.4507"
],
[
"0.0801",
"0.455",
"0.468",
"0.455",
"0.468",
"0.4685",
"0.0801",
"0.4685"
],
[
"0.5034",
"0.452",
"0.892",
"0.452",
"0.892",
"0.4653",
"0.5034",
"0.4653"
],
[
"0.0801",
"0.4697",
"0.4673",
"0.4697",
"0.4673",
"0.483",
"0.0801",
"0.483"
],
[
"0.5034",
"0.4673",
"0.892",
"0.4673",
"0.892",
"0.4817",
"0.5034",
"0.4817"
],
[
"0.0801",
"0.4849",
"0.4592",
"0.4849",
"0.4592",
"0.5",
"0.0801",
"0.5"
],
[
"0.5024",
"0.4817",
"0.892",
"0.4817",
"0.892",
"0.4956",
"0.5024",
"0.4956"
],
[
"0.5024",
"0.4968",
"0.714",
"0.4968",
"0.714",
"0.5103",
"0.5024",
"0.5103"
],
[
"0.0809",
"0.512",
"0.4028",
"0.512",
"0.4028",
"0.5264",
"0.0809",
"0.5264"
],
[
"0.523",
"0.511",
"0.892",
"0.511",
"0.892",
"0.5244",
"0.523",
"0.5244"
],
[
"0.5034",
"0.526",
"0.892",
"0.526",
"0.892",
"0.5435",
"0.5034",
"0.5435"
],
[
"0.0801",
"0.536",
"0.4673",
"0.536",
"0.4673",
"0.55",
"0.0801",
"0.55"
],
[
"0.5024",
"0.5415",
"0.8906",
"0.5415",
"0.8906",
"0.556",
"0.5024",
"0.556"
],
[
"0.0801",
"0.552",
"0.4673",
"0.552",
"0.4673",
"0.5664",
"0.0801",
"0.5664"
],
[
"0.5024",
"0.5566",
"0.892",
"0.5566",
"0.892",
"0.571",
"0.5024",
"0.571"
],
[
"0.0784",
"0.5674",
"0.468",
"0.565",
"0.468",
"0.58",
"0.0784",
"0.582"
],
[
"0.501",
"0.5713",
"0.892",
"0.5713",
"0.892",
"0.5854",
"0.501",
"0.5854"
],
[
"0.0801",
"0.5835",
"0.468",
"0.5835",
"0.468",
"0.5967",
"0.0801",
"0.5967"
],
[
"0.501",
"0.586",
"0.8906",
"0.586",
"0.8906",
"0.601",
"0.501",
"0.601"
],
[
"0.0801",
"0.5977",
"0.4697",
"0.5977",
"0.4697",
"0.6113",
"0.0801",
"0.6113"
],
[
"0.5034",
"0.604",
"0.8105",
"0.604",
"0.8105",
"0.6143",
"0.5034",
"0.6143"
],
[
"0.0784",
"0.6113",
"0.4697",
"0.6123",
"0.4697",
"0.627",
"0.0784",
"0.6255"
],
[
"0.5215",
"0.617",
"0.892",
"0.617",
"0.892",
"0.632",
"0.5215",
"0.632"
],
[
"0.0801",
"0.6274",
"0.468",
"0.6274",
"0.468",
"0.6416",
"0.0801",
"0.6416"
],
[
"0.501",
"0.631",
"0.8906",
"0.632",
"0.8906",
"0.6465",
"0.501",
"0.646"
],
[
"0.0784",
"0.643",
"0.4697",
"0.6416",
"0.4697",
"0.655",
"0.0784",
"0.657"
],
[
"0.5024",
"0.648",
"0.8823",
"0.6484",
"0.8823",
"0.662",
"0.5024",
"0.661"
],
[
"0.0801",
"0.657",
"0.468",
"0.6577",
"0.468",
"0.6724",
"0.0801",
"0.672"
],
[
"0.5024",
"0.6636",
"0.786",
"0.662",
"0.786",
"0.677",
"0.5024",
"0.678"
],
[
"0.0784",
"0.672",
"0.259",
"0.6724",
"0.259",
"0.6875",
"0.0784",
"0.686"
],
[
"0.0972",
"0.688",
"0.468",
"0.688",
"0.468",
"0.7026",
"0.0972",
"0.7026"
],
[
"0.5186",
"0.6934",
"0.8735",
"0.6934",
"0.8735",
"0.7217",
"0.5186",
"0.7217"
],
[
"0.0784",
"0.702",
"0.4697",
"0.7007",
"0.4697",
"0.7183",
"0.0784",
"0.7197"
],
[
"0.0809",
"0.7183",
"0.468",
"0.7197",
"0.468",
"0.7344",
"0.0809",
"0.733"
],
[
"0.581",
"0.7266",
"0.8267",
"0.7266",
"0.8267",
"0.7485",
"0.581",
"0.7485"
],
[
"0.0768",
"0.735",
"0.4697",
"0.732",
"0.4697",
"0.75",
"0.0768",
"0.7524"
],
[
"0.0801",
"0.7524",
"0.4248",
"0.7524",
"0.4248",
"0.7666",
"0.0801",
"0.7666"
],
[
"0.5786",
"0.752",
"0.6006",
"0.752",
"0.6006",
"0.755",
"0.5786",
"0.755"
],
[
"0.8716",
"0.759",
"0.8896",
"0.759",
"0.8896",
"0.768",
"0.8716",
"0.768"
],
[
"0.10046",
"0.7666",
"0.4656",
"0.7666",
"0.4656",
"0.78",
"0.10046",
"0.78"
],
[
"0.5024",
"0.771",
"0.8906",
"0.773",
"0.8906",
"0.787",
"0.5024",
"0.785"
],
[
"0.0784",
"0.782",
"0.468",
"0.7827",
"0.468",
"0.7974",
"0.0784",
"0.797"
],
[
"0.5034",
"0.789",
"0.6763",
"0.789",
"0.6763",
"0.8",
"0.5034",
"0.8"
],
[
"0.0801",
"0.8003",
"0.09644",
"0.8003",
"0.09644",
"0.81",
"0.0801",
"0.81"
],
[
"0.0989",
"0.8184",
"0.2263",
"0.8174",
"0.2263",
"0.832",
"0.0989",
"0.8335"
],
[
"0.5024",
"0.8145",
"0.8433",
"0.815",
"0.8433",
"0.83",
"0.5024",
"0.829"
],
[
"0.1299",
"0.832",
"0.4207",
"0.8306",
"0.4207",
"0.8486",
"0.1299",
"0.8496"
],
[
"0.5186",
"0.8384",
"0.892",
"0.839",
"0.892",
"0.857",
"0.5186",
"0.856"
],
[
"0.2917",
"0.851",
"0.304",
"0.851",
"0.304",
"0.855",
"0.2917",
"0.855"
],
[
"0.1708",
"0.8594",
"0.2607",
"0.8594",
"0.2607",
"0.8774",
"0.1708",
"0.8774"
],
[
"0.2876",
"0.8594",
"0.3694",
"0.8594",
"0.3694",
"0.877",
"0.2876",
"0.877"
],
[
"0.3857",
"0.86",
"0.4453",
"0.86",
"0.4453",
"0.875",
"0.3857",
"0.875"
],
[
"0.505",
"0.856",
"0.8936",
"0.856",
"0.8936",
"0.869",
"0.505",
"0.869"
],
[
"0.5024",
"0.8706",
"0.8945",
"0.8687",
"0.8945",
"0.886",
"0.5024",
"0.888"
],
[
"0.2876",
"0.881",
"0.308",
"0.88",
"0.308",
"0.884",
"0.2876",
"0.8853"
],
[
"0.4468",
"0.887",
"0.4697",
"0.887",
"0.4697",
"0.901",
"0.4468",
"0.901"
],
[
"0.5024",
"0.886",
"0.892",
"0.886",
"0.892",
"0.901",
"0.5024",
"0.901"
]
] | [
"ual frames. In contrast, we use an energy-based framework",
"In (1), the likelihood p(cta) is estimated as",
"with the following differences. First, in training, we min-",
"imize the total energy of valid paths in the segmentation",
"p(at/a) α p(art)",
"(2)",
"graph rather than optimize labeling probabilities of each",
"p(a)",
"frame. Second, instead of considering a single optimal valid",
"where p(alct) is the GRU's softmax score for action a E A",
"path in the segmentation graph, we specify a loss function",
"at frame t, and the prior distribution of action classes p(a) is",
"in terms of all valid paths.Hence,the Viterbi-initialized",
"an normalized frame frequency of action occurrences in the",
"training on pseudo-labels of frames [22] represents a spe-",
"training dataset. The likelihood of action length is modeled",
"cial case of our training done only for one valid path. In ad-",
"as a class-dependent Poisson distribution",
"dition, our loss enforces discriminative training by account-",
"ing for invalid paths in the segmentation graph. Unlike [3]",
"Xae-a",
"p(la) =",
"(3)",
"thatrandomly selects invalidpaths,we efficiently account",
"1",
"for all hard invalid paths in training. Finally, our training",
"is not iterative as in [12, 20], and does not require iterative",
"where a is the mean length for class a E A. Finally, the",
"refinement of action boundaries as in [7].",
"joint prior p(ai:N) is a constant if the transcript ai:N ex-",
"Our CDFL extends the loss used for training of the Graph",
"ists in the training set; otherwise, p(ai:N) = 0. The same",
"TransformerNetwork(GTN)[15.13,2,51.Tothebestof",
"modeling formulation was well-motivated and used in state",
"our knowledge, the GTN has been used only for text pars-",
"of the art [22].",
"ing,and never for action segmentation. In comparison with",
"Constrained Viterbi Algorithm: Given a training",
"the GTN training, we significantly reduce complexity by",
"video, we first find an optimal valid action segmentation",
"building the video's segmentation graph. Also, while the",
"(ai:N,lt:) by maximizing (1) with a constrained Viterbi",
"loss used for training the GTN accounts for both valid and",
"algorithm, which ensures that ay. is equal to the annotated",
"invalid textparses.it cannothandlethe special casewhen",
"transcript, ai. = ai:N. Similarly, for inference on a test",
"valid parses have lower scores than invalid ones. In con-",
"video, we first perform the constrained Viterbi algorithm",
"trast, our CDFL effectively accounts for the energy of valid",
"against all transcripts fai:N] seen in training, i.e., ensure",
"and invalid paths, even when valid paths have significantly",
"that the predicted ar: has been seen at least once in train-",
"lower energy than invalid paths in the segmentation graph",
"ing. Thus, the initial step of our inference on a training or",
"test video is the same as in [22].",
"3. Our Model for Action Segmentation",
"Our key difference from [22], is that we use the initial",
"(ai.N,i1:N) to effciently build a fully connected segmenta-",
"Problem Setup: For each training video of length T",
"tion graph of the video, as explained in Sec. 4. Importantly,",
"we are given unsupervised frame-level features, &i:T =",
"in training, the segmentation graph is not constructed to find",
"[ri, 2,.., ], and the ground-truth ordering of action",
"a more optimal video segmentation that improves upon the",
"classes ai:N =[ai, a2, ..., an], also referred to as the tran-",
"initial prediction. Instead, the graph is used to efficiently",
"script. N is the length of the annotation sequence, and an is",
"accountforallvalidandinvalidsegmentations.",
"nth action class in ai:N that belongs to the set of K action",
"Given a video i:T and a transcript ai:N, the constrained",
"classes, an E A = [1, 2, .., K]. Note that T and N may",
"Viterbi algorithm recursively maximizes the posterior in ()",
"vary across thetraining set, and that there may bemore than",
"such that the first n action labels of the transcript ai:n",
"one occurrences of the same action class spread out in ai:N",
"[ai,.., an] C ai:N are respected at time t:",
"(butof courseanan+1).",
"In inference, given frame features i:T of a video, our",
"p(a1:n,lin/l1:t) = (max p(ain-1,lin-1l1:e)",
"goal is to find an optimal segmentation (a1:N,ii.N), where",
"N is the predicted length of the action sequence, and",
"I p(aslan(s) ) - p(ln[an) p(a1:n)",
"i1.N = [i,i2,... ,ix] includes the predicted number of",
"video frames in occupied by the predicted action an.",
"",
"(4)",
"The Model: We use an HMM to model the posterior dis-",
"where ln = t - t'. We set p(ai:o) = 1, and p(a1:n) = k,",
"tribution of a video segmentation (ai:N,li:N) given Ci:T",
"where k > 0is a constant.",
"as",
"p(a1:N,II:NCI:T)",
"4.Constructing the Segmentation Graph",
"α p(1:Ta1:N,I:N)p(lI:Na1:N)p(a1:N),",
"Given a video i:T, we first run the constrained",
"",
"Ip(rtlan(t)",
"II p(inlan",
"p(ai:N)",
"Viterbi algorithm to obtain an initial video segmentation",
"(ai:N,ii.N). For simplicity, in the following, we ignore",
"",
"(1)",
"the symbol . This initial segmentation is characterized by"
] | {
"arxiv_id": "1909.13155",
"language": "en",
"page_idx": 2
} |
||
1909.13155-en-3 | 1909.13155-en-3.jpg | <FigureHere>
Figure 2. Building the segmentation graph $G$ (best seen in
color). The initial nodes of $G$ represent segmentation cuts $b_n$ obtained in the Constrained Viterbi (the predicted action classes are marked with different colors). Each $b_n$ generates additional vertices $b_n=\{v_{ns}\}$ representing neighboring video frames within a window centered at $b_n$ (the black rectangles), and corresponding new edges $(v_{ns},v_{n^{\prime}s^{\prime}})$ (the dashed lines) between all temporally ordered pairs of vertices in $G.$ For clarity, we show only a few edges. $G$ has exponential many paths, each representing a candidate action segmentation.
$N+1$ cuts, $b_{1:N+1}=[b_{1},\ldots,b_{N+1}]$, i.e., video frames where previous action ends and the next one starts including the very first frame $b_1$ and last frame $b_N+1$ at time $T.$
We use these cuts to anchor our construction of the fully connected segmentation graph, $G=(\mathcal{V},\mathcal{E},\mathcal{W})$, where $\mathcal{V}=$ $\{\boldsymbol{b}_{1:N+1}\}$ is the set of nodes, $\mathcal{E}$ is the set of directed edges linking every two temporally ordered nodes, and $\mathcal{W}$ are the corresponding edge weights.
Some of the estimated cuts in $b_{1:N+1}$ may be false positives or may not exactly coincide with the true cuts. To improve action boundary detection, we augment the initial $\mathcal{V}$ with nodes representing neighboring video frames of each cut $b_n$ within a temporal window of length $\Delta$ centered at $b_n$, as illustrated in Fig.2. For the first and last frames, we set $\Delta=1.$ Thus, each $b_n$ can be viewed as a hyper-node comprising additional vertices in $G,\mathcal{V}=\{b_n=$ $\{v_{n1},\cdots,v_{ni},\cdots,v_{n\Delta}\}:n=1,\ldots,N+1\}$,and accordingly additional edges $\mathcal{E} = \{ ( v_{ni}, v_{n^{\prime }i^{\prime }}) : n\leq n^{\prime }$, $i< i^{\prime }\} .$ In the following, we simplify notation for vertices $v_{ni}\to$ $v_i\in\mathcal{V}$,and edges $(v_ni,v_{n^{\prime}i^{\prime}})\to e_{ii^{\prime}}=(v_i,v_{i^{\prime}}).$
Each edge $e_{ii^{\prime}}$ is assigned a weight vector $w_{ii^{\prime}}=$ $[w_{ii^{\prime}}(a)]$, where $w_ii^{\prime}(a)$ is defined as the energy of labeling the video segment $(v_i,v_{i^{\prime}})$ with action class $a\in\mathcal{A}:$
(5)
$$w_{ii'}(a)=\sum_{t\in(v_i,v_{i'})}-\log p(a|x_t),$$
where $p(a|x_t)$ is the GRU's softmax score for action $a$ at
frame $t.$
$G$ comprises exponentially many directed paths $\mathcal{P}=$ $\{\pi\}$, where each $\pi$ represents a particular video segmentation. In each $\pi$, every edge $e_ii^{\prime}$ gets assigned only one action class $a_{ii^{\pi}}\in\mathcal{A}.$ Thus, the very same edge with $K$ different class assignments belongs to $K$ distinct paths in $\mathcal{P}.$ We compute the energy of a path as
(6)
$$E_{\pi}=\sum_{e_{ii'}\in\pi}w_{ii'}(a_{ii'}^{\pi}).$$
A subset of valid paths $\mathcal{P}^V\subset\mathcal{P}$ satisfies the given tran-
script. The other paths are invalid, $\mathcal{P}^I=\mathcal{P}\setminus\mathcal{P}^{V}.$
In the next section, we explain how to efficiently compute a total energy score of the exponentially many paths in $\mathcal{P}$ for estimating our loss in training.
5. Constrained Discriminative Forward Loss
In this paper, we study three distinct loss functions, defined in terms of a total energy score of paths in $G.$ As there are exponentially many paths in $G$, our key contribution is the algorithm for efficiently estimating their total energy. Below, we specify our three loss functions ordered by their complexity. As we will show in Sec. 6, we obtain the best performance when using the CDFL in training.
## $5. 1. \textbf{Forward Loss}$
We define a forward loss, $L_\mathrm{F}$, in terms of a total energy
of all valid paths using the standard logadd function as
(7)
$$L_{\mathrm{F}}=-\log(\sum_{\pi\in P^{V}}\exp(-E_{\pi})),$$
where energy of a path $E_\pi$ is given by (6). As there are exponentially many paths in $\mathcal{P}^V$,we cannot directly compute $L_{\mathrm{F}}$ as specified in (7). Therefore, we derive a novel recursive algorithm for accumulating the energy scores of edges along multiple paths, as specified below.
We begin by defining the logadd function as
$$\mathrm{logadd}(a,b)=-\log(\exp(-a)+\exp(-b)).$$
(8)
Note that the logadd function is commutative and associative, so it can be defined on a set $S$ in a recursive manner:
(9)
$$\mathrm{logadd}(S)=\mathrm{logadd}(S\backslash\{x\},x),$$
where $x$ is an element in $S.$ Therefore, the forward loss
given by (7) can be expressed as
(10)
$$L_{\mathrm{F}}=\mathrm{logadd}(\{E_{\pi}:\pi\in\mathcal{P}^{V}\}).$$
Below, we simplify notation as $L_{\mathrm{F}}=$logadd$(\mathcal{P}^V).$
We recursively compute the energy score $\ell_{i^{\prime}}(a_{1:n})$ of a path that ends at node $i^\prime$ and covers first $n$ labels of the ground truth $a_1:n=[a_1,...,a_n]\subseteq a_{1:N}$ in terms of the logadd scores $\ell_i(\boldsymbol{a}_{1:n-1})$ of all valid paths that end at node $i,i<i^{\prime}$,and cover first $n-1$ labels as
$$\ell_{i'}(\boldsymbol{a}_{1:n})=\mathrm{logadd}(\{\ell_{i}(\boldsymbol{a}_{1:n-1})+w_{ii^{\prime}}(a_{n}):i<i^{\prime}\}).$$
(11)
To prove (11),suppose that
$$\begin{aligned}\ell_{i}(\boldsymbol{a}_{1:n-1})&=\mathrm{logadd}(\{E_{\pi_{i}}:\pi_{i}\in\mathcal{P}^{V}\})\\&=-\log(\sum_{\pi_{i}\in\mathcal{P}^{V}}\exp(-E_{\pi_{i}})),\end{aligned}$$
(12) | [
[
"0.5186",
"0.0884",
"0.8936",
"0.089",
"0.8936",
"0.1067",
"0.5186",
"0.1061"
],
[
"0.5024",
"0.1067",
"0.8228",
"0.1048",
"0.8228",
"0.1193",
"0.5024",
"0.1218"
],
[
"0.5215",
"0.1225",
"0.8906",
"0.1225",
"0.8906",
"0.1364",
"0.5215",
"0.1364"
],
[
"0.5034",
"0.1383",
"0.892",
"0.1383",
"0.892",
"0.1515",
"0.5034",
"0.1515"
],
[
"0.4453",
"0.1445",
"0.468",
"0.1445",
"0.468",
"0.1483",
"0.4453",
"0.1483"
],
[
"0.0801",
"0.1572",
"0.4697",
"0.1572",
"0.4697",
"0.1704",
"0.0801",
"0.1704"
],
[
"0.5024",
"0.1515",
"0.7485",
"0.1528",
"0.7485",
"0.1674",
"0.5024",
"0.1666"
],
[
"0.0801",
"0.1704",
"0.468",
"0.1704",
"0.468",
"0.1843",
"0.0801",
"0.1843"
],
[
"0.0801",
"0.1849",
"0.4697",
"0.1849",
"0.4697",
"0.1989",
"0.0801",
"0.1989"
],
[
"0.5034",
"0.1812",
"0.875",
"0.1812",
"0.875",
"0.1945",
"0.5034",
"0.1945"
],
[
"0.0784",
"0.1976",
"0.4673",
"0.1989",
"0.4673",
"0.2122",
"0.0784",
"0.2108"
],
[
"0.5215",
"0.2058",
"0.8906",
"0.2058",
"0.8906",
"0.2197",
"0.5215",
"0.2197"
],
[
"0.0784",
"0.2134",
"0.4697",
"0.2134",
"0.4697",
"0.2267",
"0.0784",
"0.2267"
],
[
"0.501",
"0.2197",
"0.892",
"0.2203",
"0.892",
"0.2349",
"0.501",
"0.2343"
],
[
"0.0801",
"0.2267",
"0.468",
"0.2267",
"0.468",
"0.2405",
"0.0801",
"0.2405"
],
[
"0.0801",
"0.2424",
"0.4656",
"0.2424",
"0.4656",
"0.2524",
"0.0801",
"0.2524"
],
[
"0.505",
"0.238",
"0.8906",
"0.238",
"0.8906",
"0.2488",
"0.505",
"0.2488"
],
[
"0.0801",
"0.2551",
"0.468",
"0.2551",
"0.468",
"0.2683",
"0.0801",
"0.2683"
],
[
"0.5024",
"0.252",
"0.8906",
"0.252",
"0.8906",
"0.2651",
"0.5024",
"0.2651"
],
[
"0.0801",
"0.2683",
"0.468",
"0.2683",
"0.468",
"0.2822",
"0.0801",
"0.2822"
],
[
"0.5024",
"0.2664",
"0.8906",
"0.2664",
"0.8906",
"0.2798",
"0.5024",
"0.2798"
],
[
"0.0801",
"0.2827",
"0.232",
"0.2827",
"0.232",
"0.2935",
"0.0801",
"0.2935"
],
[
"0.5024",
"0.2822",
"0.892",
"0.2822",
"0.892",
"0.2954",
"0.5024",
"0.2954"
],
[
"0.5024",
"0.2974",
"0.8105",
"0.2974",
"0.8105",
"0.3113",
"0.5024",
"0.3113"
],
[
"0.0809",
"0.324",
"0.4697",
"0.324",
"0.4697",
"0.3384",
"0.0809",
"0.3384"
],
[
"0.5034",
"0.3208",
"0.6445",
"0.3208",
"0.6445",
"0.3352",
"0.5034",
"0.3352"
],
[
"0.0809",
"0.339",
"0.468",
"0.339",
"0.468",
"0.3528",
"0.0809",
"0.3528"
],
[
"0.5195",
"0.3416",
"0.8906",
"0.3435",
"0.8906",
"0.358",
"0.5195",
"0.356"
],
[
"0.0801",
"0.3542",
"0.4558",
"0.3542",
"0.4558",
"0.3687",
"0.0801",
"0.3687"
],
[
"0.5024",
"0.3591",
"0.863",
"0.3591",
"0.863",
"0.3726",
"0.5024",
"0.3726"
],
[
"0.10046",
"0.3694",
"0.4673",
"0.3694",
"0.4673",
"0.3833",
"0.10046",
"0.3833"
],
[
"0.0801",
"0.3838",
"0.4656",
"0.3838",
"0.4656",
"0.3977",
"0.0801",
"0.3977"
],
[
"0.5923",
"0.3838",
"0.8013",
"0.385",
"0.8013",
"0.4028",
"0.5923",
"0.4016"
],
[
"0.869",
"0.387",
"0.892",
"0.387",
"0.892",
"0.4016",
"0.869",
"0.4016"
],
[
"0.0825",
"0.401",
"0.468",
"0.401",
"0.468",
"0.4153",
"0.0825",
"0.4153"
],
[
"0.6772",
"0.406",
"0.7134",
"0.4016",
"0.716",
"0.4124",
"0.6807",
"0.4167"
],
[
"0.0809",
"0.4172",
"0.4673",
"0.4172",
"0.4673",
"0.428",
"0.0809",
"0.428"
],
[
"0.0801",
"0.4312",
"0.2688",
"0.4312",
"0.2688",
"0.4443",
"0.0801",
"0.4443"
],
[
"0.5034",
"0.4255",
"0.8906",
"0.4255",
"0.8906",
"0.4395",
"0.5034",
"0.4395"
],
[
"0.5024",
"0.44",
"0.892",
"0.4395",
"0.892",
"0.4539",
"0.5024",
"0.455"
],
[
"0.0972",
"0.4463",
"0.4697",
"0.4463",
"0.4697",
"0.461",
"0.0972",
"0.461"
],
[
"0.0801",
"0.461",
"0.4697",
"0.461",
"0.4697",
"0.4749",
"0.0801",
"0.4749"
],
[
"0.5034",
"0.4558",
"0.8906",
"0.4558",
"0.8906",
"0.4697",
"0.5034",
"0.4697"
],
[
"0.0801",
"0.4766",
"0.468",
"0.4766",
"0.468",
"0.4905",
"0.0801",
"0.4905"
],
[
"0.5024",
"0.4717",
"0.892",
"0.4717",
"0.892",
"0.4849",
"0.5024",
"0.4849"
],
[
"0.5034",
"0.486",
"0.7695",
"0.486",
"0.7695",
"0.5",
"0.5034",
"0.5"
],
[
"0.0801",
"0.4924",
"0.4697",
"0.4924",
"0.4697",
"0.506",
"0.0801",
"0.506"
],
[
"0.0801",
"0.507",
"0.468",
"0.507",
"0.468",
"0.521",
"0.0801",
"0.521"
],
[
"0.523",
"0.502",
"0.814",
"0.502",
"0.814",
"0.515",
"0.523",
"0.515"
],
[
"0.0801",
"0.523",
"0.4697",
"0.523",
"0.4697",
"0.536",
"0.0801",
"0.536"
],
[
"0.5483",
"0.526",
"0.851",
"0.5264",
"0.851",
"0.5415",
"0.5483",
"0.5405"
],
[
"0.8677",
"0.528",
"0.8906",
"0.528",
"0.8906",
"0.5405",
"0.8677",
"0.5405"
],
[
"0.0801",
"0.537",
"0.4697",
"0.537",
"0.4697",
"0.5503",
"0.0801",
"0.5503"
],
[
"0.0801",
"0.552",
"0.468",
"0.552",
"0.468",
"0.5664",
"0.0801",
"0.5664"
],
[
"0.5034",
"0.553",
"0.8896",
"0.553",
"0.8896",
"0.5664",
"0.5034",
"0.5664"
],
[
"0.0809",
"0.5684",
"0.4673",
"0.5664",
"0.4673",
"0.5806",
"0.0809",
"0.5835"
],
[
"0.5024",
"0.5684",
"0.8823",
"0.5684",
"0.8823",
"0.582",
"0.5024",
"0.582"
],
[
"0.0801",
"0.5835",
"0.4673",
"0.5835",
"0.4673",
"0.5967",
"0.0801",
"0.5967"
],
[
"0.5884",
"0.5947",
"0.8057",
"0.5947",
"0.8057",
"0.6094",
"0.5884",
"0.6094"
],
[
"0.871",
"0.5947",
"0.892",
"0.5947",
"0.892",
"0.608",
"0.871",
"0.608"
],
[
"0.0801",
"0.613",
"0.4004",
"0.613",
"0.4004",
"0.6274",
"0.0801",
"0.6274"
],
[
"0.0801",
"0.5977",
"0.468",
"0.5977",
"0.468",
"0.6113",
"0.0801",
"0.6113"
],
[
"0.5034",
"0.6206",
"0.892",
"0.6206",
"0.892",
"0.634",
"0.5034",
"0.634"
],
[
"0.0989",
"0.629",
"0.4478",
"0.629",
"0.4478",
"0.6426",
"0.0989",
"0.6426"
],
[
"0.5034",
"0.636",
"0.7197",
"0.636",
"0.7197",
"0.6494",
"0.5034",
"0.6494"
],
[
"0.0801",
"0.643",
"0.468",
"0.6426",
"0.468",
"0.657",
"0.0801",
"0.6577"
],
[
"0.0784",
"0.659",
"0.4436",
"0.6577",
"0.4436",
"0.6724",
"0.0784",
"0.674"
],
[
"0.591",
"0.659",
"0.8",
"0.6577",
"0.8",
"0.6753",
"0.591",
"0.677"
],
[
"0.863",
"0.662",
"0.892",
"0.662",
"0.892",
"0.675",
"0.863",
"0.675"
],
[
"0.1626",
"0.6914",
"0.2361",
"0.6914",
"0.2361",
"0.7056",
"0.1626",
"0.7056"
],
[
"0.29",
"0.6914",
"0.3857",
"0.6914",
"0.3857",
"0.7056",
"0.29",
"0.7056"
],
[
"0.501",
"0.686",
"0.8325",
"0.685",
"0.8325",
"0.6997",
"0.501",
"0.7007"
],
[
"0.4468",
"0.6924",
"0.4697",
"0.6924",
"0.4697",
"0.705",
"0.4468",
"0.705"
],
[
"0.523",
"0.702",
"0.8906",
"0.702",
"0.8906",
"0.7163",
"0.523",
"0.7163"
],
[
"0.232",
"0.7104",
"0.29",
"0.7104",
"0.29",
"0.7217",
"0.232",
"0.7217"
],
[
"0.5034",
"0.7183",
"0.8906",
"0.7183",
"0.8906",
"0.732",
"0.5034",
"0.732"
],
[
"0.0809",
"0.7363",
"0.4697",
"0.7363",
"0.4697",
"0.75",
"0.0809",
"0.75"
],
[
"0.5024",
"0.733",
"0.8906",
"0.732",
"0.8906",
"0.747",
"0.5024",
"0.7476"
],
[
"0.0809",
"0.752",
"0.1323",
"0.752",
"0.1323",
"0.763",
"0.0809",
"0.763"
],
[
"0.5034",
"0.7485",
"0.892",
"0.7485",
"0.892",
"0.763",
"0.5034",
"0.763"
],
[
"0.0989",
"0.7666",
"0.46",
"0.7666",
"0.46",
"0.78",
"0.0989",
"0.78"
],
[
"0.5024",
"0.763",
"0.761",
"0.763",
"0.761",
"0.7764",
"0.5024",
"0.7764"
],
[
"0.0825",
"0.782",
"0.4673",
"0.782",
"0.4673",
"0.7954",
"0.0825",
"0.7954"
],
[
"0.517",
"0.789",
"0.8735",
"0.789",
"0.8735",
"0.8037",
"0.517",
"0.8037"
],
[
"0.0784",
"0.797",
"0.468",
"0.7974",
"0.468",
"0.8125",
"0.0784",
"0.811"
],
[
"0.8594",
"0.803",
"0.8936",
"0.803",
"0.8936",
"0.8184",
"0.8594",
"0.8184"
],
[
"0.0801",
"0.813",
"0.468",
"0.813",
"0.468",
"0.827",
"0.0801",
"0.827"
],
[
"0.5034",
"0.8193",
"0.6846",
"0.8193",
"0.6846",
"0.834",
"0.5034",
"0.834"
],
[
"0.0809",
"0.8276",
"0.4697",
"0.8276",
"0.4697",
"0.8423",
"0.0809",
"0.8423"
],
[
"0.0801",
"0.8433",
"0.339",
"0.8433",
"0.339",
"0.857",
"0.0801",
"0.857"
],
[
"0.562",
"0.8433",
"0.8286",
"0.8423",
"0.8286",
"0.86",
"0.562",
"0.8613"
],
[
"0.636",
"0.8643",
"0.835",
"0.8643",
"0.835",
"0.8853",
"0.636",
"0.8853"
],
[
"0.2009",
"0.8735",
"0.2427",
"0.8735",
"0.2427",
"0.888",
"0.2009",
"0.888"
],
[
"0.8613",
"0.8687",
"0.8906",
"0.8687",
"0.8906",
"0.881",
"0.8613",
"0.881"
],
[
"0.267",
"0.875",
"0.3489",
"0.875",
"0.3489",
"0.8896",
"0.267",
"0.8896"
],
[
"0.4453",
"0.875",
"0.4697",
"0.875",
"0.4697",
"0.888",
"0.4453",
"0.888"
],
[
"0.699",
"0.886",
"0.7363",
"0.886",
"0.7363",
"0.898",
"0.699",
"0.898"
],
[
"0.255",
"0.879",
"0.2607",
"0.879",
"0.2607",
"0.8853",
"0.255",
"0.8853"
],
[
"0.2418",
"0.8926",
"0.2793",
"0.8926",
"0.2793",
"0.904",
"0.2418",
"0.904"
]
] | [
"A subset of valid paths PV c P satisfies the given tran-",
"script. The other paths are invalid, pI = PpV",
"In the next section, we explain how to efficiently com-",
"pute a total energy score of the exponentially many paths in",
"1tim",
"Figure 2. Building the segmentation graph G (best seen in",
"P for estimating our loss in training.",
"color). The initial nodes of G represent segmentation cuts bn ob-",
"tained in the Constrained Viterbi (the predicted action classes are",
"5. Constrained Discriminative Forward Loss",
"marked with different colors). Each bn generates additional ver-",
"In this paper, we study three distinct loss functions, de-",
"tices bn = [uns] representing neighboring video frames within a",
"fined in terms of a total energy score of paths in G. As there",
"window centered at bn (the black rectangles), and corresponding",
"newedges(vns:Vn's)(thedashedlines)betweenalltemporally",
"areexponentiallymanypathsinG,ourkeycontributionis",
"ordered pairs of vertices in G. For clarity, we show only a few",
"the algorithm for efficiently estimating their total energy.",
"edges. G has exponential many paths, each representing a candi-",
"Below, we specify our three loss functions ordered by their",
"date action segmentation.",
"complexity. As we will show in Sec. 6, we obtain the best",
"performance when using the CDFL in training.",
"N + 1 cuts, biN+1 =[bi,...,bn+i], i.e., video frames",
"5.1. Forward Loss",
"where previous action ends and the next one starts includ-",
"We define a forward loss, Lr, in terms of a total energy",
"ing the very first frame bi and last frame bN+1 at time T.",
"of all valid paths using the standard logadd function as",
"We use these cuts to anchor our construction of the fully",
"connected segmentation graph, G = (V..W),where V",
"Lp = -log( exp(-Ex)),",
"(7)",
"[bi:N+i] is the set of nodes, & is the set of directed edges",
"TEPV",
"linking everytwotemporallyordered nodes,andWarethe",
"corresponding edge weights.",
"where energy of a path E is given by (6). As there are ex-",
"ponentially many paths in py, we cannot directly compute",
"Some of the estimated cuts in bi:N+i may be false pos-",
"itives or may not exactly coincide with the true cuts. To",
"L as specified in (7). Therefore, we derive a novel recur-",
"improve action boundary detection, we augment the ini-",
"sive algorithm for accumulating the energy scores of edges",
"along multiple paths, as specified below.",
"tial V with nodes representing neighboring video frames",
"of each cut bn within a temporal window of length △ cen-",
"We begin by defining the logadd function as",
"tered at bn, as illustrated in Fig. 2. For the first and last",
"logadd(a, b) = -log(exp(-a) + exp(-b).",
"(8)",
"frames, we set = 1. Thus, each bn can be viewed as a",
"hyper-node comprising additional vertices in G, V ={bn =",
"Note that the logadd function is commutative and associa-",
"[unl,.., Uni,...,Una] : n = 1,..., N+1], and accord-",
"tive, so it can be defined on a set S in a recursive manner:",
"ingly additional edges & = {(uni, Un'i): n ≤ n', i< i'}.",
"logadd(S) = logadd(S)(r), a),",
"(9)",
"Vi E V, and edges (Vni, Vn'i) → eii = (vi, V,).",
"In the following, we simplify notation for vertices nz →",
"where is an element in S. Therefore, the forward loss",
"Each edge eirr is assigned a weight vector wii",
"given by (7) can be expressed as ",
"[wi(a)], where wi (a) is defined as the energy of label-",
"ing the video segment (vi, Ui) with action class a E A:",
"Lp = logadd({E元 : π E PV).",
"(10)",
"Wi(a) =",
"-log p(act),",
"Below, we simplify notation as Le =logadd(pV)",
"(5)",
"We recursively compute the energy score l; (ai:n) of a",
"tE(Vi,V)",
"path that ends at node i' and covers first n labels of the",
"where p(at) is the GRU's softmax score for action a at",
"ground truth ai:n = [ai, .., an] C ai:N in terms of the",
"frame t.",
"logadd scores l,(ai:n-1) of all valid paths that end at node",
"G comprises exponentially many directed paths P:",
"i, i< i,and cover first n -1 labels as",
"fh,where each represents a particular video segmen-",
"l,(ai:n) = logadd((l;(ai:n-1) +wi(an) : i< i)",
"tation. In each , every edge ei gets assigned only one",
"(11)",
"action class a, E A. Thus, the very same edge with K",
"To prove (11), suppose that",
"different class assignments belongs to K distinct paths in",
"P. We compute the energy of a path as",
"l;(ai:n-1) = logadd([E :元 E PV)",
"=- log( Z exp(-Ew,),",
"E元=",
"(12)",
"Wr(an).",
"(6)",
"T,EPI",
"Y",
"eRET"
] | {
"arxiv_id": "1909.13155",
"language": "en",
"page_idx": 3
} |
||
2306.04004-en-0 | 2306.04004-en-0.jpg | $\textbf{Randomized Schur Complement Views for Graph Contrastive Learning}$
Vignesh Kothapalli Pis
## $\mathbf{Abstract}$
We introduce a randomized topological augmentor based on Schur complements for Graph Contrastive Learning (GCL). Given a graph laplacian matrix, the technique generates unbiased approximations of its Schur complements and treats the corresponding graphs as augmented views. We discuss the benefits of our approach, provide theoretical justifications and present connections with graph diffusion. Unlike previous efforts, we study the empirical effectiveness of the augmentor in a controlled fashion by varying the design choices for subsequent GCL phases, such as encoding and contrasting. Extensive experiments on node and graph classification benchmarks demonstrate that our technique consistently outperforms predefined and adaptive augmentation approaches to achieve state-of-the-art results.
## $1. \textbf{Introduction}$
Understanding the structural properties and semantics of graph data typically requires domain expertise and efficient computational tools. Advances in machine learning techniques such as Graph Neural Networks (GNN) (Gori et al., 2005; Scarselli et al., 2008; Bruna et al., 2014; Henaff et al., 2015; Welling & Kipf, 2016; Niepert et al., 2016; Bronstein et al., 2017; Xu et al., 2018; Wu et al., 2020; Zhou et al., 2020) have paved a path for representation learning on internet scale graphs and significantly alleviated the human effort. However, real-world graphs such as social networks, citation networks, supply chains and media networks are continuously evolving, which makes labeling an extremely challenging task to accomplish. Self-supervised learning (SSL) addresses this issue by optimizing pretext objectives and learning generalizable representations for downstream tasks (Jin et al., 2020; Wu et al., 2021; Liu et al., 2022; Xie et al., 2022). This learning paradigm has been quite popular
$^{1}$Courant Institute of Mathematical Sciences, New York University, New York, USA. Correspondence to: Vignesh Kothapalli <vk2115@nyu.edu>.
Proceedings of the 40$^{th}$ International Conference on Machine $Learning$,Honolulu, Hawaii, USA. PMLR 202,2023. Copyright 2023 by the author(s).
for vision (Gidaris et al., 2018; Hjelm et al., 2018; Chen et al., 2020; Jing & Tian, 2020), language domains ( Mikolov et al., 2013; Devlin et al., 2018; Radford et al., 2018; Lan et al., 2019) and is relatively new to graphs.
Contrastive learning on graphs (Velickovic et al., 2019; Sun et al., 2019; Zhu et al., 2020; You et al., 2020; Hassani & Khasahmadi, 2020) is a variant of SSL, whose pretext task is aimed at maximizing representation agreement across augmented views. A typical GNN-based GCL framework comprises three main components: 1. Data augmentors, 2. GNN encoders and 3. Contrastive objectives with cor- responding modes. Data augmentation techniques in the literature can be categorized based on: 1. Feature masking/perturbations (You et al., 2020; Thakoor et al., 2021), 2. Structural perturbations (You et al., 2020; Hassani & Khasahmadi, 2020; Zeng & Xie, 2021), 3. Sub-graph sampling (Hu et al., 2019; Qiu et al., 2020; Jiao et al., 2020; Zhu et al., 2021a), 4. Adaptive and learnable structural perturbations (Zhu et al., 2021c; Yin et al., 2022). Efforts leveraging these techniques tend to focus on the collective comparison of GCL frameworks and lack controlled experimentation pertaining to the augmentation phase.
A recent empirical study on GCL by Zhu et al. (2021b) partially addresses this issue and compares some of the widely used augmentors with a fixed encoder design and contrastive objective. In particular, the study focuses on techniques such as feature masking, node dropping, edge perturbations, sub-graph sampling based on random walks, diffusions based on Personalized Page Rank (PPR), Heat Kernel (HK) and Markov Diffusion Kernel (MDK). From a topological augmentation perspective, pre-defined stochastic techniques such as node dropping and edge perturbations tend to be faster in practice but fail to preserve structural information of the original graph (Hübler et al., 2008). Surprisingly, views obtained by such simple techniques have been shown to outperform diffusion-based strategies when contrasting solely based on node embeddings. Despite this parity, Zhu et al. (2021b) show that, when diffusion-based approaches are combined with node dropping, there is a substantial increase in the unsupervised node and graph classification performance. An attempt to avoid such trial and error based design of augmentors was made by Yin et al. (2022), where learnable augmentors are trained in an endto-end fashion to generate semantic label preserving views.
1 | [
[
"0.125",
"0.11237",
"0.847",
"0.113",
"0.847",
"0.1307",
"0.125",
"0.13"
],
[
"0.4126",
"0.178",
"0.558",
"0.1749",
"0.558",
"0.1925",
"0.4126",
"0.1957"
],
[
"0.2427",
"0.2197",
"0.3203",
"0.2216",
"0.3203",
"0.2374",
"0.2427",
"0.2349"
],
[
"0.5024",
"0.2235",
"0.884",
"0.2235",
"0.884",
"0.2343",
"0.5024",
"0.2343"
],
[
"0.1217",
"0.2412",
"0.4429",
"0.2412",
"0.4429",
"0.2551",
"0.1217",
"0.2551"
],
[
"0.5",
"0.2361",
"0.886",
"0.2374",
"0.886",
"0.252",
"0.5",
"0.2507"
],
[
"0.1209",
"0.2568",
"0.4429",
"0.2568",
"0.4429",
"0.2708",
"0.1209",
"0.2708"
],
[
"0.501",
"0.2524",
"0.886",
"0.2524",
"0.886",
"0.2664",
"0.501",
"0.2664"
],
[
"0.11926",
"0.2708",
"0.4412",
"0.2715",
"0.4412",
"0.286",
"0.11926",
"0.2854"
],
[
"0.501",
"0.2683",
"0.7876",
"0.2683",
"0.7876",
"0.2822",
"0.501",
"0.2822"
],
[
"0.11926",
"0.286",
"0.4412",
"0.286",
"0.4412",
"0.2998",
"0.11926",
"0.2998"
],
[
"0.02287",
"0.2734",
"0.06046",
"0.2734",
"0.06046",
"0.3945",
"0.02287",
"0.3945"
],
[
"0.501",
"0.2903",
"0.8857",
"0.2903",
"0.8857",
"0.305",
"0.501",
"0.305"
],
[
"0.1209",
"0.3018",
"0.4412",
"0.3018",
"0.4412",
"0.315",
"0.1209",
"0.315"
],
[
"0.501",
"0.3062",
"0.886",
"0.3062",
"0.886",
"0.3193",
"0.501",
"0.3193"
],
[
"0.1209",
"0.3176",
"0.4412",
"0.3176",
"0.4412",
"0.3308",
"0.1209",
"0.3308"
],
[
"0.5",
"0.3193",
"0.886",
"0.3208",
"0.886",
"0.3352",
"0.5",
"0.334"
],
[
"0.1217",
"0.3333",
"0.4412",
"0.3333",
"0.4412",
"0.3467",
"0.1217",
"0.3467"
],
[
"0.501",
"0.336",
"0.886",
"0.336",
"0.886",
"0.3499",
"0.501",
"0.3499"
],
[
"0.1217",
"0.348",
"0.4412",
"0.348",
"0.4412",
"0.361",
"0.1217",
"0.361"
],
[
"0.501",
"0.3516",
"0.886",
"0.3516",
"0.886",
"0.3655",
"0.501",
"0.3655"
],
[
"0.1217",
"0.363",
"0.4412",
"0.363",
"0.4412",
"0.377",
"0.1217",
"0.377"
],
[
"0.501",
"0.3674",
"0.886",
"0.3674",
"0.886",
"0.3806",
"0.501",
"0.3806"
],
[
"0.12335",
"0.3774",
"0.4395",
"0.3774",
"0.4395",
"0.3882",
"0.12335",
"0.3882"
],
[
"0.5024",
"0.3833",
"0.884",
"0.3833",
"0.884",
"0.3933",
"0.5024",
"0.3933"
],
[
"0.02287",
"0.3965",
"0.05637",
"0.3958",
"0.05963",
"0.4797",
"0.02533",
"0.481"
],
[
"0.1209",
"0.3926",
"0.4412",
"0.3926",
"0.4412",
"0.406",
"0.1209",
"0.406"
],
[
"0.5",
"0.3965",
"0.8857",
"0.3965",
"0.8857",
"0.4104",
"0.5",
"0.4104"
],
[
"0.11926",
"0.406",
"0.4412",
"0.4077",
"0.4412",
"0.4216",
"0.11926",
"0.4192"
],
[
"0.501",
"0.411",
"0.886",
"0.411",
"0.886",
"0.4248",
"0.501",
"0.4248"
],
[
"0.1217",
"0.4248",
"0.4395",
"0.4248",
"0.4395",
"0.435",
"0.1217",
"0.435"
],
[
"0.501",
"0.4268",
"0.886",
"0.4268",
"0.886",
"0.44",
"0.501",
"0.44"
],
[
"0.1209",
"0.4382",
"0.4429",
"0.4382",
"0.4429",
"0.452",
"0.1209",
"0.452"
],
[
"0.501",
"0.4414",
"0.886",
"0.4414",
"0.886",
"0.455",
"0.501",
"0.455"
],
[
"0.1209",
"0.4526",
"0.4429",
"0.4526",
"0.4429",
"0.4666",
"0.1209",
"0.4666"
],
[
"0.501",
"0.457",
"0.886",
"0.457",
"0.886",
"0.4702",
"0.501",
"0.4702"
],
[
"0.1209",
"0.4685",
"0.4429",
"0.4685",
"0.4429",
"0.4817",
"0.1209",
"0.4817"
],
[
"0.501",
"0.4717",
"0.886",
"0.4717",
"0.886",
"0.4849",
"0.501",
"0.4849"
],
[
"0.02287",
"0.4873",
"0.05963",
"0.4873",
"0.05963",
"0.6694",
"0.02287",
"0.6694"
],
[
"0.1217",
"0.4849",
"0.3245",
"0.4849",
"0.3245",
"0.4956",
"0.1217",
"0.4956"
],
[
"0.501",
"0.4873",
"0.886",
"0.4873",
"0.886",
"0.5005",
"0.501",
"0.5005"
],
[
"0.5",
"0.5005",
"0.8857",
"0.502",
"0.8857",
"0.5166",
"0.5",
"0.515"
],
[
"0.501",
"0.5176",
"0.886",
"0.5176",
"0.886",
"0.531",
"0.501",
"0.531"
],
[
"0.08905",
"0.5244",
"0.2173",
"0.5244",
"0.2173",
"0.539",
"0.08905",
"0.539"
],
[
"0.4983",
"0.531",
"0.886",
"0.532",
"0.886",
"0.547",
"0.4983",
"0.5454"
],
[
"0.08905",
"0.5503",
"0.4756",
"0.5503",
"0.4756",
"0.5645",
"0.08905",
"0.5645"
],
[
"0.5",
"0.549",
"0.7485",
"0.549",
"0.7485",
"0.5625",
"0.5",
"0.5625"
],
[
"0.08826",
"0.5664",
"0.4756",
"0.5664",
"0.4756",
"0.58",
"0.08826",
"0.58"
],
[
"0.5",
"0.5693",
"0.8857",
"0.5693",
"0.8857",
"0.5835",
"0.5",
"0.5835"
],
[
"0.08905",
"0.5835",
"0.4739",
"0.5835",
"0.4739",
"0.5933",
"0.08905",
"0.5933"
],
[
"0.5",
"0.5854",
"0.8857",
"0.5854",
"0.8857",
"0.5986",
"0.5",
"0.5986"
],
[
"0.08826",
"0.5967",
"0.4756",
"0.5967",
"0.4756",
"0.61",
"0.08826",
"0.61"
],
[
"0.501",
"0.6016",
"0.884",
"0.6016",
"0.884",
"0.6123",
"0.501",
"0.6123"
],
[
"0.08905",
"0.6113",
"0.4756",
"0.6113",
"0.4756",
"0.625",
"0.08905",
"0.625"
],
[
"0.5",
"0.6157",
"0.886",
"0.6157",
"0.886",
"0.629",
"0.5",
"0.629"
],
[
"0.0866",
"0.625",
"0.4739",
"0.6255",
"0.4739",
"0.64",
"0.0866",
"0.6396"
],
[
"0.4968",
"0.63",
"0.884",
"0.63",
"0.884",
"0.643",
"0.4968",
"0.643"
],
[
"0.0866",
"0.6396",
"0.4756",
"0.64",
"0.4756",
"0.655",
"0.0866",
"0.654"
],
[
"0.4983",
"0.646",
"0.886",
"0.6445",
"0.886",
"0.659",
"0.4983",
"0.66"
],
[
"0.02287",
"0.657",
"0.05554",
"0.656",
"0.058",
"0.7026",
"0.02533",
"0.704"
],
[
"0.0866",
"0.655",
"0.4756",
"0.656",
"0.4756",
"0.6704",
"0.0866",
"0.6694"
],
[
"0.501",
"0.66",
"0.886",
"0.66",
"0.886",
"0.674",
"0.501",
"0.674"
],
[
"0.08826",
"0.672",
"0.4756",
"0.672",
"0.4756",
"0.686",
"0.08826",
"0.686"
],
[
"0.5",
"0.6753",
"0.886",
"0.6753",
"0.886",
"0.6895",
"0.5",
"0.6895"
],
[
"0.08905",
"0.686",
"0.4739",
"0.686",
"0.4739",
"0.6997",
"0.08905",
"0.6997"
],
[
"0.4983",
"0.6914",
"0.886",
"0.6914",
"0.886",
"0.705",
"0.4983",
"0.705"
],
[
"0.08905",
"0.702",
"0.4739",
"0.702",
"0.4739",
"0.7153",
"0.08905",
"0.7153"
],
[
"0.4983",
"0.7056",
"0.884",
"0.7056",
"0.884",
"0.7197",
"0.4983",
"0.7197"
],
[
"0.08826",
"0.7173",
"0.4739",
"0.7173",
"0.4739",
"0.731",
"0.08826",
"0.731"
],
[
"0.4983",
"0.7217",
"0.886",
"0.7217",
"0.886",
"0.735",
"0.4983",
"0.735"
],
[
"0.0866",
"0.731",
"0.4739",
"0.732",
"0.4739",
"0.747",
"0.0866",
"0.7456"
],
[
"0.501",
"0.7373",
"0.8857",
"0.7373",
"0.8857",
"0.7476",
"0.501",
"0.7476"
],
[
"0.08826",
"0.7476",
"0.4756",
"0.7476",
"0.4756",
"0.761",
"0.08826",
"0.761"
],
[
"0.501",
"0.752",
"0.886",
"0.752",
"0.886",
"0.765",
"0.501",
"0.765"
],
[
"0.08826",
"0.763",
"0.4739",
"0.763",
"0.4739",
"0.7764",
"0.08826",
"0.7764"
],
[
"0.5",
"0.765",
"0.884",
"0.765",
"0.884",
"0.779",
"0.5",
"0.779"
],
[
"0.08826",
"0.7764",
"0.4739",
"0.7764",
"0.4739",
"0.7905",
"0.08826",
"0.7905"
],
[
"0.501",
"0.7827",
"0.884",
"0.7827",
"0.884",
"0.7935",
"0.501",
"0.7935"
],
[
"0.0866",
"0.7915",
"0.4756",
"0.7925",
"0.4756",
"0.8066",
"0.0866",
"0.806"
],
[
"0.5",
"0.797",
"0.886",
"0.797",
"0.886",
"0.81",
"0.5",
"0.81"
],
[
"0.10706",
"0.8145",
"0.4756",
"0.8164",
"0.4756",
"0.8306",
"0.10706",
"0.829"
],
[
"0.5",
"0.8125",
"0.8857",
"0.8125",
"0.8857",
"0.8257",
"0.5",
"0.8257"
],
[
"0.0866",
"0.829",
"0.4739",
"0.829",
"0.4739",
"0.8423",
"0.0866",
"0.8423"
],
[
"0.501",
"0.827",
"0.8857",
"0.827",
"0.8857",
"0.8403",
"0.501",
"0.8403"
],
[
"0.08905",
"0.8403",
"0.2239",
"0.8423",
"0.2239",
"0.856",
"0.08905",
"0.854"
],
[
"0.5024",
"0.8433",
"0.8857",
"0.8433",
"0.8857",
"0.854",
"0.5024",
"0.854"
],
[
"0.501",
"0.857",
"0.886",
"0.857",
"0.886",
"0.8706",
"0.501",
"0.8706"
],
[
"0.0866",
"0.8643",
"0.4756",
"0.8643",
"0.4756",
"0.882",
"0.0866",
"0.882"
],
[
"0.5",
"0.8726",
"0.888",
"0.8726",
"0.888",
"0.886",
"0.5",
"0.886"
],
[
"0.08826",
"0.88",
"0.4739",
"0.88",
"0.4739",
"0.893",
"0.08826",
"0.893"
],
[
"0.08826",
"0.8926",
"0.2214",
"0.8926",
"0.2214",
"0.906",
"0.08826",
"0.906"
],
[
"0.5",
"0.888",
"0.888",
"0.888",
"0.888",
"0.9014",
"0.5",
"0.9014"
],
[
"0.482",
"0.925",
"0.4927",
"0.925",
"0.4927",
"0.9346",
"0.482",
"0.9346"
]
] | [
"Randomized Schur Complement Views for Graph Contrastive Learning",
"Vignesh Kothapalli !",
"Abstract",
"for vision(Gidaris et al..2018:Hjelm et al.,2018:Chen",
"We introduce a randomized topological augmen-",
"et al., 2020; Jing & Tian, 2020), language domains (Mikolov",
"tor based on Schur complements for Graph Con-",
"et al., 2013; Devlin et al., 2018; Radford et al., 2018; Lan",
"trastive Learning (GCL). Given a graph laplacian",
"et al., 2019) and is relatively new to graphs.",
"matrix, the technique generates unbiased approxi-",
" 9",
"Contrastive learning on graphs (Velickovic et al., 2019; Sun",
"mations of its Schur complements and treats the",
"et al., 2019; Zhu et al., 2020; You et al., 2020; Hassani &",
"corresponding graphs as augmented views. We",
"Khasahmadi, 2020) is a variant of SSL, whose pretext task",
"discuss the benefits of our approach, provide theo-",
"is aimed at maximizing representation agreement across",
"retical justifications and present connections with",
"augmented views. A typical GNN-based GCL framework",
"graph diffusion. Unlike previous efforts, we study",
"comprises three main components: 1. Data augmentors,",
"the empirical effectiveness of the augmentor in a",
"2.GNN encoders and 3.Contrastiveobjectives with cor",
"[T'S]",
"controlled fashion by varying the design choices",
"responding modes. Data augmentation techniques in the",
"for subsequent GCL phases, such as encoding",
"literature can be categorized based on: 1. Feature mask-",
"andcontrasting.Extensiveexperimentsonnode",
"ing/perturbations (You et al., 2020; Thakoor et al., 2021),",
"and graph classification benchmarks demonstrate",
"2. Structural perturbations (You et al., 2020; Hassani &",
"that our technique consistently outperforms pre-",
"Khasahmadi, 2020; Zeng & Xie, 2021),3. Sub-graph sam-",
"defined and adaptive augmentation approaches to",
"pling (Hu et al., 2019; Qiu et al., 2020; Jiao et al., 2020; Zhu",
"00",
"achieve state-of-the-art results.",
"et al., 202la), 4. Adaptive and learnable structural perturba-",
"tions (Zhu et al., 2021c; Yin et al., 2022). Efforts leveraging",
"these techniques tend to focus on the collective comparison",
"1.Introduction",
"of GCL frameworks and lack controlled experimentation",
"Understanding the structural properties and semantics of",
"pertaining to the augmentation phase.",
"graph data typically requires domain expertise and efficient",
"A recent empirical study on GCL by Zhu et al. (2021b)",
"computationaltools.Advancesinmachinelearningtech-",
"partially addresses this issue and compares some of the",
"niques such as Graph Neural Networks (GNN) (Gori et al.,",
"widelyused augmentors witha fixed encoder design and",
"2005; Scarselli et al., 2008; Bruna et al., 2014; Henaff et al.,",
"contrastive objective. In particular, the study focuses on",
"2015; Welling & Kipf, 2016; Niepert et al., 2016; Bronstein",
" techniques such as feature masking, node dropping, edge",
"et al., 2017; Xu et al., 2018; Wu et al., 2020; Zhou et al.,",
"perturbations, sub-graph sampling based on random walks,",
"!Xe",
"2020) have paved a path for representation learning on in-",
"diffusions based on Personalized Page Rank (PPR), Heat",
"ternet scale graphs and significantly alleviated the human",
"Kernel (HK) and Markov Diffusion Kernel (MDK). From a",
"effort. However, real-world graphs such as social networks,",
"topological augmentation perspective, pre-defined stochas-",
"citation networks, supply chains and media networks are",
"tic techniques such as node dropping and edge perturbations",
"continuously evolving, which makes labeling an extremely",
"tend to be faster in practice but fail to preserve structural",
"challenging task to accomplish. Self-supervised learning",
"information of the original graph(Hubler etal.,2008).Sur-",
"(SSL) addresses this issue by optimizing pretext objectives",
"prisingly, views obtained by such simple techniques have",
"and learninggeneralizable representations for downstream",
"been shown to outperform diffusion-based strategies when",
"tasks (Jin et al., 2020; Wu et al., 2021; Liu et al., 2022; Xie",
"contrasting solelybased onnode embeddings.Despitethis",
"et al., 2022). This learning paradigm has been quite popular",
"parity, Zhu et al. (2021b) show that, when diffusion-based",
"'Courant Institute of Mathematical Sciences, New York Uni-",
"approaches are combined with node dropping, there is a",
"versity, New York, USA. Correspondence to: Vignesh Kothapalli",
"substantial increase in the unsupervised node and graph",
"<vk2115@nyu.edu>.",
"classificationperformance.An attempttoavoid suchtrial",
"and error based design of augmentors was made by Yin et al.",
"Proceedings of the 4oth International Conference on Machine",
"(2022), where learnable augmentors are trained in an end-",
"Learning, Honolulu, Hawaii, USA. PMLR 202, 2023. Copyright",
"2023 by the author(s).",
"to-end fashion to generate semantic label preserving views.",
"1"
] | {
"arxiv_id": "2306.04004",
"language": "en",
"page_idx": 0
} |
||
2306.04004-en-3 | 2306.04004-en-3.jpg | Randomized Schur Complement Views for Graph Contrastive Learning
<table>
<tbody>
<tr>
<td>$\textbf{Algorithm 1}rLap$</td>
<td>$\overline{\mathbf{\Lambda I}}$</td>
<td>$\textbf{ithm 1 }rLap$</td>
</tr>
<tr>
<td>$\| \textit{compute weight of the new edge bet}$</td>
<td> </td>
<td>$\| \textit{compute weight of the new edge between }x_l, x_q$</td>
</tr>
<tr>
<td>$\mathbf{C}=\mathbf{C}+w_{\mathbf{R}}(x_{l},x_{q})\cdot\boldsymbol{\Delta}$ $\mathbf{\Lambda}_{x_lx_q}$ end for $\mathbf{R}=\mathbf{R}-$ $\sum$ $w_{\mathbf{R}}(e_{ij})\boldsymbol{\Delta}_{i}$ -</td>
<td> </td>
<td>$\mathbf{C}=\mathbf{C}+w_{\mathbf{R}}(x_{l},x_{q})\cdot\boldsymbol{\Delta}_{x_{l}x_{q}}$ ndfor $\mathbf{c}=\mathbf{R}-$ $w_{\mathbf{R}}(e_{ij})\boldsymbol{\Delta}_{ij}$</td>
</tr>
<tr>
<td>$e_{ij}\in\mathcal{E}_{\mathbf{R}},v_{j}\in\mathcal{N}_{\mathbf{R}}(v_{i})$ $\mathbf{R}=\mathbf{R}+\mathbf{C}$ until all the $\gamma|\mathcal{V}|$ nodes are eliminated base $\textbf{return R}$</td>
<td> </td>
<td>$e_{ij}\in\mathcal{E}_{\mathbf{R}},v_{j}\in\mathcal{N}_{\mathbf{R}}(v_{i})$ $\mathbf{c}=\mathbf{R}+\mathbf{C}$ il all the $\gamma|\mathcal{V}|$ nodes are eliminated based on $o_v$ $\textbf{lrn R}$</td>
</tr>
</tbody>
</table>
Where STAR(L,vi),CLIQUE(L,vi) for any node $v_i\in\mathcal{V}$
are given by:
$$\begin{aligned}STAR(\mathbf{L},v_{i})&=\sum_{e_{ij}\in\mathcal{E},v_{j}\in\mathcal{N}(v_{i})}w(e_{ij})\boldsymbol{\Delta}_{ij}\\CLIQUE(\mathbf{L},v_{i})&=\frac{1}{2w(v_{i})}\sum_{e_{ij}\in\mathcal{E}}\sum_{e_{ik}\in\mathcal{E}}w(e_{ij})w(e_{ik})\boldsymbol{\Delta}_{jk}\end{aligned}$$
This result indicates that $SC(\mathbf{L},\mathcal{V}\backslash v_1)$ is a weighted Laplacian matrix with an associated graph. For simplicity, we denote $SC(\mathbf{L},\mathcal{V}\backslash v_1)$ as $SC(\mathcal{G},\mathcal{V}\backslash v_1)$ when the notion of a graph is suitable for the context. Schur complements hold an interesting graph theoretic property that, the random walk transition probabilities of the remaining nodes $\mathcal{V}\backslash v_1$ through the eliminated vertex $v_1$ (with respect to $\mathcal{G})$ are preserved in $SC(\mathcal{G},\mathcal{V}\backslash v_{1}).$ Intuitively, the list of nodes visited by random walks on $SC(\mathcal{G},\mathcal{V}\backslash v_1)$ is equivalent in distribution to the list of nodes in $\mathcal{V}\backslash v_1$ visited by random walks on $\mathcal{G}.$ Refer to section 3 in Durfee et al.(2019) and section 4.1 in Gao et al. (2022) for detailed discussions on this property.
# 3. Methodology
To leverage these combinatorial properties of Schur complements in our contrastive views, two issues need to be addressed: The $O(N^2)$ overhead to compute a clique and a lack of stochasticity in the gaussian elimination procedure. We address both these issues by developing a clique approximation procedure that inherently introduces randomness. Thus, by eliminating nodes via GE and computing unbiased approximations of the induced cliques, our augmentation technique generates randomized Schur complements which preserve the random walk transition probabilities of the remaining nodes in expectation.
The need for stochasticity is based on the empirical results of Zhu et al. (2021b), where introducing randomness to the views (eg: combining diffusion with node dropping) improved contrastive learning performance on downstream classification tasks. Although a rigorous analysis of such behavior is not present in the literature, we provide interesting insights on GCL performance due to randomness introduced by our approximation procedure.
## $3. 1. \textbf{ Randomized Schur Complements}$
We present our Schur complement approximation procedure in Algorithm 1.$^2$ We name it rLap to denote the laplacian nature of its output. The input to rLap is the graph $\mathcal{G}$,the fraction of nodes to eliminate $\gamma$, node elimination scheme $o_v$ and the neighbor ordering scheme $o_n.$ The output is a randomized Schur complement of $\mathcal{G}$ after eliminating $\gamma|\mathcal{V}|$
$^{2}$The code is available at: https://github.com/kvignesh1420/rlap
nodes$^3.$ The scheme $o_v$ indicates the order in which $\gamma|\mathcal{V}|$ nodes are eliminated. The possibilities for such a scheme are huge but we limit our analysis to random ordering and ordering based on node degree. In the `random' scheme, a node is randomly selected at each step of the outer loop. In the‘degree’scheme, a priority queue is maintained to eliminate nodes based on their degree, i.e. nodes with lower degrees are eliminated before the highly connected ones. Now, without loss of generality, consider the $i^{th}$ iteration of the outer loop where node $v_i$ is being eliminated. The first step is to order the neighbors $\mathcal{N}_\mathbf{R}(v_i)$ and store the result in $\mathcal{X}_{\mathbf{R}}(v_{i})^{4}.$ The order decided by $o_{n}$ affects the arrangement of edges in the approximated clique C. Specifically, while iterating over the neighbors $x_l\in\mathcal{X}_\mathbf{R}(v_i),l\in$ $\{1,\cdots,|\mathcal{X}_{\mathbf{R}}(v_i)|-1\}$ in the inner loop, we compute a conditional probability $P(x_q|x_l)$ of choosing a neighbor $x_{q}\in\mathcal{X}_{\mathbf{R}}(v_{i}),q\in\{l+1,\cdots,|\mathcal{X}_{\mathbf{R}}(v_{i})|\}$,based on which, a new edge between $x_l,x_q$ is created. The choice of $o_n$ affects this conditional probability and can lead to sparser or denser variants of approximated cliques. Now, the weighted elementary laplacian of the newly formed edge $w_{\mathbf{R}}(x_{l},x_{q})\Delta_{x_{l}x_{q}}$is computed and added to $\mathbf{C}^{5}.$ Note that, the inner loop iterates over $O(N)$ nodes for approximating the clique and avoids the $O(N^2)$ overhead for exact compu-
${\overline{{^3}}\text{Elimination}}$ in this context indicates that a node is discon-
nected from all its neighbors. The dimensions of $\mathbf{R}$ remain $\mathbb{R}^N\times N$
with corresponding row and column set to $\mathbf{0}.$
${}^{4}{\mathcal E_{\mathbf{R}},w_{\mathbf{R}},\mathcal{N_{\mathbf{R}}}\text{ are indexed by R to denote the current state.}}$
$^{5}$Here $x_l,x_q$ are indexed on $\dot{\chi}_{\mathbf{R}}(v_i)$ for notational convenience, but have a one-to-one correspondence with our original node set $\mathcal{V}$
4 | [
[
"0.2607",
"0.05682",
"0.714",
"0.05682",
"0.714",
"0.0707",
"0.2607",
"0.0707"
],
[
"0.08826",
"0.0852",
"0.4756",
"0.0852",
"0.4756",
"0.09973",
"0.08826",
"0.09973"
],
[
"0.5",
"0.0859",
"0.6284",
"0.0859",
"0.6284",
"0.09973",
"0.5",
"0.09973"
],
[
"0.08826",
"0.1029",
"0.1748",
"0.1029",
"0.1748",
"0.11426",
"0.08826",
"0.11426"
],
[
"0.5156",
"0.1042",
"0.8857",
"0.1042",
"0.8857",
"0.11743",
"0.5156",
"0.11743"
],
[
"0.2524",
"0.1218",
"0.282",
"0.1218",
"0.282",
"0.1459",
"0.2524",
"0.1459"
],
[
"0.517",
"0.1187",
"0.871",
"0.1187",
"0.871",
"0.1332",
"0.517",
"0.1332"
],
[
"0.1177",
"0.127",
"0.2197",
"0.127",
"0.2197",
"0.1414",
"0.1177",
"0.1414"
],
[
"0.312",
"0.125",
"0.3896",
"0.127",
"0.3896",
"0.1427",
"0.312",
"0.1401"
],
[
"0.513",
"0.132",
"0.589",
"0.132",
"0.589",
"0.1465",
"0.513",
"0.1465"
],
[
"0.2214",
"0.1459",
"0.3145",
"0.1459",
"0.3145",
"0.1572",
"0.2214",
"0.1572"
],
[
"0.5146",
"0.1483",
"0.5645",
"0.1483",
"0.5645",
"0.1635",
"0.5146",
"0.1635"
],
[
"0.2427",
"0.1635",
"0.25",
"0.1635",
"0.25",
"0.1704",
"0.2427",
"0.1704"
],
[
"0.2778",
"0.1635",
"0.308",
"0.1635",
"0.308",
"0.1799",
"0.2778",
"0.1799"
],
[
"0.298",
"0.161",
"0.468",
"0.1666",
"0.4673",
"0.1875",
"0.2974",
"0.1818"
],
[
"0.09644",
"0.1686",
"0.2157",
"0.1686",
"0.2157",
"0.1818",
"0.09644",
"0.1818"
],
[
"0.5317",
"0.1635",
"0.6855",
"0.1635",
"0.6855",
"0.178",
"0.5317",
"0.178"
],
[
"0.2214",
"0.1755",
"0.2754",
"0.178",
"0.2737",
"0.1925",
"0.2197",
"0.1901"
],
[
"0.286",
"0.1799",
"0.2932",
"0.1799",
"0.2932",
"0.1831",
"0.286",
"0.1831"
],
[
"0.5317",
"0.178",
"0.581",
"0.178",
"0.581",
"0.1925",
"0.5317",
"0.1925"
],
[
"0.273",
"0.1863",
"0.3447",
"0.1863",
"0.3447",
"0.1976",
"0.273",
"0.1976"
],
[
"0.5317",
"0.1932",
"0.7256",
"0.1932",
"0.7256",
"0.2084",
"0.5317",
"0.2084"
],
[
"0.451",
"0.1995",
"0.4756",
"0.1995",
"0.4756",
"0.2134",
"0.451",
"0.2134"
],
[
"0.5483",
"0.209",
"0.871",
"0.209",
"0.871",
"0.2235",
"0.5483",
"0.2235"
],
[
"0.55",
"0.2267",
"0.63",
"0.2267",
"0.63",
"0.2412",
"0.55",
"0.2412"
],
[
"0.6685",
"0.2229",
"0.7305",
"0.2247",
"0.7305",
"0.2349",
"0.6685",
"0.2329"
],
[
"0.08826",
"0.2299",
"0.4756",
"0.2299",
"0.4756",
"0.2443",
"0.08826",
"0.2443"
],
[
"0.7646",
"0.2279",
"0.8774",
"0.2267",
"0.8774",
"0.2412",
"0.7646",
"0.2424"
],
[
"0.6313",
"0.238",
"0.7715",
"0.2393",
"0.7715",
"0.254",
"0.6313",
"0.2524"
],
[
"0.0866",
"0.2443",
"0.4739",
"0.2456",
"0.4739",
"0.26",
"0.0866",
"0.2588"
],
[
"0.5474",
"0.2551",
"0.8735",
"0.2568",
"0.8735",
"0.2747",
"0.5474",
"0.2727"
],
[
"0.08826",
"0.2615",
"0.4722",
"0.2615",
"0.4722",
"0.2747",
"0.08826",
"0.2747"
],
[
"0.08905",
"0.2766",
"0.4739",
"0.2766",
"0.4739",
"0.2903",
"0.08905",
"0.2903"
],
[
"0.6465",
"0.2778",
"0.7686",
"0.2778",
"0.7686",
"0.2917",
"0.6465",
"0.2917"
],
[
"0.767",
"0.279",
"0.853",
"0.279",
"0.853",
"0.2903",
"0.767",
"0.2903"
],
[
"0.0866",
"0.2917",
"0.4739",
"0.2903",
"0.4739",
"0.305",
"0.0866",
"0.3062"
],
[
"0.5513",
"0.2917",
"0.642",
"0.2917",
"0.642",
"0.303",
"0.5513",
"0.303"
],
[
"0.546",
"0.3018",
"0.749",
"0.3062",
"0.749",
"0.3245",
"0.546",
"0.3208"
],
[
"0.7295",
"0.2974",
"0.7715",
"0.2974",
"0.7715",
"0.308",
"0.7295",
"0.308"
],
[
"0.08826",
"0.307",
"0.4739",
"0.307",
"0.4739",
"0.3208",
"0.08826",
"0.3208"
],
[
"0.08826",
"0.3213",
"0.4756",
"0.3213",
"0.4756",
"0.336",
"0.08826",
"0.336"
],
[
"0.5317",
"0.3213",
"0.589",
"0.3213",
"0.589",
"0.336",
"0.5317",
"0.336"
],
[
"0.5317",
"0.3352",
"0.601",
"0.3352",
"0.601",
"0.3499",
"0.5317",
"0.3499"
],
[
"0.648",
"0.3333",
"0.671",
"0.3333",
"0.671",
"0.3528",
"0.648",
"0.3528"
],
[
"0.716",
"0.3372",
"0.7983",
"0.3372",
"0.7983",
"0.3516",
"0.716",
"0.3516"
],
[
"0.603",
"0.3499",
"0.716",
"0.3499",
"0.716",
"0.363",
"0.603",
"0.363"
],
[
"0.08826",
"0.336",
"0.4739",
"0.336",
"0.4739",
"0.3499",
"0.08826",
"0.3499"
],
[
"0.08826",
"0.3516",
"0.4739",
"0.3516",
"0.4739",
"0.3655",
"0.08826",
"0.3655"
],
[
"0.5312",
"0.3599",
"0.621",
"0.3623",
"0.621",
"0.377",
"0.5312",
"0.375"
],
[
"0.08826",
"0.3662",
"0.4739",
"0.3662",
"0.4739",
"0.3801",
"0.08826",
"0.3801"
],
[
"0.08905",
"0.3833",
"0.4722",
"0.3833",
"0.4722",
"0.3933",
"0.08905",
"0.3933"
],
[
"0.5156",
"0.3774",
"0.8486",
"0.3774",
"0.8486",
"0.3914",
"0.5156",
"0.3914"
],
[
"0.0866",
"0.3958",
"0.468",
"0.3965",
"0.468",
"0.411",
"0.0866",
"0.4104"
],
[
"0.5146",
"0.3926",
"0.5825",
"0.3926",
"0.5825",
"0.4072",
"0.5146",
"0.4072"
],
[
"0.0866",
"0.428",
"0.2181",
"0.43",
"0.2181",
"0.4475",
"0.0866",
"0.4458"
],
[
"0.5",
"0.4463",
"0.884",
"0.4463",
"0.884",
"0.4602",
"0.5",
"0.4602"
],
[
"0.08826",
"0.457",
"0.4756",
"0.457",
"0.4756",
"0.4702",
"0.08826",
"0.4702"
],
[
"0.5",
"0.4622",
"0.8857",
"0.4622",
"0.8857",
"0.476",
"0.5",
"0.476"
],
[
"0.08826",
"0.473",
"0.4739",
"0.473",
"0.4739",
"0.486",
"0.08826",
"0.486"
],
[
"0.501",
"0.4778",
"0.8857",
"0.4778",
"0.8857",
"0.4912",
"0.501",
"0.4912"
],
[
"0.0866",
"0.486",
"0.4739",
"0.4873",
"0.4739",
"0.502",
"0.0866",
"0.5005"
],
[
"0.5",
"0.4937",
"0.886",
"0.4937",
"0.886",
"0.507",
"0.5",
"0.507"
],
[
"0.0866",
"0.502",
"0.4756",
"0.5024",
"0.4756",
"0.5166",
"0.0866",
"0.515"
],
[
"0.501",
"0.5083",
"0.886",
"0.5083",
"0.886",
"0.5215",
"0.501",
"0.5215"
],
[
"0.08826",
"0.518",
"0.4739",
"0.518",
"0.4739",
"0.532",
"0.08826",
"0.532"
],
[
"0.5",
"0.5234",
"0.886",
"0.5234",
"0.886",
"0.537",
"0.5",
"0.537"
],
[
"0.08826",
"0.5327",
"0.4756",
"0.5327",
"0.4756",
"0.547",
"0.08826",
"0.547"
],
[
"0.5",
"0.5386",
"0.886",
"0.5386",
"0.886",
"0.552",
"0.5",
"0.552"
],
[
"0.08826",
"0.5474",
"0.4739",
"0.5474",
"0.4739",
"0.561",
"0.08826",
"0.561"
],
[
"0.501",
"0.5537",
"0.8735",
"0.5537",
"0.8735",
"0.5674",
"0.501",
"0.5674"
],
[
"0.08826",
"0.5625",
"0.4722",
"0.5625",
"0.4722",
"0.5757",
"0.08826",
"0.5757"
],
[
"0.08905",
"0.58",
"0.4722",
"0.58",
"0.4722",
"0.5903",
"0.08905",
"0.5903"
],
[
"0.501",
"0.5747",
"0.886",
"0.5747",
"0.886",
"0.589",
"0.501",
"0.589"
],
[
"0.08826",
"0.5933",
"0.4739",
"0.5933",
"0.4739",
"0.607",
"0.08826",
"0.607"
],
[
"0.5",
"0.5903",
"0.886",
"0.5903",
"0.886",
"0.604",
"0.5",
"0.604"
],
[
"0.08826",
"0.6094",
"0.298",
"0.6094",
"0.298",
"0.6226",
"0.08826",
"0.6226"
],
[
"0.501",
"0.606",
"0.886",
"0.606",
"0.886",
"0.6206",
"0.501",
"0.6206"
],
[
"0.4983",
"0.619",
"0.884",
"0.619",
"0.884",
"0.634",
"0.4983",
"0.634"
],
[
"0.08826",
"0.631",
"0.4739",
"0.631",
"0.4739",
"0.6445",
"0.08826",
"0.6445"
],
[
"0.501",
"0.636",
"0.886",
"0.636",
"0.886",
"0.6494",
"0.501",
"0.6494"
],
[
"0.08826",
"0.6465",
"0.4756",
"0.6465",
"0.4756",
"0.66",
"0.08826",
"0.66"
],
[
"0.5",
"0.6514",
"0.884",
"0.6514",
"0.884",
"0.6655",
"0.5",
"0.6655"
],
[
"0.08905",
"0.661",
"0.4756",
"0.661",
"0.4756",
"0.6753",
"0.08905",
"0.6753"
],
[
"0.5024",
"0.6665",
"0.8857",
"0.6665",
"0.8857",
"0.681",
"0.5024",
"0.681"
],
[
"0.08826",
"0.677",
"0.4739",
"0.677",
"0.4739",
"0.69",
"0.08826",
"0.69"
],
[
"0.501",
"0.6816",
"0.8857",
"0.6816",
"0.8857",
"0.696",
"0.501",
"0.696"
],
[
"0.08826",
"0.6924",
"0.4756",
"0.6924",
"0.4756",
"0.7056",
"0.08826",
"0.7056"
],
[
"0.5",
"0.6978",
"0.886",
"0.696",
"0.886",
"0.7104",
"0.5",
"0.712"
],
[
"0.08905",
"0.7085",
"0.4722",
"0.7085",
"0.4722",
"0.7183",
"0.08905",
"0.7183"
],
[
"0.501",
"0.712",
"0.886",
"0.712",
"0.886",
"0.726",
"0.501",
"0.726"
],
[
"0.08826",
"0.7217",
"0.4739",
"0.7217",
"0.4739",
"0.735",
"0.08826",
"0.735"
],
[
"0.5",
"0.726",
"0.886",
"0.7266",
"0.886",
"0.741",
"0.5",
"0.7407"
],
[
"0.08826",
"0.7373",
"0.308",
"0.7373",
"0.308",
"0.7505",
"0.08826",
"0.7505"
],
[
"0.501",
"0.742",
"0.886",
"0.742",
"0.886",
"0.756",
"0.501",
"0.756"
],
[
"0.501",
"0.757",
"0.8857",
"0.757",
"0.8857",
"0.771",
"0.501",
"0.771"
],
[
"0.08826",
"0.768",
"0.353",
"0.768",
"0.353",
"0.782",
"0.08826",
"0.782"
],
[
"0.4983",
"0.7715",
"0.8896",
"0.7695",
"0.8896",
"0.786",
"0.4983",
"0.7886"
],
[
"0.0866",
"0.7905",
"0.4756",
"0.7915",
"0.4756",
"0.806",
"0.0866",
"0.8047"
],
[
"0.501",
"0.787",
"0.886",
"0.787",
"0.886",
"0.8003",
"0.501",
"0.8003"
],
[
"0.08826",
"0.8066",
"0.4739",
"0.8066",
"0.4739",
"0.821",
"0.08826",
"0.821"
],
[
"0.501",
"0.802",
"0.886",
"0.802",
"0.886",
"0.815",
"0.501",
"0.815"
],
[
"0.08826",
"0.8228",
"0.4739",
"0.8228",
"0.4739",
"0.836",
"0.08826",
"0.836"
],
[
"0.5195",
"0.8247",
"0.886",
"0.827",
"0.886",
"0.8413",
"0.5195",
"0.839"
],
[
"0.08826",
"0.8384",
"0.4739",
"0.8384",
"0.4739",
"0.8516",
"0.08826",
"0.8516"
],
[
"0.4968",
"0.8384",
"0.886",
"0.8354",
"0.886",
"0.853",
"0.4968",
"0.856"
],
[
"0.08826",
"0.853",
"0.4722",
"0.853",
"0.4722",
"0.866",
"0.08826",
"0.866"
],
[
"0.5",
"0.853",
"0.7656",
"0.853",
"0.7656",
"0.866",
"0.5",
"0.866"
],
[
"0.08826",
"0.867",
"0.4739",
"0.867",
"0.4739",
"0.881",
"0.08826",
"0.881"
],
[
"0.5195",
"0.8643",
"0.871",
"0.8657",
"0.871",
"0.88",
"0.5195",
"0.879"
],
[
"0.5215",
"0.8774",
"0.8857",
"0.88",
"0.8857",
"0.8945",
"0.5215",
"0.8926"
],
[
"0.10864",
"0.89",
"0.4756",
"0.8916",
"0.4756",
"0.906",
"0.10864",
"0.905"
],
[
"0.5",
"0.8926",
"0.8857",
"0.8926",
"0.8857",
"0.906",
"0.5",
"0.906"
],
[
"0.4805",
"0.925",
"0.4927",
"0.925",
"0.4927",
"0.9346",
"0.4805",
"0.9346"
]
] | [
"Randomized Schur Complement Views for Graph Contrastive Learning",
"Where STAR(L, vi), CLIQUE(L, vi) for any node Vi E V",
"Algorithm 1 rLap",
"are given by:",
"Input: graph G = (V, &, w), node drop fraction , node",
"Z",
"elimination scheme ou, neighbor ordering scheme on.",
"STAR(L, U) =",
"w(et)Aij",
"set R = L",
"etjEE,,EN(v)",
"repeat",
"7",
"7",
" w(et)w(eik)Ajk",
"CLIQUE(L, V) =",
"XR(U) = On(NR(Vi))",
"2w(vi)",
"",
"C=0",
"eijESeikEE",
"for l = 1 to |Xr(ui)/ - 1 do",
"(5)",
"Il choose &g based on the conditional probability",
"P(cglat) =",
"WR(Vi,Ea)",
"This result indicates that SC(L, V) ui) is a weighted Lapla-",
";CI<g EXR(V))",
"WR(U)- WR(U,EE)",
"cian matrix with an associated graph. For simplicity, we",
"Il compute weight of the new edge between Et, q",
"denote SC(L, V vi) as SC(G, V) ui) when the notion of a",
"graph is suitable for the context. Schur complements hold",
"WR(U,EI).(WR(U)-",
"ZWR(U,TR)",
"an interesting graph theoretic property that, the random walk",
"WR(CI,Ta)=",
"C=C+ WR(, Ag)·Arg",
"WR(U)",
"transition probabilities of the remaining nodes V vi through",
"the eliminated vertex ui (with respect to G) are preserved",
"end for",
"R=R-",
"Z",
"WR(er)Aij",
"eijEER,U,ENR(U))",
"in SC(g, V i). Intuitively, the list of nodes visited by",
"random walks on SC(G, V) i) is equivalent in distribution",
"R=R+C",
"to the list of nodes in V)ui visited by random walks on G.",
"Refertosection3 inDurfee etal.(2019)and section4.1 in",
"until all the /V nodes are eliminated based on ou",
"Gao et al. (2022) for detailed discussions on this property.",
"return R",
"3. Methodology",
"nodes3. The scheme o, indicates the order in which /V",
"To leverage these combinatorial properties of Schur com-",
"nodes are eliminated. The possibilities for such a scheme",
"plements in our contrastive views, two issues need to be",
"are huge but we limit our analysis to random ordering and",
"addressed: The O(N2) overhead to compute a clique and a",
"ordering based on node degree. In the 'random' scheme,",
"lackof stochasticityinthegaussianeliminationprocedure.",
"a node is randomly selected at each step of the outer loop.",
"We address both these issues by developing a clique approx-",
"In the 'degree' scheme, a priority queue is maintained to",
"imation procedure that inherently introduces randomness.",
"eliminate nodes based on their degree, i.e. nodes with lower",
"Thus, by eliminating nodes via GE and computing unbiased",
"degrees are eliminated before the highly connected ones.",
"approximations of the induced cliques, our augmentation",
"techniquegeneratesrandomizedSchurcomplementswhich",
"Now, without loss of generality, consider the ith iteration",
"preserve the random walk transition probabilities of the",
"of the outer loop where node w; is being eliminated. The",
"remaining nodes in expectation.",
"first step is to order the neighbors NR(vi) and store the",
"result in Xr(vi) 4. The order decided by On affects the",
"The need for stochasticity is based on the empirical results",
"arrangement of edges in the approximated clique C. Specif-",
"of Zhu et al. (2021b), where introducing randomness to",
"ically, while iterating over the neighbors i E Xr(vi),l ",
"the views (eg: combining diffusion with node dropping)",
"[1,.., |Xr(i)] -- 1] in the inner loop, we compute a",
"improved contrastive learning performance on downstream",
"conditional probability P(cgri) of choosing a neighbor",
"classification tasks. Although a rigorous analysis of such be-",
"Cq E r(vi), q e [l + 1, .., /r(vi)ll, based on which,",
"haviorisnotpresentintheliterature,weprovideinteresting",
"a new edge between i, Cq is created. The choice of on af-",
"insights on GCL performance due to randomness introduced",
"fects this conditional probability and can lead to sparser",
"by our approximation procedure.",
"or denser variants of approximated cliques.Now, the",
"weighted elementary laplacian of the newly formed edge",
"3.1. Randomized Schur Complements",
"WR(rt, g)Arig is computed and added to C 5. Note that,",
"We present our Schur complement approximation procedure",
"the inner loop iterates over O(N) nodes for approximating",
"in Algorithm 1.2 We name it rLap to denote the laplacian",
"the clique and avoids the O(N2) overhead for exact compu-",
"nature of its output. The input to r Lap is the graph G, the",
"3Elimination in this context indicates that a node is discon-",
"fraction of nodes to eliminate , node elimination scheme",
"nected from allits neighbors. The dimensions of R remain IRNXN",
"Ou and the neighbor ordering scheme on. The output is a",
"with corresponding row and column set to 0.",
"randomized Schur complement of G after eliminating /V",
"4ER, wR,NR are indexed by R to denote the current state.",
"5Herei,gareindexedonXr(vi)fornotationalconvenience,",
"2The code is available at: https://github.com/kvignesh1420/rlap",
"but have a one-to-one correspondence with our original node set V",
"4"
] | {
"arxiv_id": "2306.04004",
"language": "en",
"page_idx": 3
} |
||
2306.04004-en-4 | 2306.04004-en-4.jpg | Randomized Schur Complement Views for Graph Contrastive Learning
the approximation of schur complement SC(G,Vs),Vs C $\mathcal{V}$ given by rLap, D$_{\mathcal{V}_s}$ is a diagonal matrix formed by selecting entries corresponding to $\mathcal{V} _s\textit{from D and }\widetilde{\mathbf{A} } _{\mathbf{R} }, \widetilde{\mathbf{D} } _{\mathbf{R} }$ represent the adjacency and degree matrices w.r.t R.
tation. Finally, STAR(R, $v_i)$ is subtracted and C is added to
R to continue the elimination process.
Theorem 3.1. Given an undirected graph $\mathcal{G}=(\mathcal{V},\mathcal{E},w)$, node dropping fraction $\gamma$, node elimination scheme $o_v$ and neighbor ordering scheme on, the output of $rLap( \mathcal{G} , \gamma , o_v, o_n) \textit{is an unbiased estimator of the schur}$ complement $SC( \mathcal{G} , \mathcal{V} _{\mathbf{R} _{\gamma | \mathcal{V} | }}) , where$ $\mathcal{V} _{\mathbf{R} _{\gamma | \mathcal{V} | }}\subset \mathcal{V}$ denotes the set of nodes that remain after y |V| eliminations.
$Proof.$ Without loss of generality, let $\mathbf{R}_{i-1}$ be the state after eliminating $i-1$ nodes. Let $v_i$ indicate the node which is being eliminated in the $i^{th}$ iteration of the outer loop. The proof is based on the loop invariant that $\mathbb{E}[\mathbf{R}_i]=$ $SC(\mathcal{G},\mathcal{V}_{\mathbf{R}_{i-1}}\setminus v_{i})$ after the end of this iteration. With this guarantee, we continue the elimination process for $\gamma|\mathcal{V}|$ iterations and achieve the desired randomized Schur com- $\begin{array}{ll}{\text{plement which equals }SC(\mathcal{G},\mathcal{V}_{\mathbf{R}_{\gamma|\mathcal{V}|}})\mathrm{~under~expectation.}}\\{\text{plement which equals }SC(\mathcal{G},\mathcal{V}_{\mathbf{R}_{\gamma|\mathcal{V}|}})\mathrm{~under~expectation.}}\end{array}$ The proof is available in Appendix B along with the tail bounds of deviation for the laplacian matrix martingale. $\square$
## 3.2. Connections with Graph Diffusion
For a graph $\mathcal{G}=(\mathcal{V},\mathcal{E},w)$, the diffusion operator is a polynomial filter on $\mathcal{G}$ which mitigates the effect of noisy neighbors. The diffused output S is given by:
(6)
$$\mathbf{S}=\sum_{i=0}^\infty\rho_i\mathbf{T}^i$$
Where T is a transition matrix (Klicpera et al.,2019) and can be chosen as either $\mathbf{T}_{rw}=\mathbf{AD}^{-1}$ or $\mathbf{T}_{sym}=$ $\mathbf{D}^{-1/2}\mathbf{AD}^{-1/2},\mathbf{A},\mathbf{D}$ are the adjacency and degree matrices for $\mathcal{G}$ respectively, $\rho_i$ is the scaling coefficient with $\rho_{i}^{PPR}=\alpha\cdot(1-\alpha)^{i},\alpha\in(0,1)$ and $\rho_i^{HK}=e^{-t}\cdot\frac{t^{i}}{i!}$ representing the coefficients for personalized page rank (Page et al., 1999) and heat kernels with diffusion time $t$ (Kondor & Lafferty, 2002; Chung, 2007) respectively. When $\mathbf{T}=\mathbf{T}_{rw}=\mathbf{AD}^{-1}$, the closed forms of diffusions are given by:
$$\begin{aligned}\mathbf{S}^{PPR}&=\alpha(\mathbf{I}-(1-\alpha)\mathbf{A}\mathbf{D}^{-1})^{-1}\\\mathbf{S}^{HK}&=\exp(t\cdot\mathbf{A}\mathbf{D}^{-1}-t\mathbf{I})\end{aligned}$$
(7)
$\mathrm{Note~that,~for~S}^{PPR}$,computing$(\mathbf{I}-(1-\alpha)\mathbf{AD}^{-1})^{-1}$ is an expensive operation over the entire graph and leads to undesired bottlenecks when the desired augmented view during contrasting is just a relatively smaller sub-graph. We address this issue in the following theorem using $rLap.$
Theorem $3.2.Let\rho_i=c\cdot\beta^i\in\mathbb{R},\beta\to1.Thesub$-graph $with~nodes~\mathcal{V}_{s}~sampled~from~\mathbf{S}=(\mathbf{I}-(1-\alpha)\mathbf{AD}^{-1})^{-1}$ $s$ given $by$ $S_{\mathcal{V} _{s}}= \left ( \lim _{K\to \infty }\sum _{k= 0}^{K}c\cdot \beta ^{k}\cdot ( \mathbf{AD}^{- 1}) ^{k}\right ) _{\mathcal{V} _{s}}=$ in Algorithm 2 can be used for learning node and graph features via different design choices for encoders (includ-
$\mathbb{E} \left [ \operatorname* { lim} _{K\to \infty }\sum _{k= 0}^{K}c\cdot \beta ^{k}\cdot \mathbf{D} _{{\mathcal{V} }_{s}}\widetilde{\mathbf{D} } _{\mathbf{R} }^{- 1}( \widetilde{\mathbf{A} } _{\mathbf{R} }\widetilde{\mathbf{D} } _{\mathbf{R} }^{- 1}) ^{k}\right ] , where$ $\mathbf{R}$ $i.$
$Proof.$ The proof is a simple expansion of the limit:
$$\begin{aligned}&\mathbb{E}\Bigg[\operatorname*{lim}_{K\to\infty}\sum_{k=0}^{K}c\cdot\beta^{k}\mathbf{D}_{\mathcal{V}_{s}}\widetilde{\mathbf{D}}_{\mathbf{R}}^{-1}(\widetilde{\mathbf{A}}_{\mathbf{R}}\widetilde{\mathbf{D}}_{\mathbf{R}}^{-1})^{k}\Bigg]\\&=c\mathbf{D}_{\mathcal{V}_{s}}\mathbb{E}\big[\widetilde{\mathbf{D}}_{\mathbf{R}}^{-1}\operatorname*{lim}_{K\to\infty}\sum_{k=0}^{K}(\beta\widetilde{\mathbf{A}}_{\mathbf{R}}\widetilde{\mathbf{D}}_{\mathbf{R}}^{-1})^{k}\big]\\&=c\mathbf{D}_{\mathcal{V}_{s}}\mathbb{E}\big[\widetilde{\mathbf{D}}_{\mathbf{R}}^{-1}\cdot(\mathbf{I}-\beta\widetilde{\mathbf{A}}_{\mathbf{R}}\widetilde{\mathbf{D}}_{\mathbf{R}}^{-1})^{-1}\big]\\&=c\mathbf{D}_{\mathcal{V}_{s}}\mathbb{E}\big[(\widetilde{\mathbf{D}}_{\mathbf{R}}-\beta\widetilde{\mathbf{A}}_{\mathbf{R}})^{-1}\big]\\&=c\mathbf{D}_{\mathcal{V}_{s}}\cdot SC(\mathbf{D}-\beta\mathbf{A},\:\mathcal{V}_{s})^{-1}\\&=c\mathbf{D}_{\mathcal{V}_{s}}(\mathbf{D}-\beta\mathbf{A})_{\mathcal{V}_{s}}^{-1}=\mathbf{S}_{\mathcal{V}_{s}}\end{aligned}$$
The equality $\mathbb{E}[(\widetilde{\mathbf{D}}_{\mathbf{R}}-\beta\widetilde{\mathbf{A}}_{\mathbf{R}})^{-1}]=SC(\mathbf{D}-\beta\mathbf{A},\mathcal{V}_{s})^{-1}$ is based on the guarantees of Theorem 3.1 when $\beta\to1$ and $\mathbf{D}-\beta\mathbf{A}\to\mathbf{L}.$ The equality $SC(\mathbf{D}-\beta\mathbf{A},\mathcal{V}_{s})^{-1}=(\mathbf{D}-$ $\beta\mathbf{A})_{\mathcal{V}_s}^{-1}$ is based on the standard matrix inversion lemma. Refer to Gallier (2010) and Appendix C.4 in Boyd et al.
口
(2004) for further details.
When $\beta$ is not close to 1, D- βA doesn't represent a graph laplacian but instead represents a symmetric diagonally dominant matrix (SDDM). Similar to the laplacian property of Schur complements, it can be shown that Schur complements of SDDM matrices are also SDDM (refer Appendix A, Lemma A.1 in Fahrbach et al.(2020) for the proof). To better understand the implications of Theorem $\hat{3}.2$,consider a graph $\mathcal{G}=(\mathcal{V},\mathcal{E},w)$ on which the PPR diffusion operator needs to be applied, followed by a sub-graph sampling operation with respect to nodes $\mathcal{V}_s\subset\mathcal{V}.$ Such a requirement is pervasive in GCL when views are based on diffusion matrices. However, the overheads of computing S for medium-large scale graphs can be significant. Based on the empirical analysis of Klicpera et al. (2019), the optimal choice of $\alpha$ for $\rho_i=\alpha\cdot(1-\alpha)^i$ typically lies in a close range of $\alpha\in[0.05,0.2].$ This implies $\beta=1-\alpha$ is close to 1, which justifes our assumption in Theorem 3.2 for practical settings. Thus, instead of diffusion followed by sub-graph sampling, one can apply rLap on $\mathcal{G}$ to obtain the randomized Schur complement which is relatively sparse, followed by the diffusion operator (up to $\mathbf{D}_{\nu_{s}}\widetilde{\mathbf{D}}_{\mathbf{R}}^{-1}$ scaling). See Figure 2 for an illustration.
# 3.3. Augmentation with rLap
We present a generalized GCL framework with rLap as the augmentor in Algorithm 2. Given a collection of graphs $G=\{{\mathcal G}^{i},i\in\{1,\cdots,M\},M\geq1$,the blueprint presented ing projectors) and contrastive objectives. For every graph $\mathcal{G}^i\in G,rLap$ generates the randomized schur complement
5 | [
[
"0.2607",
"0.05682",
"0.714",
"0.05682",
"0.714",
"0.0707",
"0.2607",
"0.0707"
],
[
"0.08905",
"0.0859",
"0.4756",
"0.0859",
"0.4756",
"0.09973",
"0.08905",
"0.09973"
],
[
"0.5",
"0.0859",
"0.8823",
"0.0859",
"0.8823",
"0.09973",
"0.5",
"0.09973"
],
[
"0.08826",
"0.09973",
"0.3423",
"0.1016",
"0.3423",
"0.11554",
"0.08826",
"0.113"
],
[
"0.501",
"0.1016",
"0.8857",
"0.1016",
"0.8857",
"0.11554",
"0.501",
"0.11554"
],
[
"0.0907",
"0.1193",
"0.4739",
"0.1193",
"0.4739",
"0.1332",
"0.0907",
"0.1332"
],
[
"0.5",
"0.11743",
"0.8823",
"0.11743",
"0.8823",
"0.1307",
"0.5",
"0.1307"
],
[
"0.08905",
"0.1364",
"0.4739",
"0.1364",
"0.4739",
"0.1497",
"0.08905",
"0.1497"
],
[
"0.5",
"0.1332",
"0.8447",
"0.1332",
"0.8447",
"0.1465",
"0.5",
"0.1465"
],
[
"0.08905",
"0.1509",
"0.4756",
"0.1509",
"0.4756",
"0.1642",
"0.08905",
"0.1642"
],
[
"0.08826",
"0.1654",
"0.4739",
"0.1642",
"0.4739",
"0.1787",
"0.08826",
"0.1799"
],
[
"0.501",
"0.1666",
"0.839",
"0.1666",
"0.839",
"0.1799",
"0.501",
"0.1799"
],
[
"0.0866",
"0.1799",
"0.4756",
"0.1799",
"0.4756",
"0.1976",
"0.0866",
"0.1976"
],
[
"0.523",
"0.1957",
"0.5825",
"0.1957",
"0.5825",
"0.2108",
"0.523",
"0.2108"
],
[
"0.5728",
"0.1932",
"0.7817",
"0.1895",
"0.7817",
"0.2108",
"0.5728",
"0.2153"
],
[
"0.5034",
"0.1976",
"0.5293",
"0.1976",
"0.5293",
"0.209",
"0.5034",
"0.209"
],
[
"0.08826",
"0.1963",
"0.406",
"0.1963",
"0.406",
"0.2102",
"0.08826",
"0.2102"
],
[
"0.5",
"0.2184",
"0.7837",
"0.2184",
"0.7837",
"0.2405",
"0.5",
"0.2405"
],
[
"0.08905",
"0.2267",
"0.4739",
"0.2267",
"0.4739",
"0.2405",
"0.08905",
"0.2405"
],
[
"0.0907",
"0.2437",
"0.4722",
"0.2437",
"0.4722",
"0.254",
"0.0907",
"0.254"
],
[
"0.5",
"0.2412",
"0.7485",
"0.2437",
"0.7485",
"0.2615",
"0.5",
"0.2588"
],
[
"0.08826",
"0.2568",
"0.4739",
"0.2568",
"0.4739",
"0.2708",
"0.08826",
"0.2708"
],
[
"0.501",
"0.262",
"0.689",
"0.262",
"0.689",
"0.2766",
"0.501",
"0.2766"
],
[
"0.08826",
"0.2727",
"0.4722",
"0.2727",
"0.4722",
"0.286",
"0.08826",
"0.286"
],
[
"0.5",
"0.276",
"0.7036",
"0.2734",
"0.7036",
"0.2917",
"0.5",
"0.2935"
],
[
"0.0866",
"0.286",
"0.4756",
"0.2842",
"0.4756",
"0.3018",
"0.0866",
"0.3037"
],
[
"0.4983",
"0.2903",
"0.693",
"0.2922",
"0.693",
"0.3093",
"0.4983",
"0.307"
],
[
"0.0907",
"0.303",
"0.4722",
"0.303",
"0.4722",
"0.3164",
"0.0907",
"0.3164"
],
[
"0.08905",
"0.3176",
"0.4739",
"0.3176",
"0.4739",
"0.3308",
"0.08905",
"0.3308"
],
[
"0.501",
"0.318",
"0.8857",
"0.318",
"0.8857",
"0.3333",
"0.501",
"0.3333"
],
[
"0.08905",
"0.3333",
"0.4739",
"0.3333",
"0.4739",
"0.348",
"0.08905",
"0.348"
],
[
"0.4983",
"0.3333",
"0.8857",
"0.334",
"0.8857",
"0.3484",
"0.4983",
"0.348"
],
[
"0.08905",
"0.348",
"0.4756",
"0.348",
"0.4756",
"0.361",
"0.08905",
"0.361"
],
[
"0.501",
"0.3484",
"0.875",
"0.3484",
"0.875",
"0.363",
"0.501",
"0.363"
],
[
"0.08826",
"0.363",
"0.4756",
"0.363",
"0.4756",
"0.377",
"0.08826",
"0.377"
],
[
"0.5024",
"0.3643",
"0.886",
"0.3643",
"0.886",
"0.3774",
"0.5024",
"0.3774"
],
[
"0.501",
"0.3801",
"0.886",
"0.3801",
"0.886",
"0.3933",
"0.501",
"0.3933"
],
[
"0.08905",
"0.3933",
"0.3594",
"0.3933",
"0.3594",
"0.4072",
"0.08905",
"0.4072"
],
[
"0.501",
"0.3945",
"0.669",
"0.3945",
"0.669",
"0.4077",
"0.501",
"0.4077"
],
[
"0.8667",
"0.3977",
"0.8813",
"0.3914",
"0.8896",
"0.4048",
"0.876",
"0.4104"
],
[
"0.08826",
"0.4185",
"0.4722",
"0.4185",
"0.4722",
"0.432",
"0.08826",
"0.432"
],
[
"0.5",
"0.4268",
"0.886",
"0.428",
"0.886",
"0.4426",
"0.5",
"0.4414"
],
[
"0.08826",
"0.433",
"0.4739",
"0.433",
"0.4739",
"0.4463",
"0.08826",
"0.4463"
],
[
"0.0866",
"0.449",
"0.3506",
"0.449",
"0.3506",
"0.4622",
"0.0866",
"0.4622"
],
[
"0.501",
"0.4443",
"0.8857",
"0.443",
"0.8857",
"0.457",
"0.501",
"0.4583"
],
[
"0.501",
"0.459",
"0.886",
"0.459",
"0.886",
"0.473",
"0.501",
"0.473"
],
[
"0.5",
"0.4749",
"0.886",
"0.4734",
"0.886",
"0.4873",
"0.5",
"0.488"
],
[
"0.2361",
"0.483",
"0.2688",
"0.483",
"0.2688",
"0.4944",
"0.2361",
"0.4944"
],
[
"0.2656",
"0.483",
"0.3245",
"0.483",
"0.3245",
"0.4976",
"0.2656",
"0.4976"
],
[
"0.451",
"0.483",
"0.4756",
"0.483",
"0.4756",
"0.4968",
"0.451",
"0.4968"
],
[
"0.5",
"0.4893",
"0.8857",
"0.4893",
"0.8857",
"0.5024",
"0.5",
"0.5024"
],
[
"0.267",
"0.4988",
"0.2932",
"0.4988",
"0.2932",
"0.5103",
"0.267",
"0.5103"
],
[
"0.5",
"0.505",
"0.8857",
"0.505",
"0.8857",
"0.518",
"0.5",
"0.518"
],
[
"0.08826",
"0.518",
"0.4739",
"0.518",
"0.4739",
"0.532",
"0.08826",
"0.532"
],
[
"0.5",
"0.5195",
"0.886",
"0.518",
"0.886",
"0.532",
"0.5",
"0.5327"
],
[
"0.08905",
"0.5327",
"0.4739",
"0.5327",
"0.4739",
"0.5474",
"0.08905",
"0.5474"
],
[
"0.5",
"0.534",
"0.8857",
"0.534",
"0.8857",
"0.549",
"0.5",
"0.549"
],
[
"0.08826",
"0.547",
"0.4739",
"0.549",
"0.4739",
"0.563",
"0.08826",
"0.561"
],
[
"0.4983",
"0.549",
"0.884",
"0.549",
"0.884",
"0.5625",
"0.4983",
"0.5625"
],
[
"0.0866",
"0.561",
"0.4756",
"0.5625",
"0.4756",
"0.58",
"0.0866",
"0.579"
],
[
"0.4983",
"0.5645",
"0.8857",
"0.563",
"0.8857",
"0.5776",
"0.4983",
"0.579"
],
[
"0.0866",
"0.5747",
"0.4763",
"0.577",
"0.4763",
"0.5977",
"0.0866",
"0.595"
],
[
"0.5",
"0.58",
"0.886",
"0.58",
"0.886",
"0.5933",
"0.5",
"0.5933"
],
[
"0.08826",
"0.595",
"0.4739",
"0.595",
"0.4739",
"0.6094",
"0.08826",
"0.6094"
],
[
"0.5",
"0.5947",
"0.8857",
"0.595",
"0.8857",
"0.6094",
"0.5",
"0.608"
],
[
"0.08905",
"0.61",
"0.4756",
"0.61",
"0.4756",
"0.6235",
"0.08905",
"0.6235"
],
[
"0.5",
"0.61",
"0.8857",
"0.61",
"0.8857",
"0.6235",
"0.5",
"0.6235"
],
[
"0.08826",
"0.6255",
"0.4756",
"0.6255",
"0.4756",
"0.6396",
"0.08826",
"0.6396"
],
[
"0.4983",
"0.625",
"0.8857",
"0.6255",
"0.8857",
"0.6396",
"0.4983",
"0.638"
],
[
"0.08826",
"0.6396",
"0.4722",
"0.6396",
"0.4722",
"0.654",
"0.08826",
"0.654"
],
[
"0.501",
"0.64",
"0.8857",
"0.64",
"0.8857",
"0.654",
"0.501",
"0.654"
],
[
"0.08905",
"0.6577",
"0.1544",
"0.6577",
"0.1544",
"0.6694",
"0.08905",
"0.6694"
],
[
"0.501",
"0.656",
"0.886",
"0.656",
"0.886",
"0.6694",
"0.501",
"0.6694"
],
[
"0.1626",
"0.675",
"0.3906",
"0.678",
"0.3906",
"0.6987",
"0.1626",
"0.696"
],
[
"0.501",
"0.6704",
"0.886",
"0.6704",
"0.886",
"0.6846",
"0.501",
"0.6846"
],
[
"0.501",
"0.685",
"0.884",
"0.685",
"0.884",
"0.6987",
"0.501",
"0.6987"
],
[
"0.1691",
"0.6963",
"0.367",
"0.6997",
"0.367",
"0.7183",
"0.1691",
"0.7153"
],
[
"0.451",
"0.6924",
"0.4756",
"0.6924",
"0.4756",
"0.705",
"0.451",
"0.705"
],
[
"0.501",
"0.7007",
"0.884",
"0.7007",
"0.884",
"0.714",
"0.501",
"0.714"
],
[
"0.501",
"0.7173",
"0.8857",
"0.7173",
"0.8857",
"0.728",
"0.501",
"0.728"
],
[
"0.0866",
"0.731",
"0.4722",
"0.731",
"0.4722",
"0.7485",
"0.0866",
"0.7485"
],
[
"0.501",
"0.731",
"0.886",
"0.731",
"0.886",
"0.7456",
"0.501",
"0.7456"
],
[
"0.08826",
"0.7505",
"0.4739",
"0.75",
"0.4739",
"0.7646",
"0.08826",
"0.765"
],
[
"0.501",
"0.7476",
"0.709",
"0.7476",
"0.709",
"0.7583",
"0.501",
"0.7583"
],
[
"0.08826",
"0.7666",
"0.4739",
"0.7666",
"0.4739",
"0.78",
"0.08826",
"0.78"
],
[
"0.08826",
"0.781",
"0.4756",
"0.781",
"0.4756",
"0.7944",
"0.08826",
"0.7944"
],
[
"0.4983",
"0.776",
"0.7036",
"0.7764",
"0.7036",
"0.7915",
"0.4983",
"0.7905"
],
[
"0.08826",
"0.797",
"0.451",
"0.797",
"0.451",
"0.81",
"0.08826",
"0.81"
],
[
"0.5",
"0.8003",
"0.8857",
"0.8003",
"0.8857",
"0.815",
"0.5",
"0.815"
],
[
"0.08826",
"0.813",
"0.4739",
"0.8145",
"0.4739",
"0.829",
"0.08826",
"0.8276"
],
[
"0.501",
"0.8164",
"0.8857",
"0.8164",
"0.8857",
"0.8306",
"0.501",
"0.8306"
],
[
"0.08826",
"0.83",
"0.4739",
"0.83",
"0.4739",
"0.8433",
"0.08826",
"0.8433"
],
[
"0.501",
"0.8306",
"0.8857",
"0.8306",
"0.8857",
"0.8447",
"0.501",
"0.8447"
],
[
"0.2263",
"0.8496",
"0.4307",
"0.8477",
"0.4307",
"0.869",
"0.2263",
"0.872"
],
[
"0.5",
"0.845",
"0.8857",
"0.845",
"0.8857",
"0.86",
"0.5",
"0.86"
],
[
"0.08905",
"0.853",
"0.2035",
"0.853",
"0.2035",
"0.866",
"0.08905",
"0.866"
],
[
"0.501",
"0.8623",
"0.886",
"0.8623",
"0.886",
"0.8755",
"0.501",
"0.8755"
],
[
"0.1054",
"0.881",
"0.3945",
"0.8774",
"0.3945",
"0.903",
"0.1054",
"0.906"
],
[
"0.501",
"0.877",
"0.8857",
"0.877",
"0.8857",
"0.89",
"0.501",
"0.89"
],
[
"0.09235",
"0.888",
"0.1152",
"0.888",
"0.1152",
"0.8965",
"0.09235",
"0.8965"
],
[
"0.3857",
"0.8833",
"0.4739",
"0.881",
"0.4739",
"0.8984",
"0.3857",
"0.901"
],
[
"0.501",
"0.8926",
"0.886",
"0.8926",
"0.886",
"0.906",
"0.501",
"0.906"
],
[
"0.4878",
"0.922",
"0.4983",
"0.9297",
"0.4878",
"0.938",
"0.4763",
"0.9297"
]
] | [
"Randomized Schur Complement Views for Graph Contrastive Learning",
"tation. Finally, STAR(R, vs) is subtracted and C is added to",
"the approximation of schur complement SC(g, V.),Vs ",
"Rto continue theelimination process.",
"V given by rLap, Dy is a diagonal matrix formed by",
"Theorem 3.1. Given an undirected graph g = (V, &, w),",
"selecting entries corresponding to Vs from D and AR, DR",
"node dropping fraction , node elimination scheme",
"represent the adjacency and degree matrices w.rt R.",
"Ou and neighbor ordering scheme On, the output of",
"rLap(G, , ow, On) is an unbiased estimator of the schur",
"Proof. The proof is a simple expansion of the limit:",
"complement SC(G, VRy/v)), where VRyiv C V denotes the",
"lim Z",
"Ck=o C·β*Dv,Dr'(ARDR)",
"E",
"set of nodes that remain after /V eliminations.",
"=cDvE[DR Jim ZK-0(BARDR'),]",
"Proof. Without loss of generality, let R,-1 be the state",
"after eliminating z-l nodes.Let u; indicate thenode",
"= cDvE[DR1.(I-βARDR\")-1]",
"which is being eliminated in the ith iteration of the outer",
"= cDy,E[(DR - βAR)-1",
"loop. The proof is based on the loop invariant that E[R;l =",
"= cDv.. SC(D - BA,V)-1",
"SC(G, VR.-1 /vi) after the end of this iteration. With this",
"= cDv(D - βA)l = Sva",
"guarantee, we continue the elimination process for V",
"iterations and achieve the desired randomized Schur com-",
"The equality E[(DR -βAR)-1] = SC(D -βA, Vs)-1",
"plement which equals SC(G, VR/vj) under expectation.",
"is based on the guarantees of Theorem 3.1 when β-→1 and",
"The proof is available in Appendix B along with the tail",
"D- βA → L. The equality SC(D - βA, Vs)-1 = (D",
"bounds of deviation for the laplacian matrix martingale.",
"BA)-1 is based on the standard matrix inversion lemma.",
"Refer to Gallier (2010) and Appendix C.4 in Boyd et al.",
"3.2. Connections with Graph Diffusion",
"(2004) for further details.",
"△",
"For a graph G = (V, &, w), the diffusion operator is a poly",
"When β is not close to 1, D - βA doesn't represent a",
"nomial filter on g which mitigates the effect of noisy neigh-",
"bors. The diffused output S is given by:",
"graph laplacian but instead represents a symmetric diago",
"nally dominant matrix (SDDM). Similar to the laplacian",
"property of Schur complements,it can be shown that Schur",
"S=",
">p,T",
"(6)",
"complements of SDDM matrices are also SDDM (refer Ap-",
"i=0",
"pendix A, Lemma A.1 in Fahrbach et al. (2020) for the",
"Where T is a transition matrix (Klicpera et al., 2019)",
"proof).Tobetterunderstand the implications ofTheorem",
"and can be chosen as either Trw = AD-1 or Tsum =",
"3.2, consider a graph g = (V, &, w) on which the PPR diffu-",
"D-1/2 AD-1/2, A, D are the adjacency and degree ma-",
"sion operator needs to be applied, followed by a sub-graph",
"trices for g respectively, p; is the scaling coefficient with",
"sampling operation with respect to nodes V, C V. Such a",
"pPPR =Q. (1-a),αE (0,1) and pHK = e-t. rep-",
"requirement is pervasive in GCL when views are based on",
"resenting the coefficients for personalized page rank (Page",
"diffusion matrices.However, the overheads of computing",
"et al., 1999) and heat kernels with diffusion time t (Kon-",
"S for medium-large scale graphs can be significant. Based",
"dor &Lafferty,2002;Chung,2007)respectively.When",
"on the empirical analysis of Klicpera etal.(2019),theop-",
"T = Trw = AD-1, the closed forms of diffusions are",
"timal choice of α for pi = α - (1 - α) typically lies in a",
"given by:",
"close range of α E [0.05, 0.2]. This implies β = 1 - α is",
"sPPR = α(I - (1 -α)AD-1)-",
"close to 1, which justifies our assumption in Theorem 3.2",
"for practical settings. Thus, instead of diffusion followed by",
"sHK = exp(t· AD-1-tI)",
"(7)",
"sub-graph sampling, one can apply r Lap on G to obtain the",
"randomized Schurcomplement whichisrelativelysparse",
"Note that, for SPPR, computing (I - (1 - @)AD-1)-1",
"followed by the diffusion operator (up to Dv, DR scaling)",
"is an expensive operation over the entire graph and leads",
"See Figure 2 for an illustration.",
"to undesired bottlenecks when the desired augmented view",
"during contrasting is just a relatively smaller sub-graph. We",
"3.3.AugmentationwithrLap",
"address this issue in the following theorem using r Lap.",
"We present a generalized GCL framework with rLap as the",
"Theorem 3.2. Let pi = c.Bi e R, β → 1. The sub-graph",
"augmentor in Algorithm 2. Given a collection of graphs",
"with nodes Vs sampled from S = (I- (1 -α)AD-1)-1",
"G = {gi, i E {1, ..-, M], M ≥ 1, the blueprint presented",
"kim,ZK-oc. βk (AD-1)k",
"in Algorithm 2 can be used for learning node and graph",
"is given by Sy:",
"features via different design choices for encoders (includ-",
"imEKoc.Bh.Dv,Dr(ARDR)",
"ing projectors) and contrastive objectives. For every graph",
"E",
",where Ris",
"gi E G, r Lap generates the randomized schur complement",
"5"
] | {
"arxiv_id": "2306.04004",
"language": "en",
"page_idx": 4
} |
||
2306.03678-en-0 | 2306.03678-en-0.jpg | On the Difference of BERT-style and CLIP-style Text Encoders
Zhihong Chen$^{1,2*}$ Guiming Hardy Chen$^{1,2*}$ Shizhe Diao
Xiang Wan$^2$ Benyou Wang$^{1,2\dagger}$
$^{1}$The Chinese University of Hong Kong, Shenzhen
$^{2}$Shenzhen Research Institute of Big Data
$^{3}$Hong Kong University of Science and Technology
$\{$zhihongchen,guimingchen}@link.cuhk.edu.cn sdiaoaa@connect.ust.hk
wanxiang@sribd.cn wangbenyou@cuhk.edu.cn
### $\mathbf{Abstract}$
Masked language modeling (MLM) has been one of the most popular pretraining recipes in natural language processing, $e.g.$, BERT, one of the representative models. Recently, contrastive language-image pretraining (CLIP) has also attracted attention, especially its vision models that achieve excellent performance on a broad range of vision tasks. However, few studies are dedicated to studying the text encoders learned by CLIP. In this paper, we analyze the difference between BERT-style and CLIP-style text encoders from three experiments:(i) general text understanding, (ii) vision-centric text understanding, and (iii) text-to-image generation. Experimental analyses show that although CLIP-style text encoders underperform BERTstyle ones for general text understanding tasks, they are equipped with a unique ability, $i.e.$, synesthesia, for the cross-modal association, which is more similar to the senses of humans.$^{\mathrm{l}}$
fill-in-the-blank task where a model uses the context words to predict masked tokens in a sequence. For this type, BERT (Devlin et al., 2018) and its variants (Liu et al., 2019; Sanh et al., 2019; Lan et al., 2019) are the representative encoder models allowing the bidirectional perception of texts. More recently, there has been another framework to produce text encoders, $i.e.$, contrastive languageimage pretraining (Radford et al., 2021) (CLIP). It trains image and text encoders through contrastive learning on a variety of image-text pairs, and the trained vision encoders achieve great success on vision-only tasks, especially its impressive zeroshot transfer results. However, few studies investigated the trained text encoders.
In this work, we first conduct a pilot study in §3 to benchmark BERT-style and CLIP-style text encoders in a popular natural language processing benchmark (Wang et al., 2018). It shows that CLIP-style text encoders significantly underperform BERT-style text encoders.
Therefore, a natural question arises:“Are CLIP text encoders useless byproducts (or in which case can we make use of the CLIP text encoders?)”. Our hypothesis for this question is that CLIP text encoder might additionally learn visual knowledge of textual concepts, which could be complementary to the semantic alignment of textual concepts; the latter is well-captured by BERT text encoders. This reminds us of a phenomenon called ‘synesthesia’’’’
To validate our synesthesia hypothesis, we conduct a side-by-side comparison between $BERT-$ style and CLIP-style text encoders from two aspects: 1) benchmarking them in vision-centric natural language understanding tasks in §4; and 2) further probing the generated images from their encoded text representation in §5. First, we
## 1 Introduction
Text representation learning provides a feasible solution to extract generic representations from texts, allowing models to better understand and make predictions about texts. Normally, to perform this, a language model is pretrained on large-scale text corpora to learn text representation in a self-supervised manner, and it can be further used on downstream tasks, e.g., text classification and question answering (Devlin et al., 2018).
There are many recipes for pretraining text encoders$^2$ (Mikolov et al., 2013; Pennington et al., 2014; Peters et al., 2018). One of the most popular ways is masked language modeling (MLM), a
$^{*}$Equal Contribution.
$^{\dagger}$Corresponding authors.
$^{1}$Our code is released at https://github.com/
zhjohnchan/bert-clip-synesthesia.
$^{2}$Similar exploration can be extended to decoder-based
models as well
Synesthesia is a phenomenon that stimulation in a sensory or cognitive modality might unintentionally activate the perception in another sensory or cognitive modality. $e.g.$, we might“see”shapes(vision) when listening to music(audition). | [
[
"0.1763",
"0.0861",
"0.8237",
"0.0861",
"0.8237",
"0.10095",
"0.1763",
"0.10095"
],
[
"0.236",
"0.111",
"0.7573",
"0.10986",
"0.7573",
"0.1312",
"0.236",
"0.1318"
],
[
"0.3645",
"0.13",
"0.638",
"0.1289",
"0.638",
"0.1467",
"0.3645",
"0.1478"
],
[
"0.2957",
"0.1455",
"0.708",
"0.1478",
"0.708",
"0.165",
"0.2957",
"0.1633"
],
[
"0.3315",
"0.1644",
"0.671",
"0.165",
"0.671",
"0.1799",
"0.3315",
"0.1787"
],
[
"0.2896",
"0.1787",
"0.708",
"0.1812",
"0.708",
"0.199",
"0.2896",
"0.1965"
],
[
"0.1704",
"0.1996",
"0.833",
"0.1996",
"0.833",
"0.2131",
"0.1704",
"0.2131"
],
[
"0.293",
"0.2162",
"0.7095",
"0.2162",
"0.7095",
"0.2299",
"0.293",
"0.2299"
],
[
"0.2612",
"0.253",
"0.34",
"0.253",
"0.34",
"0.2673",
"0.2612",
"0.2673"
],
[
"0.51",
"0.2542",
"0.882",
"0.2542",
"0.882",
"0.2673",
"0.51",
"0.2673"
],
[
"0.511",
"0.2708",
"0.8833",
"0.2708",
"0.8833",
"0.284",
"0.511",
"0.284"
],
[
"0.1453",
"0.284",
"0.4592",
"0.284",
"0.4592",
"0.2976",
"0.1453",
"0.2976"
],
[
"0.513",
"0.2861",
"0.882",
"0.2861",
"0.882",
"0.2998",
"0.513",
"0.2998"
],
[
"0.1436",
"0.2986",
"0.4575",
"0.2986",
"0.4575",
"0.3123",
"0.1436",
"0.3123"
],
[
"0.509",
"0.3018",
"0.882",
"0.3005",
"0.882",
"0.3154",
"0.509",
"0.3164"
],
[
"0.02351",
"0.3176",
"0.06046",
"0.3176",
"0.06046",
"0.376",
"0.02351",
"0.376"
],
[
"0.1436",
"0.3123",
"0.4592",
"0.3113",
"0.4592",
"0.326",
"0.1436",
"0.3267"
],
[
"0.511",
"0.3184",
"0.882",
"0.3184",
"0.882",
"0.332",
"0.511",
"0.332"
],
[
"0.1453",
"0.3267",
"0.4592",
"0.3267",
"0.4592",
"0.3403",
"0.1453",
"0.3403"
],
[
"0.511",
"0.3342",
"0.881",
"0.3342",
"0.881",
"0.3474",
"0.511",
"0.3474"
],
[
"0.1436",
"0.3403",
"0.4592",
"0.339",
"0.4592",
"0.354",
"0.1436",
"0.3552"
],
[
"0.1453",
"0.3552",
"0.4592",
"0.3552",
"0.4592",
"0.3682",
"0.1453",
"0.3682"
],
[
"0.511",
"0.3516",
"0.882",
"0.3516",
"0.882",
"0.3652",
"0.511",
"0.3652"
],
[
"0.02351",
"0.3635",
"0.06046",
"0.3635",
"0.06046",
"0.4312",
"0.02351",
"0.4312"
],
[
"0.1486",
"0.3706",
"0.4575",
"0.3706",
"0.4575",
"0.3813",
"0.1486",
"0.3813"
],
[
"0.511",
"0.3677",
"0.882",
"0.3677",
"0.882",
"0.3813",
"0.511",
"0.3813"
],
[
"0.147",
"0.383",
"0.4592",
"0.383",
"0.4592",
"0.3967",
"0.147",
"0.3967"
],
[
"0.51",
"0.383",
"0.882",
"0.3818",
"0.882",
"0.3967",
"0.51",
"0.398"
],
[
"0.1453",
"0.398",
"0.4575",
"0.398",
"0.4575",
"0.411",
"0.1453",
"0.411"
],
[
"0.51",
"0.3997",
"0.881",
"0.3997",
"0.881",
"0.4133",
"0.51",
"0.4133"
],
[
"0.147",
"0.411",
"0.4575",
"0.411",
"0.4575",
"0.4246",
"0.147",
"0.4246"
],
[
"0.511",
"0.415",
"0.882",
"0.415",
"0.882",
"0.4287",
"0.511",
"0.4287"
],
[
"0.1453",
"0.4258",
"0.4575",
"0.4258",
"0.4575",
"0.4395",
"0.1453",
"0.4395"
],
[
"0.511",
"0.4316",
"0.882",
"0.4316",
"0.882",
"0.4453",
"0.511",
"0.4453"
],
[
"0.02351",
"0.4382",
"0.06046",
"0.4382",
"0.06046",
"0.5103",
"0.02351",
"0.5103"
],
[
"0.1436",
"0.4395",
"0.4592",
"0.4407",
"0.4592",
"0.4536",
"0.1436",
"0.4531"
],
[
"0.511",
"0.4478",
"0.8833",
"0.4478",
"0.8833",
"0.4614",
"0.511",
"0.4614"
],
[
"0.147",
"0.456",
"0.4575",
"0.456",
"0.4575",
"0.466",
"0.147",
"0.466"
],
[
"0.1453",
"0.4685",
"0.4575",
"0.4685",
"0.4575",
"0.4822",
"0.1453",
"0.4822"
],
[
"0.511",
"0.4631",
"0.8833",
"0.4631",
"0.8833",
"0.4768",
"0.511",
"0.4768"
],
[
"0.147",
"0.484",
"0.4575",
"0.484",
"0.4575",
"0.4946",
"0.147",
"0.4946"
],
[
"0.511",
"0.4797",
"0.744",
"0.4797",
"0.744",
"0.4934",
"0.511",
"0.4934"
],
[
"0.1453",
"0.4963",
"0.4592",
"0.4963",
"0.4592",
"0.511",
"0.1453",
"0.511"
],
[
"0.529",
"0.4976",
"0.882",
"0.4976",
"0.882",
"0.511",
"0.529",
"0.511"
],
[
"0.147",
"0.511",
"0.461",
"0.511",
"0.461",
"0.525",
"0.147",
"0.525"
],
[
"0.51",
"0.5127",
"0.8833",
"0.514",
"0.8833",
"0.528",
"0.51",
"0.5254"
],
[
"0.1453",
"0.5254",
"0.461",
"0.5254",
"0.461",
"0.539",
"0.1453",
"0.539"
],
[
"0.01764",
"0.5156",
"0.05878",
"0.5156",
"0.0638",
"0.7236",
"0.02184",
"0.7236"
],
[
"0.511",
"0.53",
"0.882",
"0.53",
"0.882",
"0.5444",
"0.511",
"0.5444"
],
[
"0.1436",
"0.539",
"0.4575",
"0.539",
"0.4575",
"0.5527",
"0.1436",
"0.5527"
],
[
"0.51",
"0.546",
"0.882",
"0.5444",
"0.882",
"0.5586",
"0.51",
"0.56"
],
[
"0.1436",
"0.554",
"0.4568",
"0.554",
"0.4568",
"0.568",
"0.1436",
"0.568"
],
[
"0.513",
"0.5625",
"0.8833",
"0.5625",
"0.8833",
"0.576",
"0.513",
"0.576"
],
[
"0.513",
"0.5776",
"0.748",
"0.5776",
"0.748",
"0.5913",
"0.513",
"0.5913"
],
[
"0.12006",
"0.597",
"0.1335",
"0.6016",
"0.1293",
"0.61",
"0.11584",
"0.6055"
],
[
"0.1301",
"0.597",
"0.2603",
"0.597",
"0.2603",
"0.611",
"0.1301",
"0.611"
],
[
"0.531",
"0.5957",
"0.882",
"0.5957",
"0.882",
"0.6094",
"0.531",
"0.6094"
],
[
"0.511",
"0.6123",
"0.882",
"0.6123",
"0.882",
"0.626",
"0.511",
"0.626"
],
[
"0.11755",
"0.6235",
"0.4902",
"0.6235",
"0.4902",
"0.638",
"0.11755",
"0.638"
],
[
"0.513",
"0.6274",
"0.8833",
"0.6274",
"0.8833",
"0.641",
"0.513",
"0.641"
],
[
"0.11755",
"0.641",
"0.4895",
"0.641",
"0.4895",
"0.655",
"0.11755",
"0.655"
],
[
"0.513",
"0.6445",
"0.882",
"0.6445",
"0.882",
"0.6577",
"0.513",
"0.6577"
],
[
"0.11755",
"0.6567",
"0.4895",
"0.6567",
"0.4895",
"0.6704",
"0.11755",
"0.6704"
],
[
"0.513",
"0.661",
"0.881",
"0.661",
"0.881",
"0.675",
"0.513",
"0.675"
],
[
"0.11755",
"0.6733",
"0.4895",
"0.6733",
"0.4895",
"0.687",
"0.11755",
"0.687"
],
[
"0.511",
"0.677",
"0.881",
"0.677",
"0.881",
"0.6904",
"0.511",
"0.6904"
],
[
"0.11584",
"0.6895",
"0.4902",
"0.6895",
"0.4902",
"0.703",
"0.11584",
"0.703"
],
[
"0.511",
"0.6924",
"0.882",
"0.6924",
"0.882",
"0.706",
"0.511",
"0.706"
],
[
"0.1142",
"0.705",
"0.4878",
"0.7036",
"0.4878",
"0.7188",
"0.1142",
"0.7197"
],
[
"0.511",
"0.708",
"0.882",
"0.708",
"0.882",
"0.7217",
"0.511",
"0.7217"
],
[
"0.11584",
"0.7217",
"0.4878",
"0.7217",
"0.4878",
"0.7354",
"0.11584",
"0.7354"
],
[
"0.509",
"0.7246",
"0.8833",
"0.7227",
"0.8833",
"0.737",
"0.509",
"0.7393"
],
[
"0.11584",
"0.737",
"0.4895",
"0.737",
"0.4895",
"0.7505",
"0.11584",
"0.7505"
],
[
"0.529",
"0.742",
"0.8833",
"0.742",
"0.8833",
"0.756",
"0.529",
"0.756"
],
[
"0.11584",
"0.752",
"0.2988",
"0.752",
"0.2988",
"0.7656",
"0.11584",
"0.7656"
],
[
"0.511",
"0.7583",
"0.8833",
"0.7583",
"0.8833",
"0.7715",
"0.511",
"0.7715"
],
[
"0.1351",
"0.7686",
"0.4902",
"0.7695",
"0.4902",
"0.7837",
"0.1351",
"0.783"
],
[
"0.51",
"0.775",
"0.885",
"0.775",
"0.885",
"0.788",
"0.51",
"0.788"
],
[
"0.11584",
"0.785",
"0.4902",
"0.786",
"0.4902",
"0.8003",
"0.11584",
"0.8"
],
[
"0.511",
"0.7915",
"0.882",
"0.7915",
"0.882",
"0.8047",
"0.511",
"0.8047"
],
[
"0.11584",
"0.8003",
"0.4895",
"0.802",
"0.4895",
"0.8164",
"0.11584",
"0.8154"
],
[
"0.51",
"0.808",
"0.882",
"0.808",
"0.882",
"0.822",
"0.51",
"0.822"
],
[
"0.11584",
"0.8184",
"0.486",
"0.8184",
"0.486",
"0.832",
"0.11584",
"0.832"
],
[
"0.513",
"0.8237",
"0.882",
"0.8237",
"0.882",
"0.8374",
"0.513",
"0.8374"
],
[
"0.1394",
"0.843",
"0.2688",
"0.843",
"0.2688",
"0.8555",
"0.1394",
"0.8555"
],
[
"0.513",
"0.839",
"0.882",
"0.839",
"0.882",
"0.8525",
"0.513",
"0.8525"
],
[
"0.1411",
"0.8555",
"0.2896",
"0.8555",
"0.2896",
"0.869",
"0.1411",
"0.869"
],
[
"0.1377",
"0.866",
"0.4878",
"0.8677",
"0.4878",
"0.882",
"0.1377",
"0.8813"
],
[
"0.5356",
"0.8716",
"0.881",
"0.8716",
"0.881",
"0.882",
"0.5356",
"0.882"
],
[
"0.11584",
"0.88",
"0.3652",
"0.88",
"0.3652",
"0.8936",
"0.11584",
"0.8936"
],
[
"0.511",
"0.883",
"0.881",
"0.883",
"0.881",
"0.8936",
"0.511",
"0.8936"
],
[
"0.1351",
"0.891",
"0.4895",
"0.8926",
"0.4895",
"0.9067",
"0.1351",
"0.9062"
],
[
"0.513",
"0.8955",
"0.881",
"0.8955",
"0.881",
"0.9062",
"0.513",
"0.9062"
],
[
"0.11755",
"0.9062",
"0.2124",
"0.9062",
"0.2124",
"0.9175",
"0.11755",
"0.9175"
],
[
"0.513",
"0.9067",
"0.881",
"0.9067",
"0.881",
"0.9175",
"0.513",
"0.9175"
]
] | [
"On the Difference of BERT-style and CLIP-style Text Encoders",
"Zhihong Chenl,2*Guiming Hardy Chenl,2* Shizhe Diao3",
"Xiang Wan?Benyou Wangl,2t",
"1The Chinese University of Hong Kong, Shenzhen",
"2Shenzhen Research Institute of Big Data",
"3Hong Kong University of Science and Technology",
"zhihongchen,guimingchenj@link.cuhk.edu.cnsdiaoaa@connect.ust.hk",
"wanxiang@sribd.cn wangbenyou@cuhk.edu.cn",
"Abstract",
"fill-in-the-blank task where a model uses the con-",
"text words to predict masked tokens in a sequence.",
"Masked language modeling (MLM) has been",
"For this type, BERT (Devlin et al., 2018) and its",
"one of the most popular pretraining recipes in",
"variants (Liu et al., 2019; Sanh et al., 2019; Lan",
"",
"natural language processing, e.g., BERT, one",
"et al.,2019) are the representative encoder mod-",
"of the representative models. Recently, con-",
"els allowing the bidirectional perception of texts",
"trastive language-image pretraining (CLIP) has",
"also attracted attention, especially its vision",
"More recently, there has been another framework",
"unr9",
"modelsthat achieveexcellentperformanceona",
"to produce text encoders, i.e., contrastive language-",
"broad range of vision tasks. However, few stud-",
"image pretraining (Radford et al., 2021) (CLIP). It",
"ies are dedicated to studying the text encoders",
"trains image and text encoders through contrastive",
"learned by CLIP. In this paper, we analyze the",
"learning on a variety of image-text pairs, and the",
"difference between BERT-style and CLIP-style",
"trained vision encoders achieve great success on",
"['S]",
"text encoders from three experiments: (i) gen-",
"vision-only tasks, especially its impressive zero-",
"eral text understanding,(ii)vision-centrictext",
"understanding, and (ii) text-to-image genera-",
"shot transfer results. However, few studies investi-",
"tion. Experimental analyses show that although",
"gated the trained text encoders.",
"CLIP-style text encoders underperform BERT-",
"In this work, we first conduct a pilot study in",
"style ones for general text understanding tasks,",
"$3tobenchmarkBERT-style andCLIP-styletext",
"they are equipped with a unique ability, i.e.,",
"89007",
"encoders in a popular natural language process-",
"synesthesia, for the cross-modal association,",
"ing benchmark (Wang et al., 2018). It shows that",
"which is more similar to the senses of humans.I",
"CLIP-style text encoders significantly underper-",
"form BERT-style text encoders.",
"1",
"Introduction",
"Therefore, a natural question arises: “Are CLIP",
"text encoders useless byproducts (or in which case",
"Text representation learning provides a feasible so-",
"can we make use of the CLIP text encoders?)\".",
"lution to extract generic representations from texts,",
"Our hypothesis for this question is that CLIP text",
"allowing models to better understand and make pre-",
"encoder might additionally learn visual knowledge",
"dictions about texts. Normally, to perform this, a",
"of textual concepts, which could be complementary",
"language model is pretrained on large-scale text cor-",
"to the semantic alignment of textual concepts; the",
" pora to learn text representation in a self-supervised",
"latter is well-captured by BERT text encoders.This",
"manner, and it can be further used on downstream",
"reminds us of a phenomenon called 'synesthesia'3.",
"tasks, e.g., text classification and question answer-",
"To validate our synesthesia hypothesis, we con-",
"ing (Devlin et al., 2018).",
"duct a side-by-side comparison between BERT-",
"There are many recipes for pretraining text en-",
"style and CLIP-style text encoders from two as-",
"coders? (Mikolov et al., 2013; Pennington et al.,",
"pects: 1) benchmarking them in vision-centric",
"2014; Peters et al., 2018). One of the most popu-",
"natural language understanding tasks in s4; and",
"lar ways is masked language modeling (MLM), a",
"2) further probing the generated images from",
"\"Equal Contribution.",
"their encoded text representation in s5. First, we",
"Corresponding authors.",
"'Our code is released at https://github.com/",
"'Synesthesia isaphenomenonthat stimulation in a sen-",
"zhjohnchan/bert-clip-synesthesia.",
"sory or cognitive modalitymightunintentionally activate the",
"2similar exploration can be extended to decoder-based",
"perceptioninanothersensoryorcognitivemodality.e.g.,we",
"models as well.",
"might\"see\"shapes (vision)when listeningtomusic (audition)."
] | {
"arxiv_id": "2306.03678",
"language": "en",
"page_idx": 0
} |
||
2306.03678-en-2 | 2306.03678-en-2.jpg | 3 Pilot study: general text understanding
3.1 Experimental settings
For the datasets, we adopt two text classification tasks (CoLA (Warstadt et al.,2019) and SST2 (Socher et al., 2013)), three text similarity tasks (MRPC (Dolan and Brockett, 2005), QQP (Shankar et al., 2017), and STS-B (Cer et al., 2017)), and three inference tasks (MNLI (Williams et al., 2017), QNLI (Rajpurkar et al., 2016), and RTE (Bentivogli et al.,2009)) of GLUE.$^{5}$ For BERT-style text encoders, we adopt the base and large versions of BERT and RoBERTa; For CLIPstyle text encoders, we adopt the text encoders of four versions of CLIP $(i.e.$, ViT-B/32, ViT-B/16, ViT-L/14, and ViT-L/14@336px), where the first two text encoders share the same architecture but have different parameters (same for the last two) due to the difference of the vision branches. We adopt the commonly used metric for each dataset.
# 3.2 Empirical findings
CLIP text encoders perform poorly in GLUE We report the results in Table 1. The four CLIP-style text encoders consistently underperform BERT-style text encoders on all the datasets, where the CLIP-style text encoders achieve around 85% of the scores of BERT-style ones on average. Comparing among different datasets, the most significant performance gap occurs in the CoLA dataset. The reason behind this is that CoLA is an English acceptability dataset that requires a model to identify whether a sequence of words is a grammatical English sentence. This demonstrates that ITC is worse than MLM on $\underline{\textbf{grammatical or syntactic properties}}.$This fnd- $\overline{\text{ing is consistent with Yuksekgonul et al.(2022),}}$ which shows that the vision-and-language models trained by ITC ignore word orders and therefore lack understanding of the compositional structure in the images and captions.
# 4 CLIP-style text encoders capture visual perception for concept similarity
Motivation The aforementioned experiments show that BERT-style text encoders outperform CLIP-style text encoders on pure text tasks. Therefore, a question is “Is there any text task for us to testify the superior of CLIP-style text encoders?".
5We exclude the WNLI dataset (Levesque et al., 2012) due
to the large variance of the results on it.
<table>
<tbody>
<tr>
<th rowspan="2">Model</th>
<th rowspan="2">$Param.$</th>
<th colspan="2">$STS-L$</th>
<th colspan="2">STS- V</th>
</tr>
<tr>
<th>SpCorr</th>
<th>PCorr</th>
<th>SpCorr</th>
<th>$P$ $'Corr$</th>
</tr>
<tr>
<td>$BERT-style$ $:Text$ $:E_{l}$</td>
<td>vcoders</td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
</tr>
<tr>
<td>BERT- base</td>
<td>$110M$</td>
<td>67.60</td>
<td>68.16</td>
<td>39.67</td>
<td>39.75</td>
</tr>
<tr>
<td>BERT- large</td>
<td>$340M$</td>
<td>69.99</td>
<td>70.61</td>
<td>42.12</td>
<td>42.32</td>
</tr>
<tr>
<td>RoBERTa- base</td>
<td>$125M$</td>
<td>67.47</td>
<td>68.13</td>
<td>39.28</td>
<td>39.54</td>
</tr>
<tr>
<td>$\operatorname{RoBERTa-large}$</td>
<td>$355M$</td>
<td>70.17</td>
<td>70.68</td>
<td>43.54</td>
<td>43.48</td>
</tr>
<tr>
<td>$CLIP- style$ Text $Er$</td>
<td>vcoders</td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
</tr>
<tr>
<td>ViT- B/ 32</td>
<td>$63M$</td>
<td>66.62</td>
<td>66.30</td>
<td>44.36</td>
<td>44.65</td>
</tr>
<tr>
<td>ViT- B/ 16</td>
<td>$63M$</td>
<td>67.85</td>
<td>67.70</td>
<td>44.85</td>
<td>45.17</td>
</tr>
<tr>
<td>ViT- L/ 14</td>
<td>$123M$</td>
<td>68.71</td>
<td>69.00</td>
<td>45.03</td>
<td>45.37</td>
</tr>
<tr>
<td>ViT- L/ 14@ 336px</td>
<td>$123M$</td>
<td>68.72</td>
<td>68.95</td>
<td>45.02</td>
<td>45.38</td>
</tr>
</tbody>
</table>
Table 2: Comparisons of BERT-style text encoders and CLIP-style text encoders on the language-based textual similarity (STS-L) and vision-based textual similarity (STS-V), with the number of parameters (Param.).
CLIP-style text encoders are trained under multimodal settings and intuitively, they are better at associating a text with a real-life scenario. To find $\overline{\text{the answer}}$, we designed a vision-centric text understanding task (described as follows).
# 4.1 Experimental settings
To design the task, we start from the CxC dataset (Parekh et al., 2020), an extension of the MS-COCO Caption dataset (Lin et al., 2014). CxC contains human ratings for caption pairs, which we name STS-L (Semantic Textual Similarity from the Language perspective). Afterwards, we construct a new dataset to conduct a vision-centric text task. Specifically, we label each caption pair in STS-L by identifying whether it is from the same image or not (1 for the former and 0 for the latter). This new dataset is referred to as STS-V (Semantic Textual Similarity from the Vision perspective).$^{6}$ Therefore, we have the STS-L and STS-V ratings for every caption pair.$^{7}$ We evaluate the same models as in §3. We adopt Spearman Correlation scores (Sp Corr) and Pearson Correlation coefficient (P Corr) as the evaluation metrics.
# 4.2 Empirical findings
The results are reported in Table 2. We have two
observations.
CLIP-style text encoders learn better visual per-
ception CLIP-style text encoders underperform BERT-style text encoders on STS-L (same as verified in §3) but outperform them with respect to STSV, demonstrating that CLIP-style text encoders are
$^{6}$In MS-COCO, there are five captions for each image.
We provide some example in Appendix B for further
illustration. | [
[
"0.11584",
"0.0849",
"0.486",
"0.0849",
"0.486",
"0.0998",
"0.11584",
"0.0998"
],
[
"0.712",
"0.0873",
"0.7524",
"0.0873",
"0.7524",
"0.0986",
"0.712",
"0.0986"
],
[
"0.817",
"0.0873",
"0.8564",
"0.0873",
"0.8564",
"0.0974",
"0.817",
"0.0974"
],
[
"0.6304",
"0.09143",
"0.6753",
"0.09143",
"0.6753",
"0.1057",
"0.6304",
"0.1057"
],
[
"0.5205",
"0.09265",
"0.5615",
"0.09265",
"0.5615",
"0.10394",
"0.5205",
"0.10394"
],
[
"0.686",
"0.0986",
"0.7314",
"0.0986",
"0.7314",
"0.10986",
"0.686",
"0.10986"
],
[
"0.733",
"0.0986",
"0.7827",
"0.0986",
"0.7827",
"0.10925",
"0.733",
"0.10925"
],
[
"0.7925",
"0.0986",
"0.8423",
"0.0986",
"0.8423",
"0.10986",
"0.7925",
"0.10986"
],
[
"0.843",
"0.0986",
"0.889",
"0.0986",
"0.889",
"0.10986",
"0.843",
"0.10986"
],
[
"0.11584",
"0.111",
"0.3333",
"0.11224",
"0.3333",
"0.1271",
"0.11584",
"0.1259"
],
[
"0.519",
"0.114",
"0.6675",
"0.114",
"0.6675",
"0.1277",
"0.519",
"0.1277"
],
[
"0.638",
"0.1271",
"0.6753",
"0.1271",
"0.6753",
"0.1383",
"0.638",
"0.1383"
],
[
"0.69",
"0.1271",
"0.7256",
"0.1271",
"0.7256",
"0.1383",
"0.69",
"0.1383"
],
[
"0.744",
"0.1271",
"0.781",
"0.1271",
"0.781",
"0.1383",
"0.744",
"0.1383"
],
[
"0.797",
"0.1271",
"0.833",
"0.1271",
"0.833",
"0.1383",
"0.797",
"0.1383"
],
[
"0.8506",
"0.1271",
"0.888",
"0.1271",
"0.888",
"0.1383",
"0.8506",
"0.1383"
],
[
"0.11755",
"0.1342",
"0.4895",
"0.1342",
"0.4895",
"0.1478",
"0.11755",
"0.1478"
],
[
"0.5225",
"0.1277",
"0.588",
"0.1277",
"0.588",
"0.1383",
"0.5225",
"0.1383"
],
[
"0.638",
"0.1383",
"0.6753",
"0.1383",
"0.6753",
"0.1509",
"0.638",
"0.1509"
],
[
"0.744",
"0.1383",
"0.7793",
"0.1383",
"0.7793",
"0.1509",
"0.744",
"0.1509"
],
[
"0.797",
"0.1383",
"0.833",
"0.1383",
"0.833",
"0.1509",
"0.797",
"0.1509"
],
[
"0.8506",
"0.1383",
"0.888",
"0.1383",
"0.888",
"0.1509",
"0.8506",
"0.1509"
],
[
"0.69",
"0.1395",
"0.7256",
"0.1395",
"0.7256",
"0.1509",
"0.69",
"0.1509"
],
[
"0.5225",
"0.1395",
"0.5884",
"0.1395",
"0.5884",
"0.1509",
"0.5225",
"0.1509"
],
[
"0.11755",
"0.1497",
"0.4878",
"0.1497",
"0.4878",
"0.1633",
"0.11755",
"0.1633"
],
[
"0.519",
"0.1478",
"0.6113",
"0.1497",
"0.6113",
"0.1644",
"0.519",
"0.1621"
],
[
"0.6396",
"0.1509",
"0.6753",
"0.1509",
"0.6753",
"0.1621",
"0.6396",
"0.1621"
],
[
"0.692",
"0.1509",
"0.7256",
"0.1509",
"0.7256",
"0.1621",
"0.692",
"0.1621"
],
[
"0.744",
"0.1509",
"0.781",
"0.1509",
"0.781",
"0.1621",
"0.744",
"0.1621"
],
[
"0.797",
"0.1509",
"0.8345",
"0.1509",
"0.8345",
"0.1621",
"0.797",
"0.1621"
],
[
"0.8486",
"0.1509",
"0.8867",
"0.1509",
"0.8867",
"0.1621",
"0.8486",
"0.1621"
],
[
"0.638",
"0.1621",
"0.6753",
"0.1621",
"0.6753",
"0.1733",
"0.638",
"0.1733"
],
[
"0.69",
"0.1621",
"0.7256",
"0.1621",
"0.7256",
"0.1733",
"0.69",
"0.1733"
],
[
"0.742",
"0.1621",
"0.7793",
"0.1621",
"0.7793",
"0.1733",
"0.742",
"0.1733"
],
[
"0.796",
"0.1621",
"0.833",
"0.1621",
"0.833",
"0.1733",
"0.796",
"0.1733"
],
[
"0.8506",
"0.1621",
"0.888",
"0.1621",
"0.888",
"0.1746",
"0.8506",
"0.1746"
],
[
"0.5225",
"0.1633",
"0.6094",
"0.1633",
"0.6094",
"0.1746",
"0.5225",
"0.1746"
],
[
"0.11755",
"0.165",
"0.4895",
"0.165",
"0.4895",
"0.1787",
"0.11755",
"0.1787"
],
[
"0.11755",
"0.1818",
"0.4878",
"0.1818",
"0.4878",
"0.1953",
"0.11755",
"0.1953"
],
[
"0.5205",
"0.1776",
"0.6665",
"0.1776",
"0.6665",
"0.1912",
"0.5205",
"0.1912"
],
[
"0.5225",
"0.1912",
"0.5767",
"0.1912",
"0.5767",
"0.2031",
"0.5225",
"0.2031"
],
[
"0.644",
"0.1912",
"0.6753",
"0.1912",
"0.6753",
"0.2031",
"0.644",
"0.2031"
],
[
"0.69",
"0.1912",
"0.7256",
"0.1912",
"0.7256",
"0.2031",
"0.69",
"0.2031"
],
[
"0.744",
"0.1912",
"0.781",
"0.1912",
"0.781",
"0.2031",
"0.744",
"0.2031"
],
[
"0.797",
"0.1912",
"0.833",
"0.1912",
"0.833",
"0.2031",
"0.797",
"0.2031"
],
[
"0.8506",
"0.1912",
"0.8867",
"0.1912",
"0.8867",
"0.2031",
"0.8506",
"0.2031"
],
[
"0.11755",
"0.199",
"0.4895",
"0.199",
"0.4895",
"0.212",
"0.11755",
"0.212"
],
[
"0.5225",
"0.2031",
"0.5786",
"0.2031",
"0.5786",
"0.2144",
"0.5225",
"0.2144"
],
[
"0.644",
"0.2031",
"0.677",
"0.2031",
"0.677",
"0.2144",
"0.644",
"0.2144"
],
[
"0.69",
"0.2031",
"0.7256",
"0.2031",
"0.7256",
"0.2156",
"0.69",
"0.2156"
],
[
"0.742",
"0.2031",
"0.7827",
"0.2031",
"0.7827",
"0.2144",
"0.742",
"0.2144"
],
[
"0.796",
"0.2031",
"0.8345",
"0.2031",
"0.8345",
"0.2144",
"0.796",
"0.2144"
],
[
"0.8486",
"0.2031",
"0.888",
"0.2031",
"0.888",
"0.2144",
"0.8486",
"0.2144"
],
[
"0.11755",
"0.2144",
"0.4878",
"0.2144",
"0.4878",
"0.228",
"0.11755",
"0.228"
],
[
"0.638",
"0.2144",
"0.6753",
"0.2144",
"0.6753",
"0.2257",
"0.638",
"0.2257"
],
[
"0.6885",
"0.2131",
"0.7256",
"0.2156",
"0.7246",
"0.228",
"0.687",
"0.2257"
],
[
"0.744",
"0.2144",
"0.7827",
"0.2144",
"0.7827",
"0.2257",
"0.744",
"0.2257"
],
[
"0.796",
"0.2144",
"0.8345",
"0.2144",
"0.8345",
"0.2257",
"0.796",
"0.2257"
],
[
"0.8486",
"0.2144",
"0.888",
"0.2144",
"0.888",
"0.2257",
"0.8486",
"0.2257"
],
[
"0.797",
"0.2239",
"0.8345",
"0.2257",
"0.833",
"0.2382",
"0.796",
"0.2363"
],
[
"0.5225",
"0.2156",
"0.5767",
"0.2156",
"0.5767",
"0.2257",
"0.5225",
"0.2257"
],
[
"0.744",
"0.2245",
"0.781",
"0.2268",
"0.7793",
"0.2382",
"0.742",
"0.2363"
],
[
"0.8506",
"0.2257",
"0.888",
"0.2257",
"0.888",
"0.2382",
"0.8506",
"0.2382"
],
[
"0.638",
"0.2268",
"0.6753",
"0.2268",
"0.6753",
"0.2382",
"0.638",
"0.2382"
],
[
"0.5225",
"0.2268",
"0.623",
"0.2268",
"0.623",
"0.2382",
"0.5225",
"0.2382"
],
[
"0.69",
"0.2268",
"0.727",
"0.2268",
"0.727",
"0.2382",
"0.69",
"0.2382"
],
[
"0.11755",
"0.2465",
"0.4878",
"0.2465",
"0.4878",
"0.26",
"0.11755",
"0.26"
],
[
"0.11584",
"0.2299",
"0.4878",
"0.2286",
"0.4878",
"0.2434",
"0.11584",
"0.2446"
],
[
"0.509",
"0.2505",
"0.882",
"0.2517",
"0.882",
"0.2666",
"0.509",
"0.2654"
],
[
"0.11584",
"0.2612",
"0.4878",
"0.2625",
"0.4878",
"0.2766",
"0.11584",
"0.2756"
],
[
"0.51",
"0.2666",
"0.881",
"0.2666",
"0.881",
"0.2798",
"0.51",
"0.2798"
],
[
"0.11584",
"0.2778",
"0.4895",
"0.2766",
"0.4895",
"0.2915",
"0.11584",
"0.2922"
],
[
"0.511",
"0.282",
"0.881",
"0.282",
"0.881",
"0.2957",
"0.511",
"0.2957"
],
[
"0.1184",
"0.2964",
"0.486",
"0.2964",
"0.486",
"0.3071",
"0.1184",
"0.3071"
],
[
"0.51",
"0.2957",
"0.852",
"0.2957",
"0.852",
"0.3088",
"0.51",
"0.3088"
],
[
"0.11755",
"0.31",
"0.4878",
"0.31",
"0.4878",
"0.3247",
"0.11755",
"0.3247"
],
[
"0.11584",
"0.326",
"0.4844",
"0.326",
"0.4844",
"0.339",
"0.11584",
"0.339"
],
[
"0.513",
"0.335",
"0.882",
"0.335",
"0.882",
"0.3486",
"0.513",
"0.3486"
],
[
"0.11755",
"0.3433",
"0.4878",
"0.3433",
"0.4878",
"0.357",
"0.11755",
"0.357"
],
[
"0.511",
"0.3516",
"0.882",
"0.3516",
"0.882",
"0.3652",
"0.511",
"0.3652"
],
[
"0.11584",
"0.3582",
"0.4895",
"0.3594",
"0.4895",
"0.3735",
"0.11584",
"0.3723"
],
[
"0.511",
"0.3682",
"0.882",
"0.3682",
"0.882",
"0.3818",
"0.511",
"0.3818"
],
[
"0.11755",
"0.3772",
"0.486",
"0.3772",
"0.486",
"0.3872",
"0.11755",
"0.3872"
],
[
"0.511",
"0.3855",
"0.8833",
"0.3855",
"0.8833",
"0.3984",
"0.511",
"0.3984"
],
[
"0.11584",
"0.3914",
"0.4836",
"0.3914",
"0.4836",
"0.405",
"0.11584",
"0.405"
],
[
"0.511",
"0.401",
"0.802",
"0.401",
"0.802",
"0.4146",
"0.511",
"0.4146"
],
[
"0.11755",
"0.4192",
"0.3064",
"0.4192",
"0.3064",
"0.434",
"0.11755",
"0.434"
],
[
"0.511",
"0.4246",
"0.727",
"0.4258",
"0.727",
"0.4407",
"0.511",
"0.4395"
],
[
"0.11755",
"0.4395",
"0.4878",
"0.4407",
"0.4878",
"0.4548",
"0.11755",
"0.4536"
],
[
"0.509",
"0.4465",
"0.882",
"0.4465",
"0.882",
"0.4614",
"0.509",
"0.4614"
],
[
"0.11584",
"0.4573",
"0.4895",
"0.4573",
"0.4895",
"0.4714",
"0.11584",
"0.4714"
],
[
"0.51",
"0.4631",
"0.881",
"0.4631",
"0.881",
"0.4768",
"0.51",
"0.4768"
],
[
"0.11755",
"0.4739",
"0.4902",
"0.4739",
"0.4902",
"0.4868",
"0.11755",
"0.4868"
],
[
"0.511",
"0.4797",
"0.881",
"0.4797",
"0.881",
"0.4934",
"0.511",
"0.4934"
],
[
"0.11755",
"0.4893",
"0.4895",
"0.4893",
"0.4895",
"0.503",
"0.11755",
"0.503"
],
[
"0.51",
"0.4963",
"0.882",
"0.4963",
"0.882",
"0.5103",
"0.51",
"0.5103"
],
[
"0.11584",
"0.506",
"0.486",
"0.506",
"0.486",
"0.5195",
"0.11584",
"0.5195"
],
[
"0.51",
"0.5127",
"0.882",
"0.5127",
"0.882",
"0.5254",
"0.51",
"0.5254"
],
[
"0.11755",
"0.5215",
"0.4902",
"0.5215",
"0.4902",
"0.536",
"0.11755",
"0.536"
],
[
"0.513",
"0.5293",
"0.882",
"0.5293",
"0.882",
"0.542",
"0.513",
"0.542"
],
[
"0.1142",
"0.538",
"0.4895",
"0.5376",
"0.4895",
"0.552",
"0.1142",
"0.5527"
],
[
"0.511",
"0.546",
"0.881",
"0.546",
"0.881",
"0.5557",
"0.511",
"0.5557"
],
[
"0.11584",
"0.554",
"0.4878",
"0.554",
"0.4878",
"0.5684",
"0.11584",
"0.5684"
],
[
"0.51",
"0.5586",
"0.881",
"0.5586",
"0.881",
"0.5723",
"0.51",
"0.5723"
],
[
"0.11755",
"0.571",
"0.4895",
"0.571",
"0.4895",
"0.5845",
"0.11755",
"0.5845"
],
[
"0.511",
"0.5767",
"0.8833",
"0.5767",
"0.8833",
"0.5903",
"0.511",
"0.5903"
],
[
"0.11755",
"0.586",
"0.4895",
"0.586",
"0.4895",
"0.5996",
"0.11755",
"0.5996"
],
[
"0.513",
"0.5933",
"0.881",
"0.5933",
"0.881",
"0.604",
"0.513",
"0.604"
],
[
"0.11755",
"0.6025",
"0.4878",
"0.6025",
"0.4878",
"0.616",
"0.11755",
"0.616"
],
[
"0.513",
"0.608",
"0.882",
"0.608",
"0.882",
"0.6216",
"0.513",
"0.6216"
],
[
"0.11584",
"0.6177",
"0.486",
"0.6177",
"0.486",
"0.6313",
"0.11584",
"0.6313"
],
[
"0.513",
"0.6235",
"0.8833",
"0.6235",
"0.8833",
"0.638",
"0.513",
"0.638"
],
[
"0.11755",
"0.6343",
"0.4878",
"0.6343",
"0.4878",
"0.648",
"0.11755",
"0.648"
],
[
"0.511",
"0.64",
"0.8833",
"0.64",
"0.8833",
"0.654",
"0.511",
"0.654"
],
[
"0.11755",
"0.6494",
"0.4895",
"0.6484",
"0.4895",
"0.663",
"0.11755",
"0.6646"
],
[
"0.513",
"0.6567",
"0.882",
"0.6567",
"0.882",
"0.6704",
"0.513",
"0.6704"
],
[
"0.11584",
"0.6665",
"0.4878",
"0.665",
"0.4878",
"0.6797",
"0.11584",
"0.681"
],
[
"0.51",
"0.6714",
"0.882",
"0.673",
"0.882",
"0.687",
"0.51",
"0.686"
],
[
"0.11755",
"0.6816",
"0.4878",
"0.6816",
"0.4878",
"0.6953",
"0.11755",
"0.6953"
],
[
"0.51",
"0.6885",
"0.881",
"0.6885",
"0.881",
"0.7017",
"0.51",
"0.7017"
],
[
"0.11755",
"0.6987",
"0.4878",
"0.6987",
"0.4878",
"0.7134",
"0.11755",
"0.7134"
],
[
"0.511",
"0.705",
"0.7446",
"0.705",
"0.7446",
"0.7188",
"0.511",
"0.7188"
],
[
"0.11584",
"0.7153",
"0.4878",
"0.7153",
"0.4878",
"0.7285",
"0.11584",
"0.7285"
],
[
"0.11584",
"0.7324",
"0.318",
"0.7324",
"0.318",
"0.745",
"0.11584",
"0.745"
],
[
"0.513",
"0.731",
"0.702",
"0.731",
"0.702",
"0.745",
"0.513",
"0.745"
],
[
"0.51",
"0.752",
"0.882",
"0.752",
"0.882",
"0.7666",
"0.51",
"0.7666"
],
[
"0.12177",
"0.763",
"0.1318",
"0.763",
"0.1318",
"0.771",
"0.12177",
"0.771"
],
[
"0.1351",
"0.764",
"0.1427",
"0.764",
"0.1427",
"0.7695",
"0.1351",
"0.7695"
],
[
"0.1436",
"0.7603",
"0.486",
"0.7603",
"0.486",
"0.774",
"0.1436",
"0.774"
],
[
"0.513",
"0.7695",
"0.6094",
"0.7695",
"0.6094",
"0.781",
"0.513",
"0.781"
],
[
"0.147",
"0.778",
"0.4282",
"0.778",
"0.4282",
"0.7915",
"0.147",
"0.7915"
],
[
"0.513",
"0.792",
"0.8833",
"0.792",
"0.8833",
"0.8057",
"0.513",
"0.8057"
],
[
"0.11584",
"0.802",
"0.4878",
"0.8027",
"0.4878",
"0.818",
"0.11584",
"0.8164"
],
[
"0.51",
"0.8086",
"0.881",
"0.8086",
"0.881",
"0.8223",
"0.51",
"0.8223"
],
[
"0.11755",
"0.8193",
"0.4895",
"0.8193",
"0.4895",
"0.833",
"0.11755",
"0.833"
],
[
"0.511",
"0.8247",
"0.8833",
"0.8247",
"0.8833",
"0.8384",
"0.511",
"0.8384"
],
[
"0.11584",
"0.8345",
"0.4902",
"0.835",
"0.4902",
"0.8496",
"0.11584",
"0.8486"
],
[
"0.513",
"0.8413",
"0.8833",
"0.8413",
"0.8833",
"0.855",
"0.513",
"0.855"
],
[
"0.11584",
"0.851",
"0.4878",
"0.8516",
"0.4878",
"0.866",
"0.11584",
"0.865"
],
[
"0.511",
"0.857",
"0.882",
"0.857",
"0.882",
"0.8716",
"0.511",
"0.8716"
],
[
"0.11584",
"0.868",
"0.4878",
"0.868",
"0.4878",
"0.882",
"0.11584",
"0.882"
],
[
"0.531",
"0.879",
"0.8613",
"0.88",
"0.8613",
"0.8945",
"0.531",
"0.8936"
],
[
"0.1351",
"0.891",
"0.4878",
"0.8936",
"0.4878",
"0.9067",
"0.1351",
"0.905"
],
[
"0.531",
"0.891",
"0.8833",
"0.8936",
"0.8833",
"0.9077",
"0.531",
"0.9062"
],
[
"0.1142",
"0.904",
"0.3586",
"0.905",
"0.3586",
"0.9194",
"0.1142",
"0.9185"
],
[
"0.513",
"0.9062",
"0.5845",
"0.9062",
"0.5845",
"0.9175",
"0.513",
"0.9175"
]
] | [
"3Pilot study: general text understanding",
"STS-L",
"STS-V",
"Param.",
"Model",
"SpCorr",
"PCorr",
"Sp Corr",
"PCorr",
"3.1 Experimental settings",
"BERT-style Text Encoders",
"110M",
"67.60",
"68.16",
"39.67",
"39.75",
"For the datasets, we adopt two text classifica-",
"BERT-base",
"340M",
"70.61",
"42.12",
"42.32",
"69.99",
"BERT-large",
"tion tasks (CoLA (Warstadt et al., 2019) and",
"RoBERTa-base",
"125M",
"67.47",
"68.13",
"39.28",
"39.54",
"355M",
"70.17",
"70.68",
"43.54",
"43.48",
"RoBERTa-large",
"SST2 (Socher et al., 2013), three text similar-",
"ity tasks (MRPC (Dolan and Brockett, 2005),",
"CLIP-style Text Encoders",
"ViT-B/32",
"63M",
"66.62",
"66.30",
"44.36",
"44.65",
"QQP (Shankar et al., 2017), and STS-B (Cer et al.,",
"ViT-B/16",
"63M",
"67.85",
"67.70",
"44.85",
"45.17",
"2017)), and three inference tasks (MNLI (Williams",
"123M",
"68.71",
"69.00",
"45.03",
"45.37",
"45.02",
"ViT-L/14",
"68.95",
"45.38",
"123M",
"ViT-L/14@336px",
"68.72",
"RTE (Bentivogli et al., 2009)) of GLUE.5 For",
"et al., 2017), QNLI (Rajpurkar et al., 2016), and",
"Table 2: Comparisons of BERT-style text encoders and",
"BERT-style text encoders, we adopt the base and",
"CLIP-style text encoders on the language-based textual",
"large versions of BERT and RoBERTa; For CLIP-",
"similarity (STS-L) and vision-based textual similarity",
"styletextencoders,we adoptthetext encoders of",
"(STS-V), with the number of parameters (Param.).",
"four versions of CLIP (i.e., ViT-B/32, ViT-B/16,",
"ViT-L/14, and ViT-L/14@336px), where the first",
"CLIP-style text encoders are trained under multi",
"two text encoders share the same architecture but",
"modal settings and intuitively, they are better at",
"have different parameters (same for the last two)",
"associating a text with a real-life scenario. To find",
"due to the difference of the vision branches.We",
"the answer, we designed a vision-centric text un-",
"adopt the commonly used metric for each dataset.",
"derstanding task (described as follows)",
"3.2Empirical findings",
"4.1 Experimental settings",
"CLIP text encoders perform poorly in GLUE",
"To design the task, we start from the CxC",
"We report the results in Table 1..The four",
"dataset (Parekh et al., 2020), an extension of the",
"CLIP-style text encoders consistently underper-",
"MS-COCO Caption dataset (Lin et al., 2014). CxC",
"form BERT-style text encoders on all the datasets,",
"contains human ratings for caption pairs, which we",
"where the CLIP-style text encoders achieve around",
"name STS-L (Semantic Textual Similarity from the",
"85% of the scores of BERT-style ones on aver-",
"Language perspective). Afterwards, we construct",
"age. Comparing among different datasets, the",
"a new dataset to conduct a vision-centric text task",
"most significant performance gap occurs in the",
"Specifically, we label each caption pair in STS-L",
"CoLA dataset.The reason behind this is that",
"by identifying whether it is from the same image or",
"CoLA is an English acceptability dataset that re-",
"not (1 for the former and O for the latter). This new",
"quires a model to identify whether a sequence of",
"dataset is referred to as STS-V (Semantic Textual",
"words is a grammatical English sentence. This",
"Similarity from the Vision perspective).6 There-",
"demonstrates that ITC is worse than MLM on",
"fore, we have the STS-L and STS-V ratings for",
"grammatical or syntactic properties. This find-",
"every caption pair.7 We evaluate the same models",
"ing is consistent with Yuksekgonul et al. (2022),",
"as in g3. We adopt Spearman Correlation scores",
"which shows that the vision-and-language models",
"(Sp Corr) and Pearson Correlation coefficient (P",
"trained by ITC ignore word orders and therefore",
"Corr) as the evaluation metrics.",
"lack understanding of the compositional structure",
"in the images and captions.",
"4.2 Empirical findings",
"The results are reported in Table 2. We have two",
"4",
"",
"CLIP-style text encoders capture visual",
"observations.",
"perception for concept similarity",
"CLIP-style text encoders learn better visual per-",
"Motivation The aforementioned experiments",
"ception CLIP-style text encoders underperform",
"show that BERT-style text encoders outperform",
"BERT-style text encoders on STS-L (same as veri-",
"CLIP-style text encoders on pure text tasks. There-",
"fied in s3) but outperform them with respect to STS-",
"fore, a question is “Is there any text task for us to",
"V, demonstrating that CLIP-style text encoders are",
"testify the superior of CLIP-style text encoders?\"",
"In MS-COCO, there are five captions for each image",
"5we exclude theWNLI dataset (Levesque et al.,2012) due",
"7we provide some example in Appendix B for further",
"to the large variance of the results on it.",
"illustration."
] | {
"arxiv_id": "2306.03678",
"language": "en",
"page_idx": 2
} |
||
2306.03678-en-3 | 2306.03678-en-3.jpg | <table>
<tbody>
<tr>
<th>Model</th>
<th>Param.</th>
<th>IS</th>
<th>CLIP- S</th>
<th>CLIP- S$\left ( \mathrm{GT}\right )$</th>
</tr>
<tr>
<td>$BERT- style\textit{ Text Er}$</td>
<td>lcoders</td>
<td> </td>
<td> </td>
<td> </td>
</tr>
<tr>
<td>BERT- base</td>
<td>$110M$</td>
<td>1.01</td>
<td>$22.46\pm0.004$</td>
<td> </td>
</tr>
<tr>
<td>BERT- large</td>
<td>$340M$</td>
<td>1.01</td>
<td>$22.43\pm0.021$</td>
<td> </td>
</tr>
<tr>
<td>RoBERTa- base</td>
<td>$125M$</td>
<td>1.01</td>
<td>$22.37\pm0.049$</td>
<td> </td>
</tr>
<tr>
<td>$\operatorname{RoBERTa-large}$</td>
<td>$355M$</td>
<td>1.01</td>
<td>$22.41\pm0.020$</td>
<td>7571</td>
</tr>
<tr>
<td>$CLIP- style$ Text $En$</td>
<td>coders</td>
<td> </td>
<td> </td>
<td>$w_{i}$سل</td>
</tr>
<tr>
<td>ViT- B/ 32</td>
<td>$63M$</td>
<td>1.01</td>
<td>$22.57\pm0.043$</td>
<td> </td>
</tr>
<tr>
<td>ViT- B/ 16</td>
<td>$63M$</td>
<td>1.01</td>
<td>$22.59\pm0.049$</td>
<td> </td>
</tr>
<tr>
<td>ViT- L/ 14</td>
<td>$123M$</td>
<td>1.01</td>
<td>$22.70\pm0.032$</td>
<td> </td>
</tr>
<tr>
<td>ViT- L/ 14@ 336px</td>
<td>$123M$</td>
<td>1.01</td>
<td>$22.67\pm0.037$</td>
<td> </td>
</tr>
<tr>
<td>Random</td>
<td>1/</td>
<td>1.01</td>
<td>$22.13\pm0.029$</td>
<td> </td>
</tr>
</tbody>
</table>
Table 3: Comparison of different models regarding the IS and CLIP-S metrics on CelebAHQ, where CLIP-S (GT) denote the CLIP-S for the ground-truth pairs. We only report the standard deviation of CLIP-S. All IS have the same mean and their standard deviations are all less than 0.003, indicating that they are statistically the same.
better at associating texts with images, more similar to human, which is consistent with the findings of Bielawski et al. (2022) and Zhang et al. (2022).
Larger BERT learns better visual perception Although BERT-style text encoders do not achieve promising results on STS-V, we find that largesize models $(i.e.$, BERT-large and RoBERTa-large) achieve consistently better performance than their counterparts do on STS-V, probably due to the better generalization derived from their larger scale.
## 5 CLIP-style text encoders capture visual perception for image generation
Motivation In previous sections, we verify the superiority of CLIP-style text encoders by associating a text with an image on textural tasks. Next, we consider a question “Why don' t we directly translate a learned text representation to an image to compare their visual perception ability in a more $straightforwardway?”.$ To this end, we design a text-to-image generation pipeline to probe the association ability.
5.1 Pipeline of text-to-image generation
First, we assume the association ability is sourced from the overlap of the image representation space and the text representation space, which means that these two spaces share similar concepts. Subsequently, under such a restricted condition, we achieve the overlap of the two spaces by introducing a single linear transformation layer to project the text space onto the image space. In formal, given a text encoder $\mathcal{E}(\cdot)$ and an (unconditional)
(generative) image decoder $D(\cdot)=p(\cdot)$, we can use the former to encode a text $T$ to text representations $\mathcal{E}(T)$ or use the latter the measure the probability $\mathcal{D}(I)~=~p(I)~$of a generated image $I.$ We then denote the linear transformation as $\mathcal{T}.$ Therefore, the whole probing pipeline is described as follows:
(3)
$$I=\arg\max_{I}\mathcal{D}(I|\mathcal{T}(\mathcal{E}(T)))$$
where we use the linearly transformed text representations $\mathcal{T}(\mathcal{E}(T))$ as the condition to prompt the generation of the image $I.$ As mentioned, we only tune the parameters of the linear transformation $\mathcal{T}$, and freeze the text encoder $\mathcal{E}(\cdot)$ and the image decoder $\mathcal{D}(\cdot).$
# 5.2 Experimental settings
For the unconditional image decoder, we adopt the VQGAN-Transformer model (Esser et al., 2021) pretrained on the images of CelebA-HQ (Karras et al., 2017). The auto-regressive Transformer can generate discrete image tokens, which can be further decoded into images through VQGAN. We train and evaluate our model on the Multi-Modal CelebA-HQ dataset (Xia et al., 2021) with 30,000 text-image pairs. The same text encoders as in §3 and §4 are adopted. We also include a random baseline, where we use random embeddings as input to the linear transformation $\mathcal{T}.$ All experiments are run 3 times with different random seeds. We use Inception Score (Salimans et al., 2016) (IS) and CLIP Score (CLIP-S) as metrics.
### 5.3 Empirical findings
CLIP text embedding generates better images We report the results in Table 3.$^{8}$ The IS metric measures the realism of generated images, and it can be seen that the two types of models all achieve similar performance on the IS metric. This owes to the fact that we start from a pretrained image decoder, which guarantees the generation of highquality images and makes tuning a linear layer feasible.
Statistical Significance: The grouped means and standard deviations of BERT (B), CLIP (C) and Random ( R) are 22. 414$\pm$0. 041, 22. 631$\pm$0. 066, $22.130\pm0.029.$ Applying the pooled t-test between B& B&R, C&R yields respective p-values 2.250e-9, 5.179e-8, 1.175e-8, which indicate that
$^{8}$We showcase the generated images in Appendix C. | [
[
"0.12427",
"0.0861",
"0.1663",
"0.0861",
"0.1663",
"0.0974",
"0.12427",
"0.0974"
],
[
"0.2351",
"0.0861",
"0.2747",
"0.0861",
"0.2747",
"0.0974",
"0.2351",
"0.0974"
],
[
"0.339",
"0.0861",
"0.3862",
"0.0861",
"0.3862",
"0.0974",
"0.339",
"0.0974"
],
[
"0.413",
"0.0861",
"0.486",
"0.0861",
"0.486",
"0.0974",
"0.413",
"0.0974"
],
[
"0.511",
"0.0861",
"0.882",
"0.0861",
"0.882",
"0.0998",
"0.511",
"0.0998"
],
[
"0.12427",
"0.10156",
"0.2678",
"0.1027",
"0.2678",
"0.114",
"0.12427",
"0.1134"
],
[
"0.2896",
"0.0873",
"0.3064",
"0.0873",
"0.3064",
"0.0968",
"0.2896",
"0.0968"
],
[
"0.511",
"0.1027",
"0.882",
"0.1027",
"0.882",
"0.1164",
"0.511",
"0.1164"
],
[
"0.1276",
"0.114",
"0.1914",
"0.114",
"0.1914",
"0.1259",
"0.1276",
"0.1259"
],
[
"0.2393",
"0.114",
"0.3167",
"0.114",
"0.3167",
"0.1259",
"0.2393",
"0.1259"
],
[
"0.3208",
"0.114",
"0.403",
"0.114",
"0.403",
"0.1247",
"0.3208",
"0.1247"
],
[
"0.513",
"0.11816",
"0.882",
"0.11816",
"0.882",
"0.1318",
"0.513",
"0.1318"
],
[
"0.1276",
"0.1259",
"0.1948",
"0.1259",
"0.1948",
"0.1372",
"0.1276",
"0.1372"
],
[
"0.241",
"0.1259",
"0.2778",
"0.1259",
"0.2778",
"0.1372",
"0.241",
"0.1372"
],
[
"0.2837",
"0.1259",
"0.3132",
"0.1259",
"0.3132",
"0.1372",
"0.2837",
"0.1372"
],
[
"0.3225",
"0.1259",
"0.4014",
"0.1259",
"0.4014",
"0.136",
"0.3225",
"0.136"
],
[
"0.3225",
"0.136",
"0.403",
"0.136",
"0.403",
"0.1478",
"0.3225",
"0.1478"
],
[
"0.513",
"0.1342",
"0.882",
"0.1342",
"0.882",
"0.1484",
"0.513",
"0.1484"
],
[
"0.2854",
"0.1372",
"0.3132",
"0.1372",
"0.3132",
"0.1484",
"0.2854",
"0.1484"
],
[
"0.1276",
"0.1383",
"0.2107",
"0.1383",
"0.2107",
"0.1484",
"0.1276",
"0.1484"
],
[
"0.241",
"0.1372",
"0.2761",
"0.1372",
"0.2761",
"0.1484",
"0.241",
"0.1484"
],
[
"0.126",
"0.1478",
"0.2141",
"0.1484",
"0.2141",
"0.1603",
"0.126",
"0.1592"
],
[
"0.241",
"0.1484",
"0.2761",
"0.1484",
"0.2761",
"0.1603",
"0.241",
"0.1603"
],
[
"0.287",
"0.1484",
"0.3132",
"0.1484",
"0.3132",
"0.1603",
"0.287",
"0.1603"
],
[
"0.3225",
"0.1497",
"0.4048",
"0.1497",
"0.4048",
"0.1603",
"0.3225",
"0.1603"
],
[
"0.511",
"0.1484",
"0.882",
"0.1509",
"0.882",
"0.1644",
"0.511",
"0.1621"
],
[
"0.4324",
"0.155",
"0.465",
"0.155",
"0.465",
"0.1663",
"0.4324",
"0.1663"
],
[
"0.12427",
"0.1621",
"0.266",
"0.1644",
"0.266",
"0.1776",
"0.12427",
"0.1758"
],
[
"0.51",
"0.1663",
"0.882",
"0.1663",
"0.882",
"0.1812",
"0.51",
"0.1812"
],
[
"0.1276",
"0.177",
"0.1814",
"0.177",
"0.1814",
"0.1882",
"0.1276",
"0.1882"
],
[
"0.2468",
"0.177",
"0.2761",
"0.177",
"0.2761",
"0.1882",
"0.2468",
"0.1882"
],
[
"0.2837",
"0.1776",
"0.3105",
"0.1776",
"0.3105",
"0.1859",
"0.2837",
"0.1859"
],
[
"0.3225",
"0.177",
"0.403",
"0.177",
"0.403",
"0.187",
"0.3225",
"0.187"
],
[
"0.2468",
"0.187",
"0.2778",
"0.187",
"0.2778",
"0.199",
"0.2468",
"0.199"
],
[
"0.2837",
"0.187",
"0.3123",
"0.187",
"0.3123",
"0.199",
"0.2837",
"0.199"
],
[
"0.511",
"0.1841",
"0.596",
"0.1841",
"0.596",
"0.1953",
"0.511",
"0.1953"
],
[
"0.3225",
"0.1882",
"0.403",
"0.1882",
"0.403",
"0.199",
"0.3225",
"0.199"
],
[
"0.1276",
"0.1882",
"0.1814",
"0.1882",
"0.1814",
"0.199",
"0.1276",
"0.199"
],
[
"0.1276",
"0.1996",
"0.1814",
"0.1996",
"0.1814",
"0.2114",
"0.1276",
"0.2114"
],
[
"0.241",
"0.1996",
"0.3242",
"0.1996",
"0.3242",
"0.2114",
"0.241",
"0.2114"
],
[
"0.3198",
"0.1996",
"0.4014",
"0.1996",
"0.4014",
"0.2114",
"0.3198",
"0.2114"
],
[
"0.1276",
"0.2114",
"0.2241",
"0.2114",
"0.2241",
"0.2227",
"0.1276",
"0.2227"
],
[
"0.2393",
"0.2114",
"0.2761",
"0.2114",
"0.2761",
"0.2227",
"0.2393",
"0.2227"
],
[
"0.287",
"0.2114",
"0.3123",
"0.2114",
"0.3123",
"0.2227",
"0.287",
"0.2227"
],
[
"0.3225",
"0.212",
"0.403",
"0.212",
"0.403",
"0.2227",
"0.3225",
"0.2227"
],
[
"0.592",
"0.2131",
"0.8037",
"0.2131",
"0.8037",
"0.228",
"0.592",
"0.228"
],
[
"0.858",
"0.2131",
"0.8833",
"0.2131",
"0.8833",
"0.2268",
"0.858",
"0.2268"
],
[
"0.12427",
"0.2268",
"0.1763",
"0.2268",
"0.1763",
"0.2382",
"0.12427",
"0.2382"
],
[
"0.2627",
"0.2257",
"0.403",
"0.2257",
"0.403",
"0.2393",
"0.2627",
"0.2393"
],
[
"0.1142",
"0.2517",
"0.4878",
"0.253",
"0.4878",
"0.2673",
"0.1142",
"0.2666"
],
[
"0.51",
"0.2465",
"0.882",
"0.2477",
"0.882",
"0.2625",
"0.51",
"0.2612"
],
[
"0.11755",
"0.2673",
"0.4878",
"0.2673",
"0.4878",
"0.2808",
"0.11755",
"0.2808"
],
[
"0.51",
"0.2632",
"0.881",
"0.2632",
"0.881",
"0.2766",
"0.51",
"0.2766"
],
[
"0.1184",
"0.2832",
"0.4878",
"0.2832",
"0.4878",
"0.2935",
"0.1184",
"0.2935"
],
[
"0.513",
"0.2808",
"0.882",
"0.2808",
"0.882",
"0.2944",
"0.513",
"0.2944"
],
[
"0.11755",
"0.2957",
"0.4878",
"0.2957",
"0.4878",
"0.3088",
"0.11755",
"0.3088"
],
[
"0.509",
"0.2957",
"0.881",
"0.2957",
"0.881",
"0.3088",
"0.509",
"0.3088"
],
[
"0.11755",
"0.31",
"0.4895",
"0.31",
"0.4895",
"0.3237",
"0.11755",
"0.3237"
],
[
"0.513",
"0.3123",
"0.882",
"0.3123",
"0.882",
"0.3267",
"0.513",
"0.3267"
],
[
"0.11584",
"0.3247",
"0.4878",
"0.3247",
"0.4878",
"0.3384",
"0.11584",
"0.3384"
],
[
"0.511",
"0.3247",
"0.615",
"0.3267",
"0.615",
"0.3445",
"0.511",
"0.3425"
],
[
"0.11584",
"0.339",
"0.1823",
"0.339",
"0.1823",
"0.3508",
"0.11584",
"0.3508"
],
[
"0.511",
"0.3552",
"0.7256",
"0.3557",
"0.7256",
"0.3718",
"0.511",
"0.3706"
],
[
"0.11584",
"0.3777",
"0.4902",
"0.3777",
"0.4902",
"0.3914",
"0.11584",
"0.3914"
],
[
"0.486",
"0.3818",
"0.4902",
"0.3818",
"0.4902",
"0.3872",
"0.486",
"0.3872"
],
[
"0.513",
"0.379",
"0.882",
"0.379",
"0.882",
"0.3938",
"0.513",
"0.3938"
],
[
"0.11584",
"0.3926",
"0.4878",
"0.3943",
"0.4878",
"0.408",
"0.11584",
"0.4062"
],
[
"0.511",
"0.3943",
"0.882",
"0.3938",
"0.882",
"0.408",
"0.511",
"0.4092"
],
[
"0.11584",
"0.4104",
"0.4878",
"0.4104",
"0.4878",
"0.4233",
"0.11584",
"0.4233"
],
[
"0.511",
"0.411",
"0.881",
"0.411",
"0.881",
"0.4246",
"0.511",
"0.4246"
],
[
"0.51",
"0.4258",
"0.882",
"0.427",
"0.882",
"0.4412",
"0.51",
"0.4407"
],
[
"0.11584",
"0.434",
"0.486",
"0.434",
"0.486",
"0.4478",
"0.11584",
"0.4478"
],
[
"0.511",
"0.4436",
"0.882",
"0.4436",
"0.882",
"0.4573",
"0.511",
"0.4573"
],
[
"0.11755",
"0.4507",
"0.486",
"0.4507",
"0.486",
"0.4644",
"0.11755",
"0.4644"
],
[
"0.511",
"0.459",
"0.882",
"0.459",
"0.882",
"0.4739",
"0.511",
"0.4739"
],
[
"0.11755",
"0.466",
"0.4902",
"0.466",
"0.4902",
"0.481",
"0.11755",
"0.481"
],
[
"0.509",
"0.4744",
"0.881",
"0.4744",
"0.881",
"0.488",
"0.509",
"0.488"
],
[
"0.11584",
"0.4827",
"0.486",
"0.4827",
"0.486",
"0.4963",
"0.11584",
"0.4963"
],
[
"0.513",
"0.4912",
"0.882",
"0.4912",
"0.882",
"0.505",
"0.513",
"0.505"
],
[
"0.11584",
"0.4995",
"0.4878",
"0.4995",
"0.4878",
"0.513",
"0.11584",
"0.513"
],
[
"0.511",
"0.5083",
"0.882",
"0.5083",
"0.882",
"0.5215",
"0.511",
"0.5215"
],
[
"0.11584",
"0.5156",
"0.4902",
"0.5156",
"0.4902",
"0.5293",
"0.11584",
"0.5293"
],
[
"0.511",
"0.524",
"0.882",
"0.524",
"0.882",
"0.5376",
"0.511",
"0.5376"
],
[
"0.1142",
"0.53",
"0.476",
"0.53",
"0.476",
"0.5444",
"0.1142",
"0.5444"
],
[
"0.51",
"0.5405",
"0.882",
"0.5405",
"0.882",
"0.554",
"0.51",
"0.554"
],
[
"0.1453",
"0.558",
"0.4878",
"0.558",
"0.4878",
"0.572",
"0.1453",
"0.572"
],
[
"0.509",
"0.5547",
"0.882",
"0.5557",
"0.882",
"0.571",
"0.509",
"0.5693"
],
[
"0.1427",
"0.561",
"0.1486",
"0.561",
"0.1486",
"0.5664",
"0.1427",
"0.5664"
],
[
"0.51",
"0.5723",
"0.882",
"0.5723",
"0.882",
"0.586",
"0.51",
"0.586"
],
[
"0.1184",
"0.56",
"0.1436",
"0.56",
"0.1436",
"0.5684",
"0.1184",
"0.5684"
],
[
"0.1453",
"0.576",
"0.4248",
"0.5747",
"0.4248",
"0.589",
"0.1453",
"0.5903"
],
[
"0.511",
"0.5884",
"0.882",
"0.5884",
"0.882",
"0.6016",
"0.511",
"0.6016"
],
[
"0.11584",
"0.5996",
"0.4895",
"0.5996",
"0.4895",
"0.6133",
"0.11584",
"0.6133"
],
[
"0.51",
"0.6025",
"0.7573",
"0.604",
"0.7573",
"0.618",
"0.51",
"0.6177"
],
[
"0.11584",
"0.616",
"0.4895",
"0.616",
"0.4895",
"0.63",
"0.11584",
"0.63"
],
[
"0.11584",
"0.632",
"0.4895",
"0.632",
"0.4895",
"0.6455",
"0.11584",
"0.6455"
],
[
"0.511",
"0.6313",
"0.7036",
"0.632",
"0.7036",
"0.6465",
"0.511",
"0.6455"
],
[
"0.11755",
"0.6484",
"0.4902",
"0.6484",
"0.4902",
"0.662",
"0.11755",
"0.662"
],
[
"0.513",
"0.654",
"0.882",
"0.654",
"0.882",
"0.6685",
"0.513",
"0.6685"
],
[
"0.11584",
"0.665",
"0.4878",
"0.665",
"0.4878",
"0.6787",
"0.11584",
"0.6787"
],
[
"0.51",
"0.6685",
"0.8833",
"0.6675",
"0.8833",
"0.685",
"0.51",
"0.686"
],
[
"0.11584",
"0.681",
"0.4895",
"0.681",
"0.4895",
"0.695",
"0.11584",
"0.695"
],
[
"0.51",
"0.687",
"0.882",
"0.687",
"0.882",
"0.7007",
"0.51",
"0.7007"
],
[
"0.11584",
"0.697",
"0.4878",
"0.697",
"0.4878",
"0.7114",
"0.11584",
"0.7114"
],
[
"0.51",
"0.703",
"0.882",
"0.703",
"0.882",
"0.7163",
"0.51",
"0.7163"
],
[
"0.11584",
"0.7134",
"0.4878",
"0.7134",
"0.4878",
"0.727",
"0.11584",
"0.727"
],
[
"0.511",
"0.7173",
"0.881",
"0.7173",
"0.881",
"0.731",
"0.511",
"0.731"
],
[
"0.1142",
"0.728",
"0.2551",
"0.7285",
"0.2551",
"0.7437",
"0.1142",
"0.742"
],
[
"0.511",
"0.7363",
"0.881",
"0.7363",
"0.881",
"0.75",
"0.511",
"0.75"
],
[
"0.11755",
"0.755",
"0.4417",
"0.755",
"0.4417",
"0.7695",
"0.11755",
"0.7695"
],
[
"0.513",
"0.752",
"0.882",
"0.752",
"0.882",
"0.7656",
"0.513",
"0.7656"
],
[
"0.511",
"0.767",
"0.882",
"0.767",
"0.882",
"0.781",
"0.511",
"0.781"
],
[
"0.11584",
"0.775",
"0.486",
"0.775",
"0.486",
"0.788",
"0.11584",
"0.788"
],
[
"0.511",
"0.7837",
"0.575",
"0.7837",
"0.575",
"0.796",
"0.511",
"0.796"
],
[
"0.11584",
"0.792",
"0.4878",
"0.792",
"0.4878",
"0.8057",
"0.11584",
"0.8057"
],
[
"0.51",
"0.799",
"0.882",
"0.8",
"0.882",
"0.814",
"0.51",
"0.8135"
],
[
"0.11584",
"0.808",
"0.4878",
"0.808",
"0.4878",
"0.822",
"0.11584",
"0.822"
],
[
"0.509",
"0.8154",
"0.881",
"0.8154",
"0.881",
"0.83",
"0.509",
"0.83"
],
[
"0.11755",
"0.8247",
"0.4902",
"0.8247",
"0.4902",
"0.8384",
"0.11755",
"0.8384"
],
[
"0.513",
"0.832",
"0.882",
"0.832",
"0.882",
"0.8457",
"0.513",
"0.8457"
],
[
"0.1142",
"0.8413",
"0.4895",
"0.8403",
"0.4895",
"0.854",
"0.1142",
"0.855"
],
[
"0.51",
"0.8486",
"0.8833",
"0.8486",
"0.8833",
"0.8623",
"0.51",
"0.8623"
],
[
"0.1142",
"0.857",
"0.4895",
"0.857",
"0.4895",
"0.8706",
"0.1142",
"0.8706"
],
[
"0.509",
"0.8643",
"0.881",
"0.8643",
"0.881",
"0.8774",
"0.509",
"0.8774"
],
[
"0.11584",
"0.8726",
"0.4878",
"0.8726",
"0.4878",
"0.8857",
"0.11584",
"0.8857"
],
[
"0.513",
"0.88",
"0.8833",
"0.88",
"0.8833",
"0.8936",
"0.513",
"0.8936"
],
[
"0.11584",
"0.8896",
"0.4895",
"0.8896",
"0.4895",
"0.903",
"0.11584",
"0.903"
],
[
"0.11755",
"0.905",
"0.4895",
"0.905",
"0.4895",
"0.9185",
"0.11755",
"0.9185"
],
[
"0.531",
"0.904",
"0.848",
"0.905",
"0.848",
"0.9194",
"0.531",
"0.9185"
]
] | [
"Model",
"Param.",
"CLIP-S",
"CLIP-S (GT)",
"(generative) image decoder D() = p(), we can",
"BERT-styleTextEncoders",
"IS",
"use the former to encode a text T to text repre-",
"BERT-base",
"110M 1.01",
"22.46 ± 0.004",
"sentations &(T) or use the latter the measure the",
"BERT-large",
"340M",
"1.01",
"22.43±0.021",
"22.37 ± 0.049",
"probability D(I) = p(I) of a generated image",
"1.01",
"RoBERTa-base",
"125M",
"RoBERTa-large",
"355M",
"1.01",
"22.41±0.020",
"I. We then denote the linear transformation as T",
"25.21",
"CLIP-style Text Encoders",
"Therefore, the whole probing pipeline is described",
"ViT-B/32",
"63M",
"1.01",
"22.57±0.043",
"63M",
"1.01",
"as follows:",
"22.59±0.049",
"ViT-B/16",
"ViT-L/14",
"123M1.01",
"22.70 ±0.032",
"ViT-L/14@336px",
"123M",
"1.01",
"22.67±0.037",
"I = arg max D(I|T(ε(T)))",
"(3)",
"Random",
"/1.0122.13±0.029",
"Table 3: Comparison of different models regarding the",
"where we use the linearly transformed text repre-",
"IS and CLIP-S metrics on CelebAHQ, where CLIP-S",
"sentations T(ε(T)) as the condition to prompt the",
"(GT)denotetheCLIP-Sfortheground-truthpairs.We",
"generation of the image I. As mentioned, we only",
"only report the standard deviation of CLIP-S. All IS",
"tune the parameters of the linear transformation",
"have the same mean and their standard deviations are",
"T, and freeze the text encoder &() and the image",
"all less than 0.003, indicating that they are statistically",
"decoder D(-).",
"the same.",
"5.2Experimental settings",
"better at associating texts with images, more simi-",
".",
"For the unconditional image decoder, we adopt the",
"lar to human, which is consistent with the findings",
"VQGAN-Transformer model (Esser et al., 2021)",
"of Bielawski et al. (2022) and Zhang et al. (2022).",
"pretrained on the images of CelebA-HQ (Karras",
"et al., 2017). The auto-regressive Transformer can",
"Larger BERT learns better visual perception",
"generate discrete image tokens, which can be fur-",
"Although BERT-style text encoders do not achieve",
"ther decoded into images through VQGAN. We",
"promising results on STS-V, we find that large-",
"train and evaluate our model on the Multi-Modal",
"size models (i.e., BERT-large and RoBERTa-large)",
"CelebA-HQ dataset (Xia et al., 2021) with 30,000",
"achieve consistently better performance than their",
"text-image pairs. The same text encoders as in s3",
"counterparts do on STS-V, probably due to the bet-",
"and s4 are adopted. We also include a random base-",
"ter generalization derived from their larger scale.",
"line, where we use random embeddings as input to",
"CLIP-style text encoders capture visual",
"the linear transformation T. All experiments are",
"",
"run 3 times with different random seeds. We use",
"5",
"perception for image generation",
"Inception Score (Salimans et al., 2016) (IS) and",
"Motivation In previous sections, we verify the",
"CLIP Score (CLIP-S) as metrics.",
" superiority of CLIP-style text encoders by associat-",
"ing a text with an image on textural tasks. Next, we",
"5.3Empiricalfindings",
"consider a question “\"Why don't we directly trans-",
"CLIP text embedding generates better images",
"late a learned text representation to an image to",
"We report the results in Table 3.8 The IS metric",
"compare their visual perception ability in a more",
"measures the realism of generated images, and it",
"straightforward way?\". To this end, we design",
"can be seen that the two types of models all achieve",
"a text-to-image generation pipeline to probe the",
"similar performance on the IS metric. This owes",
"association ability.",
"to the fact that we start from a pretrained image",
"5.1 Pipeline of text-to-image generation",
"decoder, which guarantees the generation of high",
"quality images and makes tuning a linear layer",
"First, we assume the association ability is sourced",
"feasible.",
"from the overlap of the image representation space",
"Statistical Significance: The grouped means and",
"and the text representation space, which means",
"standard deviations of BERT (B), CLIP (C) and",
"that these two spaces share similar concepts. Sub-",
"Random (R) are 22.414 ± 0.041, 22.631 ± 0.066",
"sequently, under such a restricted condition, we",
"22.130 ± 0.029. Applying the pooled t-test be-",
" achieve the overlap of the two spaces by introduc-",
"tween B&C, B&R,C&R yields respective p-values",
"ing a single linear transformation layer to project",
"2.250e-9, 5.179e-8, 1.175e-8, which indicate that",
"the text space onto the image space. In formal,",
"given a text encoder E() and an (unconditional)",
"we showcase the generated images in Appendix C."
] | {
"arxiv_id": "2306.03678",
"language": "en",
"page_idx": 3
} |
||
2306.03678-en-4 | 2306.03678-en-4.jpg | CLIP-S metrics for each of the three groups are significantly different from one another's. Therefore, after stitching text encoders and the image decoder, CLIP-style text encoders achieve higher scores on the CLIP-S metric that measures the matching of an image-text pair. This demonstrates the effectiveness of CLIP-style text encoders on the association ability.$^{9}$
## 6 Conclusion
Human interaction is multi-modal. Starting from the conjecture that text encoders learned from multi-modal data have unique abilities, in this paper, we study the behavioral difference between BERT-style and CLIP-style text encoders. We compare them from three aspects systematically: (i) general (pure) text understanding; (ii) visioncentric text understanding; and (iii) text-to-image generation. Experimental analyses show that although CLIP-style text encoders underperform BERT-style text encoders on general text understanding tasks, they have a unique ability, $i.e.$, synesthesia, to associate a text and its visual appearance, which is more similar to human perception.
## Acknowledgments
This work is supported by Chinese KeyArea Research and Development Program of Guangdong Province (2020B0101350001), the Shenzhen Science and Technology Program (JCYJ20220818103001002), the Guangdong Provincial Key Laboratory of Big Data Computing, The Chinese University of Hong Kong, Shenzhen, Shenzhen Key Research Project (C10120230151) and Shenzhen Doctoral Startup
We
Funding (RCBS20221008093330065).
also thank Tiannan Wang for offering valuable suggestions on the description of experiments in Section 4.
## Limitations
We highlight two limitations of our work. First, the empirical comparisons are not conducted under fully controlled conditions, e.g., the sizes of models. Limited by computational resources, we did not replicate different types of text encoders with the same number of parameters. Instead, we show the results of different encoders of various sizes
$^{9}$Both types of text encoders are not exposed to the repre-
sentation space of the image decoder.
to reduce this effect. Second, for the last experiment, we adopted the CLIP score to evaluate the matching between a text and its generated image. This might raise an issue: “Do images prompted by CLIP-style text representations guarantee a higher CLIP-S score owing to the fact that they are the same models?". To answer this, we point out the reason why we adopted it. The frozen CLIP-style (BERT-style) text encoders are only used to generate prompts for image generation and the linear layer is trained to maximize the likelihood of generated images. Yet, the CLIP score is used to measure the matching between images and texts. Therefore, the uses of the CLIP text encoders and the CLIP score are disentangled.
## Ethics Statement
There are no ethics-related issues in this paper. The data and other related resources in this work are open-source and commonly used by many existing studies.
## References
Luisa Bentivogli, Peter Clark, Ido Dagan, and Danilo Giampiccolo. 2009. The fifth pascal recognizing textual entailment challenge. In $TAC.$
Romain Bielawski, Benjamin Devillers, Tim Van de Cruys, and Rufin Vanrullen. 2022. When does clip generalize better than unimodal models? when judging human-centric concepts. In Proceedings of the 7th Workshop on Representation Learning for NLP, pages 29-38.
Daniel Cer, Mona Diab, Eneko Agirre, Inigo Lopez-
Gazpio, and Lucia Specia. 2017. Semeval-2017 task 1: Semantic textual similarity-multilingual and cross-lingual focused evaluation. arXivpreprint $arXiv{:}1708.00055.$
Ido Dagan, Oren Glickman, and Bernardo Magnini. 2006. The pascal recognising textual entailment challenge. In Machine learning challenges workshop, pages 177-190. Springer.
Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv: 1810.04805.
Bill Dolan and Chris Brockett.2005. Automati-
cally constructing a corpus of sentential paraphrases. In Third International Workshop on Paraphrasing $(IWP2005).$
Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias | [
[
"0.11584",
"0.0849",
"0.492",
"0.0849",
"0.492",
"0.0986",
"0.11584",
"0.0986"
],
[
"0.509",
"0.0849",
"0.881",
"0.0849",
"0.881",
"0.0986",
"0.509",
"0.0986"
],
[
"0.11584",
"0.10156",
"0.4902",
"0.1027",
"0.4902",
"0.1164",
"0.11584",
"0.1152"
],
[
"0.51",
"0.10156",
"0.882",
"0.10156",
"0.882",
"0.1152",
"0.51",
"0.1152"
],
[
"0.11755",
"0.11816",
"0.4902",
"0.11816",
"0.4902",
"0.1318",
"0.11755",
"0.1318"
],
[
"0.51",
"0.11816",
"0.882",
"0.11816",
"0.882",
"0.1318",
"0.51",
"0.1318"
],
[
"0.11584",
"0.133",
"0.4895",
"0.1342",
"0.4895",
"0.1478",
"0.11584",
"0.1467"
],
[
"0.509",
"0.133",
"0.881",
"0.1342",
"0.881",
"0.1484",
"0.509",
"0.1478"
],
[
"0.11755",
"0.1509",
"0.486",
"0.1509",
"0.486",
"0.1609",
"0.11755",
"0.1609"
],
[
"0.513",
"0.1509",
"0.882",
"0.1509",
"0.882",
"0.1644",
"0.513",
"0.1644"
],
[
"0.11584",
"0.165",
"0.492",
"0.165",
"0.492",
"0.1787",
"0.11584",
"0.1787"
],
[
"0.511",
"0.165",
"0.882",
"0.1663",
"0.882",
"0.1799",
"0.511",
"0.1787"
],
[
"0.1142",
"0.1812",
"0.4895",
"0.1818",
"0.4895",
"0.1965",
"0.1142",
"0.1953"
],
[
"0.51",
"0.1829",
"0.882",
"0.1829",
"0.882",
"0.1965",
"0.51",
"0.1965"
],
[
"0.1142",
"0.1978",
"0.1797",
"0.1978",
"0.1797",
"0.212",
"0.1142",
"0.212"
],
[
"0.51",
"0.199",
"0.881",
"0.199",
"0.881",
"0.212",
"0.51",
"0.212"
],
[
"0.51",
"0.2131",
"0.882",
"0.2144",
"0.882",
"0.2286",
"0.51",
"0.228"
],
[
"0.1184",
"0.2299",
"0.1335",
"0.2299",
"0.1335",
"0.2393",
"0.1184",
"0.2393"
],
[
"0.1427",
"0.2322",
"0.1486",
"0.2322",
"0.1486",
"0.2382",
"0.1427",
"0.2382"
],
[
"0.1453",
"0.228",
"0.2477",
"0.228",
"0.2477",
"0.2423",
"0.1453",
"0.2423"
],
[
"0.511",
"0.2322",
"0.882",
"0.2322",
"0.882",
"0.2452",
"0.511",
"0.2452"
],
[
"0.51",
"0.2477",
"0.882",
"0.2477",
"0.882",
"0.2612",
"0.51",
"0.2612"
],
[
"0.1318",
"0.2328",
"0.1427",
"0.2328",
"0.1427",
"0.2363",
"0.1318",
"0.2363"
],
[
"0.11584",
"0.2559",
"0.4878",
"0.2559",
"0.4878",
"0.2695",
"0.11584",
"0.2695"
],
[
"0.509",
"0.2625",
"0.881",
"0.2625",
"0.881",
"0.2756",
"0.509",
"0.2756"
],
[
"0.11584",
"0.2725",
"0.4895",
"0.2725",
"0.4895",
"0.2861",
"0.11584",
"0.2861"
],
[
"0.509",
"0.279",
"0.882",
"0.279",
"0.882",
"0.2922",
"0.509",
"0.2922"
],
[
"0.11584",
"0.288",
"0.4902",
"0.288",
"0.4902",
"0.3018",
"0.11584",
"0.3018"
],
[
"0.513",
"0.2944",
"0.882",
"0.2944",
"0.882",
"0.3088",
"0.513",
"0.3088"
],
[
"0.1142",
"0.304",
"0.4878",
"0.3027",
"0.4878",
"0.3176",
"0.1142",
"0.3184"
],
[
"0.509",
"0.3123",
"0.681",
"0.31",
"0.681",
"0.3247",
"0.509",
"0.3267"
],
[
"0.11755",
"0.3196",
"0.4878",
"0.3196",
"0.4878",
"0.3342",
"0.11755",
"0.3342"
],
[
"0.11584",
"0.3362",
"0.4902",
"0.335",
"0.4902",
"0.3499",
"0.11584",
"0.3508"
],
[
"0.513",
"0.3384",
"0.6616",
"0.3384",
"0.6616",
"0.3528",
"0.513",
"0.3528"
],
[
"0.11755",
"0.3528",
"0.4902",
"0.3528",
"0.4902",
"0.3665",
"0.11755",
"0.3665"
],
[
"0.51",
"0.364",
"0.882",
"0.364",
"0.882",
"0.3777",
"0.51",
"0.3777"
],
[
"0.1184",
"0.3706",
"0.486",
"0.3706",
"0.486",
"0.3813",
"0.1184",
"0.3813"
],
[
"0.11755",
"0.3855",
"0.4902",
"0.3855",
"0.4902",
"0.3984",
"0.11755",
"0.3984"
],
[
"0.509",
"0.3801",
"0.881",
"0.3801",
"0.881",
"0.3938",
"0.509",
"0.3938"
],
[
"0.11584",
"0.401",
"0.486",
"0.401",
"0.486",
"0.4146",
"0.11584",
"0.4146"
],
[
"0.51",
"0.3967",
"0.882",
"0.3967",
"0.882",
"0.4104",
"0.51",
"0.4104"
],
[
"0.11755",
"0.4175",
"0.4902",
"0.4175",
"0.4902",
"0.4312",
"0.11755",
"0.4312"
],
[
"0.511",
"0.4146",
"0.5693",
"0.4146",
"0.5693",
"0.4258",
"0.511",
"0.4258"
],
[
"0.11584",
"0.4329",
"0.4895",
"0.4329",
"0.4895",
"0.4465",
"0.11584",
"0.4465"
],
[
"0.11584",
"0.4495",
"0.4895",
"0.4495",
"0.4895",
"0.4631",
"0.11584",
"0.4631"
],
[
"0.513",
"0.4548",
"0.6094",
"0.4548",
"0.6094",
"0.4697",
"0.513",
"0.4697"
],
[
"0.11584",
"0.4656",
"0.482",
"0.4656",
"0.482",
"0.4785",
"0.11584",
"0.4785"
],
[
"0.511",
"0.4768",
"0.881",
"0.4768",
"0.881",
"0.4905",
"0.511",
"0.4905"
],
[
"0.5283",
"0.4893",
"0.882",
"0.4905",
"0.882",
"0.505",
"0.5283",
"0.5044"
],
[
"0.1184",
"0.4954",
"0.2778",
"0.4954",
"0.2778",
"0.509",
"0.1184",
"0.509"
],
[
"0.529",
"0.5044",
"0.784",
"0.5044",
"0.784",
"0.517",
"0.529",
"0.517"
],
[
"0.11584",
"0.5225",
"0.4895",
"0.5225",
"0.4895",
"0.5376",
"0.11584",
"0.5376"
],
[
"0.511",
"0.5254",
"0.882",
"0.5254",
"0.882",
"0.5405",
"0.511",
"0.5405"
],
[
"0.11755",
"0.538",
"0.4895",
"0.538",
"0.4895",
"0.5527",
"0.11755",
"0.5527"
],
[
"0.5283",
"0.538",
"0.882",
"0.539",
"0.882",
"0.554",
"0.5283",
"0.5527"
],
[
"0.11584",
"0.554",
"0.486",
"0.554",
"0.486",
"0.568",
"0.11584",
"0.568"
],
[
"0.531",
"0.5527",
"0.882",
"0.5527",
"0.882",
"0.5664",
"0.531",
"0.5664"
],
[
"0.11584",
"0.5693",
"0.4878",
"0.571",
"0.4878",
"0.585",
"0.11584",
"0.5845"
],
[
"0.529",
"0.5664",
"0.882",
"0.5664",
"0.882",
"0.58",
"0.529",
"0.58"
],
[
"0.11584",
"0.5845",
"0.4878",
"0.5874",
"0.4878",
"0.6016",
"0.11584",
"0.5986"
],
[
"0.529",
"0.58",
"0.882",
"0.58",
"0.882",
"0.5933",
"0.529",
"0.5933"
],
[
"0.5283",
"0.593",
"0.623",
"0.5903",
"0.623",
"0.605",
"0.5283",
"0.607"
],
[
"0.11755",
"0.6025",
"0.4895",
"0.6025",
"0.4895",
"0.616",
"0.11755",
"0.616"
],
[
"0.11584",
"0.618",
"0.4878",
"0.619",
"0.4878",
"0.6343",
"0.11584",
"0.633"
],
[
"0.51",
"0.6133",
"0.882",
"0.6143",
"0.882",
"0.629",
"0.51",
"0.6274"
],
[
"0.11755",
"0.6353",
"0.4878",
"0.6353",
"0.4878",
"0.6484",
"0.11755",
"0.6484"
],
[
"0.531",
"0.63",
"0.881",
"0.63",
"0.881",
"0.64",
"0.531",
"0.64"
],
[
"0.529",
"0.641",
"0.882",
"0.641",
"0.882",
"0.655",
"0.529",
"0.655"
],
[
"0.11755",
"0.652",
"0.486",
"0.652",
"0.486",
"0.665",
"0.11755",
"0.665"
],
[
"0.529",
"0.655",
"0.8833",
"0.655",
"0.8833",
"0.6685",
"0.529",
"0.6685"
],
[
"0.11755",
"0.6675",
"0.424",
"0.6675",
"0.424",
"0.681",
"0.11755",
"0.681"
],
[
"0.4575",
"0.6675",
"0.486",
"0.6675",
"0.486",
"0.6797",
"0.4575",
"0.6797"
],
[
"0.5283",
"0.6675",
"0.659",
"0.6675",
"0.659",
"0.681",
"0.5283",
"0.681"
],
[
"0.11584",
"0.684",
"0.4878",
"0.684",
"0.4878",
"0.6978",
"0.11584",
"0.6978"
],
[
"0.513",
"0.6904",
"0.8833",
"0.6904",
"0.8833",
"0.7036",
"0.513",
"0.7036"
],
[
"0.1142",
"0.6997",
"0.4878",
"0.6987",
"0.4878",
"0.7134",
"0.1142",
"0.7144"
],
[
"0.529",
"0.703",
"0.885",
"0.705",
"0.885",
"0.7188",
"0.529",
"0.7163"
],
[
"0.11584",
"0.7144",
"0.1931",
"0.7163",
"0.1931",
"0.7285",
"0.11584",
"0.727"
],
[
"0.5283",
"0.7163",
"0.885",
"0.7173",
"0.885",
"0.7324",
"0.5283",
"0.731"
],
[
"0.5283",
"0.73",
"0.702",
"0.73",
"0.702",
"0.7437",
"0.5283",
"0.7437"
],
[
"0.11755",
"0.745",
"0.2217",
"0.745",
"0.2217",
"0.7603",
"0.11755",
"0.7603"
],
[
"0.513",
"0.753",
"0.882",
"0.753",
"0.882",
"0.7666",
"0.513",
"0.7666"
],
[
"0.1142",
"0.7715",
"0.4895",
"0.7725",
"0.4895",
"0.7876",
"0.1142",
"0.786"
],
[
"0.529",
"0.7666",
"0.882",
"0.7666",
"0.882",
"0.78",
"0.529",
"0.78"
],
[
"0.529",
"0.78",
"0.8833",
"0.78",
"0.8833",
"0.7935",
"0.529",
"0.7935"
],
[
"0.11755",
"0.789",
"0.4878",
"0.789",
"0.4878",
"0.804",
"0.11755",
"0.804"
],
[
"0.5283",
"0.792",
"0.7944",
"0.7915",
"0.7944",
"0.8057",
"0.5283",
"0.807"
],
[
"0.11755",
"0.8057",
"0.4878",
"0.8057",
"0.4878",
"0.8193",
"0.11755",
"0.8193"
],
[
"0.513",
"0.8154",
"0.796",
"0.8154",
"0.796",
"0.83",
"0.513",
"0.83"
],
[
"0.11584",
"0.822",
"0.4878",
"0.822",
"0.4878",
"0.835",
"0.11584",
"0.835"
],
[
"0.8076",
"0.818",
"0.882",
"0.818",
"0.882",
"0.829",
"0.8076",
"0.829"
],
[
"0.529",
"0.83",
"0.8833",
"0.83",
"0.8833",
"0.8433",
"0.529",
"0.8433"
],
[
"0.1142",
"0.8374",
"0.4878",
"0.836",
"0.4878",
"0.851",
"0.1142",
"0.8516"
],
[
"0.5283",
"0.8413",
"0.882",
"0.843",
"0.882",
"0.857",
"0.5283",
"0.8555"
],
[
"0.11755",
"0.854",
"0.4878",
"0.854",
"0.4878",
"0.8677",
"0.11755",
"0.8677"
],
[
"0.529",
"0.857",
"0.6094",
"0.857",
"0.6094",
"0.868",
"0.529",
"0.868"
],
[
"0.11755",
"0.869",
"0.486",
"0.869",
"0.486",
"0.883",
"0.11755",
"0.883"
],
[
"0.513",
"0.8774",
"0.6733",
"0.8774",
"0.6733",
"0.891",
"0.513",
"0.891"
],
[
"0.6665",
"0.879",
"0.797",
"0.879",
"0.797",
"0.8926",
"0.6665",
"0.8926"
],
[
"0.806",
"0.88",
"0.881",
"0.88",
"0.881",
"0.89",
"0.806",
"0.89"
],
[
"0.1351",
"0.891",
"0.4878",
"0.8936",
"0.4878",
"0.9077",
"0.1351",
"0.9062"
],
[
"0.5283",
"0.89",
"0.882",
"0.891",
"0.882",
"0.9062",
"0.5283",
"0.905"
],
[
"0.1142",
"0.904",
"0.346",
"0.905",
"0.346",
"0.9194",
"0.1142",
"0.9185"
],
[
"0.529",
"0.905",
"0.881",
"0.905",
"0.881",
"0.9185",
"0.529",
"0.9185"
]
] | [
"CLIP-S metrics for each of the three groups are sig-",
"to reduce this effect. Second, for the last experi",
"nificantly different from one another's. Therefore,",
"ment, we adopted the CLIP score to evaluate the",
"after stitching text encoders and the image decoder,",
"matching between a text and its generated image",
"CLIP-style text encoders achieve higher scores on",
"This might raise an issue: \"Do images prompted by",
"the CLIP-S metric that measures the matching of",
"CLIP-style text representations guarantee a higher",
"an image-text pair. This demonstrates the effective-",
"CLIP-S score owing to the fact that they are the",
"ness of CLIP-style text encoders on the association",
"same models?\". To answer this, we point out the",
"ability.9",
"reason why we adopted it. The frozen CLIP-style",
"(BERT-style) text encoders are only used to gen+",
"6",
"",
"Conclusion",
"erate prompts for image generation and the linear",
"layer is trained to maximize the likelihood of gener",
"",
"Human interaction is multi-modal. Starting from",
"ated images. Yet, the CLIP score is used to measure",
"the conjecture that text encoders learned from",
"the matching between images and texts. Therefore",
"multi-modal data have unique abilities, in this pa-",
"the uses of the CLIP text encoders and the CLIP",
"per, we study the behavioral difference between",
" score are disentangled.",
"BERT-style and CLIP-style text encoders.We",
"compare them from three aspects systematically:",
"Ethics Statement",
"() general (pure) text understanding; (ii) vision-",
"There are no ethics-related issues in this paper. The",
"centric text understanding; and (iii)text-to-image",
"generation.Experimental analyses show that al-",
" data and other related resources in this work are",
"though CLIP-style text encoders underperform",
"open-source and commonly used by many existing",
"BERT-style text encoders on general text under-",
"studies.",
"standing tasks, they have a unique ability, i.e.,",
"synesthesia, to associate a text and its visual appear-",
"References",
"ance, which is more similar to human perception.",
"Luisa Bentivogli, Peter Clark, Ido Dagan, and Danilo",
"Giampiccolo.2009.The fifth pascal recognizing",
"Acknowledgments",
"textual entailment challenge. In TAC.",
"This work is supported by Chinese Key-",
"Romain Bielawski, Benjamin Devillers, Tim Van de",
"Area Research and Development Program of",
"Cruys, and Rufin Vanrullen. 2022. When does clip",
"Guangdong Province (2020B0101350001), the",
"generalize better than unimodal models? when judg",
"Shenzhen Science and Technology Program",
"ing human-centric concepts. In Proceedings of the",
"(JCYJ20220818103001002),theGuangdong",
"7th Workshop on Representation Learning for NLP",
"pages 29-38.",
"Provincial Key Laboratory of Big Data Com-",
"puting, The Chinese University of Hong Kong,",
"Daniel Cer, Mona Diab, Eneko Agirre, Inigo Lopez-",
"Shenzhen, Shenzhen Key Research Project",
"Gazpio,and Lucia Specia.2017.Semeval-2017",
"task I: Semantic textual similarity-multilingual and",
"(C10120230151) and Shenzhen Doctoral Startup",
"cross-lingual focused evaluation.arXiv preprint",
"Funding(RCBS20221008093330065).",
"We",
"arXiv:1708.00055.",
"also thank Tiannan Wang for offering valuable",
"Ido Dagan, Oren Glickman, and Bernardo Magnini.",
"suggestions on the description of experiments in",
"2006.Thepascal recognisingtextual entailment chal-",
"Section 4.",
"lenge. In Machine learning challenges workshop,",
"pages 177-190. Springer.",
"Limitations",
"Jacob Devlin, Ming-Wei Chang, Kenton Lee, and",
"We highlight two limitations of our work. First,",
"Kristina Toutanova. 2018. Bert: Pre-training of deep",
"bidirectional transformers for language understand-",
"the empirical comparisons are not conducted under",
"ing. arXiv preprint arXiv:1810.04805.",
"fully controlled conditions, e.g., the sizes of mod-",
"Bill Dolan and Chris Brockett. 2005.",
"els. Limited by computational resources, we did",
"Automati-",
"cally constructing a corpus of sentential paraphrases.",
"not replicate different types of text encoders with",
"In Third International Workshop on Paraphrasing",
"the same number of parameters. Instead, we show",
"(IWP2005)",
"the results of different encoders of various sizes",
"AlexeyDosovitskiy,",
"LucasBeyer,",
"Alexander",
"°Both types of text encoders are not exposed to the repre-",
"Kolesnikov, Dirk Weissenborn, Xiaohua Zhai.",
"sentation space of the image decoder.",
"Thomas Unterthiner, Mostafa Dehghani, Matthias"
] | {
"arxiv_id": "2306.03678",
"language": "en",
"page_idx": 4
} |
||
2308.07223-en-0 | 2308.07223-en-0.jpg | Distance Matters For Improving Performance Estimation Under Covariate Shift
Mélanie Roschewitz
Imperial College London
mb121@ic.ac.uk
$ଝୁ$
$\pm$
## Abstract
Performance estimation under covariate shift is a crucial
$\begin{array}{ccc}{\mathrm{b0}}&{component\:of\:safe\:AI\:model\:deployment,\:especially\:for\:sen-\\{\mathrm{S}}}{sitive\:use-cases.\:Recently,\:several\:solutions\:were\:proposed}\\{\mathrm{S}}&{to\:tackle\:this\:problem,\:most\:leveraging\:model\:predictions}\end{array} $
to tackle this problem, most leveraging model predictions or softmax confidence to derive accuracy estimates. However, under dataset shifts confidence scores may become illcalibrated if samples are too far from the training distribu-
$\sum_{\begin{aligned}&tion.~In~this~work,~we~show~that~taking~into~account~dis-}\\&tances~oftest~samples~to~their~expected~training~distribution\\&can~signiffcantly~improve~performance~estimation~under~co-}\end{aligned}$
to avoid relying on their untrustworthy model outputs in the accuracy estimation step. We demonstrate the effectiveness of this method on 13 image classification tasks, across a wide-range of natural and synthetic distribution shifts and hundreds of models, with a median relative MAE improvement of 27% over the best baseline across all tasks, and SOTA performance on 10 out of 13 tasks. Our code is publicly available at https: //github. com/melanibe/ distance\_matters\_performance\_estimation.
Ben Glocker
Imperial College London
b.glocker@imperial.ac.uk
<FigureHere>
Figure 1. Performance estimation under covariate shift needs to take into account different sources of errors. Distance to the source distribution in the embedding space matters as confidence estimates become unreliable with increased distance.
### $1. \textbf{Introduction}$
Machine learning models are sensitive to variations in their deployment environments [19,63,40,26,57,38]. Due to the unavailability of ground truth labels for continuous performance monitoring at deployment time, real-time and accurate performance estimation is crucial to detect any unexpected behavior or model failure, particularly in distribution-shifted settings. This is especially important for sensitive use cases such as clinical decision making.
The difficulty in estimating model performance arises from the lack of reliability of model outputs under covariate shift [40, 28]. Recently, several attempts have been made at addressing this problem, many of them based on confidence estimates [16, 15, 32]. For example, Average Thresholded Confidence (ATC) [15] leverages softmax outputs for esti-
mating classification accuracy, considering that all outputs whose confidence do not reach a certain threshold are incorrectly classifed. While this method has shown to be effective at estimating the performance under mild shifts (e.g. on synthetically corrupted images), experiments show that the method under-performs in more substantial shifts such as natural sub-population shifts. In particular, current approaches tend to overestimate accuracy in natural realworld distribution shift settings [15,23]. This can notably be explained by a deterioration of model calibration when going further from the training distribution [25], with softmax outputs becoming over-confident and unreliable [40]. | [
[
"0.08496",
"0.1338",
"0.8857",
"0.1338",
"0.8857",
"0.1483",
"0.08496",
"0.1483"
],
[
"0.2656",
"0.1831",
"0.429",
"0.1831",
"0.429",
"0.1976",
"0.2656",
"0.1976"
],
[
"0.5664",
"0.1831",
"0.6685",
"0.1831",
"0.6685",
"0.1976",
"0.5664",
"0.1976"
],
[
"0.2484",
"0.2008",
"0.4468",
"0.2008",
"0.4468",
"0.2153",
"0.2484",
"0.2153"
],
[
"0.5156",
"0.2008",
"0.7173",
"0.1995",
"0.7173",
"0.214",
"0.5156",
"0.2153"
],
[
"0.2834",
"0.2197",
"0.4102",
"0.2197",
"0.4102",
"0.2329",
"0.2834",
"0.2329"
],
[
"0.509",
"0.2197",
"0.7236",
"0.2197",
"0.7236",
"0.2329",
"0.509",
"0.2329"
],
[
"0.0237",
"0.2583",
"0.05963",
"0.2583",
"0.05963",
"0.3225",
"0.0237",
"0.3225"
],
[
"0.2378",
"0.2695",
"0.3137",
"0.2695",
"0.3137",
"0.2842",
"0.2378",
"0.2842"
],
[
"0.761",
"0.2842",
"0.826",
"0.2842",
"0.826",
"0.2954",
"0.761",
"0.2954"
],
[
"0.0989",
"0.3018",
"0.468",
"0.3018",
"0.468",
"0.315",
"0.0989",
"0.315"
],
[
"0.7534",
"0.2974",
"0.8594",
"0.2974",
"0.8594",
"0.308",
"0.7534",
"0.308"
],
[
"0.03105",
"0.308",
"0.0645",
"0.3113",
"0.05637",
"0.3687",
"0.02287",
"0.3655"
],
[
"0.761",
"0.3113",
"0.8667",
"0.3113",
"0.8667",
"0.324",
"0.761",
"0.324"
],
[
"0.0801",
"0.3176",
"0.4673",
"0.3176",
"0.4673",
"0.3308",
"0.0801",
"0.3308"
],
[
"0.0801",
"0.3333",
"0.4697",
"0.3333",
"0.4697",
"0.3467",
"0.0801",
"0.3467"
],
[
"0.7573",
"0.3384",
"0.8447",
"0.3384",
"0.8447",
"0.3499",
"0.7573",
"0.3499"
],
[
"0.0784",
"0.3467",
"0.4697",
"0.348",
"0.4697",
"0.3623",
"0.0784",
"0.361"
],
[
"0.02696",
"0.3643",
"0.05392",
"0.3643",
"0.05392",
"0.3926",
"0.02696",
"0.3926"
],
[
"0.0801",
"0.363",
"0.4673",
"0.363",
"0.4673",
"0.377",
"0.0801",
"0.377"
],
[
"0.528",
"0.3687",
"0.649",
"0.3687",
"0.649",
"0.379",
"0.528",
"0.379"
],
[
"0.6787",
"0.3694",
"0.74",
"0.3694",
"0.74",
"0.3801",
"0.6787",
"0.3801"
],
[
"0.0784",
"0.3774",
"0.468",
"0.377",
"0.468",
"0.3901",
"0.0784",
"0.3914"
],
[
"0.55",
"0.3833",
"0.584",
"0.3833",
"0.584",
"0.3945",
"0.55",
"0.3945"
],
[
"0.652",
"0.3806",
"0.74",
"0.3833",
"0.74",
"0.3945",
"0.652",
"0.3926"
],
[
"0.0801",
"0.3926",
"0.468",
"0.3926",
"0.468",
"0.406",
"0.0801",
"0.406"
],
[
"0.5557",
"0.3965",
"0.6274",
"0.3977",
"0.6274",
"0.4092",
"0.5557",
"0.4077"
],
[
"0.6504",
"0.3977",
"0.7305",
"0.3977",
"0.7305",
"0.4092",
"0.6504",
"0.4092"
],
[
"0.02533",
"0.406",
"0.05554",
"0.406",
"0.05554",
"0.4507",
"0.02533",
"0.4507"
],
[
"0.0801",
"0.4077",
"0.4673",
"0.4077",
"0.4673",
"0.4216",
"0.0801",
"0.4216"
],
[
"0.585",
"0.4136",
"0.8",
"0.4136",
"0.8",
"0.4236",
"0.585",
"0.4236"
],
[
"0.0809",
"0.4248",
"0.468",
"0.4248",
"0.468",
"0.435",
"0.0809",
"0.435"
],
[
"0.5186",
"0.4287",
"0.572",
"0.4312",
"0.572",
"0.4426",
"0.5186",
"0.44"
],
[
"0.701",
"0.4287",
"0.7524",
"0.4312",
"0.7524",
"0.4426",
"0.701",
"0.44"
],
[
"0.0801",
"0.4382",
"0.468",
"0.4382",
"0.468",
"0.452",
"0.0801",
"0.452"
],
[
"0.5215",
"0.44",
"0.6094",
"0.44",
"0.6094",
"0.452",
"0.5215",
"0.452"
],
[
"0.701",
"0.44",
"0.7837",
"0.4395",
"0.7837",
"0.4507",
"0.701",
"0.452"
],
[
"0.03186",
"0.449",
"0.058",
"0.449",
"0.058",
"0.481",
"0.03186",
"0.481"
],
[
"0.523",
"0.452",
"0.581",
"0.452",
"0.581",
"0.4634",
"0.523",
"0.4634"
],
[
"0.7017",
"0.452",
"0.7607",
"0.452",
"0.7607",
"0.4634",
"0.7017",
"0.4634"
],
[
"0.523",
"0.4634",
"0.5864",
"0.4634",
"0.5864",
"0.4734",
"0.523",
"0.4734"
],
[
"0.0809",
"0.455",
"0.4673",
"0.455",
"0.4673",
"0.4653",
"0.0809",
"0.4653"
],
[
"0.0784",
"0.4685",
"0.4673",
"0.4685",
"0.4673",
"0.4817",
"0.0784",
"0.4817"
],
[
"0.7036",
"0.4634",
"0.767",
"0.4634",
"0.767",
"0.4734",
"0.7036",
"0.4734"
],
[
"0.0801",
"0.4841",
"0.468",
"0.4841",
"0.468",
"0.4976",
"0.0801",
"0.4976"
],
[
"0.5728",
"0.4956",
"0.5864",
"0.4956",
"0.5864",
"0.504",
"0.5728",
"0.504"
],
[
"0.759",
"0.4924",
"0.773",
"0.4924",
"0.773",
"0.5005",
"0.759",
"0.5005"
],
[
"0.02287",
"0.5005",
"0.06046",
"0.5005",
"0.06046",
"0.5986",
"0.02287",
"0.5986"
],
[
"0.0801",
"0.4988",
"0.4697",
"0.4988",
"0.4697",
"0.512",
"0.0801",
"0.512"
],
[
"0.0784",
"0.5146",
"0.4697",
"0.5146",
"0.4697",
"0.528",
"0.0784",
"0.528"
],
[
"0.0801",
"0.529",
"0.4714",
"0.529",
"0.4714",
"0.5425",
"0.0801",
"0.5425"
],
[
"0.0784",
"0.5435",
"0.4673",
"0.5444",
"0.4673",
"0.558",
"0.0784",
"0.5566"
],
[
"0.5415",
"0.5537",
"0.653",
"0.5537",
"0.653",
"0.5645",
"0.5415",
"0.5645"
],
[
"0.71",
"0.5537",
"0.8667",
"0.5537",
"0.8667",
"0.5645",
"0.71",
"0.5645"
],
[
"0.5415",
"0.5645",
"0.6543",
"0.5645",
"0.6543",
"0.5776",
"0.5415",
"0.5776"
],
[
"0.722",
"0.5645",
"0.8667",
"0.5645",
"0.8667",
"0.5776",
"0.722",
"0.5776"
],
[
"0.0809",
"0.5737",
"0.468",
"0.5737",
"0.468",
"0.587",
"0.0809",
"0.587"
],
[
"0.0801",
"0.5596",
"0.4697",
"0.5596",
"0.4697",
"0.5728",
"0.0801",
"0.5728"
],
[
"0.5435",
"0.579",
"0.614",
"0.579",
"0.614",
"0.589",
"0.5435",
"0.589"
],
[
"0.706",
"0.579",
"0.7935",
"0.579",
"0.7935",
"0.589",
"0.706",
"0.589"
],
[
"0.0801",
"0.589",
"0.468",
"0.589",
"0.468",
"0.603",
"0.0801",
"0.603"
],
[
"0.5376",
"0.5996",
"0.664",
"0.5996",
"0.664",
"0.61",
"0.5376",
"0.61"
],
[
"0.0196",
"0.584",
"0.058",
"0.584",
"0.06046",
"0.712",
"0.02287",
"0.712"
],
[
"0.0825",
"0.605",
"0.468",
"0.605",
"0.468",
"0.6187",
"0.0825",
"0.6187"
],
[
"0.7017",
"0.5996",
"0.8594",
"0.5996",
"0.8594",
"0.61",
"0.7017",
"0.61"
],
[
"0.528",
"0.6123",
"0.6733",
"0.6123",
"0.6733",
"0.6226",
"0.528",
"0.6226"
],
[
"0.705",
"0.613",
"0.8594",
"0.613",
"0.8594",
"0.6206",
"0.705",
"0.6206"
],
[
"0.5024",
"0.63",
"0.892",
"0.63",
"0.892",
"0.643",
"0.5024",
"0.643"
],
[
"0.0809",
"0.6494",
"0.2075",
"0.6494",
"0.2075",
"0.6636",
"0.0809",
"0.6636"
],
[
"0.5034",
"0.646",
"0.8906",
"0.646",
"0.8906",
"0.656",
"0.5034",
"0.656"
],
[
"0.5024",
"0.6577",
"0.892",
"0.6577",
"0.892",
"0.672",
"0.5024",
"0.672"
],
[
"0.0972",
"0.675",
"0.468",
"0.675",
"0.468",
"0.688",
"0.0972",
"0.688"
],
[
"0.5",
"0.6704",
"0.814",
"0.6704",
"0.814",
"0.6846",
"0.5",
"0.6846"
],
[
"0.0784",
"0.688",
"0.4673",
"0.69",
"0.4673",
"0.704",
"0.0784",
"0.702"
],
[
"0.0801",
"0.705",
"0.468",
"0.705",
"0.468",
"0.7183",
"0.0801",
"0.7183"
],
[
"0.0801",
"0.72",
"0.4697",
"0.72",
"0.4697",
"0.7344",
"0.0801",
"0.7344"
],
[
"0.501",
"0.72",
"0.8936",
"0.7197",
"0.8936",
"0.7344",
"0.501",
"0.735"
],
[
"0.0801",
"0.735",
"0.4697",
"0.735",
"0.4697",
"0.7485",
"0.0801",
"0.7485"
],
[
"0.5034",
"0.7363",
"0.8906",
"0.7363",
"0.8906",
"0.75",
"0.5034",
"0.75"
],
[
"0.0784",
"0.7505",
"0.468",
"0.7505",
"0.468",
"0.7646",
"0.0784",
"0.7646"
],
[
"0.5034",
"0.7524",
"0.8906",
"0.7524",
"0.8906",
"0.763",
"0.5034",
"0.763"
],
[
"0.0801",
"0.7666",
"0.468",
"0.7666",
"0.468",
"0.78",
"0.0801",
"0.78"
],
[
"0.5024",
"0.7666",
"0.892",
"0.7666",
"0.892",
"0.78",
"0.5024",
"0.78"
],
[
"0.0801",
"0.781",
"0.4673",
"0.781",
"0.4673",
"0.7944",
"0.0801",
"0.7944"
],
[
"0.5024",
"0.782",
"0.892",
"0.782",
"0.892",
"0.7954",
"0.5024",
"0.7954"
],
[
"0.5024",
"0.797",
"0.892",
"0.797",
"0.892",
"0.81",
"0.5024",
"0.81"
],
[
"0.10046",
"0.811",
"0.4697",
"0.811",
"0.4697",
"0.8247",
"0.10046",
"0.8247"
],
[
"0.5024",
"0.811",
"0.892",
"0.811",
"0.892",
"0.8247",
"0.5024",
"0.8247"
],
[
"0.0801",
"0.827",
"0.4697",
"0.827",
"0.4697",
"0.8403",
"0.0801",
"0.8403"
],
[
"0.501",
"0.827",
"0.8896",
"0.827",
"0.8896",
"0.8403",
"0.501",
"0.8403"
],
[
"0.0801",
"0.8413",
"0.4697",
"0.8413",
"0.4697",
"0.855",
"0.0801",
"0.855"
],
[
"0.5024",
"0.8413",
"0.8906",
"0.8413",
"0.8906",
"0.855",
"0.5024",
"0.855"
],
[
"0.0801",
"0.857",
"0.4697",
"0.857",
"0.4697",
"0.8706",
"0.0801",
"0.8706"
],
[
"0.5024",
"0.857",
"0.892",
"0.857",
"0.892",
"0.8706",
"0.5024",
"0.8706"
],
[
"0.0801",
"0.872",
"0.4697",
"0.872",
"0.4697",
"0.8853",
"0.0801",
"0.8853"
],
[
"0.501",
"0.872",
"0.892",
"0.8706",
"0.892",
"0.8853",
"0.501",
"0.886"
],
[
"0.0801",
"0.887",
"0.468",
"0.887",
"0.468",
"0.901",
"0.0801",
"0.901"
],
[
"0.5024",
"0.887",
"0.892",
"0.887",
"0.892",
"0.901",
"0.5024",
"0.901"
]
] | [
"Distance Matters For Improving Performance Estimation Under Covariate Shift",
"Melanie Roschewitz",
"Ben Glocker",
"Imperial College London",
"Imperial College London",
"mb121@ic.ac.uk",
"b.glocker@imperial.ac.uk",
"2",
"Abstract",
"Toofar from",
"Performance estimation under covariate shift is a crucial",
"trainingdistribution",
"Sn",
"unreliable confidence",
"component of safe AI model deployment, especially for sen-",
"sitive use-cases. Recently, several solutions were proposed",
"Lowconfidence",
"to tackle this problem, most leveraging model predictions",
"4",
"or softmax confidence to derive accuracy estimates. How-",
"OOD(test)ID(train)",
"Misclassified",
"ever, underdataset shifts confidence scores may become ill-",
"Class 0",
"·Learned Boundary",
"calibrated if samples are too far from the training distribu-",
"Well calibrated",
"Poorly calibrated",
"LAD",
"tion. In this work, we show that taking into account dis-",
"(a)SourcesoferrorsforOODgeneralisation",
"tancesoftestsamplestotheirexpectedtrainingdistribution",
"Accuracy",
"Accuracy",
"can significantly improve performance estimation under co-",
"Predicted:12/15",
"Predicted:7/15",
"SO",
"True:8/15",
"True:8/15",
"Error:4/15",
"variate shift. Preciselywe introduce a\"distance-check\"to",
"flag samples that lie too far from the expected distribution,",
"Error:1/15",
"to avoid relying on their untrustworthy model outputs in the",
"",
"",
"",
"accuracy estimation step. We demonstrate the effectiveness",
"of this method on 13 image classification tasks, across a",
"wide-range of natural and synthetic distribution shifts and",
"hundreds of models, with a median relative MAE improve-",
"Passconfidencecheck",
"Pass confid.+distancecheck",
"Fail confidence check",
"Fail confid. + distance check",
"SOTA performance on 10 out of 13 tasks. Our code is pub-",
"ment of 27% over the best baseline across all tasks, and",
"Misclassified",
"Misclassified",
"liclyavailableathttps://github.com/melanibe/",
"(b)Confidence-basedonly",
"00",
"distance_matters_performance_estimation.",
"(c)Confidence-based+distance",
"accuracy estimation (e.g.[14])",
"basedaccuracyestimation(ours)",
"Figure 1. Performance estimation under covariate shift needs",
"1. Introduction",
"to take into account different sources of errors. Distance to the",
"source distribution in the embedding space matters as confidence",
"Machine learning models are sensitive to variations in",
"estimates become unreliable with increased distance.",
"their deployment environments [19, 63, 40, 26, 57, 38].",
"Due to the unavailability of ground truth labels for continu-",
"ous performance monitoring at deployment time, real-time",
"mating classification accuracy, considering that all outputs",
"and accurate performance estimation is crucial to detect",
"whose confidence do not reach a certain threshold are in-",
"any unexpected behavior or model failure, particularly",
"correctlyclassified.While this method has shownto be",
"in distribution-shifted settings. This is especially impor-",
"effective at estimating the performance under mild shifts",
"tant for sensitive use cases such as clinical decision making.",
"(e.g. on synthetically corrupted images), experiments show",
"that the method under-performs in more substantial shifts",
"The difficulty in estimating model performance arises",
"such as natural sub-population shifts. In particular, current",
"from the lack of reliability of model outputs under covariate",
"approaches tend to overestimate accuracy in natural real-",
"shift [40, 28]. Recently, several attempts have been made at",
"world distribution shift settings [15, 23]. This can notably",
"addressing this problem, many of them based on confidence",
"be explained by a deterioration of model calibration when",
"estimates [16, 15, 32]. For example, Average Thresholded",
"going further from the training distribution [25], with soft-",
"Confidence (ATC) [15] leverages softmax outputs for esti-",
"max outputs becoming over-confident and unreliable [40]."
] | {
"arxiv_id": "2308.07223",
"language": "en",
"page_idx": 0
} |
||
2308.07223-en-1 | 2308.07223-en-1.jpg | If test samples are too far from training samples, relying on the output of the classification layer for performance estimation is insufficient. From an uncertainty point of view, softmax outputs can been seen as capturing aleatoric uncertainty, arising from overlapping class boundaries [24,9]. However, under dataset shifts, errors may also arise from the fact that the model has never seen this type of input data and does not know how to respond to such inputs. This is referred to as epistemic uncertainty [9] and is not well captured by softmax outputs [24,53], as demonstrated by its poor performance on the related out-of-distribution (OOD) detection task[40,28,48,53,30]. Note that in OOD detection, the goal is to separate on separating ID from OOD inputs, regardless of the downstream classification performance, often considering inputs completely unrelated to the task. This differs from performance estimation under covariate shift, where we assume that the classification task still applies to the shifted inputs and we focus on estimating performance, not on detecting shifts.
Methodological contributions In this paper, we argue that performance estimators should identify samples far away from the training set in the embedding space, for which softmax estimates are most likely unreliable. By measuring the distance in the embedding space, we are able to measure how well the model“understood”the sample when projecting the input to the classification space. This idea is illustrated in fig. 1. Following this intuition, we propose a simple yet effective method to improve the quality of current SOTA performance estimators. Specifically, we use nearest-neighbours distance in the embedding space to reject samples that lay too far from the training distribution. We then only use confidence-based performance estimators on the remaining samples, considering all previously rejected samples as mis-classified. Our distance check approach is versatile and can be used to improve the quality of various existing performance estimators (e.g. [15,23]).
Main results We evaluate our approach on 13 classification tasks ranging from cancer cell to animal classification. The nature of the distribution shifts studied covers a widerange of shifts: from synthetic corruption, acquisition shift, real-world population shift to representation shift. For each task we evaluate between 18 and 259 models, covering various training strategies and network architectures. These experiments demonstrate that integrating distance into accuracy estimators significantly improves the quality of the estimation. For example, our proposed estimator ATC-DistCS is significantly better than previous SOTA ATC[15] on all but one task, with a median relative MAE improvement of 30% across all tasks. Furthermore, comparing to the most recent COT method[34], we demonstrate a 27% median relative performance improvement across all tasks, with new
SOTA performances on 10 out of 13 tasks. We also demonstrate significant improvements across all datasets for agreement based accuracy estimation when integrating our distance check. Ablation studies yield further insights in the method and its limitations. Finally, to the best of our knowledge, we provide the first comprehensive publicly available codebase of current SOTA baselines for accuracy estimation, along with the complete code to reproduce our experiments.
## $2. \textbf{ Background}$
$2. 1. \textbf{ Performance estimation without ground truth}$
Current methods for performance estimation under co-
variate shift can be broadly grouped in 4 categories:
Estimating performance via auxiliary task performance Modifies the main classification model to incorporate a(sufficiently informative) auxiliary task for which ground truth labels are available at test time: accuracy on the main task is then approximated by the computed accuracy on the auxiliary task. For example, [11] trains a multi-task model for predicting the class at hand as well as the rotation applied to the input image. The main limitation of this line of work is the requirement to build a multi-task model, making it unusable as a post-hoc tool.
Training a regressor between ID and OOD accuracy This regressor can be trained based on model outputs or on measures of distance between datasets [13,47,16,35,12]. One major drawback of this class of estimators is their requirement for having access to labelled OOD data for training the regression model. This is not always available in practice, in particular in data-scarce domains such that healthcare. In absence of such OOD datasets, regressor are sometimes trained using corrupted versions of the original validation set as“OOD” sets. $\hat{\text{However,thiscannotguaran-}}$ tee the robustness of this estimator against other shifts e.g. natural subpopulation shift[46].
Agreement-based estimators Are based on the idea that agreement between member of model ensembles correlate with model accuracy. For example, generalised disagreement equality (GDE) [23] use pairs of models trained with different random seeds to compute disagreement. Others use more intricate methods for training specialised models to align disagreement and accuracy further [10,6]. However, these procedures often require expensive additional training steps to derive the siblings models and are not applicable to post-hoc scenarios where only the final model is available to the end user. In [1], the authors go as far as training dozens models to fit a regressor between agreement | [
[
"0.0809",
"0.09216",
"0.4697",
"0.09216",
"0.4697",
"0.1061",
"0.0809",
"0.1061"
],
[
"0.5034",
"0.09155",
"0.892",
"0.09155",
"0.892",
"0.1061",
"0.5034",
"0.1061"
],
[
"0.0801",
"0.108",
"0.468",
"0.108",
"0.468",
"0.1218",
"0.0801",
"0.1218"
],
[
"0.5034",
"0.1092",
"0.8906",
"0.1092",
"0.8906",
"0.1193",
"0.5034",
"0.1193"
],
[
"0.0801",
"0.1237",
"0.4697",
"0.1237",
"0.4697",
"0.137",
"0.0801",
"0.137"
],
[
"0.5034",
"0.125",
"0.8906",
"0.125",
"0.8906",
"0.1351",
"0.5034",
"0.1351"
],
[
"0.0801",
"0.1383",
"0.468",
"0.1383",
"0.468",
"0.1515",
"0.0801",
"0.1515"
],
[
"0.5024",
"0.1383",
"0.892",
"0.1383",
"0.892",
"0.1515",
"0.5024",
"0.1515"
],
[
"0.0801",
"0.154",
"0.468",
"0.154",
"0.468",
"0.1674",
"0.0801",
"0.1674"
],
[
"0.5024",
"0.1528",
"0.8906",
"0.1528",
"0.8906",
"0.1666",
"0.5024",
"0.1666"
],
[
"0.0784",
"0.1666",
"0.4697",
"0.1674",
"0.4697",
"0.1818",
"0.0784",
"0.1812"
],
[
"0.501",
"0.1674",
"0.8906",
"0.1674",
"0.8906",
"0.1812",
"0.501",
"0.1812"
],
[
"0.0801",
"0.1831",
"0.4697",
"0.1831",
"0.4697",
"0.1963",
"0.0801",
"0.1963"
],
[
"0.5024",
"0.1831",
"0.8906",
"0.1831",
"0.8906",
"0.1963",
"0.5024",
"0.1963"
],
[
"0.0809",
"0.1995",
"0.468",
"0.1995",
"0.468",
"0.2102",
"0.0809",
"0.2102"
],
[
"0.501",
"0.1989",
"0.8896",
"0.1989",
"0.8896",
"0.2122",
"0.501",
"0.2122"
],
[
"0.0801",
"0.2134",
"0.468",
"0.2134",
"0.468",
"0.2267",
"0.0801",
"0.2267"
],
[
"0.5024",
"0.214",
"0.5513",
"0.214",
"0.5513",
"0.226",
"0.5024",
"0.226"
],
[
"0.0809",
"0.2299",
"0.4673",
"0.2299",
"0.4673",
"0.2405",
"0.0809",
"0.2405"
],
[
"0.0809",
"0.2437",
"0.468",
"0.2437",
"0.468",
"0.2568",
"0.0809",
"0.2568"
],
[
"0.5034",
"0.2437",
"0.6274",
"0.2437",
"0.6274",
"0.2583",
"0.5034",
"0.2583"
],
[
"0.0801",
"0.2583",
"0.4673",
"0.2583",
"0.4673",
"0.2715",
"0.0801",
"0.2715"
],
[
"0.0801",
"0.2734",
"0.468",
"0.2734",
"0.468",
"0.2874",
"0.0801",
"0.2874"
],
[
"0.5024",
"0.2676",
"0.884",
"0.2676",
"0.884",
"0.2822",
"0.5024",
"0.2822"
],
[
"0.0801",
"0.289",
"0.4697",
"0.289",
"0.4697",
"0.303",
"0.0801",
"0.303"
],
[
"0.523",
"0.2922",
"0.8906",
"0.2922",
"0.8906",
"0.307",
"0.523",
"0.307"
],
[
"0.0801",
"0.3037",
"0.4697",
"0.3037",
"0.4697",
"0.3176",
"0.0801",
"0.3176"
],
[
"0.5034",
"0.3093",
"0.8433",
"0.3093",
"0.8433",
"0.3225",
"0.5034",
"0.3225"
],
[
"0.0801",
"0.318",
"0.4673",
"0.318",
"0.4673",
"0.332",
"0.0801",
"0.332"
],
[
"0.0809",
"0.3352",
"0.4673",
"0.3352",
"0.4673",
"0.3452",
"0.0809",
"0.3452"
],
[
"0.0784",
"0.348",
"0.468",
"0.3484",
"0.468",
"0.363",
"0.0784",
"0.3623"
],
[
"0.5034",
"0.3447",
"0.892",
"0.3447",
"0.892",
"0.358",
"0.5034",
"0.358"
],
[
"0.0801",
"0.363",
"0.3203",
"0.363",
"0.3203",
"0.377",
"0.0801",
"0.377"
],
[
"0.5034",
"0.3599",
"0.8906",
"0.3599",
"0.8906",
"0.3738",
"0.5034",
"0.3738"
],
[
"0.5034",
"0.375",
"0.892",
"0.375",
"0.892",
"0.3882",
"0.5034",
"0.3882"
],
[
"0.5034",
"0.3901",
"0.892",
"0.3901",
"0.892",
"0.404",
"0.5034",
"0.404"
],
[
"0.0801",
"0.3977",
"0.468",
"0.3984",
"0.468",
"0.4136",
"0.0801",
"0.4124"
],
[
"0.5024",
"0.406",
"0.892",
"0.406",
"0.892",
"0.4192",
"0.5024",
"0.4192"
],
[
"0.0801",
"0.414",
"0.468",
"0.414",
"0.468",
"0.428",
"0.0801",
"0.428"
],
[
"0.5034",
"0.4204",
"0.892",
"0.4204",
"0.892",
"0.4343",
"0.5034",
"0.4343"
],
[
"0.0784",
"0.43",
"0.468",
"0.428",
"0.468",
"0.4414",
"0.0784",
"0.443"
],
[
"0.5034",
"0.435",
"0.892",
"0.435",
"0.892",
"0.449",
"0.5034",
"0.449"
],
[
"0.0809",
"0.4458",
"0.4656",
"0.4458",
"0.4656",
"0.4558",
"0.0809",
"0.4558"
],
[
"0.5034",
"0.4507",
"0.8906",
"0.4507",
"0.8906",
"0.464",
"0.5034",
"0.464"
],
[
"0.0784",
"0.459",
"0.468",
"0.4583",
"0.468",
"0.473",
"0.0784",
"0.4734"
],
[
"0.5024",
"0.4666",
"0.892",
"0.4666",
"0.892",
"0.4797",
"0.5024",
"0.4797"
],
[
"0.0801",
"0.4749",
"0.468",
"0.4749",
"0.468",
"0.488",
"0.0801",
"0.488"
],
[
"0.5024",
"0.481",
"0.6855",
"0.481",
"0.6855",
"0.4944",
"0.5024",
"0.4944"
],
[
"0.0809",
"0.4912",
"0.468",
"0.4912",
"0.468",
"0.502",
"0.0809",
"0.502"
],
[
"0.0801",
"0.505",
"0.468",
"0.505",
"0.468",
"0.518",
"0.0801",
"0.518"
],
[
"0.0809",
"0.5215",
"0.4673",
"0.5215",
"0.4673",
"0.532",
"0.0809",
"0.532"
],
[
"0.5034",
"0.5176",
"0.8906",
"0.5176",
"0.8906",
"0.532",
"0.5034",
"0.532"
],
[
"0.0801",
"0.535",
"0.4697",
"0.535",
"0.4697",
"0.549",
"0.0801",
"0.549"
],
[
"0.5024",
"0.5327",
"0.892",
"0.5327",
"0.892",
"0.547",
"0.5024",
"0.547"
],
[
"0.0801",
"0.5503",
"0.468",
"0.5503",
"0.468",
"0.5645",
"0.0801",
"0.5645"
],
[
"0.5024",
"0.5474",
"0.8906",
"0.5474",
"0.8906",
"0.561",
"0.5024",
"0.561"
],
[
"0.0809",
"0.565",
"0.468",
"0.565",
"0.468",
"0.579",
"0.0809",
"0.579"
],
[
"0.5034",
"0.5645",
"0.8906",
"0.5645",
"0.8906",
"0.5747",
"0.5034",
"0.5747"
],
[
"0.0801",
"0.579",
"0.468",
"0.579",
"0.468",
"0.5923",
"0.0801",
"0.5923"
],
[
"0.5034",
"0.5776",
"0.892",
"0.5776",
"0.892",
"0.5913",
"0.5034",
"0.5913"
],
[
"0.0801",
"0.595",
"0.468",
"0.595",
"0.468",
"0.6094",
"0.0801",
"0.6094"
],
[
"0.501",
"0.5923",
"0.8906",
"0.5923",
"0.8906",
"0.606",
"0.501",
"0.606"
],
[
"0.0801",
"0.61",
"0.4673",
"0.61",
"0.4673",
"0.6235",
"0.0801",
"0.6235"
],
[
"0.5024",
"0.608",
"0.892",
"0.608",
"0.892",
"0.622",
"0.5024",
"0.622"
],
[
"0.0801",
"0.625",
"0.4697",
"0.625",
"0.4697",
"0.6396",
"0.0801",
"0.6396"
],
[
"0.501",
"0.6226",
"0.892",
"0.6235",
"0.892",
"0.637",
"0.501",
"0.636"
],
[
"0.0801",
"0.6416",
"0.4453",
"0.6416",
"0.4453",
"0.655",
"0.0801",
"0.655"
],
[
"0.5024",
"0.6396",
"0.8936",
"0.6396",
"0.8936",
"0.653",
"0.5024",
"0.653"
],
[
"0.505",
"0.655",
"0.892",
"0.655",
"0.892",
"0.6655",
"0.505",
"0.6655"
],
[
"0.501",
"0.6675",
"0.8906",
"0.6685",
"0.8906",
"0.683",
"0.501",
"0.6816"
],
[
"0.0801",
"0.675",
"0.4673",
"0.675",
"0.4673",
"0.6895",
"0.0801",
"0.6895"
],
[
"0.0801",
"0.69",
"0.468",
"0.69",
"0.468",
"0.704",
"0.0801",
"0.704"
],
[
"0.5024",
"0.6846",
"0.714",
"0.6846",
"0.714",
"0.6978",
"0.5024",
"0.6978"
],
[
"0.0809",
"0.707",
"0.4673",
"0.707",
"0.4673",
"0.7173",
"0.0809",
"0.7173"
],
[
"0.0809",
"0.7227",
"0.4673",
"0.7227",
"0.4673",
"0.733",
"0.0809",
"0.733"
],
[
"0.5024",
"0.7197",
"0.8906",
"0.7197",
"0.8906",
"0.733",
"0.5024",
"0.733"
],
[
"0.0809",
"0.7373",
"0.4673",
"0.7373",
"0.4673",
"0.7476",
"0.0809",
"0.7476"
],
[
"0.505",
"0.7373",
"0.8906",
"0.7373",
"0.8906",
"0.7476",
"0.505",
"0.7476"
],
[
"0.0801",
"0.7505",
"0.468",
"0.7505",
"0.468",
"0.7646",
"0.0801",
"0.7646"
],
[
"0.501",
"0.75",
"0.8896",
"0.7505",
"0.8896",
"0.765",
"0.501",
"0.7646"
],
[
"0.0801",
"0.7666",
"0.468",
"0.7666",
"0.468",
"0.78",
"0.0801",
"0.78"
],
[
"0.5024",
"0.7666",
"0.8906",
"0.7666",
"0.8906",
"0.78",
"0.5024",
"0.78"
],
[
"0.0801",
"0.782",
"0.468",
"0.782",
"0.468",
"0.7954",
"0.0801",
"0.7954"
],
[
"0.501",
"0.78",
"0.892",
"0.781",
"0.892",
"0.7954",
"0.501",
"0.7944"
],
[
"0.0809",
"0.797",
"0.468",
"0.797",
"0.468",
"0.81",
"0.0809",
"0.81"
],
[
"0.5024",
"0.797",
"0.892",
"0.797",
"0.892",
"0.81",
"0.5024",
"0.81"
],
[
"0.0801",
"0.811",
"0.468",
"0.811",
"0.468",
"0.8247",
"0.0801",
"0.8247"
],
[
"0.5024",
"0.811",
"0.8906",
"0.811",
"0.8906",
"0.8247",
"0.5024",
"0.8247"
],
[
"0.0784",
"0.8257",
"0.468",
"0.8247",
"0.468",
"0.839",
"0.0784",
"0.8403"
],
[
"0.5024",
"0.827",
"0.8906",
"0.827",
"0.8906",
"0.8403",
"0.5024",
"0.8403"
],
[
"0.0801",
"0.8413",
"0.4697",
"0.8413",
"0.4697",
"0.855",
"0.0801",
"0.855"
],
[
"0.5024",
"0.8413",
"0.8906",
"0.8413",
"0.8906",
"0.855",
"0.5024",
"0.855"
],
[
"0.0784",
"0.855",
"0.4697",
"0.856",
"0.4697",
"0.8706",
"0.0784",
"0.869"
],
[
"0.5024",
"0.857",
"0.892",
"0.857",
"0.892",
"0.8706",
"0.5024",
"0.8706"
],
[
"0.0801",
"0.872",
"0.4697",
"0.872",
"0.4697",
"0.8853",
"0.0801",
"0.8853"
],
[
"0.501",
"0.8706",
"0.892",
"0.872",
"0.892",
"0.886",
"0.501",
"0.8853"
],
[
"0.0784",
"0.887",
"0.4673",
"0.887",
"0.4673",
"0.901",
"0.0784",
"0.901"
],
[
"0.5024",
"0.887",
"0.8936",
"0.887",
"0.8936",
"0.901",
"0.5024",
"0.901"
]
] | [
"If test samples are too far from training samples, relying on",
"SOTA performances on 10 out of 13 tasks. We also demon-",
"the output of the classification layer for performance esti-",
"strate significantimprovementsacrossalldatasetsforagree-",
"mation is insufficient. From an uncertainty point of view,",
"ment based accuracy estimation when integrating our dis-",
"softmax outputs can been seen as capturing aleatoric un-",
"tance check. Ablation studies yield further insights in the",
"certainty, arising from overlapping class boundaries [24, 9].",
"method and its limitations. Finally, to the best of our knowl-",
"However, under dataset shifts, errors may also arise from",
"edge,we provide thefirst comprehensivepubliclyavailable",
"the fact that the model has never seen this type of input data",
"codebase of current SOTA baselines for accuracy estima-",
"anddoesnotknowhowtorespondto suchinputs.This is",
"tion, along with the complete code to reproduce our experi-",
"referred to as epistemic uncertainty [9] and is not well cap-",
"ments.",
"turedbysoftmaxoutputs[24,53l,asdemonstratedbyits",
"poor performance on the related out-of-distribution (OOD)",
"2.Background",
"detection task [40, 28, 48, 53, 30]. Note that in O0D de-",
"tection, the goal is to separate on separating ID from OOD",
"2.1. Performance estimation without ground truth",
"inputs, regardless of the downstream classification perfor-",
"Current methods for performance estimation under co-",
"mance, often considering inputs completely unrelated to the",
"variate shift can be broadly grouped in 4 categories:",
"task.This differs from performance estimation under co-",
"variate shift.wherewe assumethat the classification task",
"still applies to the shifted inputs and we focus on estimating",
"Estimating performance via auxiliary task performance",
"performance, not on detecting shifts.",
"Modifies the main classification model to incorporate a (suf-",
"ficiently informative)auxiliary task for which ground truth",
"labels are available at test time: accuracy on the main task",
"Methodological contributions In this paper, we argue",
"is then approximated by the computed accuracy on the aux-",
"that performance estimators should identify samples far",
"iliary task. For example, [11] trains a multi-task model for",
"away from the training set in the embedding space,for",
"predicting the class at hand as well as the rotation applied",
"which softmax estimates are mostlikelyunreliable.By",
"to the input image. The main limitation of this line of work",
"measuring the distance in the embedding space, we are able",
"is the requirement to build a multi-task model, making it",
"to measure how well the model “understood\" the sample",
"unusable as a post-hoc tool.",
"whenprojectingtheinputtotheclassificationspace.This",
"idea is illustrated in fig. I. Following this intuition, we pro-",
"posea simpleyeteffectivemethodtoimprovethequality",
"Training a regressor between ID and OOD accuracy",
"of current SOTA performance estimators.Specifically, we",
"This regressor can be trained based on model outputs or on",
"use nearest-neighbours distance in the embedding space to",
"measures of distance between datasets [13, 47, 16, 35, 12]",
"reject samples that lay too far from the training distribu-",
"Onemajordrawbackof this class of estimatorsis theirre-",
"tion. We then only use confidence-based performance esti-",
"quirement for having access to labelled OOD data for train-",
"mators on the remaining samples, considering all previously",
"ing the regression model.This is not always available",
"rejected samples as mis-classified. Our distance check ap-",
"in practice, in particular in data-scarce domains such that",
"proach is versatile and can be used to improve the quality of",
"healthcare. In absence of such OOD datasets,regressor are",
"various existing performance estimators (e.g. [15, 23]).",
"sometimes trained using corrupted versions of the original",
"validation setas\"OOD\"sets.However,this can notguaran-",
"tee the robustness of this estimator against other shifts e.g.",
"Main results We evaluate our approach on 13 classifica-",
"tion tasks ranging from cancer cell to animal classification.",
"natural subpopulation shift [46].",
"The natureof the distribution shifts studied covers a wide-",
"rangeof shifts:from syntheticcorruption,acquisition shift.",
"Agreement-based estimatorsAre based on the idea that",
"real-worldpopulation shifttorepresentation shift.Foreach",
"agreementbetweenmemberofmodelensemblescorrelate",
"task we evaluate between 18 and 259 models, covering vari-",
"with model accuracy. For example, generalised disagree-",
"ous training strategies and network architectures. These ex-",
"ment equality (GDE) [23] use pairs of models trained with",
"periments demonstrate that integrating distance into accu-",
"different random seeds to compute disagreement. Others",
"racy estimators significantly improves the quality of the es-",
"use more intricate methods for training specialised models",
"timation. For example, our proposed estimator ATC-DistCS",
"to align disagreement and accuracy further [10, 6]. How-",
"is significantly better than previous SOTA ATC [15] on all",
"ever, these procedures often require expensive additional",
"but one task, with a median relative MAE improvement of",
"training steps to derive the siblings models and are not ap-",
"30% across all tasks. Furthermore, comparing to the most",
"plicable to post-hoc scenarios where only the final model",
"recent COT method [34], we demonstrate a 27% median rel-",
"is available to the end user. In [I], the authors go as far as",
"ative performance improvement across all tasks, with new",
"training dozens models to fit a regressor between agreement"
] | {
"arxiv_id": "2308.07223",
"language": "en",
"page_idx": 1
} |
||
2308.07223-en-2 | 2308.07223-en-2.jpg | and accuracy, whereas [6] requires training a new ensemble for every single test set requiring performance estimation.
Confidence-based estimators These methods, contrarily
to the ones above, only require the final model's outputs to perform accuracy estimation and do not require any OOD data for calibration. As such, they are versatile and can be used with any classification model. For example, Difference of Confdence (DOC)[16] approximates the difference in accuracy between the evaluation set and the in-distribution (ID) validation set by the difference in average model confidence. ATC[15] introduces a confidence threshold such that all test samples for which the confidence is lower than this threshold are considered wrong and all samples meeting the minimum confidence requirement are considered correct (see Methods). Finally, concurrently to our work, COT [34] proposed to estimate accuracy based on based on optimal transport of model confidences. Precisely, they measure the Wasserstein distance between source label distribution and target softmax distribution to estimate the test error. Note, that this is expected to perform well if the source label distribution matches the target label distribution but might fail if this assumption breaks.
$2. 2. \textbf{ Distance- based out- of- distribution detection}$
The idea that OOD samples should lie far from the training samples in the embedding space is at the core of distance-based methods for OOD detection. For example, [30] propose to fit multi-variate Gaussians on the training embedding distribution of each class and use the Mahalanobis distance [36] to characterise how far test samples are from this expected distribution. If a sample is far from all class clusters, it is considered OOD. This method has shown some success at various OOD detection tasks [4,30] and extensions of this work have since further improved its capabilities [43]. However, this method suffers from one major limitation: it has a strong assumption that the class embeddings clusters can be accurately modelled by a Gaussian Multivariate distribution. Without any constraints on the training procedure or the embedding space at training time, this assumption may not hold [48]. This is the motivation for the work of [48] who proposed a non-parametric alternative OOD detection method. The authors still focus on the idea of using distances in the embedding space to detect OOD samples, but they leverage nearest-neighbours distances instead of the Mahalanobis distances, removing the normality assumption on the embedding. Precisely, they use the distance to the K$^{th}$ nearest-neighbour to classify samples as OOD. They derive the classification threshold for OOD versus ID task such that 95-99% of the training samples are classified as ID.
# 3. Methods
In this section, we begin by reminding the reader of the core principles of two base performance estimators which we build on top of: ATC[15] and GDE[23]. We then introduce our plug-in distance checker designed to flag untrustworthy samples, and discuss how to incorporate this distance check into these performance estimators to yield our proposed estimators “ATC-DistCS” and“GDE-DistCS”.
## 3.1. Base estimators
Average Thresholded Confidence (ATC)[15] approximates accuracy by the proportion of OOD predictions that do not exceed a certain confidence threshold (derived from the ID validation set, where confidence is defined as temperature-scaled[17] softmax confidence). Precisely, the threshold ATC is defined such that on source data $D_s$ the expected number of points that obtain a confidence less than ATC matches the error of the model on $D_{s}.$ This method has been further refined by [32], where the authors propose to apply class-wise temperature scaling and to define classwise confidence thresholds to improve the quality of the estimation, in particular for class-imbalanced problems.
Generalised Disagreement Equality (GDE)[23] Assuming access to two models $g$ and $g^\prime$ trained with different random seeds (but identical architecture and training paradigms), GDE estimates model accuracy by $\frac{1}{N}\sum_{i\in\mathrm{test~set}}[g(x_{i})=g^{\prime}(x_{i})]$,where N is the size of the OOD tat at at tho imuta tha madal and and $a(x_{n})$ danata ${\mathrm{OOD~test~set,~}x_{i}}$ the inputs to the model, and $g(x_{i})$ denotes the model prediction.
# 3.2. Integrating distance to training set
Average Distance Check Inspired by the OOD detection work in [48], we propose to improve standard performance estimators with a“distance checker”. Instead of simply rejecting samples with low confidence or model disagreement, we argue that the distance of any given sample to the in-distribution training set should also be taken into account to determine whether its confidence (and prediction) is likely to be trustworthy for estimation purposes or not. In simple terms, we “reject” samples whose penultimate-layer embeddings lie in a region “far”from the ID embedding space. The distance from a sample to the in-distribution set is determined by the average distance between the sample and all of its K-nearest-neighbours:
(1)
$$\mathrm{AD}_i=\frac{1}{K}\sum_k\left\|f_i-n_i^{(k)}\right\|_2,$$
where $K$ is the number of nearest-neighbours to consider, $f_{i}$ is the embedding of the $i^th$ test sample and $n_i^{(k)}$ the $k^{th}$ nearest neighbours to $f_i$ in the embedding space, nearest | [
[
"0.0809",
"0.09216",
"0.4697",
"0.09216",
"0.4697",
"0.1061",
"0.0809",
"0.1061"
],
[
"0.5024",
"0.0903",
"0.5986",
"0.0903",
"0.5986",
"0.1048",
"0.5024",
"0.1048"
],
[
"0.0801",
"0.108",
"0.46",
"0.108",
"0.46",
"0.1218",
"0.0801",
"0.1218"
],
[
"0.5195",
"0.11554",
"0.8906",
"0.11615",
"0.8906",
"0.1307",
"0.5195",
"0.13"
],
[
"0.5024",
"0.132",
"0.892",
"0.132",
"0.892",
"0.1459",
"0.5024",
"0.1459"
],
[
"0.5024",
"0.1465",
"0.8906",
"0.1465",
"0.8906",
"0.1604",
"0.5024",
"0.1604"
],
[
"0.0809",
"0.1604",
"0.468",
"0.1604",
"0.468",
"0.1736",
"0.0809",
"0.1736"
],
[
"0.5034",
"0.1622",
"0.8906",
"0.1622",
"0.8906",
"0.1755",
"0.5034",
"0.1755"
],
[
"0.0801",
"0.1755",
"0.468",
"0.1755",
"0.468",
"0.1895",
"0.0801",
"0.1895"
],
[
"0.5034",
"0.1768",
"0.8906",
"0.1768",
"0.8906",
"0.1901",
"0.5034",
"0.1901"
],
[
"0.0784",
"0.1901",
"0.468",
"0.1895",
"0.468",
"0.2039",
"0.0784",
"0.2052"
],
[
"0.5024",
"0.1925",
"0.8906",
"0.1925",
"0.8906",
"0.2058",
"0.5024",
"0.2058"
],
[
"0.0801",
"0.2058",
"0.4697",
"0.2058",
"0.4697",
"0.2197",
"0.0801",
"0.2197"
],
[
"0.501",
"0.207",
"0.8716",
"0.2058",
"0.8716",
"0.2203",
"0.501",
"0.2216"
],
[
"0.0809",
"0.2229",
"0.4673",
"0.2229",
"0.4673",
"0.2329",
"0.0809",
"0.2329"
],
[
"0.0801",
"0.2361",
"0.468",
"0.2361",
"0.468",
"0.25",
"0.0801",
"0.25"
],
[
"0.5034",
"0.2329",
"0.6587",
"0.2329",
"0.6587",
"0.2468",
"0.5034",
"0.2468"
],
[
"0.0801",
"0.2507",
"0.468",
"0.2507",
"0.468",
"0.2644",
"0.0801",
"0.2644"
],
[
"0.5024",
"0.2556",
"0.8267",
"0.2556",
"0.8267",
"0.2708",
"0.5024",
"0.2708"
],
[
"0.8267",
"0.2588",
"0.8896",
"0.2588",
"0.8896",
"0.2708",
"0.8267",
"0.2708"
],
[
"0.0825",
"0.2676",
"0.4673",
"0.2676",
"0.4673",
"0.2778",
"0.0825",
"0.2778"
],
[
"0.5024",
"0.2727",
"0.892",
"0.2727",
"0.892",
"0.286",
"0.5024",
"0.286"
],
[
"0.0801",
"0.281",
"0.4697",
"0.281",
"0.4697",
"0.2942",
"0.0801",
"0.2942"
],
[
"0.5024",
"0.2874",
"0.892",
"0.2874",
"0.892",
"0.3005",
"0.5024",
"0.3005"
],
[
"0.0801",
"0.2966",
"0.4697",
"0.2966",
"0.4697",
"0.31",
"0.0801",
"0.31"
],
[
"0.5024",
"0.3018",
"0.8906",
"0.3018",
"0.8906",
"0.315",
"0.5024",
"0.315"
],
[
"0.0801",
"0.3125",
"0.468",
"0.3125",
"0.468",
"0.3257",
"0.0801",
"0.3257"
],
[
"0.5024",
"0.3176",
"0.8906",
"0.3176",
"0.8906",
"0.3308",
"0.5024",
"0.3308"
],
[
"0.0801",
"0.327",
"0.468",
"0.327",
"0.468",
"0.3403",
"0.0801",
"0.3403"
],
[
"0.5034",
"0.334",
"0.886",
"0.334",
"0.886",
"0.3447",
"0.5034",
"0.3447"
],
[
"0.0801",
"0.3403",
"0.468",
"0.3416",
"0.468",
"0.356",
"0.0801",
"0.3547"
],
[
"0.5024",
"0.348",
"0.8906",
"0.348",
"0.8906",
"0.361",
"0.5024",
"0.361"
],
[
"0.0801",
"0.3567",
"0.4697",
"0.3567",
"0.4697",
"0.3706",
"0.0801",
"0.3706"
],
[
"0.5034",
"0.3623",
"0.892",
"0.3623",
"0.892",
"0.3757",
"0.5034",
"0.3757"
],
[
"0.0801",
"0.3718",
"0.4673",
"0.3718",
"0.4673",
"0.385",
"0.0801",
"0.385"
],
[
"0.5024",
"0.3774",
"0.892",
"0.3774",
"0.892",
"0.3914",
"0.5024",
"0.3914"
],
[
"0.0809",
"0.3882",
"0.4656",
"0.3882",
"0.4656",
"0.3984",
"0.0809",
"0.3984"
],
[
"0.5024",
"0.3933",
"0.8906",
"0.3933",
"0.8906",
"0.4072",
"0.5024",
"0.4072"
],
[
"0.0801",
"0.4016",
"0.4697",
"0.4016",
"0.4697",
"0.4153",
"0.0801",
"0.4153"
],
[
"0.5034",
"0.4077",
"0.8906",
"0.4077",
"0.8906",
"0.4216",
"0.5034",
"0.4216"
],
[
"0.0801",
"0.4167",
"0.4697",
"0.4167",
"0.4697",
"0.43",
"0.0801",
"0.43"
],
[
"0.5024",
"0.4236",
"0.8545",
"0.4236",
"0.8545",
"0.4375",
"0.5024",
"0.4375"
],
[
"0.0801",
"0.432",
"0.468",
"0.432",
"0.468",
"0.4458",
"0.0801",
"0.4458"
],
[
"0.0801",
"0.4475",
"0.3677",
"0.4475",
"0.3677",
"0.461",
"0.0801",
"0.461"
],
[
"0.5034",
"0.4583",
"0.8906",
"0.4583",
"0.8906",
"0.473",
"0.5034",
"0.473"
],
[
"0.501",
"0.4734",
"0.892",
"0.473",
"0.892",
"0.4873",
"0.501",
"0.488"
],
[
"0.0809",
"0.4817",
"0.4495",
"0.4817",
"0.4495",
"0.4956",
"0.0809",
"0.4956"
],
[
"0.5024",
"0.4893",
"0.892",
"0.4893",
"0.892",
"0.5024",
"0.5024",
"0.5024"
],
[
"0.0972",
"0.507",
"0.468",
"0.5083",
"0.468",
"0.523",
"0.0972",
"0.5215"
],
[
"0.5",
"0.5024",
"0.8936",
"0.502",
"0.8936",
"0.5195",
"0.5",
"0.521"
],
[
"0.0801",
"0.5244",
"0.4697",
"0.5244",
"0.4697",
"0.5386",
"0.0801",
"0.5386"
],
[
"0.509",
"0.5244",
"0.5215",
"0.5244",
"0.5215",
"0.531",
"0.509",
"0.531"
],
[
"0.513",
"0.5195",
"0.8936",
"0.5166",
"0.8936",
"0.534",
"0.513",
"0.537"
],
[
"0.0784",
"0.5386",
"0.468",
"0.539",
"0.468",
"0.5537",
"0.0784",
"0.553"
],
[
"0.5024",
"0.5327",
"0.892",
"0.534",
"0.892",
"0.549",
"0.5024",
"0.5474"
],
[
"0.0801",
"0.5547",
"0.4697",
"0.5547",
"0.4697",
"0.5684",
"0.0801",
"0.5684"
],
[
"0.5034",
"0.55",
"0.6445",
"0.55",
"0.6445",
"0.563",
"0.5034",
"0.563"
],
[
"0.0801",
"0.5693",
"0.468",
"0.5693",
"0.468",
"0.5835",
"0.0801",
"0.5835"
],
[
"0.5024",
"0.5737",
"0.8022",
"0.5757",
"0.8022",
"0.5903",
"0.5024",
"0.5884"
],
[
"0.0801",
"0.5854",
"0.4697",
"0.5854",
"0.4697",
"0.5986",
"0.0801",
"0.5986"
],
[
"0.0801",
"0.5996",
"0.468",
"0.5996",
"0.468",
"0.613",
"0.0801",
"0.613"
],
[
"0.5024",
"0.5977",
"0.892",
"0.5986",
"0.892",
"0.613",
"0.5024",
"0.6123"
],
[
"0.0809",
"0.6157",
"0.4697",
"0.6157",
"0.4697",
"0.629",
"0.0809",
"0.629"
],
[
"0.5034",
"0.616",
"0.8906",
"0.616",
"0.8906",
"0.627",
"0.5034",
"0.627"
],
[
"0.0801",
"0.63",
"0.468",
"0.63",
"0.468",
"0.643",
"0.0801",
"0.643"
],
[
"0.5024",
"0.63",
"0.8906",
"0.63",
"0.8906",
"0.643",
"0.5024",
"0.643"
],
[
"0.0809",
"0.6465",
"0.468",
"0.6465",
"0.468",
"0.657",
"0.0809",
"0.657"
],
[
"0.501",
"0.6445",
"0.8896",
"0.643",
"0.8896",
"0.6577",
"0.501",
"0.659"
],
[
"0.0801",
"0.66",
"0.4697",
"0.66",
"0.4697",
"0.674",
"0.0801",
"0.674"
],
[
"0.5024",
"0.66",
"0.892",
"0.66",
"0.892",
"0.674",
"0.5024",
"0.674"
],
[
"0.0809",
"0.677",
"0.468",
"0.677",
"0.468",
"0.6875",
"0.0809",
"0.6875"
],
[
"0.5024",
"0.675",
"0.8896",
"0.675",
"0.8896",
"0.688",
"0.5024",
"0.688"
],
[
"0.0801",
"0.69",
"0.468",
"0.69",
"0.468",
"0.704",
"0.0801",
"0.704"
],
[
"0.5034",
"0.6914",
"0.8906",
"0.6914",
"0.8906",
"0.702",
"0.5034",
"0.702"
],
[
"0.0801",
"0.705",
"0.468",
"0.705",
"0.468",
"0.7183",
"0.0801",
"0.7183"
],
[
"0.5024",
"0.705",
"0.8936",
"0.705",
"0.8936",
"0.7197",
"0.5024",
"0.7197"
],
[
"0.0801",
"0.7197",
"0.468",
"0.7197",
"0.468",
"0.7344",
"0.0801",
"0.7344"
],
[
"0.5034",
"0.7217",
"0.8906",
"0.7217",
"0.8906",
"0.732",
"0.5034",
"0.732"
],
[
"0.0809",
"0.7373",
"0.4673",
"0.7373",
"0.4673",
"0.7476",
"0.0809",
"0.7476"
],
[
"0.5024",
"0.735",
"0.8906",
"0.735",
"0.8906",
"0.7485",
"0.5024",
"0.7485"
],
[
"0.0809",
"0.7524",
"0.468",
"0.7524",
"0.468",
"0.763",
"0.0809",
"0.763"
],
[
"0.5024",
"0.7505",
"0.892",
"0.7505",
"0.892",
"0.7646",
"0.5024",
"0.7646"
],
[
"0.0809",
"0.7676",
"0.468",
"0.7676",
"0.468",
"0.778",
"0.0809",
"0.778"
],
[
"0.5024",
"0.7666",
"0.892",
"0.7666",
"0.892",
"0.78",
"0.5024",
"0.78"
],
[
"0.0801",
"0.782",
"0.4697",
"0.782",
"0.4697",
"0.7954",
"0.0801",
"0.7954"
],
[
"0.5024",
"0.781",
"0.738",
"0.781",
"0.738",
"0.7944",
"0.5024",
"0.7944"
],
[
"0.0801",
"0.797",
"0.468",
"0.797",
"0.468",
"0.81",
"0.0801",
"0.81"
],
[
"0.0784",
"0.81",
"0.468",
"0.811",
"0.468",
"0.8247",
"0.0784",
"0.8237"
],
[
"0.6045",
"0.8125",
"0.6562",
"0.8145",
"0.6543",
"0.83",
"0.603",
"0.8276"
],
[
"0.675",
"0.8145",
"0.7656",
"0.8125",
"0.7656",
"0.8276",
"0.675",
"0.83"
],
[
"0.8677",
"0.813",
"0.8945",
"0.813",
"0.8945",
"0.83",
"0.8677",
"0.83"
],
[
"0.0801",
"0.827",
"0.468",
"0.827",
"0.468",
"0.8403",
"0.0801",
"0.8403"
],
[
"0.653",
"0.8237",
"0.6733",
"0.8237",
"0.6733",
"0.832",
"0.653",
"0.832"
],
[
"0.773",
"0.827",
"0.7837",
"0.827",
"0.7837",
"0.83",
"0.773",
"0.83"
],
[
"0.0801",
"0.8403",
"0.4673",
"0.8403",
"0.4673",
"0.855",
"0.0801",
"0.855"
],
[
"0.0809",
"0.858",
"0.468",
"0.858",
"0.468",
"0.8687",
"0.0809",
"0.8687"
],
[
"0.5034",
"0.854",
"0.888",
"0.854",
"0.888",
"0.867",
"0.5034",
"0.867"
],
[
"0.0784",
"0.869",
"0.4697",
"0.872",
"0.4697",
"0.886",
"0.0784",
"0.884"
],
[
"0.501",
"0.869",
"0.8906",
"0.867",
"0.8906",
"0.8853",
"0.501",
"0.887"
],
[
"0.0801",
"0.887",
"0.267",
"0.887",
"0.267",
"0.901",
"0.0801",
"0.901"
],
[
"0.501",
"0.887",
"0.8906",
"0.887",
"0.8906",
"0.901",
"0.501",
"0.901"
]
] | [
"and accuracy, whereas [6] requires training a new ensemble",
"3.Methods",
"for every single test set requiring performance estimation.",
"In this section, we begin by reminding the reader of the",
"core principles of two base performance estimators which",
"We build on top of: ATC [15] and GDE [23]. We then intro-",
"Confidence-based estimatorsThese methods, contrarily",
"duce our plug-in distance checker designed to flag untrust-",
"to the ones above, only require the final model's outputs to",
"worthy samples, and discuss how to incorporate this dis-",
"perform accuracy estimation and do not require any OOD",
"tance check into these performance estimators to yield our",
"data for calibration. As such, they are versatile and can be",
"proposed estimators\"ATC-DistCS\" and \"GDE-DistCS\"",
"usedwithanyclassificationmodel.Forexample,Difference",
"of Confidence (DOC) [16] approximates the difference in",
"3.1. Base estimators",
"accuracy between the evaluation set and the in-distribution",
"Average Thresholded Confidence (ATC) [15]",
"approx-",
"(ID)validationsetbythedifferenceinaveragemodelcon-",
"imates accuracy by the proportion of OOD predictions",
"fidence. ATC [15] introduces a confidence threshold such",
"that do not exceed a certain confidence threshold (derived",
"that all test samples for which the confidence is lower than",
"from the ID validation set, where confidence is defined as",
"this threshold are considered wrong and all samples meet-",
"temperature-scaled [17] softmax confidence). Precisely, the",
"ing the minimum confidence requirement are considered",
"thresholdATCis defined suchthat on sourcedataDtheex",
"correct (see Methods). Finally, concurrently to our work,",
"pected number of points that obtain a confidence less than",
"COT [34] proposed to estimate accuracy based on based",
"ATC matches the error of the model on D,. This method",
"on optimal transport of model confidences. Precisely, they",
"has been further refined by [32], where the authors propose",
"measure the Wasserstein distancebetween sourcelabel dis-",
"to apply class-wise temperature scaling and to define class-",
"tribution and target softmax distribution to estimate the test",
"wise confidence thresholds to improve the quality of the es-",
"error. Note, that this is expected to perform well if the",
"timation, in particular for class-imbalanced problems.",
"source label distribution matches the target label distribu-",
"tion but might fail if this assumption breaks.",
"Generalised Disagreement Equality (GDE) [23] As-",
"suming access to two models g and g' trained with",
"2.2. Distance-based out-of-distribution detection",
"different random seeds (but identical architecture and",
"The idea that OOD samples should lie far from the",
"training paradigms), GDE estimates model accuracy by",
"training samples in the embedding space is at the core of",
"A",
"ietest set [g(ri) = g'(ri)], where N is the size of the",
"distance-based methods for OOD detection. For example,",
"OOD test set, &; the inputs to the model, and g(r) denotes",
"[30] propose to fit multi-variate Gaussians on the training",
"the model prediction.",
"embedding distribution of each class and use the Maha-",
"3.2. Integrating distance to training set",
"lanobis distance [36] to characterise how far test samples",
"are from this expected distribution. If a sample is far from",
"Average Distance Check Inspired by the OOD detection",
"all class clusters, it is considered OOD. This method has",
"workin[48l,weproposetoimprovestandardperformance",
"shown some success at various O0D detection tasks [4, 30]",
"estimators with a “distance checker\". Instead of simply",
"and extensionsofthisworkhavesincefurtherimprovedits",
"rejecting samples with low confidence or model disagree-",
"capabilities [43]. However, this method suffers from one",
"ment, we argue that the distance of any given sample to",
"major limitation: it has a strong assumption that the class",
"the in-distribution training set should also be taken into ac-",
"embeddings clusters can be accurately modelled by a Gaus-",
"countto determine whetherits confidence (and prediction)",
"sian Multivariate distribution. Without any constraints on",
"is likely to be trustworthy for estimation purposes or not. In",
"the training procedure or the embedding space at training",
"simple terms,we\"reject\"samples whose penultimate-layer",
"time.this assumption maynothold48l.This is the moti-",
"embeddings lie in a region “far\" from the ID embedding",
"vationfortheworkof[48lwhoproposedanon-parametric",
"space. The distance from a sample to the in-distribution set",
"alternativeOOD detectionmethod.Theauthorsstillfocus",
"is determined by the average distance between the sample",
"on the idea of using distances in the embedding space to",
"and all of its K-nearest-neighbours:",
"detect OOD samples, but they leverage nearest-neighbours",
"distances instead of the Mahalanobis distances, removing",
"AD; =",
"fi-n(k",
"(1)",
"the normality assumption on the embedding. Precisely, they",
"K",
"2",
"use the distance to the Kth nearest-neighbour to classify",
"samples as OOD.They derive the classification threshold",
"where K is the number of nearest-neighbours to consider",
"for O0D versus ID task such that 95-99% of the training",
"f; is the embedding of the ith test sample and n(k) the kth",
"samples are classified as ID.",
"nearest neighbours to fi in the embedding space, nearest"
] | {
"arxiv_id": "2308.07223",
"language": "en",
"page_idx": 2
} |
||
2204.12749-en-1 | 2204.12749-en-1.jpg | mainly focus on the sequential contextual information, ignoring the hierarchical relationships between the global cause, local psychological intention and dialog history, which leads to a weak ability of emotional support. An example in Figure 1 is illustrated to explain the above process. Globally, the system needs to capture the global cause (The help-seeker is anxious because of credit card bills and needs a second job) to enhance the semantic information of the whole conversation. Locally, the system should respond with the captured intention to understand the help-seeker's mental state during different turn. For example, in turn two, the help-seeker indicates that he/she wants to work more to earn more money. And the system generates an appropriate and encouraging response It will take dedication and hard work, but you can get more money. Hence, 1) how to capture the global cause and local psychological intentions and $\hat{2})$ how to model the relationships between them are two important challenges in the emotional support conversation.
To address these problems, we propose a novel Global-toLocal Hierarchical Graph Network (GLHG) for emotional support modeling. For the first issue, a multi-source encoder is presented to capture the global cause and local psychological intention (with COMET, a pre-trained generative commonsense reasoning model[Bosselut et al.,2019]). As for the second issue, a hierarchical graph reasoner is designed to model the hierarchical relationships between the global cause (dialog-level), local psychological intention (sentence-level) and dialog history. In addition, a novel training objective is designed in global-guide decoder to monitor semantic information of the global cause, so that the global cause will not be biased by the local intention.
The contributions can be summarized as follows:
$\cdot$ We propose a Global-to-Local Hierarchical Graph Network (GLHG) for emotional support conversation from the global-to-local perspective.
$\bullet$ To capture the global cause and local intention, the multisource encoder utilizes the information of situation and incorporates psychological intention with COMET.
$\bullet$ To model the different level relationships, the hierarchical graph reasoner makes an interaction between the global cause, local psychological intention and dialog history.
$\bullet$ Experiments on the dataset show that the GLHG achieves the state-of-the-art performance in terms of both automatic evaluation metrics and human evaluations.
## 2 $\textbf{Related Work}$
2.1 $\textbf{Emotional }\&$ Empathetic Conversation
In recent years, research on emotional conversation has embraced a booming in dialog systems [Zhou et al., 2018; Huang $etal.,2020;$ Peng $etal.,2022].$ [Zhou $etal.,2018]$ propose Emotional Chatting Machine to accurately generate emotional responses conditioning on a manually specified label. Different from the task of emotional chatting conversation, the task of empathetic dialog generation [Lin et al., 2019; Rashkin $etal.,2019]$ aim to respond to a speaker accordingly on the emotional situation in an empathetic way. Unlike the above tasks, emotional support conversation [Liu $etal.,2021]$
focuses on exploring the help-seeker's problems and generat-
ing more supportive responses.
# 2.2 Graph Modeling in Dialog
Motivated by impressive advancements and structured modeling of GCN[Guo et al., 2021; Sun et al., 2021], efforts towards better performance and construction utilizing GCNs have sprung up on dialog systems [Ghosal et al., 2019; Qin et al. , 2021; Xu et al. , 2021] . DialogueGCN [ Ghosal et al. , 2019] leverages self and inter-speaker dependency with GCN to model conversational context.[Qin et al., 2021] propose a Co-Interactive Graph Attention Network for dialog act recognition and sentiment classification task. DVAE-GNN [Xu et $al.,2021]$ focuses on discovering dialog structure in the opendomain dialogue, and EGAE[Qin et al., 2021] captures the dialogue schema in the task-oriented dialogue with the graph network. Unlike the previous work, we propose a Hierarchical Graph Reasoner for modeling the relationships between the global cause, local psychological intentions and dialog history, and to facilitate information flow over the graph.
2.3 Commonsense Knowledge
ATOMIC[Sap et al., 2019] is a commonsense knowledge[Yu $etal.,2020;$ Xing $etal.,2020]$ graph, which focuses on inferential knowledge organized as typed if-then relations with variables (e.g., if“PersonX drinks coffee", then“PersonX needed to brew coffee"). The type of relations contains IfEvent- Then- Mental- State, If- Event- Then- Event and If- Event- Then-Persona. In this paper, to capture the mental state of the help-seeker, we focus on the If-Event-Then-Mental-State which infers three commonsense relations: PersonX's reaction to the event $(xReact)$, PersonX's intention before the event $(xIntent)$ and PersonY's reaction to the event $(oReact).$ Since oReact refers to the other person (e.g., system), which is not considered, we ignore it in this paper. To obtain the psychological intention (xIntent) of the help-seeker, we employ COMET [Bosselut $etal.,2019]$, a pre-trained generative commonsense reasoning model, to generate rich commonsense descriptions in natural language (e.g., input “I need another job to pay more bills", then output “to be financially stable”).
# 3 $\mathbf{Approach}$
As shown in Figure 2, the proposed GLHG consists of three parts. Firstly, Multi-source Encoder captures the global cause information and local psychological intentions, as well as encoding the dialog history. Secondly, Hierarchical Graph Reasoner incorporates the different level representa- $\dot{\text{tions for graph reasoning from the global-to-local perspective.}}$ Finally, Global-guide Decoder generates the supportive responses and monitors semantic information of the global cause $\hat{\text{by the novel designed training objective.}}$
## 3.1 Problem Formulation
The problem formulation of our task could be formulated in the following. Given a dialog history $C=(u^1,u^2,\ldots,u^{N-1})$, a set of $N-1$ utterances, where $u^n=\left(x_{1}^{n},x_{2}^{n},\ldots,x_{M}^{n}\right)$ that consists of $M$ words, with $s=(s_1,s_2,\ldots,s_P)$ being the corresponding global cause information of the help-seeker | [
[
"0.0866",
"0.0682",
"0.4885",
"0.0682",
"0.4885",
"0.0821",
"0.0866",
"0.0821"
],
[
"0.513",
"0.06946",
"0.9126",
"0.06946",
"0.9126",
"0.0833",
"0.513",
"0.0833"
],
[
"0.0866",
"0.084",
"0.4885",
"0.084",
"0.4885",
"0.09784",
"0.0866",
"0.09784"
],
[
"0.513",
"0.084",
"0.7173",
"0.084",
"0.7173",
"0.09784",
"0.513",
"0.09784"
],
[
"0.0866",
"0.09784",
"0.486",
"0.09784",
"0.486",
"0.1111",
"0.0866",
"0.1111"
],
[
"0.511",
"0.1067",
"0.7563",
"0.1092",
"0.7563",
"0.1237",
"0.511",
"0.1218"
],
[
"0.0866",
"0.113",
"0.486",
"0.113",
"0.486",
"0.1237",
"0.0866",
"0.1237"
],
[
"0.08496",
"0.1256",
"0.4878",
"0.1256",
"0.4878",
"0.1395",
"0.08496",
"0.1395"
],
[
"0.5146",
"0.1276",
"0.9126",
"0.1276",
"0.9126",
"0.1414",
"0.5146",
"0.1414"
],
[
"0.08496",
"0.1395",
"0.4878",
"0.1395",
"0.4878",
"0.1528",
"0.08496",
"0.1528"
],
[
"0.513",
"0.1414",
"0.9126",
"0.1414",
"0.9126",
"0.1547",
"0.513",
"0.1547"
],
[
"0.08496",
"0.1528",
"0.4878",
"0.1528",
"0.4878",
"0.1666",
"0.08496",
"0.1666"
],
[
"0.513",
"0.1559",
"0.9126",
"0.1559",
"0.9126",
"0.1692",
"0.513",
"0.1692"
],
[
"0.08826",
"0.1686",
"0.4878",
"0.1686",
"0.4878",
"0.1787",
"0.08826",
"0.1787"
],
[
"0.513",
"0.1692",
"0.914",
"0.1692",
"0.914",
"0.1831",
"0.513",
"0.1831"
],
[
"0.08826",
"0.1818",
"0.4878",
"0.1818",
"0.4878",
"0.1925",
"0.08826",
"0.1925"
],
[
"0.513",
"0.1831",
"0.9126",
"0.1831",
"0.9126",
"0.1963",
"0.513",
"0.1963"
],
[
"0.0866",
"0.1957",
"0.486",
"0.1957",
"0.486",
"0.2058",
"0.0866",
"0.2058"
],
[
"0.513",
"0.1963",
"0.914",
"0.1963",
"0.914",
"0.2108",
"0.513",
"0.2108"
],
[
"0.0866",
"0.2102",
"0.486",
"0.2102",
"0.486",
"0.2203",
"0.0866",
"0.2203"
],
[
"0.513",
"0.2102",
"0.914",
"0.2102",
"0.914",
"0.2235",
"0.513",
"0.2235"
],
[
"0.08496",
"0.2229",
"0.4878",
"0.2229",
"0.4878",
"0.2361",
"0.08496",
"0.2361"
],
[
"0.511",
"0.2229",
"0.914",
"0.2247",
"0.914",
"0.238",
"0.511",
"0.2361"
],
[
"0.08496",
"0.2361",
"0.4878",
"0.2361",
"0.4878",
"0.25",
"0.08496",
"0.25"
],
[
"0.513",
"0.238",
"0.914",
"0.238",
"0.914",
"0.252",
"0.513",
"0.252"
],
[
"0.08826",
"0.252",
"0.4836",
"0.252",
"0.4836",
"0.262",
"0.08826",
"0.262"
],
[
"0.5146",
"0.254",
"0.911",
"0.254",
"0.911",
"0.2644",
"0.5146",
"0.2644"
],
[
"0.08496",
"0.2644",
"0.486",
"0.2644",
"0.486",
"0.2778",
"0.08496",
"0.2778"
],
[
"0.513",
"0.2664",
"0.9126",
"0.2664",
"0.9126",
"0.2798",
"0.513",
"0.2798"
],
[
"0.08826",
"0.2778",
"0.486",
"0.2778",
"0.486",
"0.2917",
"0.08826",
"0.2917"
],
[
"0.511",
"0.2798",
"0.9126",
"0.2798",
"0.9126",
"0.2935",
"0.511",
"0.2935"
],
[
"0.08826",
"0.2935",
"0.4844",
"0.2935",
"0.4844",
"0.3037",
"0.08826",
"0.3037"
],
[
"0.513",
"0.2935",
"0.9126",
"0.2935",
"0.9126",
"0.307",
"0.513",
"0.307"
],
[
"0.08496",
"0.3062",
"0.2303",
"0.3062",
"0.2303",
"0.3193",
"0.08496",
"0.3193"
],
[
"0.513",
"0.308",
"0.9126",
"0.308",
"0.9126",
"0.3213",
"0.513",
"0.3213"
],
[
"0.10297",
"0.3208",
"0.4885",
"0.3208",
"0.4885",
"0.334",
"0.10297",
"0.334"
],
[
"0.511",
"0.3208",
"0.914",
"0.3213",
"0.914",
"0.336",
"0.511",
"0.3352"
],
[
"0.0866",
"0.3333",
"0.4844",
"0.3333",
"0.4844",
"0.3467",
"0.0866",
"0.3467"
],
[
"0.511",
"0.3352",
"0.831",
"0.3352",
"0.831",
"0.3484",
"0.511",
"0.3484"
],
[
"0.08496",
"0.348",
"0.486",
"0.348",
"0.486",
"0.361",
"0.08496",
"0.361"
],
[
"0.0866",
"0.3623",
"0.4878",
"0.3623",
"0.4878",
"0.3757",
"0.0866",
"0.3757"
],
[
"0.511",
"0.358",
"0.755",
"0.3599",
"0.755",
"0.375",
"0.511",
"0.3726"
],
[
"0.08417",
"0.3738",
"0.4878",
"0.375",
"0.4878",
"0.3894",
"0.08417",
"0.3882"
],
[
"0.5146",
"0.379",
"0.911",
"0.379",
"0.911",
"0.3926",
"0.5146",
"0.3926"
],
[
"0.08496",
"0.3894",
"0.4878",
"0.3894",
"0.4878",
"0.4028",
"0.08496",
"0.4028"
],
[
"0.511",
"0.3914",
"0.9126",
"0.3933",
"0.9126",
"0.4077",
"0.511",
"0.406"
],
[
"0.0866",
"0.4048",
"0.486",
"0.4048",
"0.486",
"0.4153",
"0.0866",
"0.4153"
],
[
"0.513",
"0.4072",
"0.9126",
"0.4072",
"0.9126",
"0.4204",
"0.513",
"0.4204"
],
[
"0.08826",
"0.4185",
"0.4844",
"0.4185",
"0.4844",
"0.4287",
"0.08826",
"0.4287"
],
[
"0.5107",
"0.4204",
"0.9126",
"0.4192",
"0.9126",
"0.4343",
"0.5107",
"0.435"
],
[
"0.08496",
"0.43",
"0.4878",
"0.4312",
"0.4878",
"0.4443",
"0.08496",
"0.443"
],
[
"0.513",
"0.435",
"0.914",
"0.435",
"0.914",
"0.449",
"0.513",
"0.449"
],
[
"0.08826",
"0.4463",
"0.4878",
"0.4463",
"0.4878",
"0.457",
"0.08826",
"0.457"
],
[
"0.511",
"0.449",
"0.914",
"0.449",
"0.914",
"0.4622",
"0.511",
"0.4622"
],
[
"0.08826",
"0.4602",
"0.4878",
"0.4602",
"0.4878",
"0.4702",
"0.08826",
"0.4702"
],
[
"0.5146",
"0.4634",
"0.914",
"0.4634",
"0.914",
"0.4766",
"0.5146",
"0.4766"
],
[
"0.08826",
"0.473",
"0.486",
"0.473",
"0.486",
"0.486",
"0.08826",
"0.486"
],
[
"0.513",
"0.4766",
"0.9126",
"0.4766",
"0.9126",
"0.4905",
"0.513",
"0.4905"
],
[
"0.08826",
"0.488",
"0.2793",
"0.488",
"0.2793",
"0.4988",
"0.08826",
"0.4988"
],
[
"0.513",
"0.4905",
"0.9126",
"0.4905",
"0.9126",
"0.504",
"0.513",
"0.504"
],
[
"0.10455",
"0.502",
"0.4248",
"0.502",
"0.4248",
"0.515",
"0.10455",
"0.515"
],
[
"0.5146",
"0.506",
"0.9126",
"0.506",
"0.9126",
"0.5166",
"0.5146",
"0.5166"
],
[
"0.1054",
"0.5166",
"0.486",
"0.5166",
"0.486",
"0.53",
"0.1054",
"0.53"
],
[
"0.513",
"0.518",
"0.9126",
"0.518",
"0.9126",
"0.532",
"0.513",
"0.532"
],
[
"0.11926",
"0.531",
"0.4878",
"0.531",
"0.4878",
"0.5444",
"0.11926",
"0.5444"
],
[
"0.511",
"0.532",
"0.914",
"0.532",
"0.914",
"0.5454",
"0.511",
"0.5454"
],
[
"0.1177",
"0.5444",
"0.3203",
"0.5444",
"0.3203",
"0.558",
"0.1177",
"0.558"
],
[
"0.513",
"0.5454",
"0.914",
"0.5454",
"0.914",
"0.5596",
"0.513",
"0.5596"
],
[
"0.10455",
"0.558",
"0.486",
"0.558",
"0.486",
"0.5713",
"0.10455",
"0.5713"
],
[
"0.513",
"0.5596",
"0.9126",
"0.5596",
"0.9126",
"0.5728",
"0.513",
"0.5728"
],
[
"0.1177",
"0.5713",
"0.486",
"0.5713",
"0.486",
"0.5854",
"0.1177",
"0.5854"
],
[
"0.511",
"0.5713",
"0.9126",
"0.5728",
"0.9126",
"0.587",
"0.511",
"0.586"
],
[
"0.1177",
"0.586",
"0.4534",
"0.5854",
"0.4534",
"0.5986",
"0.1177",
"0.5996"
],
[
"0.513",
"0.587",
"0.9126",
"0.587",
"0.9126",
"0.601",
"0.513",
"0.601"
],
[
"0.10455",
"0.5996",
"0.486",
"0.5996",
"0.486",
"0.613",
"0.10455",
"0.613"
],
[
"0.513",
"0.601",
"0.9126",
"0.601",
"0.9126",
"0.6157",
"0.513",
"0.6157"
],
[
"0.1177",
"0.6143",
"0.4878",
"0.6123",
"0.4878",
"0.6255",
"0.1177",
"0.6274"
],
[
"0.513",
"0.6157",
"0.806",
"0.6157",
"0.806",
"0.629",
"0.513",
"0.629"
],
[
"0.11926",
"0.629",
"0.4763",
"0.629",
"0.4763",
"0.6396",
"0.11926",
"0.6396"
],
[
"0.10455",
"0.6416",
"0.486",
"0.6416",
"0.486",
"0.655",
"0.10455",
"0.655"
],
[
"0.517",
"0.6484",
"0.527",
"0.6484",
"0.527",
"0.654",
"0.517",
"0.654"
],
[
"0.539",
"0.6445",
"0.6274",
"0.6445",
"0.6274",
"0.659",
"0.539",
"0.659"
],
[
"0.1177",
"0.655",
"0.4878",
"0.655",
"0.4878",
"0.6685",
"0.1177",
"0.6685"
],
[
"0.513",
"0.664",
"0.9126",
"0.6636",
"0.9126",
"0.678",
"0.513",
"0.6787"
],
[
"0.11926",
"0.6704",
"0.437",
"0.6704",
"0.437",
"0.681",
"0.11926",
"0.681"
],
[
"0.511",
"0.6787",
"0.9126",
"0.678",
"0.9126",
"0.6924",
"0.511",
"0.6934"
],
[
"0.511",
"0.6934",
"0.9126",
"0.6934",
"0.9126",
"0.707",
"0.511",
"0.707"
],
[
"0.08496",
"0.7026",
"0.2345",
"0.7026",
"0.2345",
"0.7173",
"0.08496",
"0.7173"
],
[
"0.511",
"0.7056",
"0.911",
"0.7056",
"0.911",
"0.7197",
"0.511",
"0.7197"
],
[
"0.0866",
"0.728",
"0.4248",
"0.728",
"0.4248",
"0.742",
"0.0866",
"0.742"
],
[
"0.5146",
"0.7227",
"0.9126",
"0.7227",
"0.9126",
"0.733",
"0.5146",
"0.733"
],
[
"0.513",
"0.735",
"0.914",
"0.735",
"0.914",
"0.7485",
"0.513",
"0.7485"
],
[
"0.0866",
"0.7505",
"0.4885",
"0.7505",
"0.4885",
"0.7646",
"0.0866",
"0.7646"
],
[
"0.513",
"0.7485",
"0.9126",
"0.7485",
"0.9126",
"0.762",
"0.513",
"0.762"
],
[
"0.08496",
"0.765",
"0.4878",
"0.765",
"0.4878",
"0.779",
"0.08496",
"0.779"
],
[
"0.513",
"0.763",
"0.914",
"0.763",
"0.914",
"0.7764",
"0.513",
"0.7764"
],
[
"0.0866",
"0.779",
"0.486",
"0.779",
"0.486",
"0.7925",
"0.0866",
"0.7925"
],
[
"0.513",
"0.776",
"0.7793",
"0.776",
"0.7793",
"0.789",
"0.513",
"0.789"
],
[
"0.0866",
"0.7925",
"0.486",
"0.7925",
"0.486",
"0.806",
"0.0866",
"0.806"
],
[
"0.0866",
"0.806",
"0.4878",
"0.806",
"0.4878",
"0.8193",
"0.0866",
"0.8193"
],
[
"0.513",
"0.8003",
"0.721",
"0.8003",
"0.721",
"0.8145",
"0.513",
"0.8145"
],
[
"0.08905",
"0.8213",
"0.4878",
"0.8213",
"0.4878",
"0.832",
"0.08905",
"0.832"
],
[
"0.5146",
"0.8193",
"0.9126",
"0.8193",
"0.9126",
"0.8335",
"0.5146",
"0.8335"
],
[
"0.0866",
"0.834",
"0.4844",
"0.834",
"0.4844",
"0.8477",
"0.0866",
"0.8477"
],
[
"0.5107",
"0.832",
"0.914",
"0.8306",
"0.914",
"0.8486",
"0.5107",
"0.8496"
],
[
"0.0866",
"0.8477",
"0.4844",
"0.8477",
"0.4844",
"0.8613",
"0.0866",
"0.8613"
],
[
"0.511",
"0.8477",
"0.9126",
"0.8477",
"0.9126",
"0.8623",
"0.511",
"0.8623"
],
[
"0.0866",
"0.863",
"0.4878",
"0.863",
"0.4878",
"0.8735",
"0.0866",
"0.8735"
],
[
"0.511",
"0.863",
"0.91",
"0.863",
"0.91",
"0.8735",
"0.511",
"0.8735"
],
[
"0.08496",
"0.8755",
"0.4844",
"0.8755",
"0.4844",
"0.8896",
"0.08496",
"0.8896"
],
[
"0.511",
"0.8755",
"0.9126",
"0.8755",
"0.9126",
"0.8896",
"0.511",
"0.8896"
]
] | [
"mainly focus on the sequential contextual information, ignor-",
"focuses on exploring the help-seeker's problems and generat-",
"ing the hierarchical relationships between the global cause,",
"ing more supportive responses.",
"local psychological intention and dialog history, which leads",
"2.2Graph Modeling in Dialog",
"toaweakabilityof emotional support.AnexampleinFigure",
"1 is illustrated to explain the above process. Globally, the",
"Motivated by impressive advancements and structured mod-",
"system needs to capture the global cause (The help-seeker is",
"eling of GCN [Guo et al., 2021; Sun et al., 2021], efforts",
"anxious because of credit card bills and needs a second job) to",
"towards better performance and construction utilizing GCNs",
"enhancethesemanticinformationofthewholeconversation.",
"have sprung up on dialog systems [Ghosal et al., 2019;",
"Locally,thesystemshouldrespond withthecapturedintention",
"Qin et al., 2021; Xu et al., 2021]. DialogueGCN [Ghosal et al.,",
"to understandthehelp-seeker'smental state duringdifferent",
"2019] leverages self and inter-speaker dependency with GCN",
"turn.For example,in turn two,the help-seekerindicates that",
"to model conversational context. [Qin et al., 2021l propose a",
"he/she wants to work more to earn more money.And the",
"Co-Interactive Graph Attention Network for dialog act recog-",
" system generates an appropriate and encouraging response",
"nition and sentiment classification task. DVAE-GNN [Xu et",
"Itwill take dedication and hard work, butyou can get more",
"al.,2021lfocusesondiscoveringdialogstructureintheopen",
"money. Hence, 1) how to capture the global cause and local",
"domain dialogue, and EGAE [Qin et al., 2021l captures the",
"psychological intentions and 2) how to model the relationships",
"dialogue schema in the task-oriented dialogue with the graph",
"betweenthemaretwoimportantchallengesintheemotional",
"network. Unlike the previous work, we propose a Hierarchical",
"support conversation.",
"Graph Reasoner for modeling the relationships between the",
"To address these problems, we propose a novel Global-to-",
"global cause, local psychological intentions and dialog history,",
"Local Hierarchical Graph Network (GLHG) for emotional",
"and to facilitate information flow over the graph.",
" support modeling. For the first issue, a multi-source encoder is",
"presented to capture the global cause and local psychological",
"2.3CommonsenseKnowledge",
"intention (with COMET, a pre-trained generative common-",
"ATOMIC [Sap et al., 2019] is a commonsense knowledge [Yu",
"sense reasoning model [Bosselut et al., 2019]). As for the",
"et al., 2020; Xing et al., 2020l graph, which focuses on in-",
"second issue, a hierarchical graph reasoner is designed to",
"ferential knowledge organized as typed if-then relations with",
"modelthehierarchicalrelationshipsbetweentheglobalcause",
"variables (e.g,if\"PersonX drinks coffee\",then\"Personx",
"(dialog-level),local psychological intention (sentence-level)",
"needed to brew coffee\"). The type of relations contains If-",
"and dialoghistory.In addition,anovel training objectiveis",
"Event-Then-Mental-State, If-Event-Then-Event and If-Event-",
"designedinglobal-guidedecodertomonitor semanticinfor-",
"Then-Persona. In this paper, to capture the mental state of",
"mation of the global cause, so that the global cause will not be",
"the help-seeker, we focus on the If-Event-Then-Mental-State",
"biasedbythelocal intention.",
"which infers three commonsense relations: PersonX's reaction",
"The contributions can be summarized as follows:",
"tothe event (xReact),PersonX's intention beforetheevent",
": We propose a Global-to-Local Hierarchical Graph Net-",
"(xlntent) and PersonY's reaction to the event (oReact). Since",
"work (GLHG) for emotional support conversation from",
"oReact refers to the other person (e.g., system), which is not",
"the global-to-local perspective.",
"considered, we ignore it in this paper. To obtain the psycholog-",
": To capture the global cause and local intention, the multi-",
"ical intention (xIntent) of the help-seeker, we employ COMET",
"source encoder utilizes the information of situation and",
"[Bosselut et al., 2019], a pre-trained generative commonsense",
"incorporatespsychological intentionwithCOMET",
"reasoning model, to generate rich commonsense descriptions",
". To model the different level relationships, the hierarchical",
"in natural language (e.g., input \"I need another job to pay more",
"graph reasonermakes an interaction between the global",
"bills\", then output \"to be financially stable\")",
"cause,localpsychologicalintentionanddialoghistory",
". Experiments on the dataset show that the GLHG achieves",
"1",
"Approach",
"the state-of-the-art performance in terms of both auto-",
"As shown in Figure 2, the proposed GLHG consists of three",
"matic evaluationmetrics and human evaluations.",
"parts. Firstly, Multi-source Encoder captures the global",
"cause information and local psychological intentions, as",
"2Related Work",
"well as encoding the dialog history. Secondly, Hierarchical",
"2.1Emotional &Empathetic Conversation",
"GraphReasonerincorporatesthedifferentlevelrepresenta-",
"tions for graph reasoning from the global-to-local perspective.",
"In recent years, research on emotional conversation has em-",
"Finally, Global-guide Decoder generates the supportive re-",
"braced a booming in dialog systems [Zhou et al., 2018;",
"sponses and monitors semantic information of the global cause",
"Huang et al., 2020; Peng et al., 2022]. [Zhou et al., 2018]",
"by the novel designed training objective.",
"propose Emotional Chatting Machine to accurately generate",
"emotional responses conditioning on a manually specified",
"3.1 Problem Formulation",
"label.Differentfromthetaskofemotionalchattingconversa-",
"The problem formulation of our task could be formulated in the",
"tion, the task of empathetic dialog generation [Lin et al., 2019:",
"following. Given a dialog history C = (u', u?, .., uN-1), a",
"Rashkin et al., 2019] aim to respond to a speaker accordingly",
"set of N - I utterances, where un = (n, r2,..., rm) that",
"on the emotional situation in an empathetic way.Unlike the",
"consists of M words, with s = (s1,s2....,sp)being the",
"above tasks, emotional support conversation [Liu et al., 2021]",
"corresponding global cause information of the help-seeker"
] | {
"arxiv_id": "2204.12749",
"language": "en",
"page_idx": 1
} |
||
2204.12749-en-2 | 2204.12749-en-2.jpg | <FigureHere>
Figure 2: The overview of GLHG, which consists of Multi-source Encoder, Hierarchical Graph Reasoner and Global-guide Decoder. Green
box indicates the pre-trained generative commonsense reasoning model COMET.
(describes the cause of the emotional problem), a sequence of $P$ words. GLHG generates a supportive response $Y$ conditioned on the help-seeker's global cause, local psychological intention and dialog history.
with the extra special relation tokens (e.g.,“〈head〉I need another job to pay more bills $\langle$ /head > / relation > xIntent $\langle$ /relation > [GEN]") and generates corresponding mental state inferences $(ms_1,\ldots,ms_L)$ for XIntent relation, where $L$ is the length of inferences sequence. Similarly, the module $\mathrm{Enc}_{loc}$ obtains the representation of the local intention by:
## 3.2 $\textbf{Multi- Source Encoder}$
(3)
$$l=\text{Max-pooling}(\mathrm{Enc}_{loc}(ms_1,\ldots,ms_L))$$
where $l\in\mathbb{R}^d.$ (Note, We only use XIntent in this paper).
## 33 Hierarchical Graph Reasoner
Multi-Source Encoder considers three sources information, including dialog history, global cause and local intention. The contextual encoder, global encoder and local encoder have the same architecture with the encoder of BlenderBot [Roller $et$ $al. , 2021] $, which is an open-domain conversational agent pre-trained on large-scale dialog corpora.
Firstly, representations of dialog history are obtained. Following the previous work [Liu et al., 2021; Roller et al., 2021], we separate each utterance with [SEP] and prepend a special token [CLS] at the beginning of the dialog history to reconstruct the context input $\tilde{C}=$ $([CLS],\breve{u}^{1},[SE\tilde{P}],\ldots,u^{N-1},[SEP]).$To obtain the contextual representations, the contextual encoder Enc$_ctx$ is utilized to encode each word $x$, leading to a series of contextaware hidden states $(\boldsymbol{h}_1,\boldsymbol{h}_2,\ldots,\boldsymbol{h}_T)$,as:
The graph reasoner is proposed to organize the multi-source information and model the interactions between the global cause, local intention and dialog history, which enables the model to better build situational understanding of the conversation
Following [Scarselli $et\textit{al. , 2009] , after obtaining the repre- }$ sentations of the global cause, local psychological intentions and dialog history, we build a graph network [Scarselli $et$ $al. $, 2009] by connecting nodes with edges. And then we integrate relevant information for each graph node with a graph attention network (GAT) [Velickovic et $al.,2017]$, which propagates features from other neighbourhood's information to the current node and has the advantage of determining the importance and relevance between nodes. In this section, we first define the components of our graph $\mathcal{G}=(\mathcal{V},\mathcal{E})$ and then introduce the reasoning process.
Vertices. The representation of the global cause $g$, each token $h_t$ in the dialog history and local intention $l$ are represented as the vertices in graph $\mathcal{G}$, which are initialized with the corresponding encoded feature vectors. We represent the vertex features as $\mathcal{V}=\{\boldsymbol{g},\boldsymbol{h}_1,\ldots,\boldsymbol{h}_T,\boldsymbol{l}\}~=$
(1)
$$\boldsymbol{h}_{t}=\mathrm{Enc}_{ctx}([CLS],x_{1}^{1},\ldots,x_{M}^{1},[SEP],\ldots,[SEP])$$
where $T$ is the max length of the input sequence, $\boldsymbol h_t\in\mathbb{R}^d$ is
the $t$-th token in the $C$, and $d$ is the hidden size of the encoder. Then, the global encoder Enc$_{glo}$ obtains the global cause with the situation information, the max-pooling operation is performed to output representation of the whole sequence:
(2)
$$g=\mathrm{Max-pooling}(\mathrm{Enc}_{glo}(s_{1},\ldots,s_{P}))$$
Finally, to capture the psychological intention of the helpseeker, COMET utilizes the last utterance of the help-seeker | [
[
"0.7344",
"0.0821",
"0.7563",
"0.0821",
"0.7563",
"0.089",
"0.7344",
"0.089"
],
[
"0.7715",
"0.0852",
"0.7876",
"0.0852",
"0.7876",
"0.0884",
"0.7715",
"0.0884"
],
[
"0.846",
"0.0852",
"0.864",
"0.0852",
"0.864",
"0.089",
"0.846",
"0.089"
],
[
"0.7607",
"0.1042",
"0.7817",
"0.1042",
"0.7817",
"0.108",
"0.7607",
"0.108"
],
[
"0.8022",
"0.1042",
"0.831",
"0.1042",
"0.831",
"0.108",
"0.8022",
"0.108"
],
[
"0.7563",
"0.1111",
"0.7646",
"0.1111",
"0.7646",
"0.11426",
"0.7563",
"0.11426"
],
[
"0.773",
"0.125",
"0.8203",
"0.125",
"0.8203",
"0.1364",
"0.773",
"0.1364"
],
[
"0.746",
"0.1572",
"0.7817",
"0.1572",
"0.7817",
"0.161",
"0.746",
"0.161"
],
[
"0.794",
"0.1572",
"0.8364",
"0.1572",
"0.8364",
"0.161",
"0.794",
"0.161"
],
[
"0.7715",
"0.1768",
"0.8203",
"0.1768",
"0.8203",
"0.1875",
"0.7715",
"0.1875"
],
[
"0.1136",
"0.1831",
"0.1177",
"0.1818",
"0.1209",
"0.1901",
"0.1168",
"0.1913"
],
[
"0.7686",
"0.207",
"0.8228",
"0.207",
"0.8228",
"0.2184",
"0.7686",
"0.2184"
],
[
"0.7715",
"0.2291",
"0.822",
"0.2291",
"0.822",
"0.2393",
"0.7715",
"0.2393"
],
[
"0.761",
"0.2556",
"0.83",
"0.2556",
"0.83",
"0.2664",
"0.761",
"0.2664"
],
[
"0.1969",
"0.3516",
"0.2607",
"0.3516",
"0.2607",
"0.3623",
"0.1969",
"0.3623"
],
[
"0.2688",
"0.351",
"0.282",
"0.3547",
"0.2778",
"0.363",
"0.2646",
"0.3591"
],
[
"0.1642",
"0.3706",
"0.2688",
"0.3706",
"0.2688",
"0.3806",
"0.1642",
"0.3806"
],
[
"0.4412",
"0.3726",
"0.562",
"0.3726",
"0.562",
"0.3801",
"0.4412",
"0.3801"
],
[
"0.7417",
"0.3726",
"0.8447",
"0.3726",
"0.8447",
"0.3801",
"0.7417",
"0.3801"
],
[
"0.0866",
"0.3914",
"0.9126",
"0.3914",
"0.9126",
"0.4048",
"0.0866",
"0.4048"
],
[
"0.08417",
"0.4048",
"0.569",
"0.404",
"0.569",
"0.4185",
"0.08417",
"0.4192"
],
[
"0.0866",
"0.443",
"0.4878",
"0.443",
"0.4878",
"0.457",
"0.0866",
"0.457"
],
[
"0.513",
"0.443",
"0.9126",
"0.443",
"0.9126",
"0.457",
"0.513",
"0.457"
],
[
"0.0866",
"0.457",
"0.4878",
"0.457",
"0.4878",
"0.4702",
"0.0866",
"0.4702"
],
[
"0.513",
"0.4583",
"0.9126",
"0.4583",
"0.9126",
"0.4717",
"0.513",
"0.4717"
],
[
"0.08496",
"0.4717",
"0.4878",
"0.4717",
"0.4878",
"0.4849",
"0.08496",
"0.4849"
],
[
"0.513",
"0.4717",
"0.9126",
"0.4717",
"0.9126",
"0.4849",
"0.513",
"0.4849"
],
[
"0.08417",
"0.4841",
"0.2712",
"0.4849",
"0.2712",
"0.5",
"0.08417",
"0.4988"
],
[
"0.511",
"0.4849",
"0.9126",
"0.4849",
"0.9126",
"0.4988",
"0.511",
"0.4988"
],
[
"0.08496",
"0.506",
"0.2974",
"0.507",
"0.2974",
"0.5215",
"0.08496",
"0.521"
],
[
"0.5146",
"0.5005",
"0.911",
"0.5005",
"0.911",
"0.511",
"0.5146",
"0.511"
],
[
"0.513",
"0.513",
"0.8945",
"0.513",
"0.8945",
"0.5264",
"0.513",
"0.5264"
],
[
"0.08496",
"0.5234",
"0.4878",
"0.5244",
"0.4878",
"0.539",
"0.08496",
"0.5386"
],
[
"0.08496",
"0.539",
"0.4878",
"0.539",
"0.4878",
"0.553",
"0.08496",
"0.553"
],
[
"0.5645",
"0.5405",
"0.8594",
"0.5425",
"0.8594",
"0.5566",
"0.5645",
"0.5547"
],
[
"0.8906",
"0.5425",
"0.911",
"0.5425",
"0.911",
"0.5547",
"0.8906",
"0.5547"
],
[
"0.08496",
"0.5537",
"0.486",
"0.5537",
"0.486",
"0.5674",
"0.08496",
"0.5674"
],
[
"0.5107",
"0.56",
"0.8813",
"0.561",
"0.8813",
"0.5757",
"0.5107",
"0.5747"
],
[
"0.08826",
"0.5684",
"0.4844",
"0.5684",
"0.4844",
"0.579",
"0.08826",
"0.579"
],
[
"0.08496",
"0.5806",
"0.4885",
"0.5806",
"0.4885",
"0.5947",
"0.08496",
"0.5947"
],
[
"0.5107",
"0.5835",
"0.785",
"0.5854",
"0.785",
"0.603",
"0.5107",
"0.601"
],
[
"0.08496",
"0.5947",
"0.3555",
"0.5933",
"0.3555",
"0.607",
"0.08496",
"0.608"
],
[
"0.10297",
"0.608",
"0.4885",
"0.608",
"0.4885",
"0.622",
"0.10297",
"0.622"
],
[
"0.511",
"0.605",
"0.911",
"0.605",
"0.911",
"0.6187",
"0.511",
"0.6187"
],
[
"0.08496",
"0.622",
"0.4885",
"0.622",
"0.4885",
"0.636",
"0.08496",
"0.636"
],
[
"0.5146",
"0.622",
"0.911",
"0.622",
"0.911",
"0.632",
"0.5146",
"0.632"
],
[
"0.0866",
"0.637",
"0.4844",
"0.637",
"0.4844",
"0.648",
"0.0866",
"0.648"
],
[
"0.513",
"0.634",
"0.914",
"0.634",
"0.914",
"0.648",
"0.513",
"0.648"
],
[
"0.08496",
"0.6494",
"0.486",
"0.6494",
"0.486",
"0.6636",
"0.08496",
"0.6636"
],
[
"0.511",
"0.6484",
"0.888",
"0.6484",
"0.888",
"0.659",
"0.511",
"0.659"
],
[
"0.08496",
"0.6636",
"0.4656",
"0.662",
"0.4656",
"0.677",
"0.08496",
"0.678"
],
[
"0.527",
"0.66",
"0.914",
"0.661",
"0.914",
"0.6753",
"0.527",
"0.675"
],
[
"0.2573",
"0.6753",
"0.4878",
"0.677",
"0.4878",
"0.69",
"0.2573",
"0.6895"
],
[
"0.513",
"0.6753",
"0.9126",
"0.6753",
"0.9126",
"0.6895",
"0.513",
"0.6895"
],
[
"0.2133",
"0.681",
"0.2646",
"0.678",
"0.2656",
"0.688",
"0.2141",
"0.6914"
],
[
"0.08496",
"0.677",
"0.2197",
"0.6753",
"0.2197",
"0.69",
"0.08496",
"0.6914"
],
[
"0.0866",
"0.6924",
"0.4885",
"0.6924",
"0.4885",
"0.7056",
"0.0866",
"0.7056"
],
[
"0.513",
"0.69",
"0.914",
"0.69",
"0.914",
"0.704",
"0.513",
"0.704"
],
[
"0.08496",
"0.7056",
"0.4902",
"0.7056",
"0.4902",
"0.7197",
"0.08496",
"0.7197"
],
[
"0.5146",
"0.705",
"0.9126",
"0.705",
"0.9126",
"0.7153",
"0.5146",
"0.7153"
],
[
"0.08496",
"0.7197",
"0.3635",
"0.7197",
"0.3635",
"0.7344",
"0.08496",
"0.7344"
],
[
"0.5146",
"0.7173",
"0.9126",
"0.7173",
"0.9126",
"0.731",
"0.5146",
"0.731"
],
[
"0.513",
"0.731",
"0.9126",
"0.731",
"0.9126",
"0.744",
"0.513",
"0.744"
],
[
"0.5146",
"0.744",
"0.914",
"0.744",
"0.914",
"0.7583",
"0.5146",
"0.7583"
],
[
"0.0948",
"0.7505",
"0.486",
"0.7505",
"0.486",
"0.765",
"0.0948",
"0.765"
],
[
"0.5146",
"0.7603",
"0.911",
"0.7603",
"0.911",
"0.771",
"0.5146",
"0.771"
],
[
"0.08496",
"0.768",
"0.486",
"0.7676",
"0.486",
"0.782",
"0.08496",
"0.7827"
],
[
"0.511",
"0.773",
"0.914",
"0.773",
"0.914",
"0.786",
"0.511",
"0.786"
],
[
"0.0866",
"0.7827",
"0.4885",
"0.7827",
"0.4885",
"0.797",
"0.0866",
"0.797"
],
[
"0.511",
"0.786",
"0.914",
"0.785",
"0.914",
"0.8",
"0.511",
"0.8003"
],
[
"0.10297",
"0.7974",
"0.4878",
"0.7974",
"0.4878",
"0.811",
"0.10297",
"0.811"
],
[
"0.511",
"0.8003",
"0.726",
"0.8003",
"0.726",
"0.8145",
"0.511",
"0.8145"
],
[
"0.08496",
"0.811",
"0.486",
"0.811",
"0.486",
"0.8247",
"0.08496",
"0.8247"
],
[
"0.5146",
"0.8193",
"0.9126",
"0.8193",
"0.9126",
"0.8335",
"0.5146",
"0.8335"
],
[
"0.08496",
"0.8257",
"0.468",
"0.8257",
"0.468",
"0.839",
"0.08496",
"0.839"
],
[
"0.513",
"0.834",
"0.9126",
"0.834",
"0.9126",
"0.8477",
"0.513",
"0.8477"
],
[
"0.152",
"0.8433",
"0.4224",
"0.8433",
"0.4224",
"0.858",
"0.152",
"0.858"
],
[
"0.464",
"0.845",
"0.4844",
"0.845",
"0.4844",
"0.856",
"0.464",
"0.856"
],
[
"0.5107",
"0.8477",
"0.9126",
"0.8477",
"0.9126",
"0.8613",
"0.5107",
"0.8613"
],
[
"0.10297",
"0.8613",
"0.486",
"0.8613",
"0.486",
"0.875",
"0.10297",
"0.875"
],
[
"0.5107",
"0.8613",
"0.9126",
"0.8623",
"0.9126",
"0.8755",
"0.5107",
"0.875"
],
[
"0.0866",
"0.877",
"0.4878",
"0.877",
"0.4878",
"0.887",
"0.0866",
"0.887"
],
[
"0.511",
"0.8755",
"0.911",
"0.8755",
"0.911",
"0.89",
"0.511",
"0.89"
]
] | [
"<B>",
"",
"3",
"utmn",
"oddin",
"",
"Add&Norm",
"en-to-noo",
"erossAtte",
"Add&Norm",
"",
"Feed Forward",
"Add& Norm",
"Linear&Soffmax",
"[SEP]Token",
"M",
"Multi-Source Encoder",
"HierarchicalGraphReaso",
"Global-guideDecoder",
"Figure 2: The overview of GLHG, which consists of Multi-source Encoder, Hierarchical Graph Reasoner and Global-guide Decoder. Green",
"box indicates the pre-trained generative commonsense reasoning model COMET.",
"(describes the cause of the emotional problem), a sequence",
"with the extra special relation tokens (e.g., \"( head ) I need",
"of P words. GLHG generates a supportive response Y condi-",
"another job to pay more bills (/head )( relation ) xlntent",
"tioned on the help-seeker's global cause, local psychological",
"《/relation】[GEN]\")and generates corresponding mental",
"intention and dialog history.",
"state inferences (msi, ..:,msL)for XIntent relation, where",
"3.2 Multi-Source Encoder",
"L is the length of inferences sequence.Similarly,the module",
"Encloc obtains the representation of the local intention by:",
"Multi-Source Encoder considers three sources information,",
"including dialog history, global cause and local intention. The",
"l=Max-pooling(Encloc(ms1,...,msL))",
"(3)",
"contextual encoder, global encoder and local encoder have",
"where I e IRd. (Note, We only use XIntent in this paper)",
"thesamearchitecturewiththeencoderofBlenderBot[Roller",
"et al., 2021l, which is an open-domain conversational agent",
"3.3 Hierarchical Graph Reasoner",
"pre-trained on large-scale dialog corpora.",
"Firstly, representations of dialog history are obtained.",
"The graph reasoner is proposed to organize the multi-source in.",
"Following the previous work [Liu et al., 202; Roller",
"formationandmodeltheinteractionsbetweentheglobalcause",
"etal.,2021],weseparateeachutterancewith[SEP",
"local intention and dialog history, which enables the model to",
"and prepend a special token [CLS] at the beginning of",
"betterbuild situational understandingoftheconversation.",
"the dialog history to reconstruct the context input C",
"Following [Scarselli et al., 2009], after obtaining the repre-",
"N-1.[SEP]).Toobtainthecon-",
"sentations of the global cause, local psychological intentions",
".....u4",
"(ICLS], u, [SEP]",
"textual representations, the contextual encoder Encctr is uti-",
"and dialog history, we build a graph network [Scarselli et al.,",
"lized to encode each word c, leading to a series of context-",
"2009lbyconnectingnodeswithedges.Andthenweinte-",
"aware hidden states (hi, h2, ..., hr), as:",
"grate relevant information for each graph node with a graph",
"attention network (GAT) [Velickovic et al., 2017], which prop-",
"agates features from other neighbourhood's information to",
"ht = Encctr([CLSl, ri,...,&M,[SEP,...,[SEP) ()",
"the current node and has the advantage of determining the",
"where T is the max length of the input sequence, ht E Rd is",
"importance and relevance between nodes. In this section, we",
"the t-th token in the C, and d is the hidden size of the encoder.",
"first define the components of our graph g = (V, ) and then",
"Then, the global encoder Encglo obtains the global cause",
"introduce the reasoning process.",
"with the situation information, the max-pooling operation is",
"Vertices.The representation of the global cause g, each",
"performed to output representation of the whole sequence:",
"token ht in the dialog history and local intention I are rep-",
"g = Max-pooling(Encglo(s1,..., $p))",
"(2)",
"resented as the vertices in graph G, which are initialized",
"Finally, to capture the psychological intention of the help-",
"with the corresponding encoded feature vectors.We rep-",
"seeker.COMETutilizesthelastutteranceofthehelp-seeker",
"resent the vertex features as V = (g,hi,...,h,l) -"
] | {
"arxiv_id": "2204.12749",
"language": "en",
"page_idx": 2
} |
||
2204.12749-en-3 | 2204.12749-en-3.jpg | $\{\boldsymbol{v}_1,\boldsymbol{v}_2,\ldots,\boldsymbol{v}_{T+1},\boldsymbol{v}_{T+2}\}\in\mathbb{R}^{(T+2)\times d}.$Three types of vertices/nodes $v_i\in\mathcal{V}$ include global cause embedding, token embedding and local psychological intention embedding.
Edges. There are three types of edges in the graph.
global connection The global connection is constructed for controlling the whole semantic information of the conversation, where node $g$ is connected to all the other nodes to globally make the mutual interactions.
local connection The local connection is constructed for selecting most relevant context based solely on the last helpseeker's state. In this way, the node $l$ should only be connected to tokens in the last utterance and global feature.
contextual connection The contextual connection is constructed for maintaining the information flow between conversations, where node $i$ is linked to the contextual tokens to propagate features with its neighbourhood.
Graph Modeling. By doing this, we describe the methodology to transform the sequentially encoded features using the graph attention network with the global connection, local connection and contextual connection. More specifically, we need to update the nodes of the three types $g^{(k)},h_{t}^{(k)}$ and $l^{(k)}$,which are presented to global node, token nodes and local node in the $k$-th layer of the graph, respectively. Firstly, $v_{1}^{(k+1)}=g^{(k+1)}$,computed by aggregating neighbourhood information, as:
(4)
$$\boldsymbol{v}_{1}^{(k+1)}=\sigma\big(\sum_{j\in\mathcal{N}_{g}}\alpha_{j}^{k}\boldsymbol{W}_{b}^{k}\boldsymbol{v}_{j}{}^{(k)}\big)$$
where $\mathcal{N}_g$ is the neighbors of the global node in the graph, $k=\{1,\cdots,K\},\boldsymbol{W}_{b}^{k}\in\mathbb{R}^{d\times d}$ is the trainable weight matrix, and $\sigma$ represents the nonlinearity activation function.
The weight $\alpha_j^k=\mathcal{H}(\boldsymbol{v}_1^{(k)},\dot{\boldsymbol{v}}_j^{(k)})$ in Equation (4) is calculated with an attention mechanism, which models the importance of each $v_j^{(k)}$ to $v_1^{(k)}:$
$$\alpha_{j}^{k}=\frac{\exp(\mathcal{F}(\boldsymbol{v}_{1}^{(k)},\boldsymbol{v}_{j}^{(k)}))}{\sum_{j^{\prime}\in\mathcal{N}_{g}}\exp(\mathcal{F}(\boldsymbol{v}_{1}^{(k)},\boldsymbol{v}_{j^{\prime}}^{(k)}))}$$
(5)
where $\mathcal{F}$ is an attention function.
Following[Velickovic $etal.,2017]$, the attention function
can be formulated as:
$$F(\boldsymbol{v}_{1}^{(k)},\boldsymbol{v}_{j}^{(k)})=\mathrm{LeakyReLU}\left(\mathbf{a}^{\top}[\boldsymbol{W}_{b}^{k}\boldsymbol{v}_{1}^{(k)}\|\boldsymbol{W}_{b}^{k}\boldsymbol{v}_{j}^{(k)}]\right)$$
(6)
where $\mathbf{a}\in\mathbb{R}^{2d}$ is the trainable weight matrix, LeakyReLU is a nonlinearity activation function, and $\cdot^T$ represents transposition and $\parallel$ is the concatenation operation.
Similarly, for token nodes and local node, the graph interac-
tion update process can be formulated as:
(7)
$$\boldsymbol{v}_{i}^{(k+1)}=\sigma\big(\sum_{j\in\mathcal{N}_{t}}\beta_{j}^{k}\boldsymbol{W}_{d}^{k}\boldsymbol{v}_{j}{}^{(k)}\big)\\\boldsymbol{v}_{T+2}^{(k+1)}=\sigma\big(\sum_{j\in\mathcal{N}_{l}}\gamma_{j}^{k}\boldsymbol{W}_{d}^{k}\boldsymbol{v}_{j}{}^{(k)}\big)$$
(8)
where $\mathcal{N}_t$ and $\mathcal{N}_l$ are the neighbors of the token nodes and
# 3.4 Global-guide Decoder Global-guide Decoder generates the supportive responses with
the updated features of nodes. In addition, a novel training objective monitors semantic information of the global cause. $\textbf{Response Generation. }$The output $Y=(y_1,y_2,\ldots,y_Z)$ consists of $Z$ words. Following the previous work[Roller $et$ $al.,2021;$Liu $etal.,2021]$,the decoder aims at predicting the word probability distribution. At each decoding timestep $z$,it reads the word embedding $W_{y<z}$ and the outputs of the graph reasoner $v^{(K)}$ for decoding
$$\boldsymbol{p}(y_{z}|\{y_{1},\ldots,y_{z-1}\},\boldsymbol{v}^{(K)})=\mathrm{Decoder}(\boldsymbol{W}_{y<z},\boldsymbol{v}^{(K)})$$
where $W_{y<z}$ denotes the embeddings of the generated tokens. And we use the updated representations in the graph reasoner $v^{(K)}$ to make a token-to-nodes cross attention, where token is in the decoder, nodes are in the graph reasoner.
Supervision of Global Semantic Information. Considering that the global cause controls the whole conversation flow, we design a novel task to monitor global semantic information with the category (e.g., the problem type of this conversation). The probability of the category can be calculated by leveraging the representation of the global node, as:
$$\boldsymbol{p}(o)=\mathrm{Softmax}(\mathrm{MLP}(\boldsymbol{v}_{1}{}^{(K)}))$$
(10)
# 3.5 Joint Training
The standard negative log-likelihood loss and cross-entropy
loss of the two tasks are optimized as:
(11)
$$\mathcal{L}_1=-\sum\limits_{z=1}^Z\log\boldsymbol{p}(y_z|\{y_1,\dots,y_{z-1}\},\boldsymbol{v}^{(K)})$$
$$\mathcal{L}_{2}=-\hat{\boldsymbol{o}}\log\boldsymbol{p}(o)$$
(12)
where $\hat{o}$ is the true label (problem type) of the conversation. We combine the above two loss functions as the training loss in a multi-task learning manner [Peng et al., 2021] as:
(13)
$$\mathcal{L}(\theta)=\lambda_1\mathcal{L}_1+\lambda_2\mathcal{L}_2$$
where $\theta$ is the all learnable parameters, and $\lambda_1$ and $\lambda_2$ are two hyper-parameters for controlling the weight of the rest tasks.
### 4 $\mathbf{Experiments}$
### 4.1 Dataset & Evaluation Metrics
Emotional Supprot Conversation dataset contains 1,053 multiturn dialogues with 31,410 utterances. Each conversation contains the cause information, dialog contextual, problem type. We make a statistic of the problem type in the ESConv, which includes 12 categories in total. We keep the train/test partition unchanged [Liu et al., 2021].
Automatic Evaluations. The automatic metrics include perplexity (PPL), BLEU-$n$ (B-$n)$, ROUGE-L (R-L), Distinct-1 (D-1), and Distinct-2 (D-2)[Li et al.,2016]. Perplexity measures the high-level general quality of the generation model. D-1/ D-2 is the proportion of the distinct unigrams/ bigrams
local node in the graph, respectively. $i= \{ 2, \cdots , ( T+ 1) \} $, D- 1/ D- 2 is the proportion of the distinct unigrams $\beta _{j}^{k}= \mathcal{H} ( \boldsymbol{v}_{i}^{( k) }, \boldsymbol{v}_{j}^{( k) })$ and $\gamma _{j}^{k}= \mathcal{H} ( \boldsymbol{v}_{( T+ 2) }^{( k) }, \boldsymbol{v}_{j}^{( k) }) .$ in all the generated results to ind | [
[
"0.0866",
"0.0663",
"0.4927",
"0.065",
"0.4927",
"0.0833",
"0.0866",
"0.084"
],
[
"0.511",
"0.06946",
"0.722",
"0.06946",
"0.722",
"0.0833",
"0.511",
"0.0833"
],
[
"0.0866",
"0.0833",
"0.4885",
"0.0833",
"0.4885",
"0.0966",
"0.0866",
"0.0966"
],
[
"0.513",
"0.0859",
"0.911",
"0.0859",
"0.911",
"0.09973",
"0.513",
"0.09973"
],
[
"0.0866",
"0.09784",
"0.4592",
"0.09784",
"0.4592",
"0.1111",
"0.0866",
"0.1111"
],
[
"0.513",
"0.101",
"0.9126",
"0.101",
"0.9126",
"0.11426",
"0.513",
"0.11426"
],
[
"0.0866",
"0.11554",
"0.4314",
"0.11554",
"0.4314",
"0.1288",
"0.0866",
"0.1288"
],
[
"0.513",
"0.11615",
"0.9067",
"0.11615",
"0.9067",
"0.127",
"0.513",
"0.127"
],
[
"0.1013",
"0.13",
"0.4878",
"0.1276",
"0.4878",
"0.1414",
"0.1013",
"0.1433"
],
[
"0.5107",
"0.132",
"0.8716",
"0.1332",
"0.8716",
"0.1477",
"0.5107",
"0.1465"
],
[
"0.871",
"0.1364",
"0.911",
"0.1364",
"0.911",
"0.1477",
"0.871",
"0.1477"
],
[
"0.0866",
"0.1433",
"0.4885",
"0.1433",
"0.4885",
"0.1572",
"0.0866",
"0.1572"
],
[
"0.5107",
"0.1459",
"0.9126",
"0.1465",
"0.9126",
"0.161",
"0.5107",
"0.1604"
],
[
"0.0866",
"0.1578",
"0.486",
"0.1578",
"0.486",
"0.1686",
"0.0866",
"0.1686"
],
[
"0.513",
"0.161",
"0.9126",
"0.161",
"0.9126",
"0.1749",
"0.513",
"0.1749"
],
[
"0.08496",
"0.1704",
"0.339",
"0.1686",
"0.339",
"0.1818",
"0.08496",
"0.1843"
],
[
"0.511",
"0.1749",
"0.9126",
"0.1755",
"0.9126",
"0.1895",
"0.511",
"0.1881"
],
[
"0.10297",
"0.1843",
"0.486",
"0.1843",
"0.486",
"0.1976",
"0.10297",
"0.1976"
],
[
"0.511",
"0.1895",
"0.914",
"0.1895",
"0.914",
"0.2039",
"0.511",
"0.2039"
],
[
"0.0866",
"0.1995",
"0.4844",
"0.1995",
"0.4844",
"0.2102",
"0.0866",
"0.2102"
],
[
"0.5073",
"0.202",
"0.699",
"0.2026",
"0.699",
"0.2203",
"0.5073",
"0.2197"
],
[
"0.0866",
"0.2122",
"0.4878",
"0.2122",
"0.4878",
"0.226",
"0.0866",
"0.226"
],
[
"0.08496",
"0.226",
"0.4028",
"0.226",
"0.4028",
"0.2393",
"0.08496",
"0.2393"
],
[
"0.10297",
"0.2412",
"0.4878",
"0.2412",
"0.4878",
"0.252",
"0.10297",
"0.252"
],
[
"0.5186",
"0.2361",
"0.914",
"0.2349",
"0.914",
"0.2556",
"0.5186",
"0.2568"
],
[
"0.0866",
"0.2551",
"0.4844",
"0.2551",
"0.4844",
"0.2651",
"0.0866",
"0.2651"
],
[
"0.528",
"0.2615",
"0.9126",
"0.2615",
"0.9126",
"0.276",
"0.528",
"0.276"
],
[
"0.08496",
"0.2676",
"0.486",
"0.2676",
"0.486",
"0.281",
"0.08496",
"0.281"
],
[
"0.511",
"0.276",
"0.9126",
"0.276",
"0.9126",
"0.289",
"0.511",
"0.289"
],
[
"0.08496",
"0.2822",
"0.3652",
"0.281",
"0.3652",
"0.2942",
"0.08496",
"0.2954"
],
[
"0.509",
"0.2886",
"0.914",
"0.2874",
"0.914",
"0.305",
"0.509",
"0.3062"
],
[
"0.0866",
"0.2986",
"0.486",
"0.2986",
"0.486",
"0.3125",
"0.0866",
"0.3125"
],
[
"0.513",
"0.305",
"0.8784",
"0.305",
"0.8784",
"0.318",
"0.513",
"0.318"
],
[
"0.08417",
"0.3132",
"0.486",
"0.3125",
"0.486",
"0.3257",
"0.08417",
"0.327"
],
[
"0.08496",
"0.327",
"0.486",
"0.327",
"0.486",
"0.3403",
"0.08496",
"0.3403"
],
[
"0.511",
"0.3225",
"0.914",
"0.3225",
"0.914",
"0.336",
"0.511",
"0.336"
],
[
"0.08417",
"0.3403",
"0.4878",
"0.3403",
"0.4878",
"0.3542",
"0.08417",
"0.3542"
],
[
"0.511",
"0.3372",
"0.914",
"0.3372",
"0.914",
"0.351",
"0.511",
"0.351"
],
[
"0.0825",
"0.356",
"0.4885",
"0.3542",
"0.4885",
"0.3718",
"0.0825",
"0.3738"
],
[
"0.511",
"0.351",
"0.9126",
"0.351",
"0.9126",
"0.3643",
"0.511",
"0.3643"
],
[
"0.511",
"0.3655",
"0.914",
"0.3655",
"0.914",
"0.379",
"0.511",
"0.379"
],
[
"0.08496",
"0.3718",
"0.4878",
"0.3738",
"0.4878",
"0.3882",
"0.08496",
"0.3865"
],
[
"0.513",
"0.379",
"0.914",
"0.379",
"0.914",
"0.3926",
"0.513",
"0.3926"
],
[
"0.08496",
"0.3882",
"0.4885",
"0.3882",
"0.4885",
"0.4028",
"0.08496",
"0.4028"
],
[
"0.511",
"0.3914",
"0.7812",
"0.3926",
"0.7812",
"0.4072",
"0.511",
"0.406"
],
[
"0.0825",
"0.3977",
"0.4902",
"0.4016",
"0.4902",
"0.4236",
"0.0825",
"0.4192"
],
[
"0.08417",
"0.4172",
"0.1936",
"0.4185",
"0.1936",
"0.433",
"0.08417",
"0.432"
],
[
"0.601",
"0.4192",
"0.819",
"0.4172",
"0.819",
"0.432",
"0.601",
"0.4343"
],
[
"0.88",
"0.4185",
"0.915",
"0.4185",
"0.915",
"0.4343",
"0.88",
"0.4343"
],
[
"0.1584",
"0.4414",
"0.3774",
"0.443",
"0.3774",
"0.4685",
"0.1584",
"0.4666"
],
[
"0.511",
"0.443",
"0.665",
"0.443",
"0.665",
"0.4583",
"0.511",
"0.4583"
],
[
"0.4631",
"0.4495",
"0.4878",
"0.4495",
"0.4878",
"0.4634",
"0.4631",
"0.4634"
],
[
"0.5107",
"0.4583",
"0.914",
"0.4602",
"0.914",
"0.4778",
"0.5107",
"0.476"
],
[
"0.2507",
"0.4666",
"0.282",
"0.4666",
"0.282",
"0.4778",
"0.2507",
"0.4778"
],
[
"0.0866",
"0.4797",
"0.4885",
"0.4797",
"0.4885",
"0.4944",
"0.0866",
"0.4944"
],
[
"0.511",
"0.476",
"0.7646",
"0.476",
"0.7646",
"0.4893",
"0.511",
"0.4893"
],
[
"0.08496",
"0.4944",
"0.4902",
"0.4944",
"0.4902",
"0.512",
"0.08496",
"0.512"
],
[
"0.6216",
"0.5",
"0.6396",
"0.4912",
"0.649",
"0.505",
"0.6313",
"0.512"
],
[
"0.6167",
"0.507",
"0.859",
"0.505",
"0.859",
"0.526",
"0.6167",
"0.528"
],
[
"0.88",
"0.509",
"0.914",
"0.509",
"0.914",
"0.5244",
"0.88",
"0.5244"
],
[
"0.5703",
"0.511",
"0.6274",
"0.511",
"0.6274",
"0.523",
"0.5703",
"0.523"
],
[
"0.08496",
"0.511",
"0.433",
"0.511",
"0.433",
"0.5244",
"0.08496",
"0.5244"
],
[
"0.1013",
"0.5244",
"0.4902",
"0.5244",
"0.4902",
"0.5425",
"0.1013",
"0.5425"
],
[
"0.626",
"0.529",
"0.648",
"0.529",
"0.648",
"0.536",
"0.626",
"0.536"
],
[
"0.08496",
"0.5425",
"0.4885",
"0.5425",
"0.4885",
"0.556",
"0.08496",
"0.556"
],
[
"0.653",
"0.549",
"0.773",
"0.549",
"0.773",
"0.563",
"0.653",
"0.563"
],
[
"0.8813",
"0.5474",
"0.914",
"0.5474",
"0.914",
"0.563",
"0.8813",
"0.563"
],
[
"0.08496",
"0.5596",
"0.1936",
"0.5596",
"0.1936",
"0.5728",
"0.08496",
"0.5728"
],
[
"0.2091",
"0.561",
"0.2361",
"0.561",
"0.2361",
"0.5693",
"0.2091",
"0.5693"
],
[
"0.2418",
"0.558",
"0.2507",
"0.558",
"0.2507",
"0.5625",
"0.2418",
"0.5625"
],
[
"0.511",
"0.5664",
"0.903",
"0.5664",
"0.903",
"0.58",
"0.511",
"0.58"
],
[
"0.2214",
"0.579",
"0.3513",
"0.5757",
"0.3513",
"0.5933",
"0.2214",
"0.5967"
],
[
"0.528",
"0.58",
"0.911",
"0.5806",
"0.911",
"0.595",
"0.528",
"0.5947"
],
[
"0.1544",
"0.5923",
"0.1765",
"0.5923",
"0.1765",
"0.605",
"0.1544",
"0.605"
],
[
"0.464",
"0.5913",
"0.486",
"0.5913",
"0.486",
"0.605",
"0.464",
"0.605"
],
[
"0.511",
"0.5947",
"0.8965",
"0.5947",
"0.8965",
"0.608",
"0.511",
"0.608"
],
[
"0.1953",
"0.601",
"0.384",
"0.5977",
"0.384",
"0.6187",
"0.1953",
"0.622"
],
[
"0.6406",
"0.6143",
"0.785",
"0.6143",
"0.785",
"0.629",
"0.6406",
"0.629"
],
[
"0.08496",
"0.622",
"0.304",
"0.622",
"0.304",
"0.636",
"0.08496",
"0.636"
],
[
"0.8823",
"0.616",
"0.9126",
"0.616",
"0.9126",
"0.629",
"0.8823",
"0.629"
],
[
"0.10046",
"0.6353",
"0.486",
"0.634",
"0.486",
"0.6484",
"0.10046",
"0.6494"
],
[
"0.5146",
"0.637",
"0.9126",
"0.637",
"0.9126",
"0.651",
"0.5146",
"0.651"
],
[
"0.08826",
"0.651",
"0.2279",
"0.651",
"0.2279",
"0.661",
"0.08826",
"0.661"
],
[
"0.513",
"0.6514",
"0.9087",
"0.6514",
"0.9087",
"0.6655",
"0.513",
"0.6655"
],
[
"0.0866",
"0.6655",
"0.452",
"0.6655",
"0.452",
"0.683",
"0.0866",
"0.683"
],
[
"0.4575",
"0.6694",
"0.4844",
"0.6694",
"0.4844",
"0.681",
"0.4575",
"0.681"
],
[
"0.517",
"0.6816",
"0.539",
"0.6816",
"0.539",
"0.69",
"0.517",
"0.69"
],
[
"0.539",
"0.6787",
"0.652",
"0.6787",
"0.652",
"0.6934",
"0.539",
"0.6934"
],
[
"0.08496",
"0.6895",
"0.4844",
"0.6895",
"0.4844",
"0.704",
"0.08496",
"0.704"
],
[
"0.08496",
"0.705",
"0.4885",
"0.705",
"0.4885",
"0.7183",
"0.08496",
"0.7183"
],
[
"0.513",
"0.7007",
"0.7837",
"0.7007",
"0.7837",
"0.714",
"0.513",
"0.714"
],
[
"0.08417",
"0.7173",
"0.3545",
"0.7173",
"0.3545",
"0.731",
"0.08417",
"0.731"
],
[
"0.513",
"0.7173",
"0.9126",
"0.7173",
"0.9126",
"0.732",
"0.513",
"0.732"
],
[
"0.10297",
"0.733",
"0.4885",
"0.733",
"0.4885",
"0.747",
"0.10297",
"0.747"
],
[
"0.513",
"0.732",
"0.9126",
"0.732",
"0.9126",
"0.7456",
"0.513",
"0.7456"
],
[
"0.08417",
"0.7456",
"0.357",
"0.7456",
"0.357",
"0.759",
"0.08417",
"0.759"
],
[
"0.5146",
"0.7476",
"0.911",
"0.7476",
"0.911",
"0.7583",
"0.5146",
"0.7583"
],
[
"0.513",
"0.7603",
"0.914",
"0.7603",
"0.914",
"0.7734",
"0.513",
"0.7734"
],
[
"0.2091",
"0.7744",
"0.3708",
"0.773",
"0.3708",
"0.7915",
"0.2091",
"0.7935"
],
[
"0.5107",
"0.773",
"0.911",
"0.773",
"0.911",
"0.786",
"0.5107",
"0.786"
],
[
"0.464",
"0.778",
"0.4878",
"0.778",
"0.4878",
"0.7905",
"0.464",
"0.7905"
],
[
"0.1871",
"0.776",
"0.2075",
"0.776",
"0.2075",
"0.7827",
"0.1871",
"0.7827"
],
[
"0.2507",
"0.7944",
"0.2834",
"0.7944",
"0.2834",
"0.806",
"0.2507",
"0.806"
],
[
"0.513",
"0.7886",
"0.7646",
"0.7886",
"0.7646",
"0.802",
"0.513",
"0.802"
],
[
"0.1772",
"0.8096",
"0.2157",
"0.8096",
"0.2157",
"0.8193",
"0.1772",
"0.8193"
],
[
"0.2141",
"0.8096",
"0.3757",
"0.8096",
"0.3757",
"0.83",
"0.2141",
"0.83"
],
[
"0.511",
"0.8037",
"0.9126",
"0.8047",
"0.9126",
"0.8193",
"0.511",
"0.8184"
],
[
"0.1666",
"0.8164",
"0.2035",
"0.8184",
"0.2018",
"0.829",
"0.165",
"0.827"
],
[
"0.464",
"0.813",
"0.486",
"0.813",
"0.486",
"0.827",
"0.464",
"0.827"
],
[
"0.513",
"0.8193",
"0.914",
"0.8193",
"0.914",
"0.8335",
"0.513",
"0.8335"
],
[
"0.2524",
"0.83",
"0.2834",
"0.83",
"0.2834",
"0.8413",
"0.2524",
"0.8413"
],
[
"0.5146",
"0.8354",
"0.911",
"0.8354",
"0.911",
"0.845",
"0.5146",
"0.845"
],
[
"0.0866",
"0.8433",
"0.486",
"0.8433",
"0.486",
"0.857",
"0.0866",
"0.857"
],
[
"0.513",
"0.8477",
"0.914",
"0.8477",
"0.914",
"0.8613",
"0.513",
"0.8613"
],
[
"0.08496",
"0.856",
"0.492",
"0.858",
"0.492",
"0.8726",
"0.08496",
"0.8706"
],
[
"0.511",
"0.86",
"0.914",
"0.8613",
"0.914",
"0.8755",
"0.511",
"0.875"
],
[
"0.08496",
"0.869",
"0.4019",
"0.8735",
"0.4019",
"0.8945",
"0.08496",
"0.89"
],
[
"0.3245",
"0.8735",
"0.3342",
"0.8735",
"0.3342",
"0.8774",
"0.3245",
"0.8774"
],
[
"0.3774",
"0.8726",
"0.3945",
"0.8755",
"0.392",
"0.882",
"0.3743",
"0.879"
],
[
"0.513",
"0.8755",
"0.846",
"0.8755",
"0.846",
"0.8896",
"0.513",
"0.8896"
],
[
"0.1978",
"0.8755",
"0.2091",
"0.8755",
"0.2091",
"0.879",
"0.1978",
"0.879"
]
] | [
"[1,V2,..., T+1, VT+2] E R(T+2)xd. Three types of ver-",
"3.4Global-guide Decoder",
"tices/nodes i E V include global cause embedding, token",
"Global-guide Decoder generates the supportive responses with",
"embedding and local psychological intention embedding.",
"the updated features of nodes. In addition, a novel training",
"Edges.There are three types of edges in the graph.",
"objectivemonitorssemanticinformationoftheglobalcause",
"global connection The global connection is constructed",
"Response Generation.The output Y =(yi, y2,.:",
":.yz",
"for controlling the whole semantic information of the conver-",
"consists of Z words. Following the previous work [Roller et",
"sation,wherenodegisconnected to all the othernodes to",
"al., 2021; Liu et al., 2021l, the decoder aims at predicting the",
"globally make the mutual interactions.",
"word probability distribution.At each decoding timestep z, it",
"local connection The local connection is constructed for",
"reads the word embedding W y<z and the outputs of the graph",
"selectingmostrelevantcontextbasedsolelyonthelasthelp",
"reasoner (K) for decoding.",
"seeker's state. In this way, the node I should only be connected",
"to tokens in the last utterance and global feature.",
"contextual connection The contextual connection is con-",
"p(yz/(y1,..,ya-1), (K)) = Decoder(Wy<2, (K) (9)",
"structedformaintainingtheinformationflow between con",
"where Wy<z denotes the embeddings of the generated",
"versations, where node i is linked to the contextual tokens to",
"tokens. And we use the updated representations in the graph",
"propagate features with its neighbourhood",
"reasoner (K) to make a token-to-nodes cross attention, where",
"Graph Modeling.By doing this, we describe the method-",
"token is in the decoder, nodes are in the graph reasoner.",
"ology to transform the sequentiallyencoded features using",
"the graph attention network with the global connection, local",
"Supervision of Global Semantic Information.Consider-",
"connection and contextual connection. More specifically, we",
"ing that the global cause controls the whole conversation flow,",
"need to update the nodes of the three types g(k), h(*) and",
"we design a novel task to monitor global semantic information",
"with the category (e.g., the problem type of this conversation).",
"(k), which are presented to global node, token nodes and lo-",
"The probability of the category can be calculated by leveraging",
"cal node in the k-th layer of the graph, respectively. Firstly,",
"the representation of the global node, as:",
"(k+1) = g(k+1), computed by aggregating neighbourhood",
"information, as:",
"p(o) = Softmax(MLP(ui(K)))",
"(10)",
"(++1) =0(Z awtu,()",
"3.5Joint Training",
"(4)",
"The standard negative log-likelihood loss and cross-entropy",
"jeN",
"where N. is the neighbors of the global node in the graph,",
"loss of the two tasks are optimized as:",
"k =- {1, .. , K), Wb e Rdxd is the trainable weight matrix,",
"",
"Elog p(yz/(y1, .., z--), (K)",
"(11)",
"C= -",
"and represents the nonlinearity activation function.",
"The weight ah = H(u(k), u(k)) in Equation (4) is calcu-",
"2=1",
"lated with an attention mechanism, which models the impor-",
"L2 = -olog p(o)",
"(12)",
"tance of each v;",
"tow",
"",
"where o is the true label (problem type)of the conversation.",
"exp(F(u(k), v(k))",
"We combine the above two loss functions as the training",
"at",
"(5)",
"loss in a multi-task learning manner [Peng et al., 2021] as:",
"E'eN, exp(F(u(h),u()",
"C(0)=L + 22",
"where F is an attention function.",
"(13)",
"Following [Velickovic et al., 2017], the attention function",
"where is the all learnable parameters, and Ai and >2 are two",
"can be formulated as:",
"hyper-parameters for controlling the weight of the rest tasks.",
"F(u(*), u(k)) = LeakyReLU (aT[Wbu() Wbu(*)])",
"(6",
"4",
"Experiments",
"where a E R2d is the trainable weight matrix, LeakyReLU is",
"a nonlinearity activation function, and .T represents transposi-",
"4.1 Dataset & Evaluation Metrics",
"tion and Il is the concatenation operation.",
"Emotional Supprot Conversation dataset contains 1,053 multi",
"Similarly, for token nodes and local node, the graph interac-",
"turn dialogues with 31,410 utterances. Each conversation",
"tion update process can be formulated as:",
"containsthecauseinformation,dialog contextual,problem",
"type. We make a statistic of the problem type in the ESConv,",
"=o(Z Bfwau,(k)",
"which includes 12 categories in total. We keep the train/test",
"(7)",
"+1",
"jENt",
"partition unchanged [Liu et al., 2021].",
"(k+1)",
"=o( wiu;()",
"Automatic Evaluations.The automatic metrics include per",
"VT+2",
"(8)",
"plexity (PPL), BLEU-n (B-n), ROUGE-L (R-L), Distinct-1",
"jeN",
"(D-1),andDistinct-2(D-2)[Liet al.,2016l.Perplexitymea-",
"where Nt and Ni are the neighbors of the token nodes and",
"sures the high-level general quality of the generation model.",
"local node in the graph, respectively. i = {2, ..., (T + 1)),",
"D-1 / D-2 is the proportion of the distinct unigrams / bigrams",
"β = H(v(), u) and = H((+2),).",
"",
"k)",
"in all the generated results to indicate the diversity.",
""
] | {
"arxiv_id": "2204.12749",
"language": "en",
"page_idx": 3
} |
||
2310.14528-en-1 | 2310.14528-en-1.jpg | knowledge retrieval (Madotto et al., 2020; Huang et al., 2022). These approaches typically integrate the processes of knowledge retrieval and response generation and train them under the supervision of reference responses. However, this approach suffers from two notable shortcomings. Firstly, the system response often comprises both pure language tokens and KB-related tokens (e.g., hotel names and addresses), making it challenging to train an effective retriever using weak supervision signals from reference responses. Secondly, the efficiency of these systems may decrease as the knowledge base expands in size.
Unlike the aforementioned works, we employ a retriever-generator architecture that explicitly separates the retrieval process from response generation. The retriever is responsible for identifying relevant information from the KB, while the generator utilizes the dialogue context and retrieved information to generate the response. Although this architecture seems straightforward, constructing an effective retriever presents significant challenges, particularly in two key aspects. Firstly, TOD systems inherently possess ground-truth responses to train the generator, but lack annotations for training the retriever. Thus, the retriever can only be trained using weak supervision signals derived from the response generator. Secondly, within a specific domain, different entities often exhibit structural and content similarities, making it challenging for the generator to learn which entities are truly relevant. Consequently, the weak supervision signals from the generator may not be reliable. Figure 1 provides a visual illustration of these challenges.
To tackle these challenges, we propose a dualfeedback mechanism for the retriever, consisting of positive feedback and negative feedback. Positive feedback is constructed based on the conditional generation probabilities of responses corresponding to different retrieved entities. Intuitively, if the relevance between an entity and the response is higher, the conditional generation probability of the response corresponding to this entity will also be higher. Utilizing positive feedback allows us to train the retriever based on the knowledge learned by the generator from reference responses. In order to prevent the retriever from being misled by inaccurate information, we contend that calibration is necessary. Calibration involves the initial identification and resolution of errors, which we accomplish by sampling negative samples derived
from the generator's hypothesis responses. Then, we construct negative feedback based on these negative samples to calibrate the positive feedback.
We conducted evaluations of our system on three well-established TOD datasets: Multi-WOZ $2. 1\allowbreak ( {\mathrm{mod}} $MWOZ) (Eric et al., 2020), Stanford MultiDomain dataset (SMD) (Eric et al., 2017), and CamRest (Wen et al., 2017). The experimental results demonstrate that our model outperforms the baseline methods, particularly in scenarios involving large-scale KB. Additionally, through extensive analysis, we have made several notable findings. Firstly, our retriever exhibits clear advantages over the baseline methods as the size of KB increases. Secondly, our dual-feedback mechanism effectively mitigates the problem of incorrect knowledge learned solely with positive feedback from the generator. Lastly, our method exhibits good compatibility with LLMs like ChatGPT.
### 2 $\textbf{Related Work}$
2.1 End-to-End Task-Oriented Dialogue
E2E-TOD systems employ different strategies to incorporate KB information for response generation. Firstly, memory networks are utilized to store knowledge, using multi-hop interactions to aggregate relevant information. For instance, Mem2seq (Madotto et al., 2018) employs multi-hop attention over memory cells to select KB-related tokens during response generation. GLMP (Wu et al., 2019) introduces a global-to-local memory pointer network to retrieve relevant triplets and complete the response template. CD-NET (Raghu et al., 2021) retrieves relevant KB records by computing a distillation distribution based on dialogue context.
Secondly, the entire linearized KB is encoded by pre-trained language models and taken as input to a generator to generate the final system response. For instance, UnifedSKG (Xie et al., 2022) uses a unified text-to-text framework, while Q-TOD (Tian et al., 2022) rewrites the dialogue context into a natural language query for knowledge retrieval. MAKER (Wan et al., 2023) introduces a multi-grained retrival with both entity and attribute selection.
Thirdly, knowledge bases are stored in model parameters for implicit retrieval during response generation. GPT-KE (Madotto et al., 2020) embeds the KB into pre-trained model parameters through data augmentation. Following GPT-KE, ECO (Huang et al., 2022) first generates the most relevant entity | [
[
"0.11755",
"0.0849",
"0.4895",
"0.0849",
"0.4895",
"0.0998",
"0.11755",
"0.0998"
],
[
"0.511",
"0.0861",
"0.882",
"0.0861",
"0.882",
"0.0998",
"0.511",
"0.0998"
],
[
"0.11755",
"0.10156",
"0.4895",
"0.10156",
"0.4895",
"0.1152",
"0.11755",
"0.1152"
],
[
"0.509",
"0.10156",
"0.882",
"0.1027",
"0.882",
"0.1164",
"0.509",
"0.1152"
],
[
"0.11584",
"0.11816",
"0.486",
"0.11816",
"0.486",
"0.1318",
"0.11584",
"0.1318"
],
[
"0.51",
"0.11816",
"0.8623",
"0.11816",
"0.8623",
"0.1318",
"0.51",
"0.1318"
],
[
"0.11584",
"0.1354",
"0.4878",
"0.1342",
"0.4878",
"0.1478",
"0.11584",
"0.1484"
],
[
"0.529",
"0.133",
"0.882",
"0.1342",
"0.882",
"0.1484",
"0.529",
"0.1478"
],
[
"0.11755",
"0.1509",
"0.4878",
"0.1509",
"0.4878",
"0.1644",
"0.11755",
"0.1644"
],
[
"0.511",
"0.1509",
"0.882",
"0.1509",
"0.882",
"0.1644",
"0.511",
"0.1644"
],
[
"0.11755",
"0.1663",
"0.4895",
"0.1663",
"0.4895",
"0.1799",
"0.11755",
"0.1799"
],
[
"0.511",
"0.165",
"0.881",
"0.165",
"0.881",
"0.1787",
"0.511",
"0.1787"
],
[
"0.11755",
"0.1829",
"0.4902",
"0.1829",
"0.4902",
"0.1965",
"0.11755",
"0.1965"
],
[
"0.51",
"0.1818",
"0.881",
"0.1818",
"0.881",
"0.1953",
"0.51",
"0.1953"
],
[
"0.11755",
"0.199",
"0.4878",
"0.199",
"0.4878",
"0.212",
"0.11755",
"0.212"
],
[
"0.511",
"0.199",
"0.879",
"0.199",
"0.879",
"0.212",
"0.511",
"0.212"
],
[
"0.1142",
"0.2156",
"0.486",
"0.2156",
"0.486",
"0.2286",
"0.1142",
"0.2286"
],
[
"0.509",
"0.2156",
"0.882",
"0.2156",
"0.882",
"0.2286",
"0.509",
"0.2286"
],
[
"0.11584",
"0.2322",
"0.4895",
"0.2322",
"0.4895",
"0.2452",
"0.11584",
"0.2452"
],
[
"0.51",
"0.2322",
"0.882",
"0.2322",
"0.882",
"0.2452",
"0.51",
"0.2452"
],
[
"0.11584",
"0.2477",
"0.4878",
"0.2477",
"0.4878",
"0.2612",
"0.11584",
"0.2612"
],
[
"0.509",
"0.2477",
"0.882",
"0.2477",
"0.882",
"0.2612",
"0.509",
"0.2612"
],
[
"0.11584",
"0.2625",
"0.4878",
"0.2625",
"0.4878",
"0.2756",
"0.11584",
"0.2756"
],
[
"0.51",
"0.2632",
"0.882",
"0.2632",
"0.882",
"0.2766",
"0.51",
"0.2766"
],
[
"0.11755",
"0.2798",
"0.3586",
"0.2798",
"0.3586",
"0.2935",
"0.11755",
"0.2935"
],
[
"0.51",
"0.2798",
"0.8833",
"0.2798",
"0.8833",
"0.2944",
"0.51",
"0.2944"
],
[
"0.1335",
"0.3005",
"0.4878",
"0.3018",
"0.4878",
"0.3164",
"0.1335",
"0.3154"
],
[
"0.51",
"0.2957",
"0.8833",
"0.2944",
"0.8833",
"0.3088",
"0.51",
"0.31"
],
[
"0.511",
"0.31",
"0.8833",
"0.3113",
"0.8833",
"0.326",
"0.511",
"0.3247"
],
[
"0.11584",
"0.3184",
"0.4878",
"0.3184",
"0.4878",
"0.332",
"0.11584",
"0.332"
],
[
"0.513",
"0.328",
"0.882",
"0.328",
"0.882",
"0.3416",
"0.513",
"0.3416"
],
[
"0.11755",
"0.335",
"0.4895",
"0.335",
"0.4895",
"0.3486",
"0.11755",
"0.3486"
],
[
"0.511",
"0.3445",
"0.879",
"0.3445",
"0.879",
"0.3552",
"0.511",
"0.3552"
],
[
"0.11755",
"0.3508",
"0.4878",
"0.3508",
"0.4878",
"0.364",
"0.11755",
"0.364"
],
[
"0.511",
"0.3599",
"0.882",
"0.3599",
"0.882",
"0.3735",
"0.511",
"0.3735"
],
[
"0.11755",
"0.3665",
"0.4895",
"0.3665",
"0.4895",
"0.3801",
"0.11755",
"0.3801"
],
[
"0.511",
"0.376",
"0.8506",
"0.3748",
"0.8506",
"0.3896",
"0.511",
"0.3901"
],
[
"0.11755",
"0.3818",
"0.4895",
"0.3818",
"0.4895",
"0.3955",
"0.11755",
"0.3955"
],
[
"0.11584",
"0.3984",
"0.4895",
"0.3984",
"0.4895",
"0.412",
"0.11584",
"0.412"
],
[
"0.51",
"0.4038",
"0.665",
"0.4038",
"0.665",
"0.4187",
"0.51",
"0.4187"
],
[
"0.11584",
"0.415",
"0.4878",
"0.415",
"0.4878",
"0.4287",
"0.11584",
"0.4287"
],
[
"0.11584",
"0.4312",
"0.4954",
"0.4312",
"0.4954",
"0.4448",
"0.11584",
"0.4448"
],
[
"0.513",
"0.4312",
"0.8403",
"0.4312",
"0.8403",
"0.4453",
"0.513",
"0.4453"
],
[
"0.11584",
"0.4478",
"0.4902",
"0.4478",
"0.4902",
"0.4614",
"0.11584",
"0.4614"
],
[
"0.511",
"0.4507",
"0.881",
"0.4531",
"0.881",
"0.4673",
"0.511",
"0.4656"
],
[
"0.11584",
"0.4631",
"0.4878",
"0.4631",
"0.4878",
"0.4768",
"0.11584",
"0.4768"
],
[
"0.51",
"0.4685",
"0.881",
"0.4697",
"0.881",
"0.4827",
"0.51",
"0.4822"
],
[
"0.11584",
"0.4797",
"0.4878",
"0.4797",
"0.4878",
"0.4934",
"0.11584",
"0.4934"
],
[
"0.51",
"0.4868",
"0.879",
"0.4868",
"0.879",
"0.4976",
"0.51",
"0.4976"
],
[
"0.11755",
"0.4954",
"0.4878",
"0.4954",
"0.4878",
"0.509",
"0.11755",
"0.509"
],
[
"0.511",
"0.502",
"0.882",
"0.502",
"0.882",
"0.5156",
"0.511",
"0.5156"
],
[
"0.1142",
"0.511",
"0.4878",
"0.5103",
"0.4878",
"0.525",
"0.1142",
"0.5254"
],
[
"0.513",
"0.517",
"0.882",
"0.517",
"0.882",
"0.531",
"0.513",
"0.531"
],
[
"0.11755",
"0.528",
"0.4878",
"0.528",
"0.4878",
"0.5415",
"0.11755",
"0.5415"
],
[
"0.511",
"0.532",
"0.882",
"0.533",
"0.882",
"0.5474",
"0.511",
"0.5464"
],
[
"0.11755",
"0.5435",
"0.4878",
"0.5435",
"0.4878",
"0.557",
"0.11755",
"0.557"
],
[
"0.51",
"0.549",
"0.881",
"0.549",
"0.881",
"0.5625",
"0.51",
"0.5625"
],
[
"0.11755",
"0.56",
"0.4878",
"0.56",
"0.4878",
"0.5737",
"0.11755",
"0.5737"
],
[
"0.509",
"0.5654",
"0.882",
"0.564",
"0.882",
"0.579",
"0.509",
"0.58"
],
[
"0.11584",
"0.5747",
"0.4878",
"0.5747",
"0.4878",
"0.5884",
"0.11584",
"0.5884"
],
[
"0.51",
"0.583",
"0.882",
"0.583",
"0.882",
"0.5967",
"0.51",
"0.5967"
],
[
"0.11755",
"0.5933",
"0.486",
"0.5933",
"0.486",
"0.604",
"0.11755",
"0.604"
],
[
"0.511",
"0.5986",
"0.882",
"0.5986",
"0.882",
"0.6123",
"0.511",
"0.6123"
],
[
"0.11755",
"0.608",
"0.4878",
"0.608",
"0.4878",
"0.6226",
"0.11755",
"0.6226"
],
[
"0.51",
"0.6143",
"0.882",
"0.6143",
"0.882",
"0.6274",
"0.51",
"0.6274"
],
[
"0.11584",
"0.6235",
"0.476",
"0.6235",
"0.476",
"0.638",
"0.11584",
"0.638"
],
[
"0.509",
"0.629",
"0.8833",
"0.63",
"0.8833",
"0.6445",
"0.509",
"0.6436"
],
[
"0.1368",
"0.6465",
"0.4902",
"0.6465",
"0.4902",
"0.661",
"0.1368",
"0.661"
],
[
"0.51",
"0.6465",
"0.8613",
"0.6465",
"0.8613",
"0.66",
"0.51",
"0.66"
],
[
"0.11584",
"0.662",
"0.4878",
"0.662",
"0.4878",
"0.676",
"0.11584",
"0.676"
],
[
"0.5283",
"0.661",
"0.882",
"0.662",
"0.882",
"0.677",
"0.5283",
"0.676"
],
[
"0.11755",
"0.6797",
"0.4895",
"0.6797",
"0.4895",
"0.6934",
"0.11755",
"0.6934"
],
[
"0.509",
"0.6787",
"0.882",
"0.6777",
"0.882",
"0.6924",
"0.509",
"0.6934"
],
[
"0.11755",
"0.6953",
"0.4878",
"0.6953",
"0.4878",
"0.709",
"0.11755",
"0.709"
],
[
"0.51",
"0.6953",
"0.8833",
"0.6953",
"0.8833",
"0.709",
"0.51",
"0.709"
],
[
"0.11755",
"0.7114",
"0.4902",
"0.7104",
"0.4902",
"0.7246",
"0.11755",
"0.7256"
],
[
"0.51",
"0.7104",
"0.882",
"0.7104",
"0.882",
"0.7236",
"0.51",
"0.7236"
],
[
"0.11755",
"0.728",
"0.4895",
"0.728",
"0.4895",
"0.741",
"0.11755",
"0.741"
],
[
"0.51",
"0.727",
"0.881",
"0.727",
"0.881",
"0.7407",
"0.51",
"0.7407"
],
[
"0.11755",
"0.7446",
"0.4844",
"0.7446",
"0.4844",
"0.755",
"0.11755",
"0.755"
],
[
"0.51",
"0.741",
"0.882",
"0.742",
"0.882",
"0.7573",
"0.51",
"0.756"
],
[
"0.11755",
"0.7603",
"0.4895",
"0.7603",
"0.4895",
"0.774",
"0.11755",
"0.774"
],
[
"0.51",
"0.759",
"0.882",
"0.759",
"0.882",
"0.7725",
"0.51",
"0.7725"
],
[
"0.11755",
"0.7754",
"0.4895",
"0.7754",
"0.4895",
"0.789",
"0.11755",
"0.789"
],
[
"0.51",
"0.7754",
"0.882",
"0.7754",
"0.882",
"0.789",
"0.51",
"0.789"
],
[
"0.11584",
"0.7915",
"0.4878",
"0.7915",
"0.4878",
"0.8047",
"0.11584",
"0.8047"
],
[
"0.509",
"0.7915",
"0.882",
"0.7905",
"0.882",
"0.8047",
"0.509",
"0.8057"
],
[
"0.11755",
"0.808",
"0.4878",
"0.808",
"0.4878",
"0.822",
"0.11755",
"0.822"
],
[
"0.511",
"0.8086",
"0.581",
"0.8086",
"0.581",
"0.821",
"0.511",
"0.821"
],
[
"0.11584",
"0.8247",
"0.4895",
"0.8247",
"0.4895",
"0.8384",
"0.11584",
"0.8384"
],
[
"0.529",
"0.8237",
"0.8833",
"0.8237",
"0.8833",
"0.8384",
"0.529",
"0.8384"
],
[
"0.11584",
"0.8403",
"0.4878",
"0.839",
"0.4878",
"0.854",
"0.11584",
"0.855"
],
[
"0.51",
"0.8413",
"0.8833",
"0.8413",
"0.8833",
"0.855",
"0.51",
"0.855"
],
[
"0.11584",
"0.857",
"0.4902",
"0.857",
"0.4902",
"0.8706",
"0.11584",
"0.8706"
],
[
"0.51",
"0.855",
"0.882",
"0.8555",
"0.882",
"0.8706",
"0.51",
"0.869"
],
[
"0.1142",
"0.8726",
"0.4878",
"0.8726",
"0.4878",
"0.8857",
"0.1142",
"0.8857"
],
[
"0.511",
"0.8726",
"0.882",
"0.8726",
"0.882",
"0.8857",
"0.511",
"0.8857"
],
[
"0.1142",
"0.887",
"0.4895",
"0.888",
"0.4895",
"0.903",
"0.1142",
"0.902"
],
[
"0.509",
"0.888",
"0.881",
"0.888",
"0.881",
"0.903",
"0.509",
"0.903"
],
[
"0.11584",
"0.905",
"0.4895",
"0.905",
"0.4895",
"0.9185",
"0.11584",
"0.9185"
],
[
"0.51",
"0.905",
"0.882",
"0.905",
"0.882",
"0.9185",
"0.51",
"0.9185"
]
] | [
"knowledge retrieval (Madotto et al., 2020; Huang",
"from the generator's hypothesis responses. Then.",
"et al., 2022). These approaches typically integrate",
"we construct negative feedback based on these neg-",
"the processes of knowledge retrieval and response",
"ative samples to calibrate the positive feedback.",
"generation and train them underthe supervision",
"We conducted evaluations of our system on",
"of reference responses. However, this approach",
"three well-established TOD datasets: Multi-WOZ",
"suffers from two notable shortcomings. Firstly,",
"2.1 (MWOZ) (Eric et al., 2020). Stanford Multi-",
"the system response often comprises both pure lan-",
"Domain dataset (SMD) (Eric et al., 2017), and",
"guage tokens and KB-related tokens (e.g., hotel",
"CamRest (Wen et al., 2017). The experimental",
"names and addresses), making it challenging to",
"results demonstrate that our model outperforms",
"train an effective retriever using weak supervision",
"the baseline methods, particularly in scenarios in-",
"signals from reference responses. Secondly, the",
"volving large-scale KB. Additionally, through ex-",
"efficiency of these systems may decrease as the",
"tensive analysis, we have made several notable",
"knowledge base expands in size.",
"findings. Firstly, our retriever exhibits clear ad-",
"Unlike the aforementioned works, we employ a",
"vantages over the baseline methods as the size of",
"KB increases. Secondly, our dual-feedback mecha-",
"retriever-generator architecture that explicitly sep-",
"nism effectively mitigates the problem of incorrect",
"arates the retrieval process from response genera-",
"knowledgelearned solely withpositivefeedback",
"tion. The retriever is responsible for identifying",
"from the generator. Lastly, our method exhibits",
"relevant information from the KB, while the gen-",
"good compatibility with LLMs like ChatGPT.",
"erator utilizes the dialogue context and retrieved",
"information to generate the response. Although this",
"2Related Work",
"architecture seems straightforward, constructing an",
"effective retriever presents significant challenges,",
"2.1 End-to-End Task-Oriented Dialogue",
"particularly in two key aspects. Firstly, TOD sys-",
"E2E-TOD systems employ different strategies to",
"tems inherently possess ground-truth responses to",
"incorporate KB information forresponse genera",
"train the generator, but lack annotations for training",
"tion.Firstly,memorynetworks areutilizedto store",
"the retriever. Thus, the retriever can only be trained",
"knowledge, using multi-hop interactions to aggre-",
"using weak supervision signals derived from the",
"gate relevant information. For instance, Mem2seq",
"response generator. Secondly, within a specific",
"(Madotto et al., 2018) employs multi-hop attention",
"domain, different entities often exhibit structural",
" over memory cells to select KB-related tokens dur",
"and content similarities, making it challenging for",
"ing response generation. GLMP (Wu et al., 2019)",
"the generator to learn which entities are truly rele-",
"introduces a global-to-local memory pointer net",
"vant. Consequently,the weak supervision signals",
"work to retrieve relevant triplets and complete the",
"from the generator may not be reliable. Figure 1",
"response template. CD-NET (Raghu et al., 2021)",
"provides a visual illustration of these challenges.",
"retrieves relevant KB records by computing a dis-",
"To tackle these challenges, we propose a dual-",
"tillation distribution based on dialogue context.",
"feedback mechanism for the retriever, consisting of",
"Secondly, the entire linearized KB is encoded by",
"positive feedback and negative feedback. Positive",
" pre-trained language models and taken as input to",
"feedback is constructed based on the conditional",
"a generator to generate the final system response.",
"generation probabilities of responses correspond-",
"For instance, UnifiedSKG (Xie et al., 2022) uses",
"ing to different retrieved entities. Intuitively, if the",
"a unified text-to-text framework, while Q-TOD",
"relevance between an entity and the response is",
"(Tian et al., 2022) rewrites the dialogue context",
"higher, the conditional generation probability of",
"into a natural language query for knowledge re",
"the response corresponding to this entity will also",
"trieval. MAKER (Wan et al., 2023) introduces a",
"be higher. Utilizing positive feedback allows us to",
"multi-grained retrival with both entity and attribute",
"train the retriever based on the knowledge learned",
"selection",
"by the generator from reference responses. In or-",
"Thirdly, knowledge bases are stored in model pa-",
"der to prevent the retriever from being misled by",
"rameters for implicit retrieval during response gen-",
"inaccurate information, we contend that calibra-",
"eration. GPT-KE (Madotto et al., 2020) embeds the",
" tion is necessary. Calibration involves the initial",
"KB into pre-trained model parameters through data",
"identification and resolution of errors, which we",
" augmentation. Following GPT-KE, ECO (Huang",
"accomplish by sampling negative samples derived",
"et al., 2022) first generates the most relevant entity"
] | {
"arxiv_id": "2310.14528",
"language": "en",
"page_idx": 1
} |
||
2310.14528-en-2 | 2310.14528-en-2.jpg | with trie constraint before response generation to
ensure entity consistency in the response.
## 2.2 Knowledge Retrieval
Previous research has extensively explored methods for knowledge retrieval across various tasks, including question answering (Chen et al., 2017; Kwiatkowski et al., 2019), fact checking (Thorne et al., 2018), and dialogue systems (Dinan et al., 2019). Recently, neural network-based dense retrievers have become popular. These retrievers commonly use a dual-encoder architecture (Yih et al., 2011), where queries and passages are encoded as separate vectors, and relevance is determined through inner product or Euclidean distance.
Supervised retrievers, such as DPR (Karpukhin et al., 2020), have been developed for open-domain question answering. These retrievers are trained using labeled question-document pairs. To overcome the need for costly query-document annotations, researchers have explored alternative approaches that leverage signals from the answer generator. REALM (Guu et al., 2020) and RAG (Lewis et al., 2020) propose joint training of the retriever and the generator by treating documents as latent variables. FiD-KD (Izacard and Grave, 2021a) employs crossattention scores as supervision through knowledge distillation. EMDR$^2$ (Sachan et al., 2021) models retrieval decisions as latent variables and employs an expectation-maximization algorithm to approximate the computation. Unsupervised approaches have also been explored. ICT (Lee et al., 2019) introduces the inverse cloze task for unsupervised pre-training of dense retrievers. Izacard et al. (2022) investigate contrastive learning methods for training retrievers, while Spider (Ram et al., 2022) utilizes recurring spans within a document to create pseudo-positive query-document pairs.
## 3 $\textbf{Methods}$
As illustrated in Figure 2, our system comprises a knowledge retriever and a response generator. The retriever fırst retrieves top-$K$ relevant entities from a knowledge base. These retrieved entities, along with the dialogue context, are then fed into the generator model to generate the response.
## 3.1 Notations
Given a dialogue $\mathcal{D}=\{u_1,y_1,...,u_T,y_T\}$ consisting of $T$ turns, where $u_t$ and $y_t$ represent the user utterance and system response at the $t$-th turn,
respectively, we denote the dialogue context of the $t$-th turn as $c_t$ = $\{ u_1, y_1, . . . , u_{t- 1}, y_{t- 1}, u_t\} .$ An external KB is provided, represented as a set of entities, i.e., $\mathcal{E}=\{e_1,e_2,...,e_B\}$, where each entity $e_i$ consists of $N$ attribute-value pairs, i.e., $e_{i}=\{a^{1},v_{i}^{1},...,a^{N},v_{i}^{N}\}.$E2E-TOD takes the dialogue context $c_t$ and the KB as input and directly generates a natural language response $y_t.$
### 3.2 Knowledge Retriever
Our knowledge retriever comprises an encoder Enc$_r$ that maps any input to a $d$-dimensional vector. The user utterances and system responses in the dialogue context $c_t$ are first concatenated and encoded by Enc$_r$ as the query. To encode an entity, we concatenate the attribute-value pairs of the entity into a sequence and pass it to Enc$_r.$ The similarity score between $c_t$ and $e_i$ is computed by taking the dot product of their respective vectors:
$$s_{t,i}=\mathrm{Enc}_r(c_t)^T\mathrm{Enc}_r(e_i).$$
(1)
Based on these similarity scores, we identify the top-$K$ entities from the entity set $\mathcal{E}$ as the candidate entities for response generation. This set of candidates is represented as $\hat{\mathcal{E}}=\{e_1,e_2,...,e_K\}.$
We employ BERT (Devlin et al., 2019) as the encoder and extract the representations of the[CLS] token to represent $c_t$ and $e_i.$ Prior research has highlighted that initializing the encoder directly with BERT weights can lead to collapsed representations and impact retrieval performance. Consequently, we initialize the weights via pre-training with distant supervision (Qin et al., 2019).$^{1}$
## 3.3 Response Generator
Our generator is built upon the Fusion-in-Decoder ( FiD) ( Izacard and Grave, 2021b) model, which is based on the pre-trained T5 (Raffel et al., 2020). The FiD model consists of an encoder Enc$_g$ and a decoder Dec$_g.$ The encoder is responsible for processing $K$ different text inputs independently, where each input $x_{t,i}$ is formed by concatenating the dialogue context $c_t$ and a candidate entity $e_i.$ The output representations of the encoder are then concatenated to create a global representation $X_t=$ $[\mathrm{Enc}_{g}(x_{t,1}),\ldots,\mathrm{Enc}_{g}(x_{t,K})]$for the current turn.
The decoder takes $X_t$ as input and generates the response autoregressively. During this process, the decoder utilizes causal attention to incorporate information from previously generated tokens, as
$^{\mathrm{l}}$Details of the pre-training are available in Appendix D. | [
[
"0.11755",
"0.0849",
"0.4902",
"0.0861",
"0.4902",
"0.0998",
"0.11755",
"0.0986"
],
[
"0.511",
"0.0861",
"0.882",
"0.0861",
"0.882",
"0.0998",
"0.511",
"0.0998"
],
[
"0.11755",
"0.1027",
"0.424",
"0.1027",
"0.424",
"0.1164",
"0.11755",
"0.1164"
],
[
"0.511",
"0.10095",
"0.882",
"0.1027",
"0.882",
"0.11755",
"0.511",
"0.1152"
],
[
"0.511",
"0.1164",
"0.882",
"0.11755",
"0.882",
"0.1318",
"0.511",
"0.1312"
],
[
"0.11755",
"0.13",
"0.3257",
"0.13",
"0.3257",
"0.1437",
"0.11755",
"0.1437"
],
[
"0.513",
"0.1342",
"0.882",
"0.1342",
"0.882",
"0.1484",
"0.513",
"0.1484"
],
[
"0.11755",
"0.1526",
"0.4937",
"0.1526",
"0.4937",
"0.1663",
"0.11755",
"0.1663"
],
[
"0.509",
"0.1484",
"0.882",
"0.1497",
"0.882",
"0.1644",
"0.509",
"0.1633"
],
[
"0.11755",
"0.1687",
"0.4902",
"0.1687",
"0.4902",
"0.1818",
"0.11755",
"0.1818"
],
[
"0.51",
"0.165",
"0.8833",
"0.165",
"0.8833",
"0.1799",
"0.51",
"0.1799"
],
[
"0.11584",
"0.1853",
"0.4902",
"0.1853",
"0.4902",
"0.199",
"0.11584",
"0.199"
],
[
"0.509",
"0.1812",
"0.881",
"0.1818",
"0.881",
"0.1965",
"0.509",
"0.1953"
],
[
"0.11755",
"0.2007",
"0.4878",
"0.2007",
"0.4878",
"0.2144",
"0.11755",
"0.2144"
],
[
"0.509",
"0.1978",
"0.817",
"0.199",
"0.817",
"0.2131",
"0.509",
"0.212"
],
[
"0.1184",
"0.2173",
"0.4878",
"0.2173",
"0.4878",
"0.231",
"0.1184",
"0.231"
],
[
"0.511",
"0.2227",
"0.721",
"0.2239",
"0.721",
"0.2382",
"0.511",
"0.2369"
],
[
"0.11755",
"0.2328",
"0.4878",
"0.2328",
"0.4878",
"0.2465",
"0.11755",
"0.2465"
],
[
"0.11755",
"0.2494",
"0.492",
"0.2494",
"0.492",
"0.2632",
"0.11755",
"0.2632"
],
[
"0.513",
"0.2452",
"0.8833",
"0.2452",
"0.8833",
"0.2588",
"0.513",
"0.2588"
],
[
"0.11755",
"0.2654",
"0.492",
"0.2654",
"0.492",
"0.279",
"0.11755",
"0.279"
],
[
"0.511",
"0.2625",
"0.882",
"0.2625",
"0.882",
"0.2756",
"0.511",
"0.2756"
],
[
"0.11755",
"0.2808",
"0.4902",
"0.2808",
"0.4902",
"0.2957",
"0.11755",
"0.2957"
],
[
"0.51",
"0.279",
"0.882",
"0.279",
"0.882",
"0.2922",
"0.51",
"0.2922"
],
[
"0.11584",
"0.2976",
"0.4902",
"0.2976",
"0.4902",
"0.3113",
"0.11584",
"0.3113"
],
[
"0.513",
"0.2957",
"0.881",
"0.2957",
"0.881",
"0.306",
"0.513",
"0.306"
],
[
"0.11755",
"0.3135",
"0.4475",
"0.3135",
"0.4475",
"0.328",
"0.11755",
"0.328"
],
[
"0.511",
"0.3088",
"0.881",
"0.3088",
"0.881",
"0.3225",
"0.511",
"0.3225"
],
[
"0.1351",
"0.33",
"0.4878",
"0.33",
"0.4878",
"0.3433",
"0.1351",
"0.3433"
],
[
"0.513",
"0.3267",
"0.882",
"0.3267",
"0.882",
"0.3403",
"0.513",
"0.3403"
],
[
"0.11755",
"0.3467",
"0.4895",
"0.3467",
"0.4895",
"0.3599",
"0.11755",
"0.3599"
],
[
"0.509",
"0.3425",
"0.882",
"0.3416",
"0.882",
"0.3557",
"0.509",
"0.357"
],
[
"0.11755",
"0.3623",
"0.492",
"0.3623",
"0.492",
"0.3772",
"0.11755",
"0.3772"
],
[
"0.511",
"0.3594",
"0.882",
"0.3594",
"0.882",
"0.3723",
"0.511",
"0.3723"
],
[
"0.11755",
"0.379",
"0.4895",
"0.379",
"0.4895",
"0.3926",
"0.11755",
"0.3926"
],
[
"0.511",
"0.3748",
"0.876",
"0.3748",
"0.876",
"0.3884",
"0.511",
"0.3884"
],
[
"0.11584",
"0.3943",
"0.4878",
"0.3943",
"0.4878",
"0.408",
"0.11584",
"0.408"
],
[
"0.596",
"0.401",
"0.7944",
"0.3997",
"0.7944",
"0.4187",
"0.596",
"0.4192"
],
[
"0.858",
"0.4026",
"0.8833",
"0.4026",
"0.8833",
"0.4175",
"0.858",
"0.4175"
],
[
"0.11755",
"0.411",
"0.4878",
"0.411",
"0.4878",
"0.4246",
"0.11755",
"0.4246"
],
[
"0.11584",
"0.4258",
"0.4895",
"0.427",
"0.4895",
"0.4412",
"0.11584",
"0.4407"
],
[
"0.513",
"0.4316",
"0.881",
"0.4316",
"0.881",
"0.4453",
"0.513",
"0.4453"
],
[
"0.11755",
"0.4424",
"0.4878",
"0.4424",
"0.4878",
"0.456",
"0.11755",
"0.456"
],
[
"0.51",
"0.449",
"0.882",
"0.449",
"0.882",
"0.462",
"0.51",
"0.462"
],
[
"0.11755",
"0.459",
"0.4878",
"0.459",
"0.4878",
"0.4727",
"0.11755",
"0.4727"
],
[
"0.511",
"0.4644",
"0.882",
"0.4644",
"0.882",
"0.478",
"0.511",
"0.478"
],
[
"0.12006",
"0.4768",
"0.4878",
"0.4768",
"0.4878",
"0.4868",
"0.12006",
"0.4868"
],
[
"0.509",
"0.4768",
"0.8774",
"0.4785",
"0.8774",
"0.4963",
"0.509",
"0.4946"
],
[
"0.11755",
"0.4912",
"0.4902",
"0.4912",
"0.4902",
"0.505",
"0.11755",
"0.505"
],
[
"0.529",
"0.4963",
"0.882",
"0.4963",
"0.882",
"0.5103",
"0.529",
"0.5103"
],
[
"0.11755",
"0.5073",
"0.4878",
"0.5073",
"0.4878",
"0.521",
"0.11755",
"0.521"
],
[
"0.51",
"0.5127",
"0.881",
"0.5127",
"0.881",
"0.5254",
"0.51",
"0.5254"
],
[
"0.11755",
"0.5225",
"0.4902",
"0.5225",
"0.4902",
"0.536",
"0.11755",
"0.536"
],
[
"0.511",
"0.5293",
"0.882",
"0.5293",
"0.882",
"0.542",
"0.511",
"0.542"
],
[
"0.11755",
"0.539",
"0.4902",
"0.539",
"0.4902",
"0.5527",
"0.11755",
"0.5527"
],
[
"0.511",
"0.5435",
"0.881",
"0.5444",
"0.881",
"0.5586",
"0.511",
"0.558"
],
[
"0.11584",
"0.5557",
"0.4878",
"0.5557",
"0.4878",
"0.5693",
"0.11584",
"0.5693"
],
[
"0.51",
"0.561",
"0.882",
"0.561",
"0.882",
"0.5747",
"0.51",
"0.5747"
],
[
"0.11755",
"0.5723",
"0.4902",
"0.5723",
"0.4902",
"0.586",
"0.11755",
"0.586"
],
[
"0.511",
"0.579",
"0.881",
"0.579",
"0.881",
"0.589",
"0.511",
"0.589"
],
[
"0.11755",
"0.5884",
"0.4902",
"0.5884",
"0.4902",
"0.6016",
"0.11755",
"0.6016"
],
[
"0.513",
"0.5933",
"0.881",
"0.5933",
"0.881",
"0.607",
"0.513",
"0.607"
],
[
"0.11755",
"0.604",
"0.4895",
"0.604",
"0.4895",
"0.6177",
"0.11755",
"0.6177"
],
[
"0.51",
"0.6094",
"0.8286",
"0.607",
"0.8286",
"0.6216",
"0.51",
"0.6235"
],
[
"0.11584",
"0.619",
"0.4878",
"0.618",
"0.4878",
"0.633",
"0.11584",
"0.6343"
],
[
"0.1184",
"0.6357",
"0.4895",
"0.6357",
"0.4895",
"0.6494",
"0.1184",
"0.6494"
],
[
"0.511",
"0.6357",
"0.7153",
"0.6357",
"0.7153",
"0.6494",
"0.511",
"0.6494"
],
[
"0.1142",
"0.651",
"0.486",
"0.651",
"0.486",
"0.6646",
"0.1142",
"0.6646"
],
[
"0.511",
"0.655",
"0.881",
"0.655",
"0.881",
"0.6685",
"0.511",
"0.6685"
],
[
"0.11584",
"0.6685",
"0.4878",
"0.6685",
"0.4878",
"0.6816",
"0.11584",
"0.6816"
],
[
"0.513",
"0.673",
"0.882",
"0.673",
"0.882",
"0.686",
"0.513",
"0.686"
],
[
"0.1142",
"0.683",
"0.4048",
"0.684",
"0.4048",
"0.6987",
"0.1142",
"0.6978"
],
[
"0.51",
"0.687",
"0.882",
"0.6885",
"0.882",
"0.703",
"0.51",
"0.7017"
],
[
"0.51",
"0.705",
"0.8833",
"0.705",
"0.8833",
"0.7188",
"0.51",
"0.7188"
],
[
"0.1142",
"0.7134",
"0.2258",
"0.7134",
"0.2258",
"0.728",
"0.1142",
"0.728"
],
[
"0.511",
"0.72",
"0.882",
"0.72",
"0.882",
"0.734",
"0.511",
"0.734"
],
[
"0.1184",
"0.7407",
"0.4895",
"0.7407",
"0.4895",
"0.7544",
"0.1184",
"0.7544"
],
[
"0.51",
"0.7363",
"0.882",
"0.737",
"0.882",
"0.752",
"0.51",
"0.7505"
],
[
"0.11755",
"0.7573",
"0.4878",
"0.7573",
"0.4878",
"0.771",
"0.11755",
"0.771"
],
[
"0.51",
"0.752",
"0.882",
"0.753",
"0.882",
"0.767",
"0.51",
"0.7666"
],
[
"0.511",
"0.767",
"0.882",
"0.7695",
"0.882",
"0.783",
"0.511",
"0.781"
],
[
"0.11755",
"0.775",
"0.486",
"0.775",
"0.486",
"0.785",
"0.11755",
"0.785"
],
[
"0.11755",
"0.789",
"0.4878",
"0.789",
"0.4878",
"0.8027",
"0.11755",
"0.8027"
],
[
"0.511",
"0.7837",
"0.879",
"0.7837",
"0.879",
"0.7974",
"0.511",
"0.7974"
],
[
"0.11584",
"0.8047",
"0.4878",
"0.8047",
"0.4878",
"0.8184",
"0.11584",
"0.8184"
],
[
"0.5146",
"0.8027",
"0.876",
"0.8027",
"0.876",
"0.8135",
"0.5146",
"0.8135"
],
[
"0.11755",
"0.822",
"0.4265",
"0.822",
"0.4265",
"0.835",
"0.11755",
"0.835"
],
[
"0.5146",
"0.818",
"0.875",
"0.818",
"0.875",
"0.832",
"0.5146",
"0.832"
],
[
"0.5283",
"0.832",
"0.881",
"0.833",
"0.881",
"0.847",
"0.5283",
"0.8467"
],
[
"0.11584",
"0.8496",
"0.2377",
"0.8496",
"0.2377",
"0.8633",
"0.11584",
"0.8633"
],
[
"0.51",
"0.8496",
"0.881",
"0.8496",
"0.881",
"0.8633",
"0.51",
"0.8633"
],
[
"0.11755",
"0.8716",
"0.492",
"0.8716",
"0.492",
"0.8857",
"0.11755",
"0.8857"
],
[
"0.513",
"0.8677",
"0.881",
"0.8677",
"0.881",
"0.8774",
"0.513",
"0.8774"
],
[
"0.51",
"0.882",
"0.881",
"0.882",
"0.881",
"0.8955",
"0.51",
"0.8955"
],
[
"0.11584",
"0.8896",
"0.4895",
"0.8896",
"0.4895",
"0.903",
"0.11584",
"0.903"
],
[
"0.1142",
"0.905",
"0.4878",
"0.905",
"0.4878",
"0.9185",
"0.1142",
"0.9185"
],
[
"0.5356",
"0.905",
"0.875",
"0.905",
"0.875",
"0.9185",
"0.5356",
"0.9185"
]
] | [
"with trie constraint before response generation to",
"respectively, we denote the dialogue context of",
"ensure entity consistency in the response.",
"the t-th turn as ct = [ui, yi,.., ut-1, yt-1, ut]",
"An external KB is provided, represented as a set",
"2.2Knowledge Retrieval",
"of entities, i.e., & = [ei,e2, ..., eb], where each",
"Previous research has extensively explored meth-",
"entity ei consists of N attribute-value pairs, i.e.,",
"ods for knowledge retrieval across various tasks,",
"ei = {al, u],.., aN, vNI. E2E-TOD takes the dia-",
"including question answering (Chen et al., 2017;",
"logue context Ct and the KB as input and directly",
"Kwiatkowski et al., 2019), fact checking (Thorne",
" generates a natural language response yt.",
"et al., 2018), and dialogue systems (Dinan et al.,",
"3.2 Knowledge Retriever",
"2019).Recently,neural network-based dense re-",
"trievers have become popular. These retrievers com-",
"Our knowledge retriever comprises an encoder",
"monly use a dual-encoder architecture (Yih et al.,",
"Encr that maps any input to a d-dimensional vec-",
"2011), where queries and passages are encoded",
"tor. The user utterances and system responses in",
"as separate vectors, and relevance is determined",
"thedialoguecontextCtarefirstconcatenated and",
"through inner product or Euclidean distance.",
"encoded by Enc, as the query. To encode an en-",
"Supervised retrievers, such as DPR (Karpukhin",
"tity, we concatenate the attribute-value pairs of the",
"et al., 2020), have been developed for open-domain",
"entity into a sequence and pass it to Encr. The",
"question answering. These retrievers are trained us-",
"similarity score between ct and ei is computed by",
"ing labeled question-document pairs. To overcome",
"taking the dot product of their respective vectors:",
"the need for costly query-document annotations,",
"St,i = Enc, (ct)\"Encr(ei).",
"(1)",
"researchers have explored alternative approaches",
"that leverage signals from the answer generator.",
"Based on these similarity scores, we identify the",
"REALM (Guu et al., 2020) and RAG (Lewis et al..",
"top-K entities from the entity set & as the candi-",
"2020) propose joint training of the retriever and the",
"date entities for response generation. This set of",
"generatorbytreatingdocumentsaslatentvariables.",
"candidates is represented as & - [ei, e2, ., ek].",
"FiD-KD (Izacard and Grave, 2021la) employs cross-",
"We employ BERT (Devlin et al., 2019) as the en-",
"attention scores as supervision through knowledge",
"coder and extract the representations of the [CLS]",
"distillation..EMDR2 (Sachan et al.,2021)mod-",
"token to represent Ct and ei. Prior research has",
"els retrieval decisions as latent variables and em-",
"highlighted that initializing the encoder directly",
"ploys an expectation-maximization algorithm to",
"with BERT weights can lead to collapsed represen",
"approximate the computation. Unsupervised ap-",
"tations and impactretrieval performance.Conse",
"proaches have also been explored. ICT (Lee et al.,",
"quently, we initialize the weights via pre-training",
"2019) introduces the inverse cloze task for unsuper-",
"with distant supervision (Qin et al., 2019).1",
"vised pre-training of dense retrievers. Izacard et al.",
"(2022) investigate contrastive learning methods for",
"3.3 Response Generator",
"training retrievers, while Spider (Ram et al., 2022)",
"Our generator is built upon the Fusion-in-Decoder",
"utilizes recurring spans within a document to create",
"(FiD) (Izacard and Grave, 2021b) model, which is",
"pseudo-positive query-document pairs.",
"based on the pre-trained T5 (Raffel et al., 2020)",
"The FiD model consists of an encoder Enc and",
"3 Methods",
"a decoder Decg. The encoder is responsible for",
"As illustrated in Figure 2, our system comprises a",
"processing K different text inputs independently,",
"knowledge retriever and a response generator. The",
"where each input t,i is formed by concatenating",
"the dialogue context ct and a candidate entity ei.",
"retrieverfirstretrievestop-Krelevant entitiesfrom",
"a knowledge base. These retrieved entities, along",
"The output representations of the encoder are then",
"with the dialogue context, are then fed into the",
"concatenatedtocreateaglobalrepresentationXt=",
"generator model to generate the response.",
"[Encg(ct.i),.., Encg(&t,k)] for the current turn.",
"The decoder takes Xt as input and generates",
"3.1 Notations",
"the response autoregressively. During this process",
"Given a dialogue D = fui, yi, .., uT, yrl con--",
"thedecoderutilizescausalattentiontoincorporate",
" information from previously generated tokens, as",
"sisting of T turns, where ut and yt represent the",
" user utterance and system response at the t-th turn,",
"'Details of the pre-training are available in Appendix D."
] | {
"arxiv_id": "2310.14528",
"language": "en",
"page_idx": 2
} |
||
2310.14528-en-3 | 2310.14528-en-3.jpg | <FigureHere>
Figure 2: The overall architecture of our end-to-end task-oriented dialogue system, comprising a knowledge retriever
and a response generator. The retrieval component is trained using dual feedback from the response generator.
between the retrieval scores $S_t=\{s_{t,i}\}_{1\leq i\leq K}$ obtained from Eq. (1) and the pseudo-labels $G_t$ using KL-divergence as follows:
well as cross-attention to the input tokens represented by $X_t.$ This enables the decoder to consider information from both the generated tokens and the tokens of retrieved entities. The probability of the response is defined as follows:
(4)
$$\begin{aligned}L_{\mathrm{pos}}&=D_{KL}(G_{t},S_{t})\\&=\sum_{i=1}^{K}\tilde{g}_{t,i}(\log\tilde{g}_{t,i}-\log\tilde{s}_{t,i}),\end{aligned}$$
$$\begin{aligned}p(y_t|X_t)=\prod_{i=1}^{|y_t|}p(y_{t,i}|y_{t,<i},X_t),\end{aligned}$$
(2)
where $|y_t|$ represents the length of the response $y_t.$
## 3.4 End-to-End Training
Entity scoring Given the retrieved entities $\hat{\mathcal{E}}=$ $\{e_1,e_2,...,e_K\}$,we employ the aforementioned response generator to assign a score to each entity. Firstly, the dialogue context $c_t$ and each entity $e_i$ are concatenated to form $x_{t,i}$, which is then fed into the generator. The conditional log probability (length-normalized) of the response $y_t$ is utilized to score entity $e_i$ as follows:
(3)
$$\begin{aligned}g_{t,i}&=p(y_t|X_{t,i})\\&=\frac{\sum_{i=1}^{|y_t|}\log p(y_{t,i}|y_{t,<i},X_{t,i})}{|y_t|},\end{aligned}$$
where $\tilde{g}_{t,i}=$softmax$(G_t)_i,\tilde{s}_{t,i}=$softmax$(S_t)_i.$ The use of the generator's supervision to train the retriever is referred to as positive feedback. However, it is important to note that the generator may assimilate incorrect knowledge from the entities, leading to inaccurate entity relevancy scores.
Negative feedback To address the issue of positive feedback when training the retriever, we propose incorporating negative feedback from the generator to calibrate the pseudo-labels. To achieve this, we select a negative sample for the global representation $X_t$ during the generation process. A negative sample refers to a response that exhibits a high generation probability but is of low quality. Specifically, we generate a set $\{\hat{y}_t^i\}_{1\leq i\leq M}$ of responses using beam search, where $M$ is the beam search size, and rank them based on their generation probabilities. We use $R_t^g(\hat{y}_t^i)$ to represent the rank of candidate response $\hat{y}_t^i.$ Note that a high generation probability does not guarantee a highquality response. To evaluate the quality, we define a function $o(\hat{y}_t^i,y_t)$ that measures the overlap between a response and the reference response. We implement $o(\hat{y}_t^i,y_t)$ using the BLEU metric. As a result, we obtain a new sorted list of responses based on their true quality, where the rank of candidate response $\hat{y}_t^i$ is $R_t^o(\hat{y}_t^i).$ Finally, we identify
where $X_t,i=\mathrm{Enc}_g(x_{t,i}).$
The rationale behind the scoring is straightforward: if an entity is pertinent to the response, the probability of generating this response given the dialogue context and the entity should also be high. Positive feedback We utilize the entity scores $G_t$ = $\{ g_{t, i}\} _{1\leq i\leq K}$ obtained from the generator as pseudo-labels to train the retriever, with the objective of transferring the knowledge acquired by the generator from reference responses to the retriever. To achieve this, we enforce consistency | [
[
"0.511",
"0.08435",
"0.6113",
"0.08435",
"0.6113",
"0.0944",
"0.511",
"0.0944"
],
[
"0.1545",
"0.1069",
"0.2301",
"0.1069",
"0.2301",
"0.114",
"0.1545",
"0.114"
],
[
"0.3005",
"0.10394",
"0.3123",
"0.10394",
"0.3123",
"0.1134",
"0.3005",
"0.1134"
],
[
"0.1604",
"0.1194",
"0.1763",
"0.1194",
"0.1763",
"0.1271",
"0.1604",
"0.1271"
],
[
"0.1738",
"0.11816",
"0.22",
"0.11816",
"0.22",
"0.1289",
"0.1738",
"0.1289"
],
[
"0.2183",
"0.12054",
"0.2225",
"0.12054",
"0.2225",
"0.1259",
"0.2183",
"0.1259"
],
[
"0.2225",
"0.1194",
"0.241",
"0.1194",
"0.241",
"0.1277",
"0.2225",
"0.1277"
],
[
"0.2837",
"0.12354",
"0.345",
"0.12354",
"0.345",
"0.1354",
"0.2837",
"0.1354"
],
[
"0.5767",
"0.12354",
"0.644",
"0.12354",
"0.644",
"0.1354",
"0.5767",
"0.1354"
],
[
"0.1503",
"0.1318",
"0.2418",
"0.1318",
"0.2418",
"0.1426",
"0.1503",
"0.1426"
],
[
"0.7446",
"0.1289",
"0.827",
"0.1289",
"0.827",
"0.1395",
"0.7446",
"0.1395"
],
[
"0.4055",
"0.1395",
"0.4785",
"0.1395",
"0.4785",
"0.1509",
"0.4055",
"0.1509"
],
[
"0.178",
"0.1455",
"0.1973",
"0.1455",
"0.1973",
"0.1538",
"0.178",
"0.1538"
],
[
"0.1351",
"0.1342",
"0.1528",
"0.1342",
"0.1528",
"0.1414",
"0.1351",
"0.1414"
],
[
"0.1394",
"0.1478",
"0.1486",
"0.1478",
"0.1486",
"0.1526",
"0.1394",
"0.1526"
],
[
"0.1621",
"0.1478",
"0.1721",
"0.1478",
"0.1721",
"0.1526",
"0.1621",
"0.1526"
],
[
"0.2301",
"0.1467",
"0.2377",
"0.1467",
"0.2377",
"0.1526",
"0.2301",
"0.1526"
],
[
"0.5366",
"0.1484",
"0.5728",
"0.1509",
"0.571",
"0.165",
"0.5356",
"0.1633"
],
[
"0.294",
"0.1497",
"0.3267",
"0.1497",
"0.3267",
"0.1621",
"0.294",
"0.1621"
],
[
"0.2007",
"0.1497",
"0.2124",
"0.1437",
"0.2183",
"0.1484",
"0.2065",
"0.155"
],
[
"0.4048",
"0.152",
"0.4902",
"0.152",
"0.4902",
"0.1621",
"0.4048",
"0.1621"
],
[
"0.1528",
"0.1592",
"0.1738",
"0.1592",
"0.1738",
"0.1663",
"0.1528",
"0.1663"
],
[
"0.1755",
"0.1592",
"0.1973",
"0.1592",
"0.1973",
"0.1663",
"0.1755",
"0.1663"
],
[
"0.2048",
"0.1592",
"0.2217",
"0.1592",
"0.2217",
"0.1663",
"0.2048",
"0.1663"
],
[
"0.2241",
"0.1592",
"0.2393",
"0.1592",
"0.2393",
"0.1675",
"0.2241",
"0.1675"
],
[
"0.754",
"0.155",
"0.8193",
"0.155",
"0.8193",
"0.1663",
"0.754",
"0.1663"
],
[
"0.1394",
"0.1609",
"0.1486",
"0.1609",
"0.1486",
"0.1663",
"0.1394",
"0.1663"
],
[
"0.4014",
"0.1644",
"0.4937",
"0.1644",
"0.4937",
"0.1758",
"0.4014",
"0.1758"
],
[
"0.5225",
"0.1621",
"0.5845",
"0.1644",
"0.5825",
"0.1787",
"0.5205",
"0.177"
],
[
"0.5903",
"0.1644",
"0.613",
"0.1644",
"0.613",
"0.1746",
"0.5903",
"0.1746"
],
[
"0.7515",
"0.1687",
"0.8237",
"0.1687",
"0.8237",
"0.1799",
"0.7515",
"0.1799"
],
[
"0.65",
"0.1729",
"0.6826",
"0.1746",
"0.681",
"0.1882",
"0.6484",
"0.1859"
],
[
"0.309",
"0.1733",
"0.3167",
"0.1733",
"0.3167",
"0.1799",
"0.309",
"0.1799"
],
[
"0.3787",
"0.1663",
"0.3914",
"0.1644",
"0.3938",
"0.1729",
"0.382",
"0.1746"
],
[
"0.4148",
"0.1776",
"0.4685",
"0.1776",
"0.4685",
"0.1895",
"0.4148",
"0.1895"
],
[
"0.4626",
"0.1776",
"0.4902",
"0.1776",
"0.4902",
"0.1895",
"0.4626",
"0.1895"
],
[
"0.523",
"0.1787",
"0.58",
"0.1787",
"0.58",
"0.1907",
"0.523",
"0.1907"
],
[
"0.403",
"0.1799",
"0.4182",
"0.1799",
"0.4182",
"0.1882",
"0.403",
"0.1882"
],
[
"0.1318",
"0.1812",
"0.2435",
"0.1787",
"0.2435",
"0.1936",
"0.1318",
"0.1953"
],
[
"0.2747",
"0.1841",
"0.3477",
"0.1841",
"0.3477",
"0.1948",
"0.2747",
"0.1948"
],
[
"0.3855",
"0.1812",
"0.392",
"0.1812",
"0.392",
"0.1853",
"0.3855",
"0.1853"
],
[
"0.592",
"0.1924",
"0.608",
"0.1924",
"0.608",
"0.2019",
"0.592",
"0.2019"
],
[
"0.7524",
"0.1882",
"0.825",
"0.1882",
"0.825",
"0.2037",
"0.7524",
"0.2037"
],
[
"0.3787",
"0.2037",
"0.4434",
"0.2019",
"0.4434",
"0.2131",
"0.3787",
"0.2156"
],
[
"0.4685",
"0.2031",
"0.4902",
"0.206",
"0.4878",
"0.2131",
"0.465",
"0.2102"
],
[
"0.523",
"0.2037",
"0.581",
"0.2037",
"0.581",
"0.2156",
"0.523",
"0.2156"
],
[
"0.76",
"0.209",
"0.812",
"0.209",
"0.812",
"0.2203",
"0.76",
"0.2203"
],
[
"0.6465",
"0.2102",
"0.6616",
"0.2048",
"0.669",
"0.2156",
"0.6543",
"0.2203"
],
[
"0.44",
"0.206",
"0.465",
"0.206",
"0.465",
"0.2131",
"0.44",
"0.2131"
],
[
"0.147",
"0.2173",
"0.2107",
"0.2173",
"0.2107",
"0.228",
"0.147",
"0.228"
],
[
"0.294",
"0.2144",
"0.3257",
"0.2162",
"0.3242",
"0.231",
"0.293",
"0.2286"
],
[
"0.6245",
"0.2299",
"0.7095",
"0.2299",
"0.7095",
"0.2411",
"0.6245",
"0.2411"
],
[
"0.7446",
"0.2299",
"0.8296",
"0.2299",
"0.8296",
"0.2411",
"0.7446",
"0.2411"
],
[
"0.7188",
"0.2405",
"0.7314",
"0.2405",
"0.7314",
"0.2434",
"0.7188",
"0.2434"
],
[
"0.294",
"0.2542",
"0.3167",
"0.2505",
"0.3198",
"0.2625",
"0.2957",
"0.2654"
],
[
"0.7676",
"0.2488",
"0.781",
"0.2405",
"0.7925",
"0.2488",
"0.7793",
"0.257"
],
[
"0.51",
"0.2654",
"0.615",
"0.2654",
"0.615",
"0.2756",
"0.51",
"0.2756"
],
[
"0.11584",
"0.2874",
"0.882",
"0.2874",
"0.882",
"0.3005",
"0.11584",
"0.3005"
],
[
"0.11584",
"0.3027",
"0.8564",
"0.3027",
"0.8564",
"0.3164",
"0.11584",
"0.3164"
],
[
"0.11755",
"0.335",
"0.492",
"0.335",
"0.492",
"0.3457",
"0.11755",
"0.3457"
],
[
"0.51",
"0.332",
"0.882",
"0.3333",
"0.882",
"0.3474",
"0.51",
"0.3467"
],
[
"0.1184",
"0.3508",
"0.4878",
"0.3508",
"0.4878",
"0.361",
"0.1184",
"0.361"
],
[
"0.51",
"0.3499",
"0.882",
"0.3499",
"0.882",
"0.3635",
"0.51",
"0.3635"
],
[
"0.11584",
"0.3652",
"0.4895",
"0.3652",
"0.4895",
"0.379",
"0.11584",
"0.379"
],
[
"0.513",
"0.3652",
"0.7095",
"0.3652",
"0.7095",
"0.379",
"0.513",
"0.379"
],
[
"0.11755",
"0.383",
"0.486",
"0.383",
"0.486",
"0.3938",
"0.11755",
"0.3938"
],
[
"0.11584",
"0.398",
"0.3435",
"0.398",
"0.3435",
"0.411",
"0.11584",
"0.411"
],
[
"0.571",
"0.401",
"0.7256",
"0.3984",
"0.7256",
"0.4133",
"0.571",
"0.415"
],
[
"0.6274",
"0.4187",
"0.6484",
"0.4146",
"0.6543",
"0.4287",
"0.6323",
"0.4329"
],
[
"0.272",
"0.4258",
"0.297",
"0.4258",
"0.297",
"0.4382",
"0.272",
"0.4382"
],
[
"0.614",
"0.4287",
"0.825",
"0.4329",
"0.825",
"0.4548",
"0.614",
"0.4507"
],
[
"0.858",
"0.4233",
"0.885",
"0.4233",
"0.885",
"0.437",
"0.858",
"0.437"
],
[
"0.1797",
"0.4407",
"0.4224",
"0.4407",
"0.4224",
"0.4578",
"0.1797",
"0.4578"
],
[
"0.4626",
"0.4412",
"0.4878",
"0.4412",
"0.4878",
"0.4548",
"0.4626",
"0.4548"
],
[
"0.629",
"0.4531",
"0.6533",
"0.4531",
"0.6533",
"0.4614",
"0.629",
"0.4614"
],
[
"0.272",
"0.4602",
"0.2957",
"0.4602",
"0.2957",
"0.4685",
"0.272",
"0.4685"
],
[
"0.11755",
"0.4827",
"0.4895",
"0.4827",
"0.4895",
"0.4976",
"0.11755",
"0.4976"
],
[
"0.511",
"0.481",
"0.867",
"0.481",
"0.867",
"0.4954",
"0.511",
"0.4954"
],
[
"0.529",
"0.4995",
"0.881",
"0.4995",
"0.881",
"0.513",
"0.529",
"0.513"
],
[
"0.1142",
"0.509",
"0.3267",
"0.5103",
"0.3267",
"0.525",
"0.1142",
"0.524"
],
[
"0.511",
"0.5166",
"0.8833",
"0.5166",
"0.8833",
"0.53",
"0.511",
"0.53"
],
[
"0.11755",
"0.531",
"0.465",
"0.53",
"0.465",
"0.5444",
"0.11755",
"0.546"
],
[
"0.51",
"0.532",
"0.882",
"0.533",
"0.882",
"0.5464",
"0.51",
"0.546"
],
[
"0.11755",
"0.549",
"0.4902",
"0.5464",
"0.4902",
"0.561",
"0.11755",
"0.5635"
],
[
"0.511",
"0.549",
"0.8833",
"0.549",
"0.8833",
"0.5625",
"0.511",
"0.5625"
],
[
"0.11755",
"0.5654",
"0.4902",
"0.5654",
"0.4902",
"0.579",
"0.11755",
"0.579"
],
[
"0.513",
"0.5664",
"0.8423",
"0.5664",
"0.8423",
"0.5767",
"0.513",
"0.5767"
],
[
"0.11584",
"0.579",
"0.4878",
"0.58",
"0.4878",
"0.594",
"0.11584",
"0.5933"
],
[
"0.511",
"0.583",
"0.882",
"0.583",
"0.882",
"0.5967",
"0.511",
"0.5967"
],
[
"0.11584",
"0.5967",
"0.4878",
"0.5967",
"0.4878",
"0.611",
"0.11584",
"0.611"
],
[
"0.51",
"0.597",
"0.882",
"0.5996",
"0.882",
"0.6133",
"0.51",
"0.611"
],
[
"0.11755",
"0.6123",
"0.4878",
"0.6123",
"0.4878",
"0.627",
"0.11755",
"0.627"
],
[
"0.51",
"0.615",
"0.882",
"0.615",
"0.882",
"0.629",
"0.51",
"0.629"
],
[
"0.11755",
"0.629",
"0.4878",
"0.629",
"0.4878",
"0.6436",
"0.11755",
"0.6436"
],
[
"0.51",
"0.6313",
"0.882",
"0.6313",
"0.882",
"0.6445",
"0.51",
"0.6445"
],
[
"0.11584",
"0.6445",
"0.3284",
"0.6445",
"0.3284",
"0.6577",
"0.11584",
"0.6577"
],
[
"0.511",
"0.648",
"0.8833",
"0.648",
"0.8833",
"0.661",
"0.511",
"0.661"
],
[
"0.51",
"0.662",
"0.8833",
"0.663",
"0.8833",
"0.6777",
"0.51",
"0.677"
],
[
"0.1704",
"0.6733",
"0.2957",
"0.673",
"0.2957",
"0.6885",
"0.1704",
"0.6895"
],
[
"0.511",
"0.6797",
"0.882",
"0.6797",
"0.882",
"0.6934",
"0.511",
"0.6934"
],
[
"0.2217",
"0.6895",
"0.4248",
"0.695",
"0.424",
"0.7163",
"0.22",
"0.7114"
],
[
"0.4634",
"0.695",
"0.4895",
"0.695",
"0.4895",
"0.707",
"0.4634",
"0.707"
],
[
"0.511",
"0.6953",
"0.8833",
"0.6953",
"0.8833",
"0.709",
"0.511",
"0.709"
],
[
"0.3074",
"0.7153",
"0.339",
"0.7153",
"0.339",
"0.7324",
"0.3074",
"0.7324"
],
[
"0.513",
"0.712",
"0.8833",
"0.712",
"0.8833",
"0.727",
"0.513",
"0.727"
],
[
"0.511",
"0.728",
"0.882",
"0.728",
"0.882",
"0.741",
"0.511",
"0.741"
],
[
"0.11584",
"0.742",
"0.3015",
"0.742",
"0.3015",
"0.7573",
"0.11584",
"0.7573"
],
[
"0.51",
"0.742",
"0.881",
"0.742",
"0.881",
"0.7573",
"0.51",
"0.7573"
],
[
"0.1351",
"0.7583",
"0.4902",
"0.759",
"0.4902",
"0.774",
"0.1351",
"0.7725"
],
[
"0.511",
"0.7603",
"0.882",
"0.7603",
"0.882",
"0.774",
"0.511",
"0.774"
],
[
"0.1142",
"0.775",
"0.486",
"0.775",
"0.486",
"0.788",
"0.1142",
"0.788"
],
[
"0.51",
"0.7754",
"0.882",
"0.7754",
"0.882",
"0.7905",
"0.51",
"0.7905"
],
[
"0.11755",
"0.792",
"0.4878",
"0.792",
"0.4878",
"0.8057",
"0.11755",
"0.8057"
],
[
"0.511",
"0.7905",
"0.881",
"0.7905",
"0.881",
"0.8047",
"0.511",
"0.8047"
],
[
"0.1184",
"0.8086",
"0.4878",
"0.8086",
"0.4878",
"0.8193",
"0.1184",
"0.8193"
],
[
"0.511",
"0.808",
"0.882",
"0.808",
"0.882",
"0.822",
"0.511",
"0.822"
],
[
"0.11584",
"0.8223",
"0.4895",
"0.8237",
"0.4895",
"0.8384",
"0.11584",
"0.8374"
],
[
"0.51",
"0.8247",
"0.8833",
"0.8247",
"0.8833",
"0.8384",
"0.51",
"0.8384"
],
[
"0.11584",
"0.8403",
"0.4895",
"0.839",
"0.4895",
"0.854",
"0.11584",
"0.855"
],
[
"0.509",
"0.8403",
"0.882",
"0.839",
"0.882",
"0.854",
"0.509",
"0.855"
],
[
"0.11584",
"0.857",
"0.4895",
"0.857",
"0.4895",
"0.8706",
"0.11584",
"0.8706"
],
[
"0.509",
"0.8555",
"0.8833",
"0.855",
"0.8833",
"0.869",
"0.509",
"0.8706"
],
[
"0.11584",
"0.8726",
"0.4878",
"0.8726",
"0.4878",
"0.8857",
"0.11584",
"0.8857"
],
[
"0.51",
"0.8726",
"0.882",
"0.8726",
"0.882",
"0.8857",
"0.51",
"0.8857"
],
[
"0.1142",
"0.8896",
"0.4902",
"0.8896",
"0.4902",
"0.903",
"0.1142",
"0.903"
],
[
"0.511",
"0.8896",
"0.8833",
"0.8896",
"0.8833",
"0.903",
"0.511",
"0.903"
],
[
"0.11584",
"0.905",
"0.4878",
"0.905",
"0.4878",
"0.9185",
"0.11584",
"0.9185"
],
[
"0.511",
"0.905",
"0.882",
"0.905",
"0.882",
"0.9194",
"0.511",
"0.9194"
]
] | [
"Positive Feedback",
"KnowledgeBas",
"St",
"al",
"aa3",
":",
"a4",
"Retriever",
"Generator",
"",
"[Entity Scoring",
"Top-K Entities",
"v2",
"e1",
"25",
"12",
"12.",
"Encg",
"Encr",
"b",
"αl? α3 t",
"",
"",
"13",
"13",
"p(ye|Xr,1)",
"e3",
"i",
"Enc,(xt.1)",
"Xt,i.",
"p(yt|Xt2) :",
"Decg",
"+",
"e",
"吃#",
"",
"Encg(xt,2)",
"12",
"造唱唱馆",
"Inner Product",
"",
"Xt",
" p(yejxtr):",
"ek限呢",
"",
"Encg(xt,)",
"p(ytlxt)",
"P",
"",
"Dialogue Cont",
"Enc,",
"Negative Sample",
"EntityScoring",
"",
"S.L",
"E",
"Negative Feedback",
"Figure 2: The overall architecture of our end-to-end task-oriented dialogue system, comprising a knowledge retriever",
"and a response generator. The retrieval component is trained using dual feedback from the response generator.",
"well ascross-attentionto the input tokens repre-",
"between the retrieval scores St = [st,i}1<i<K ob-",
"sented by Xt.This enables the decoderto consider",
"tained from Eq. (1) and the pseudo-labels Gt using",
"information from both the generated tokens and the",
"KL-divergence as follows:",
"tokens of retrievedentities.Theprobabilityof the",
"response is defined as follows:",
"Lpos = DKL(Gt, St)",
"R",
"lytl",
"gt,;(log gt,i -log St,),",
"(4)",
"p(yt/Xt) = IIp(yt,;lyt,<i, Xt),",
"(2)",
"i=1",
"=1",
"where [yt| represents the length of the response yt.",
"where gt.i = softmax(Gt)i, st., = softmax(St)i.",
"The use of the generator's supervision to train the",
"3.4End-to-End Training",
"retriever is referred to as positive feedback. How-",
"Entity scoring Given the retrieved entities ",
"ever, it is important to note that the generator may",
"[ei, e2, .., ek ], we employ the aforementioned re",
"assimilate incorrect knowledge from the entities,",
"sponse generator to assign a score to each entity.",
"leadingtoinaccurateentityrelevancyscores.",
"Firstly, the dialogue context ct and each entity ei",
"Negative feedback To address the issue of positive",
"are concatenated to form &t,i, which is then fed",
"feedback when training the retriever, we propose",
"into the generator. The conditional log probability",
"incorporating negative feedback from the generator",
"(length-normalized) of the response yt is utilized",
"to calibrate the pseudo-labels. To achieve this, we",
"to score entity e; as follows:",
"select a negative sample for the global represen-",
"tation Xt during the generation process. A neg-",
"9t,i = p(yt/Xt,i)",
"ative sample refers to a response that exhibits a",
"Ell log p(yt,;lyt,<i, Xt,)",
"(3)",
"high generation probability but is of low quality",
"lytl",
"Specifically, we generate a set [yt)<i<M of re-",
"sponses using beam search, where M is the beam",
"where Xt,i = Encg(ct,i)",
" search size, and rank them based on their genera-",
"The rationale behind the scoring is straightfor-",
"tion probabilities. We use R (yt) to represent the",
"ward: if an entity is pertinent to the response, the",
"rank of candidate response yt. Note that a high",
"probability of generating this response given the",
"generation probability does not guarantee a high.",
"dialogue context and theentity shouldalso behigh.",
"quality response. To evaluate the quality, we define",
"Positive feedback We utilize the entity scores",
"a function o(yt, yt) that measures the overlap be-",
"Gt = {gt,J1<i<K obtained from the generator",
" tween a response and the reference response. We",
"as pseudo-labels to train the retriever, with the ob-",
"implement o(yt,yt)using theBLEU metric.As",
"jective of transferring the knowledge acquired by",
"a result, we obtain a new sorted list of responses",
"the generator from reference responses to the re-",
"based on their true quality, where the rank of can-",
"triever. To achieve this, we enforce consistency",
"didate response yt is R(yt). Finally, we identify"
] | {
"arxiv_id": "2310.14528",
"language": "en",
"page_idx": 3
} |
||
2310.14528-en-4 | 2310.14528-en-4.jpg | the negative sample as follows:
$$y_t^-=\operatorname{argmin}_{\hat{y}_t^i}(R_t^g(\hat{y}_t^i)-R_t^o(\hat{y}_t^i)).$$
(5)
Intuitively, it is desirable for the retriever to avoid incorporating the signal from the negative sample when updating its parameters. Therefore, we employ Eq. (3) to calculate the score of each entity in generating the negative sample, and utilize the scores $G_t^-=\{g_{t,i}^-\}_{1\leq i\leq K}$ for all retrieved entities to calibrate the positive feedback by introducing a margin loss:
$$L_{\mathrm{neg}}=\max\left(0,-(D_{\mathrm{KL}}(G_{t}^{-},S_{t})-D_{\mathrm{KL}}(G_{t},S_{t}))+\eta\right).$$
(6)
The training objective of the retriever comprises $L_{\mathrm{pos}}$ and $L_\mathrm{neg}.$ Additionally, we train the response generator by minimizing the negative loglikelihood loss:
$$L_{\text{NLL}}=-\sum_{i=1}^{|y_t|}\log p(y_{t,i}|y_{t,\leq i},X_t).$$
(7)
The final training objective is a combination of the
objectives of the retriever and the generator:
$$L=L_{\mathrm{NLL}}+L_{\mathrm{pos}}+L_{\mathrm{neg}}.$$
(8)
## 4 Experiments
In this section, we provide a detailed description of
the experiment and present the main results.
### 4.1 Datasets
We conduct our evaluation using three publicly available TOD datasets: Multi-WOZ 2.1 (MWOZ) (Eric et al., 2020), Stanford Multi-Domain dataset (SMD) (Eric et al., 2017), and CamRest (Wen et al., 2017). These datasets contain dialogues that are associated with relevant KBs. This is referred to as the session-level KB. To construct a comprehensive and extensive knowledge base, we merge the session-level KBs corresponding to each dialog turn, resulting in a dataset-level KB. The provided train, validation, and test splits for each dataset are utilized in our evaluation. The statistics of the three datasets are summarized in Appendix A.
### 4.2 Evaluation Metrics
We evaluate the performance of our model using two metrics. Firstly, we calculate the top-$K$ retrieval recall (Re@K),which is inspired by the widely used top-$K$ retrieval accuracy in question
answering (Karpukhin et al., 2020). Re@ $K$ measures the effectiveness of the retriever by determining the percentage of gold attribute values covered by the entities retrieved in the top-$K$ list. Secondly, we assess the overall performance of our TOD system from two aspects. To evaluate the system's capability to generate relevant entities, we employ the micro Entity-F1 metric (Eric et al., 2017). Additionally, we utilize the BLEU metric to measure the N-gram overlap between the generated response and the reference response. With these metrics, we can comprehensively evaluate the performance of our model in both retrieval and TOD systems.
## 4.3 Experimental Settings
We instantiate the knowledge retriever using the BERT-base model. As for the generator, we instantiate it with T5 of varying model sizes: T5-base and T5-large. Both the retriever and generator models are fine-tuned using the ADAM algorithm (?) with different learning rate schedulers. The retriever model employs a fixed learning rate scheduler, while the generator model uses a linear learning rate scheduler. The experiments are conducted on a single 24G NVIDIA RTX 3090 GPU. We select the checkpoint that yields the best results on the validation set. For more detailed information regarding our experimental setup, please refer to Appendix B.
# 4.4 Baselines
In our comparison, we classify the existing approaches for E2E-TOD systems into three categories based on their utilization of KB.
Memory networks: These methods store external knowledge in memory cells in the form of triplets and utilize multi-hop attention to retrieve relevant information for generating responses. Examples include DSR (Wen et al., 2017), KBRetriever (Qin et al., 2019), GLMP (Wu et al., 2019), DF-Net (Qin et al., 2020), FG2Seq (He et al., 2020), and CDNET (Raghu et al., 2021).
Linearized KB: These approaches leverage pretrained language models to encode the entire linearized KB, along with the dialogue context, as input for assisting response generation. Examples include DialoKG (Rony et al., 2022), UnifiedSKG (Xie et al., 2022), and Q-TOD (Tian et al., 2022).
Model parameters: These approaches encode the KB into model parameters through data augmentation of the dialogues with KB records, enabling implicit retrieval during response genera- | [
[
"0.11584",
"0.0849",
"0.3477",
"0.0849",
"0.3477",
"0.0986",
"0.11584",
"0.0986"
],
[
"0.513",
"0.0861",
"0.882",
"0.0861",
"0.882",
"0.0998",
"0.513",
"0.0998"
],
[
"0.513",
"0.10394",
"0.8774",
"0.10394",
"0.8774",
"0.114",
"0.513",
"0.114"
],
[
"0.1704",
"0.114",
"0.4375",
"0.1134",
"0.4375",
"0.1312",
"0.1704",
"0.1318"
],
[
"0.4626",
"0.11755",
"0.4878",
"0.11755",
"0.4878",
"0.1289",
"0.4626",
"0.1289"
],
[
"0.492",
"0.12054",
"0.4995",
"0.12354",
"0.4963",
"0.1271",
"0.4895",
"0.12354"
],
[
"0.51",
"0.11816",
"0.882",
"0.11816",
"0.882",
"0.1318",
"0.51",
"0.1318"
],
[
"0.513",
"0.1342",
"0.882",
"0.1342",
"0.882",
"0.1478",
"0.513",
"0.1478"
],
[
"0.1335",
"0.1467",
"0.4895",
"0.1467",
"0.4895",
"0.1603",
"0.1335",
"0.1603"
],
[
"0.509",
"0.1497",
"0.882",
"0.1509",
"0.882",
"0.1644",
"0.509",
"0.1633"
],
[
"0.11584",
"0.1633",
"0.4895",
"0.1633",
"0.4895",
"0.177",
"0.11584",
"0.177"
],
[
"0.509",
"0.1663",
"0.882",
"0.1663",
"0.882",
"0.1799",
"0.509",
"0.1799"
],
[
"0.11584",
"0.1787",
"0.4902",
"0.1787",
"0.4902",
"0.1924",
"0.11584",
"0.1924"
],
[
"0.511",
"0.1829",
"0.882",
"0.1829",
"0.882",
"0.1965",
"0.511",
"0.1965"
],
[
"0.11755",
"0.1953",
"0.4878",
"0.1953",
"0.4878",
"0.209",
"0.11755",
"0.209"
],
[
"0.51",
"0.199",
"0.882",
"0.199",
"0.882",
"0.212",
"0.51",
"0.212"
],
[
"0.11755",
"0.2114",
"0.4902",
"0.2114",
"0.4902",
"0.2245",
"0.11755",
"0.2245"
],
[
"0.509",
"0.2131",
"0.882",
"0.2144",
"0.882",
"0.2286",
"0.509",
"0.228"
],
[
"0.11584",
"0.2257",
"0.4895",
"0.2257",
"0.4895",
"0.2434",
"0.11584",
"0.2434"
],
[
"0.509",
"0.2299",
"0.882",
"0.231",
"0.882",
"0.2452",
"0.509",
"0.2446"
],
[
"0.11755",
"0.2434",
"0.4902",
"0.2434",
"0.4902",
"0.257",
"0.11755",
"0.257"
],
[
"0.511",
"0.2477",
"0.882",
"0.2477",
"0.882",
"0.2612",
"0.511",
"0.2612"
],
[
"0.11755",
"0.2588",
"0.2795",
"0.2588",
"0.2795",
"0.2725",
"0.11755",
"0.2725"
],
[
"0.51",
"0.2625",
"0.881",
"0.2625",
"0.881",
"0.2756",
"0.51",
"0.2756"
],
[
"0.51",
"0.2778",
"0.8506",
"0.279",
"0.8506",
"0.2935",
"0.51",
"0.2922"
],
[
"0.126",
"0.2874",
"0.4702",
"0.2861",
"0.4702",
"0.3005",
"0.126",
"0.3018"
],
[
"0.4702",
"0.3018",
"0.486",
"0.3018",
"0.486",
"0.3113",
"0.4702",
"0.3113"
],
[
"0.513",
"0.3047",
"0.729",
"0.3047",
"0.729",
"0.3196",
"0.513",
"0.3196"
],
[
"0.1368",
"0.3154",
"0.4878",
"0.3154",
"0.4878",
"0.3289",
"0.1368",
"0.3289"
],
[
"0.2393",
"0.33",
"0.4895",
"0.3308",
"0.4895",
"0.3457",
"0.2393",
"0.3445"
],
[
"0.511",
"0.326",
"0.881",
"0.326",
"0.881",
"0.339",
"0.511",
"0.339"
],
[
"0.1184",
"0.332",
"0.2451",
"0.332",
"0.2451",
"0.3467",
"0.1184",
"0.3467"
],
[
"0.511",
"0.3403",
"0.882",
"0.3416",
"0.882",
"0.3557",
"0.511",
"0.3552"
],
[
"0.11584",
"0.3467",
"0.4878",
"0.3467",
"0.4878",
"0.3599",
"0.11584",
"0.3599"
],
[
"0.1142",
"0.3623",
"0.2351",
"0.3635",
"0.2351",
"0.3777",
"0.1142",
"0.3772"
],
[
"0.511",
"0.3594",
"0.882",
"0.3594",
"0.882",
"0.3723",
"0.511",
"0.3723"
],
[
"0.511",
"0.3748",
"0.882",
"0.3748",
"0.882",
"0.3884",
"0.511",
"0.3884"
],
[
"0.2585",
"0.3901",
"0.2805",
"0.3901",
"0.2805",
"0.401",
"0.2585",
"0.401"
],
[
"0.513",
"0.3901",
"0.882",
"0.3901",
"0.882",
"0.405",
"0.513",
"0.405"
],
[
"0.1721",
"0.405",
"0.2217",
"0.405",
"0.2217",
"0.4192",
"0.1721",
"0.4192"
],
[
"0.2627",
"0.4038",
"0.4307",
"0.4038",
"0.4307",
"0.4216",
"0.2627",
"0.4216"
],
[
"0.2627",
"0.4067",
"0.2703",
"0.412",
"0.2627",
"0.4175",
"0.2551",
"0.412"
],
[
"0.4634",
"0.4067",
"0.4895",
"0.4067",
"0.4895",
"0.4187",
"0.4634",
"0.4187"
],
[
"0.511",
"0.4067",
"0.882",
"0.4067",
"0.882",
"0.4204",
"0.511",
"0.4204"
],
[
"0.2585",
"0.4246",
"0.2778",
"0.4246",
"0.2778",
"0.4316",
"0.2585",
"0.4316"
],
[
"0.51",
"0.4229",
"0.882",
"0.4216",
"0.882",
"0.4358",
"0.51",
"0.437"
],
[
"0.511",
"0.4395",
"0.8833",
"0.4395",
"0.8833",
"0.4531",
"0.511",
"0.4531"
],
[
"0.11755",
"0.4453",
"0.4878",
"0.4453",
"0.4878",
"0.459",
"0.11755",
"0.459"
],
[
"0.511",
"0.4548",
"0.8833",
"0.4548",
"0.8833",
"0.4685",
"0.511",
"0.4685"
],
[
"0.11755",
"0.462",
"0.444",
"0.462",
"0.444",
"0.4756",
"0.11755",
"0.4756"
],
[
"0.51",
"0.4702",
"0.8833",
"0.4697",
"0.8833",
"0.484",
"0.51",
"0.485"
],
[
"0.2024",
"0.488",
"0.3972",
"0.4912",
"0.3972",
"0.509",
"0.2024",
"0.506"
],
[
"0.4634",
"0.4922",
"0.4895",
"0.4922",
"0.4895",
"0.505",
"0.4634",
"0.505"
],
[
"0.513",
"0.4868",
"0.882",
"0.4868",
"0.882",
"0.5005",
"0.513",
"0.5005"
],
[
"0.513",
"0.503",
"0.882",
"0.503",
"0.882",
"0.5166",
"0.513",
"0.5166"
],
[
"0.1436",
"0.5215",
"0.2585",
"0.5215",
"0.2585",
"0.536",
"0.1436",
"0.536"
],
[
"0.51",
"0.5195",
"0.881",
"0.5195",
"0.881",
"0.533",
"0.51",
"0.533"
],
[
"0.1234",
"0.5254",
"0.1301",
"0.5254",
"0.1301",
"0.532",
"0.1234",
"0.532"
],
[
"0.51",
"0.536",
"0.608",
"0.5337",
"0.608",
"0.549",
"0.51",
"0.5503"
],
[
"0.11755",
"0.549",
"0.4878",
"0.549",
"0.4878",
"0.5625",
"0.11755",
"0.5625"
],
[
"0.11584",
"0.5654",
"0.4434",
"0.5654",
"0.4434",
"0.579",
"0.11584",
"0.579"
],
[
"0.511",
"0.561",
"0.629",
"0.561",
"0.629",
"0.576",
"0.511",
"0.576"
],
[
"0.511",
"0.583",
"0.882",
"0.583",
"0.882",
"0.5967",
"0.511",
"0.5967"
],
[
"0.1184",
"0.5933",
"0.2275",
"0.5933",
"0.2275",
"0.605",
"0.1184",
"0.605"
],
[
"0.51",
"0.5986",
"0.8833",
"0.597",
"0.8833",
"0.6123",
"0.51",
"0.6133"
],
[
"0.1142",
"0.6123",
"0.486",
"0.6133",
"0.486",
"0.6274",
"0.1142",
"0.627"
],
[
"0.51",
"0.6143",
"0.7983",
"0.6133",
"0.7983",
"0.6274",
"0.51",
"0.629"
],
[
"0.11584",
"0.63",
"0.4878",
"0.63",
"0.4878",
"0.6436",
"0.11584",
"0.6436"
],
[
"0.529",
"0.6313",
"0.882",
"0.6313",
"0.882",
"0.6445",
"0.529",
"0.6445"
],
[
"0.11755",
"0.6455",
"0.4895",
"0.6455",
"0.4895",
"0.659",
"0.11755",
"0.659"
],
[
"0.51",
"0.6465",
"0.8833",
"0.6465",
"0.8833",
"0.661",
"0.51",
"0.661"
],
[
"0.11755",
"0.662",
"0.4937",
"0.662",
"0.4937",
"0.676",
"0.11755",
"0.676"
],
[
"0.511",
"0.663",
"0.8833",
"0.663",
"0.8833",
"0.677",
"0.511",
"0.677"
],
[
"0.11755",
"0.677",
"0.4895",
"0.6777",
"0.4895",
"0.6924",
"0.11755",
"0.6914"
],
[
"0.509",
"0.6797",
"0.882",
"0.6797",
"0.882",
"0.6934",
"0.509",
"0.6934"
],
[
"0.11755",
"0.695",
"0.4895",
"0.695",
"0.4895",
"0.708",
"0.11755",
"0.708"
],
[
"0.51",
"0.6953",
"0.882",
"0.6953",
"0.882",
"0.709",
"0.51",
"0.709"
],
[
"0.11584",
"0.709",
"0.492",
"0.7104",
"0.492",
"0.7246",
"0.11584",
"0.7236"
],
[
"0.513",
"0.7114",
"0.882",
"0.7114",
"0.882",
"0.7246",
"0.513",
"0.7246"
],
[
"0.11755",
"0.727",
"0.4878",
"0.727",
"0.4878",
"0.7407",
"0.11755",
"0.7407"
],
[
"0.511",
"0.7256",
"0.882",
"0.7256",
"0.882",
"0.7393",
"0.511",
"0.7393"
],
[
"0.11755",
"0.742",
"0.4878",
"0.742",
"0.4878",
"0.756",
"0.11755",
"0.756"
],
[
"0.513",
"0.7437",
"0.812",
"0.7437",
"0.812",
"0.7573",
"0.513",
"0.7573"
],
[
"0.11584",
"0.759",
"0.4878",
"0.759",
"0.4878",
"0.7725",
"0.11584",
"0.7725"
],
[
"0.5283",
"0.7583",
"0.882",
"0.7603",
"0.882",
"0.775",
"0.5283",
"0.7725"
],
[
"0.11755",
"0.777",
"0.486",
"0.777",
"0.486",
"0.7876",
"0.11755",
"0.7876"
],
[
"0.509",
"0.775",
"0.882",
"0.775",
"0.882",
"0.788",
"0.509",
"0.788"
],
[
"0.1184",
"0.792",
"0.486",
"0.792",
"0.486",
"0.8027",
"0.1184",
"0.8027"
],
[
"0.513",
"0.7915",
"0.882",
"0.7915",
"0.882",
"0.8057",
"0.513",
"0.8057"
],
[
"0.11755",
"0.808",
"0.4182",
"0.808",
"0.4182",
"0.822",
"0.11755",
"0.822"
],
[
"0.511",
"0.808",
"0.882",
"0.808",
"0.882",
"0.822",
"0.511",
"0.822"
],
[
"0.51",
"0.8237",
"0.882",
"0.8223",
"0.882",
"0.8374",
"0.51",
"0.8384"
],
[
"0.11755",
"0.835",
"0.3123",
"0.835",
"0.3123",
"0.8486",
"0.11755",
"0.8486"
],
[
"0.51",
"0.8403",
"0.875",
"0.8403",
"0.875",
"0.854",
"0.51",
"0.854"
],
[
"0.1142",
"0.855",
"0.4878",
"0.8555",
"0.4878",
"0.8706",
"0.1142",
"0.869"
],
[
"0.5283",
"0.8555",
"0.881",
"0.8555",
"0.881",
"0.869",
"0.5283",
"0.869"
],
[
"0.11584",
"0.8726",
"0.4895",
"0.8726",
"0.4895",
"0.8857",
"0.11584",
"0.8857"
],
[
"0.509",
"0.8706",
"0.882",
"0.8726",
"0.882",
"0.887",
"0.509",
"0.8853"
],
[
"0.11584",
"0.888",
"0.4895",
"0.888",
"0.4895",
"0.902",
"0.11584",
"0.902"
],
[
"0.51",
"0.888",
"0.8833",
"0.888",
"0.8833",
"0.903",
"0.51",
"0.903"
],
[
"0.11584",
"0.905",
"0.4895",
"0.905",
"0.4895",
"0.9185",
"0.11584",
"0.9185"
],
[
"0.51",
"0.905",
"0.8833",
"0.905",
"0.8833",
"0.9185",
"0.51",
"0.9185"
]
] | [
"the negative sample as follows:",
"answering (Karpukhin et al., 2020). Re@ K mea",
"sures theeffectiveness of the retrieverby determin",
"yt = argmingr(R(gt) - R(gt).",
"(5)",
"",
"ing the percentage of gold attribute values covered",
"by the entities retrieved in the top-K list. Secondly.",
"Intuitively, it is desirable for the retriever to",
"we assess the overall performance of our TOD sys-",
"avoid incorporating the signal from the negative",
"tem from two aspects. To evaluate the system's",
"sample when updating its parameters. Therefore,",
"capability to generate relevant entities, we employ",
"we employ Eq. (3) to calculate the score of each",
"the micro Entity-F1 metric (Eric et al., 2017). Ad-",
"entity in generating the negative sample, and uti-",
"ditionally, we utilize the BLEU metric to measure",
"lize the scores Gt = (gt,J1<i<k for all retrieved",
"the N-gram overlap between the generated response",
"entities to calibrate the positive feedback by intro-",
"and the reference response. With these metrics, we",
"ducing a margin loss:",
"can comprehensively evaluate the performance of",
"our model in both retrieval and TOD systems.",
"Lneg = max (0, -(DkL(GT, St) - DkL(Gt, St) + n)",
"(6)",
"4.3Experimental Settings",
"The training objective of the retriever comprises",
"Additionally, we train the re-",
"We instantiate the knowledge retriever using the",
"Lpos and Lneg.",
"BERT-base model. As for the generator, we instan-",
"sponsegeneratorby minimizing thenegative log-",
"likelihood loss:",
"tiate it with T5 of varying model sizes: T5-base",
"and T5-large. Both the retriever and generator mod-",
"ytl",
"els are fine-tuned using the ADAM algorithm (?)",
"LNLL",
"log p(yt, lyt,<i, Xt).",
"一",
"(7)",
"with different learning rate schedulers. The re-",
"=1",
"triever model employs a fixed learning rate sched-",
"uler, while the generator model uses a linear learn-",
"The final training objective is a combination of the",
"ing rate scheduler. The experiments are conducted",
"objectives of the retriever and the generator:",
"on a single 24G NVIDIA RTX 3090 GPU. We se-",
"L = LNLL + Lpos + Lneg",
"(8)",
"lect the checkpoint that yields the best results on",
"the validation set. For more detailed information",
"Experiments",
"regarding our experimental setup, please refer to",
"1",
"Appendix B.",
"In this section, we provide a detailed description of",
"the experiment and present the main results.",
"4.4 Baselines",
"In our comparison, we classify the existing ap-",
"4.1 Datasets",
"proaches for E2E-TOD systems into three cate-",
"We conduct our evaluation using three publicly",
"gories based on their utilization of KB.",
"available TOD datasets: Multi-WOZ 2.1 (MWOZ)",
"Memory networks: These methods store ex-",
"(Eric et al., 2020), Stanford Multi-Domain dataset",
"ternal knowledge in memory cells in the form of",
"(SMD) (Eric et al., 2017), and CamRest (Wen et al.,",
"triplets and utilize multi-hop attention to retrieve",
"2017). These datasets contain dialogues that are",
"relevant information for generating responses. Ex-",
"associated with relevant KBs. This is referred to",
"amples include DSR (Wen et al., 2017), KB",
"as the session-level KB. To construct a compre-",
"Retriever (Qin et al., 2019), GLMP (Wu et al.,",
"hensive and extensive knowledge base, we merge",
"2019), DF-Net (Oin et al.. 2020), FG2Seg (He et al..",
"the session-level KBs corresponding to each dialog",
"2020), and CDNET (Raghu et al., 2021).",
"turn, resulting in a dataset-level KB. The provided",
"Linearized KB: These approaches leverage pre-",
"train,validation, and test splits for each dataset are",
"trained language models to encode the entire lin-",
"utilized in our evaluation.The statistics of the three",
"earized KB, along with the dialogue context, as",
"datasets are summarized in Appendix A.",
"input for assisting response generation. Examples",
"include DialoKG (Rony et al., 2022), UnifiedSKG",
"4.2Evaluation Metrics",
"(Xie et al., 2022), and Q-TOD (Tian et al., 2022).",
"We evaluatethe performance of our model using",
"Model parameters: These approaches encode",
"two metrics. Firstly, we calculate the top-K re-",
"the KB into model parameters through data aug-",
"trieval recall (Re@ K), which is inspired by the",
"mentation of the dialogues with KB records, en-",
"widely used top-K retrieval accuracy in question",
"abling implicit retrieval during response genera-"
] | {
"arxiv_id": "2310.14528",
"language": "en",
"page_idx": 4
} |
||
1604.01146-en-1 | 1604.01146-en-1.jpg | ## $1. \textbf{Introduction}$
Unlike traditional object classification tasks in which the training and test categories are identical, zero-shot learning aims to recognize objects from classes not seen at the training stage. It is recognized as an effective way for large scale visual classification since it alleviates the burden of collecting sufficient training data for every possible class. The key component ensuring the success of zero-shot learning is to find an intermediate semantic representation to bridge the gap between seen and unseen classes. In a nutshell, with this semantic representation we can first learn its connection with image features and then transfer this connection to unseen classes. So once the semantic representation of an unseen class is given, one can easily classify the image through the learned connection.
Attributes, which essentially represent the discriminative properties shared among both seen and unseen categories, have become the most popular semantic representation in zero-shot learning [7,8,24,12,26]. Although the recent use of attributes has led to exciting advances in zero-shot learning [10,2,27],the creation of attributes still relies on much human labour. This is inevitably discouraging since the motivation for zero-shot learning is to free large-scale recognition tasks from cumbersome annotation requirements.
To remedy this drawback and move towards the goal of fully automatic zero-shot learning, several recent works [23,9,16] have explored the possibility of using the easily accessed online information sources to create the intermediate semantic representation. One possible choice is to directly use online textual documents, such as are found in Wikipedia, to build such a representation [6,3]. This is promising because online text documents can be easily obtained and contain rich information about the object. To conduct zero-shot learning with textual documents, existing works [2,10] develop various ways to measure the similarity between text and visual features. Our work is also based on this idea but we take a step further, however, to consider one additional important factor: the document representation is much more noisy than the human specified semantic representation and ignoring this fact will lead to inferior performance. For example, when the bag-of-words model is adopted as the document representation, the occurrence of every word in a document will trigger a signal in one dimension of the document representation. However, it is clear that most words in a document are not directly relevant for identifying the object category, thus it is necessary to design a noise suppression mechanism$^{\mathrm{l}}$ to down-weight the importance of the less relevant words for zero-shot learning.
Motivated by this consideration, we propose a zero-shot
l This mechanism is closely related to feature selection but is not exactly same. As will be discussed in the following sections, the solution of our method does not discard the less relevant dimensions of the document representation but only suppress their impact for zero-shot learning.
learning method which particularly caters for the need for noise suppression. More specifically, we proposed a simplebut-effective $l_2,1$-norm based objective function which simultaneously suppresses the noisy signal within text descriptions and learns a function to match the visual and text domains. We also develop an efficient optimization algorithm to solve this problem. By conducting experiments on two large scale zero-shot learning evaluation benchmarks, we demonstrate the benefit of the proposed noise suppression mechanism as well as its superior performance over other zero-shot learning methods which also rely on online information sources. In addition, we also conduct an indepth analysis of the proposed method which provides an insight as to what kinds of information within a document are useful for zero-shot learning.
## $2. \textbf{Related work}$
Most zero-shot learning approaches rely on human specified attributes. As one of the earliest attempt in zero-shot learning, Lampert et al. [12] adopted a set of attributes obtained from a psychology study. By learning probabilistic predictors of those attributes, they developed a framework to estimate the posterior of the test class. Later, a number of works has been proposed to improve the way of learning the connnection between attributes and object categories. For example, the work in [11] addresses unreliability of attributes by exploring the idea of random forest. The work in [1] turns the zero-shot learning into a cross-domain matching problem and they propose to learn a matching function to compare the attribute and the image feature. Following the same idea, Romera [20] proposes a simpler but more effective objective function to learn the matching function. Zhang et $al.[27]$ advocates the benefits of using attributeattribute relationship, called semantic similarity, as the intermediate semantic representation and they learn a function to match the image features with the semantic similarity.
To go beyond the human specified attributes, recent works also explore the use of other form of semantic representations which can be easily obtained[14,2,9,10]. For example, the co-ocurrence statistic of words has been explored in [14, 2] to capture the semantic relateness of two concepts. The distributed word representation e.g. word2vec, has been utilized as a substitution of attributes [9] and more recently the word2vec representation has been shown to be complementary to the human specified attributes [10].
Another information source for creating the semantic representation is the online textual document, such as Wikipedia articles. In an earlier work, Berg et al.[4] attempts to discover attribute representation from a noisy web source by ranking the visualness scores of attribute candidates. Rohrbach $etal.[18,19]$ mines semantic relatedness for attribute-class association from different internet | [
[
"0.0809",
"0.09155",
"0.2075",
"0.09155",
"0.2075",
"0.1048",
"0.0809",
"0.1048"
],
[
"0.5024",
"0.09216",
"0.892",
"0.09216",
"0.892",
"0.1061",
"0.5024",
"0.1061"
],
[
"0.5024",
"0.108",
"0.892",
"0.108",
"0.892",
"0.1218",
"0.5024",
"0.1218"
],
[
"0.10046",
"0.11615",
"0.4697",
"0.11615",
"0.4697",
"0.13",
"0.10046",
"0.13"
],
[
"0.5024",
"0.1237",
"0.892",
"0.1237",
"0.892",
"0.137",
"0.5024",
"0.137"
],
[
"0.0801",
"0.132",
"0.4697",
"0.132",
"0.4697",
"0.1459",
"0.0801",
"0.1459"
],
[
"0.501",
"0.1383",
"0.8936",
"0.1383",
"0.8936",
"0.1515",
"0.501",
"0.1515"
],
[
"0.0801",
"0.1477",
"0.468",
"0.1477",
"0.468",
"0.161",
"0.0801",
"0.161"
],
[
"0.5024",
"0.1547",
"0.892",
"0.1547",
"0.892",
"0.1654",
"0.5024",
"0.1654"
],
[
"0.0801",
"0.1622",
"0.4697",
"0.1622",
"0.4697",
"0.1755",
"0.0801",
"0.1755"
],
[
"0.5024",
"0.1686",
"0.8906",
"0.1686",
"0.8906",
"0.1818",
"0.5024",
"0.1818"
],
[
"0.0801",
"0.1768",
"0.468",
"0.1768",
"0.468",
"0.1901",
"0.0801",
"0.1901"
],
[
"0.5024",
"0.1831",
"0.892",
"0.1831",
"0.892",
"0.1963",
"0.5024",
"0.1963"
],
[
"0.0801",
"0.1932",
"0.4673",
"0.1932",
"0.4673",
"0.2039",
"0.0801",
"0.2039"
],
[
"0.5024",
"0.1989",
"0.892",
"0.1989",
"0.892",
"0.2122",
"0.5024",
"0.2122"
],
[
"0.0801",
"0.207",
"0.468",
"0.207",
"0.468",
"0.2203",
"0.0801",
"0.2203"
],
[
"0.5024",
"0.2134",
"0.892",
"0.2134",
"0.892",
"0.2267",
"0.5024",
"0.2267"
],
[
"0.0801",
"0.2229",
"0.4697",
"0.2229",
"0.4697",
"0.2361",
"0.0801",
"0.2361"
],
[
"0.5024",
"0.2291",
"0.892",
"0.2291",
"0.892",
"0.2424",
"0.5024",
"0.2424"
],
[
"0.0784",
"0.2374",
"0.468",
"0.2361",
"0.468",
"0.2507",
"0.0784",
"0.252"
],
[
"0.5024",
"0.2437",
"0.892",
"0.2437",
"0.892",
"0.2568",
"0.5024",
"0.2568"
],
[
"0.0809",
"0.254",
"0.4656",
"0.254",
"0.4656",
"0.2644",
"0.0809",
"0.2644"
],
[
"0.5034",
"0.26",
"0.8906",
"0.26",
"0.8906",
"0.2708",
"0.5034",
"0.2708"
],
[
"0.0801",
"0.2676",
"0.4697",
"0.2676",
"0.4697",
"0.281",
"0.0801",
"0.281"
],
[
"0.5034",
"0.276",
"0.8906",
"0.276",
"0.8906",
"0.286",
"0.5034",
"0.286"
],
[
"0.0784",
"0.2827",
"0.4697",
"0.2827",
"0.4697",
"0.2966",
"0.0784",
"0.2966"
],
[
"0.5024",
"0.289",
"0.8936",
"0.289",
"0.8936",
"0.303",
"0.5024",
"0.303"
],
[
"0.0801",
"0.2986",
"0.468",
"0.2986",
"0.468",
"0.3125",
"0.0801",
"0.3125"
],
[
"0.501",
"0.305",
"0.7197",
"0.305",
"0.7197",
"0.318",
"0.501",
"0.318"
],
[
"0.0809",
"0.3145",
"0.2893",
"0.3145",
"0.2893",
"0.3245",
"0.0809",
"0.3245"
],
[
"0.0989",
"0.3276",
"0.4697",
"0.327",
"0.4697",
"0.3416",
"0.0989",
"0.3423"
],
[
"0.5024",
"0.3333",
"0.636",
"0.3333",
"0.636",
"0.3467",
"0.5024",
"0.3467"
],
[
"0.0801",
"0.3435",
"0.468",
"0.3435",
"0.468",
"0.3567",
"0.0801",
"0.3567"
],
[
"0.0801",
"0.358",
"0.4697",
"0.358",
"0.4697",
"0.3718",
"0.0801",
"0.3718"
],
[
"0.5215",
"0.358",
"0.892",
"0.358",
"0.892",
"0.3726",
"0.5215",
"0.3726"
],
[
"0.0801",
"0.3726",
"0.4697",
"0.3726",
"0.4697",
"0.3865",
"0.0801",
"0.3865"
],
[
"0.501",
"0.3726",
"0.8906",
"0.3726",
"0.8906",
"0.3865",
"0.501",
"0.3865"
],
[
"0.0801",
"0.3882",
"0.4697",
"0.3882",
"0.4697",
"0.4016",
"0.0801",
"0.4016"
],
[
"0.5024",
"0.3894",
"0.8936",
"0.3894",
"0.8936",
"0.4028",
"0.5024",
"0.4028"
],
[
"0.0801",
"0.4028",
"0.468",
"0.4028",
"0.468",
"0.4167",
"0.0801",
"0.4167"
],
[
"0.5024",
"0.404",
"0.892",
"0.404",
"0.892",
"0.4172",
"0.5024",
"0.4172"
],
[
"0.0784",
"0.4185",
"0.468",
"0.4185",
"0.468",
"0.432",
"0.0784",
"0.432"
],
[
"0.505",
"0.4204",
"0.8906",
"0.4204",
"0.8906",
"0.4312",
"0.505",
"0.4312"
],
[
"0.0801",
"0.4343",
"0.468",
"0.4343",
"0.468",
"0.4475",
"0.0801",
"0.4475"
],
[
"0.5024",
"0.4343",
"0.892",
"0.4343",
"0.892",
"0.4475",
"0.5024",
"0.4475"
],
[
"0.0784",
"0.4475",
"0.4314",
"0.449",
"0.4314",
"0.4622",
"0.0784",
"0.461"
],
[
"0.5024",
"0.449",
"0.892",
"0.449",
"0.892",
"0.4622",
"0.5024",
"0.4622"
],
[
"0.0972",
"0.464",
"0.468",
"0.464",
"0.468",
"0.4778",
"0.0972",
"0.4778"
],
[
"0.5024",
"0.464",
"0.8906",
"0.464",
"0.8906",
"0.4778",
"0.5024",
"0.4778"
],
[
"0.0801",
"0.4792",
"0.4714",
"0.4792",
"0.4714",
"0.4924",
"0.0801",
"0.4924"
],
[
"0.5024",
"0.4792",
"0.892",
"0.4792",
"0.892",
"0.4924",
"0.5024",
"0.4924"
],
[
"0.0825",
"0.4956",
"0.4673",
"0.4956",
"0.4673",
"0.506",
"0.0825",
"0.506"
],
[
"0.5034",
"0.4956",
"0.8906",
"0.4956",
"0.8906",
"0.506",
"0.5034",
"0.506"
],
[
"0.0801",
"0.5103",
"0.4673",
"0.5103",
"0.4673",
"0.521",
"0.0801",
"0.521"
],
[
"0.501",
"0.509",
"0.892",
"0.5083",
"0.892",
"0.523",
"0.501",
"0.5234"
],
[
"0.0801",
"0.5234",
"0.4697",
"0.5234",
"0.4697",
"0.537",
"0.0801",
"0.537"
],
[
"0.5024",
"0.5244",
"0.892",
"0.5244",
"0.892",
"0.5386",
"0.5024",
"0.5386"
],
[
"0.0801",
"0.539",
"0.4697",
"0.539",
"0.4697",
"0.553",
"0.0801",
"0.553"
],
[
"0.501",
"0.539",
"0.892",
"0.5405",
"0.892",
"0.5537",
"0.501",
"0.553"
],
[
"0.0801",
"0.5537",
"0.4714",
"0.5537",
"0.4714",
"0.5674",
"0.0801",
"0.5674"
],
[
"0.5024",
"0.5547",
"0.892",
"0.5547",
"0.892",
"0.5684",
"0.5024",
"0.5684"
],
[
"0.0784",
"0.5693",
"0.4673",
"0.5684",
"0.4673",
"0.582",
"0.0784",
"0.5835"
],
[
"0.5024",
"0.5693",
"0.892",
"0.5693",
"0.892",
"0.5835",
"0.5024",
"0.5835"
],
[
"0.0801",
"0.586",
"0.468",
"0.586",
"0.468",
"0.5967",
"0.0801",
"0.5967"
],
[
"0.501",
"0.584",
"0.8906",
"0.584",
"0.8906",
"0.5977",
"0.501",
"0.5977"
],
[
"0.0784",
"0.5996",
"0.4697",
"0.5996",
"0.4697",
"0.613",
"0.0784",
"0.613"
],
[
"0.5024",
"0.5996",
"0.892",
"0.5996",
"0.892",
"0.613",
"0.5024",
"0.613"
],
[
"0.0809",
"0.616",
"0.4673",
"0.616",
"0.4673",
"0.627",
"0.0809",
"0.627"
],
[
"0.501",
"0.6157",
"0.8936",
"0.6157",
"0.8936",
"0.629",
"0.501",
"0.629"
],
[
"0.0801",
"0.63",
"0.4697",
"0.63",
"0.4697",
"0.643",
"0.0801",
"0.643"
],
[
"0.5024",
"0.63",
"0.8735",
"0.63",
"0.8735",
"0.643",
"0.5024",
"0.643"
],
[
"0.0809",
"0.6465",
"0.4697",
"0.6465",
"0.4697",
"0.657",
"0.0809",
"0.657"
],
[
"0.5215",
"0.6445",
"0.8936",
"0.643",
"0.8936",
"0.6577",
"0.5215",
"0.659"
],
[
"0.0801",
"0.66",
"0.468",
"0.66",
"0.468",
"0.674",
"0.0801",
"0.674"
],
[
"0.5024",
"0.66",
"0.8906",
"0.66",
"0.8906",
"0.674",
"0.5024",
"0.674"
],
[
"0.0784",
"0.675",
"0.4697",
"0.675",
"0.4697",
"0.688",
"0.0784",
"0.688"
],
[
"0.5034",
"0.677",
"0.8906",
"0.677",
"0.8906",
"0.6875",
"0.5034",
"0.6875"
],
[
"0.0801",
"0.69",
"0.4697",
"0.69",
"0.4697",
"0.704",
"0.0801",
"0.704"
],
[
"0.5024",
"0.69",
"0.892",
"0.69",
"0.892",
"0.704",
"0.5024",
"0.704"
],
[
"0.0801",
"0.705",
"0.4697",
"0.705",
"0.4697",
"0.7183",
"0.0801",
"0.7183"
],
[
"0.5024",
"0.7056",
"0.8936",
"0.7056",
"0.8936",
"0.7197",
"0.5024",
"0.7197"
],
[
"0.0784",
"0.72",
"0.4714",
"0.72",
"0.4714",
"0.7344",
"0.0784",
"0.7344"
],
[
"0.501",
"0.7197",
"0.892",
"0.72",
"0.892",
"0.735",
"0.501",
"0.7344"
],
[
"0.0809",
"0.7373",
"0.4673",
"0.7373",
"0.4673",
"0.7476",
"0.0809",
"0.7476"
],
[
"0.505",
"0.7373",
"0.8906",
"0.7373",
"0.8906",
"0.7476",
"0.505",
"0.7476"
],
[
"0.0801",
"0.7505",
"0.4697",
"0.7505",
"0.4697",
"0.7646",
"0.0801",
"0.7646"
],
[
"0.505",
"0.7524",
"0.8906",
"0.7524",
"0.8906",
"0.763",
"0.505",
"0.763"
],
[
"0.0809",
"0.7676",
"0.468",
"0.7676",
"0.468",
"0.778",
"0.0809",
"0.778"
],
[
"0.5024",
"0.7666",
"0.892",
"0.7666",
"0.892",
"0.78",
"0.5024",
"0.78"
],
[
"0.0801",
"0.78",
"0.468",
"0.78",
"0.468",
"0.7935",
"0.0801",
"0.7935"
],
[
"0.5034",
"0.782",
"0.5884",
"0.782",
"0.5884",
"0.7935",
"0.5034",
"0.7935"
],
[
"0.0784",
"0.797",
"0.4697",
"0.797",
"0.4697",
"0.81",
"0.0784",
"0.81"
],
[
"0.5215",
"0.797",
"0.892",
"0.797",
"0.892",
"0.81",
"0.5215",
"0.81"
],
[
"0.0801",
"0.811",
"0.433",
"0.811",
"0.433",
"0.8247",
"0.0801",
"0.8247"
],
[
"0.5024",
"0.811",
"0.892",
"0.811",
"0.892",
"0.8247",
"0.5024",
"0.8247"
],
[
"0.0989",
"0.827",
"0.4697",
"0.827",
"0.4697",
"0.8403",
"0.0989",
"0.8403"
],
[
"0.5024",
"0.827",
"0.8906",
"0.827",
"0.8906",
"0.8403",
"0.5024",
"0.8403"
],
[
"0.5024",
"0.8413",
"0.892",
"0.8413",
"0.892",
"0.855",
"0.5024",
"0.855"
],
[
"0.1054",
"0.851",
"0.468",
"0.851",
"0.468",
"0.8643",
"0.1054",
"0.8643"
],
[
"0.5024",
"0.857",
"0.892",
"0.857",
"0.892",
"0.8706",
"0.5024",
"0.8706"
],
[
"0.0801",
"0.863",
"0.4714",
"0.863",
"0.4714",
"0.877",
"0.0801",
"0.877"
],
[
"0.0801",
"0.8755",
"0.4697",
"0.8755",
"0.4697",
"0.8896",
"0.0801",
"0.8896"
],
[
"0.501",
"0.872",
"0.892",
"0.872",
"0.892",
"0.8853",
"0.501",
"0.8853"
],
[
"0.0784",
"0.888",
"0.437",
"0.888",
"0.437",
"0.9014",
"0.0784",
"0.9014"
],
[
"0.5024",
"0.887",
"0.8936",
"0.887",
"0.8936",
"0.901",
"0.5024",
"0.901"
]
] | [
"1.Introduction",
"learning method which particularly caters for the need for",
"noise suppression. More specifically, we proposed a simple-",
"Unlike traditional object classification tasks in which the",
"but-effective l2,1-norm based objective function which si-",
"training and test categories are identical, zero-shot learning",
"multaneously suppresses the noisy signal within text de-",
"aims to recognize objects from classes not seen at the train-",
"scriptions and learns a function tomatch the visual and text",
"ing stage. It is recognized as an effective way for large scale",
"domains. We also develop an efficient optimization algo-",
"visual classification since it alleviates the burden of collect-",
"rithm to solve this problem. By conducting experiments on",
"ingsufficienttrainingdataforeverypossibleclass.Thekey",
"two large scale zero-shot learning evaluation benchmarks,",
"component ensuring the success of zero-shot learning is to",
"we demonstrate the benefit of the proposed noise suppres-",
"find an intermediate semantic representation to bridge the",
"sion mechanism as well as its superior performance over",
"gap between seen and unseen classes. In a nutshell, with",
"other zero-shot learning methods which also rely on online",
"thissemanticrepresentationwe canfirstlearn its connec-",
"information sources.In addition,wealso conduct an in-",
"tion with image features and then transfer this connection",
"depthanalysisoftheproposedmethodwhichprovides an",
"to unseen classes. So once the semantic representation of",
"insight as to what kinds of information within a document",
"an unseen class is given, one can easily classify the image",
"are useful for zero-shot learning.",
"throughthelearnedconnection.",
"Attributes, which essentially represent the discriminative",
"2.Relatedwork",
"properties shared among both seen and unseen categories,",
"have become the most popular semantic representation in",
"Most zero-shot learning approaches rely on human spec-",
"zero-shot learning [7, 8, 24, 12, 26]. Although the recent use",
"ified attributes. As one of the earliest attempt in zero-shot",
"of attributes has led to exciting advances in zero-shot learn-",
"learning, Lampert et al. [12] adopted a set of attributes ob-",
"ing [10, 2, 27], the creation of attributes still relies on much",
"tained from a psychology study. By learning probabilistic",
"human labour. This is inevitably discouraging since the mo-",
"predictorsofthoseattributes,theydevelopedaframework",
"tivation for zero-shot learning is to free large-scale recogni-",
"to estimate the posterior of the test class. Later, a number",
"tion tasks from cumbersome annotation requirements.",
"of works has been proposed to improve the way of learning",
"To remedy this drawback and move towards the goal",
"the connnection between attributes and object categories.",
"of fully automatic zero-shot learning, several recent works",
"For example, the work in [11] addresses unreliability of at-",
"[23.9.16]haveexplored thepossibilityofusingtheeas-",
"tributesbyexploringtheideaof randomforest.Theworkin",
"ily accessed online information sources to create the inter-",
"[1] turns the zero-shot learning into a cross-domain match-",
"mediate semantic representation. One possible choice is to",
"ing problem and they propose to learn a matching function",
"directly use online textual documents, such as are found in",
"to compare the attribute and the image feature.Following",
"Wikipedia, to build such a representation [6, 3]. This is",
"the same idea, Romera [20] proposes a simpler but more",
"promising because online text documents can be easily ob-",
"effective objective function to learn the matching function.",
"tained and contain rich information about the object.To",
"Zhang et al. [27] advocates the benefits of using attribute-",
"conduct zero-shot learning with textual documents, existing",
"attribute relationship, called semantic similarity, as the in-",
"works [2, 10] develop various ways to measure the similar-",
"termediate semantic representation and they learn a function",
"ity between text and visual features. Our work is also based",
"to match the image features with the semantic similarity.",
"on this idea but we take a stepfurther,however,to consider",
"To go beyond the human specified attributes, recent",
"one additional important factor: the document representa-",
"works also explore the use of other form of semantic rep-",
"tion is much more noisy than the human specified seman-",
"resentations which can be easily obtained [14, 2, 9,10].",
"tic representation and ignoring this fact will lead to inferior",
"For example, the co-ocurrence statistic of words has been",
"performance. For example, when the bag-of-words model is",
"explored in [14, 2] to capture the semantic relateness of",
" adopted as the document representation, the occurrence of",
"two concepts.The distributed word representation e.g.",
"every word in a document will triggera signal in one dimen-",
"word2vec.has been utilized as a substitution of attributes",
"sion of the document representation. However, it is clear",
"[9]andmorerecentlytheword2vecrepresentationhasbeen",
"thatmost words in adocument are not directlyrelevantfor",
"shown to be complementary to the human specified at-",
"identifying the object category, thus it is necessary to design",
"tributes [10].",
"a noise suppression mechanism to down-weight the impor-",
"Another information source for creating the seman-",
"tance of the less relevant words for zero-shot learning.",
"tic representation is the online textual document, such as",
"Motivated by this consideration, we propose a zero-shot",
"Wikipedia articles. In an earlier work, Berg et al. [4] at-",
"tempts to discover attribute representation from a noisy web",
"IThis mechanism is closely related to feature selection but is not ex-",
"source by ranking the visualness scores of attribute can-",
"actly same. As will be discussed in the following sections, the solution of",
"our method does not discard the less relevant dimensions of the document",
"didates. Rohrbach et al. [18, 19] mines semantic related-",
"representation but only suppress their impact for zero-shot learning.",
"ness for attribute-class association from different internet"
] | {
"arxiv_id": "1604.01146",
"language": "en",
"page_idx": 1
} |
||
1604.01146-en-2 | 1604.01146-en-2.jpg | sources. More recent works [6,3] directly learn a function to measure the compatibiliy between documents and visual features. However, compared with the state-of-the-art zeroshot learning methods, their performance seems to be disappointing even though some advanced technologies, such as deep learning, has been applied [3].
## 3.0ur Approach
3.1.0verview
The overview of our method is depicted in Figure 1. It starts with a raw document representation which is simply a binarized histogram of words. This representation is fed into our zero-shot learning algorithm to generate a classifier to detect relevant images. In the process of generating this classifier, the noise suppression regularizer in our method will automatically suppress the impact of less relevant words (illustrated as the red words in Figure 1).
## 3.2. Text representation
We extract our text representation based on a simple bagof-words model. We start by a preprocessing step of tokenizing the words and removing stop words and punctuations. Then a histogram of the remaining word occurrences is calculated and is subsequently binarized as the text representation. In other words, once a word appears in a document, its corresponding dimension within the text representation is set to“1”. One more commonly used choice for the text representation is based on TF-IDF as in [6,3]. However, we find it produces worse performance 2 than directly using the binarized representation. This is probably because the weighting calculated of TF-IDF is not suitable for our zero-shot learning although it is considered to be less noisy for applications like document classification. In the binarized histogram we essentialy treat each word in a document equally and this inevitably introduces a lot of noisy signals. However, thanks to our noise suppressing zero-shot learning algorithm, we can substantially down-weight the less relevant words and achieve good performance even with a noisy document representation.
3.3. Learning to match text and visual features
We frst formally define our problem and introduce the notation used in the following sections. At the training stage, both image features and document descriptions for $C$ seen categories are available. Let $\mathbf{X}\in\mathbb{R}^d\times N$ denote the image features of $N$ training examples and $\mathbf{Z}\in\{0,1\}^{\hat{d}\times C}$ the aforementioned document representations for $C$ seen classes, where $\hat{d}$ and $d$ are the dimensionality of the document representation and the image features respectively. We also define $\mathbf{Y}\in\{0,1\}^{N\times C}$ as the indicator matrix for
$^{2}$Using TD-IDF is about 7% and 5% inferior to binarized representa-
tions on AwA and CUB, respectively.
the $C$ seen classes. Each row of Y has a unique“1”indicating its corresponding class label. At the test stage, the document representations of the $\hat{C}$ unseen classes are given and our task is to assign $\hat{C}$ unseen class labels to the test $images.$
## 3.4. Formulation
Our method is inspired by a recently proposed zero-shot learning approach[20] which has demonstrated impressive performance despite a very simple learning process. More specifically, it learns a matrix V which optimizes the following objective function.
$$\min_{\mathbf{V}}\|\mathbf{X}^{\top}\mathbf{VS}-\mathbf{Y}\|_{F}^{2}+\lambda\|\mathbf{VS}\|_{F}^{2}+\gamma\|\mathbf{X}^{\top}\mathbf{V}\|_{F}^{2}+\lambda\gamma\|\mathbf{V}\|_{F}^{2}$$
(1)
where S denotes the semantic attribute matrix and it can be either a binary matrix or a real value matrix. The scalars $\gamma$ and $\lambda$ are weights controlling the prominence of the various terms. The underlying idea of this algorithm can be understood as follows. If the task is to classify X into $C$ categories, we can simply learn a linear classifier by fitting to $\mathbf{Y}$,that is, min$\mathbf{w}\|\mathbf{X}^\top\mathbf{W}-\mathbf{Y}\|_F^2.$ However, in this case W cannot be transferred to unseen classes. Thus we further impose that $\mathbf{W}=\mathbf{VS}.$ In other words, the classifier of a class is generated from its attributes. With this requirement, the classifier of an unseen class can be easily obtained and utilized to predict the category of a test image. Similarly, we can also treat $\mathbf{X}^{\top}\mathbf{V}$ as the classifier operated on the attributes S. The above understanding naturally gives rise to the regularization terms $\lambda\|\mathbf{VS}\|_F^2$ and $\gamma\|\mathbf{X}^\top\mathbf{V}\|_F^2$ which play the same role of the Frobenius norm regularizer as commonly introduced in multi-class classification or regression.
Since our document representation can also be seen as an attribute vector, the method in [20] can be readily applied to our problem by simply setting $\mathbf{S}=\mathbf{Z}.$ However, this naive solution ignores an important fact that the document representation is much more noisy than the human specified attribute vectors. To handle this issue, we introduce a noise suppression mechanism into Eq.(1). More specifically, we first decompose V into two terms:
$$\mathbf{V}=\mathbf{W_x}^\top\mathbf{W_z},$$
(2)
where $\mathbf{W_x\in\mathbb{R}^{m\times d}}$ and $\mathbf{W_z\in\mathbb{R}^{m\times\hat d}.}$ These two matrices will play different roles in our method. $\mathbf{W_z}$ is used to suppress the noisy components of Z and transform Z into a $m\times C$ intermediate representation. $\mathbf{W_x}$ is used to generate the image classifier from the noise-suppressed intermediate representation. Thus, two different regularization terms are imposed to suit these two different roles. The first term is the $l_{2,1}$-norm of $\mathbf{W_z}^\top$which achieves the noise suppression effect. The second term is the Frobenius norm of | [
[
"0.0801",
"0.09216",
"0.4697",
"0.09216",
"0.4697",
"0.1061",
"0.0801",
"0.1061"
],
[
"0.5024",
"0.09155",
"0.8906",
"0.09155",
"0.8906",
"0.1061",
"0.5024",
"0.1061"
],
[
"0.0801",
"0.108",
"0.468",
"0.108",
"0.468",
"0.1218",
"0.0801",
"0.1218"
],
[
"0.5024",
"0.108",
"0.892",
"0.108",
"0.892",
"0.1218",
"0.5024",
"0.1218"
],
[
"0.0809",
"0.125",
"0.4673",
"0.125",
"0.4673",
"0.1351",
"0.0809",
"0.1351"
],
[
"0.501",
"0.1218",
"0.892",
"0.1225",
"0.892",
"0.137",
"0.501",
"0.1364"
],
[
"0.0801",
"0.1383",
"0.4697",
"0.1383",
"0.4697",
"0.1515",
"0.0801",
"0.1515"
],
[
"0.5024",
"0.1383",
"0.892",
"0.1383",
"0.892",
"0.1515",
"0.5024",
"0.1515"
],
[
"0.0784",
"0.154",
"0.4697",
"0.154",
"0.4697",
"0.1674",
"0.0784",
"0.1674"
],
[
"0.5034",
"0.1547",
"0.558",
"0.1547",
"0.558",
"0.1666",
"0.5034",
"0.1666"
],
[
"0.0784",
"0.1686",
"0.3308",
"0.1686",
"0.3308",
"0.1818",
"0.0784",
"0.1818"
],
[
"0.5024",
"0.1755",
"0.6323",
"0.1755",
"0.6323",
"0.1901",
"0.5024",
"0.1901"
],
[
"0.0784",
"0.1945",
"0.2255",
"0.1945",
"0.2255",
"0.2122",
"0.0784",
"0.2122"
],
[
"0.523",
"0.1995",
"0.892",
"0.1995",
"0.892",
"0.2134",
"0.523",
"0.2134"
],
[
"0.0801",
"0.2197",
"0.1871",
"0.2197",
"0.1871",
"0.2343",
"0.0801",
"0.2343"
],
[
"0.501",
"0.2153",
"0.8906",
"0.2153",
"0.8906",
"0.2291",
"0.501",
"0.2291"
],
[
"0.5024",
"0.2311",
"0.892",
"0.2311",
"0.892",
"0.2443",
"0.5024",
"0.2443"
],
[
"0.0989",
"0.2437",
"0.4697",
"0.2437",
"0.4697",
"0.2568",
"0.0989",
"0.2568"
],
[
"0.5034",
"0.2456",
"0.8906",
"0.2456",
"0.8906",
"0.2588",
"0.5034",
"0.2588"
],
[
"0.0801",
"0.2588",
"0.468",
"0.2588",
"0.468",
"0.2727",
"0.0801",
"0.2727"
],
[
"0.5024",
"0.2615",
"0.6772",
"0.2615",
"0.6772",
"0.2747",
"0.5024",
"0.2747"
],
[
"0.0801",
"0.2734",
"0.468",
"0.2734",
"0.468",
"0.2874",
"0.0801",
"0.2874"
],
[
"0.5024",
"0.2827",
"0.898",
"0.281",
"0.898",
"0.3018",
"0.5024",
"0.3037"
],
[
"0.0801",
"0.289",
"0.468",
"0.289",
"0.468",
"0.303",
"0.0801",
"0.303"
],
[
"0.0784",
"0.303",
"0.468",
"0.3037",
"0.468",
"0.318",
"0.0784",
"0.3176"
],
[
"0.871",
"0.308",
"0.892",
"0.308",
"0.892",
"0.3213",
"0.871",
"0.3213"
],
[
"0.0801",
"0.3193",
"0.468",
"0.3193",
"0.468",
"0.3333",
"0.0801",
"0.3333"
],
[
"0.0784",
"0.3333",
"0.4673",
"0.334",
"0.4673",
"0.3484",
"0.0784",
"0.348"
],
[
"0.5034",
"0.3333",
"0.892",
"0.3333",
"0.892",
"0.3467",
"0.5034",
"0.3467"
],
[
"0.0801",
"0.3499",
"0.4224",
"0.3499",
"0.4224",
"0.363",
"0.0801",
"0.363"
],
[
"0.5024",
"0.3484",
"0.892",
"0.3484",
"0.892",
"0.3623",
"0.5024",
"0.3623"
],
[
"0.5024",
"0.363",
"0.892",
"0.3623",
"0.892",
"0.377",
"0.5024",
"0.3774"
],
[
"0.0784",
"0.3718",
"0.2615",
"0.3726",
"0.2615",
"0.387",
"0.0784",
"0.3865"
],
[
"0.5024",
"0.3774",
"0.892",
"0.3774",
"0.892",
"0.3914",
"0.5024",
"0.3914"
],
[
"0.0989",
"0.3945",
"0.468",
"0.3958",
"0.468",
"0.4104",
"0.0989",
"0.4092"
],
[
"0.5024",
"0.3933",
"0.892",
"0.3933",
"0.892",
"0.4072",
"0.5024",
"0.4072"
],
[
"0.0784",
"0.4104",
"0.468",
"0.411",
"0.468",
"0.4255",
"0.0784",
"0.4248"
],
[
"0.5024",
"0.4077",
"0.892",
"0.4077",
"0.892",
"0.4216",
"0.5024",
"0.4216"
],
[
"0.0801",
"0.4268",
"0.468",
"0.4268",
"0.468",
"0.44",
"0.0801",
"0.44"
],
[
"0.501",
"0.4224",
"0.8906",
"0.4224",
"0.8906",
"0.4375",
"0.501",
"0.4375"
],
[
"0.0801",
"0.4414",
"0.468",
"0.4414",
"0.468",
"0.455",
"0.0801",
"0.455"
],
[
"0.5034",
"0.4382",
"0.892",
"0.4382",
"0.892",
"0.452",
"0.5034",
"0.452"
],
[
"0.0801",
"0.457",
"0.468",
"0.457",
"0.468",
"0.4702",
"0.0801",
"0.4702"
],
[
"0.505",
"0.455",
"0.8906",
"0.455",
"0.8906",
"0.4653",
"0.505",
"0.4653"
],
[
"0.0801",
"0.473",
"0.4697",
"0.473",
"0.4697",
"0.486",
"0.0801",
"0.486"
],
[
"0.5034",
"0.4685",
"0.8906",
"0.4685",
"0.8906",
"0.4817",
"0.5034",
"0.4817"
],
[
"0.0801",
"0.4873",
"0.468",
"0.4873",
"0.468",
"0.5005",
"0.0801",
"0.5005"
],
[
"0.5024",
"0.4841",
"0.892",
"0.4841",
"0.892",
"0.4976",
"0.5024",
"0.4976"
],
[
"0.0801",
"0.5005",
"0.468",
"0.5005",
"0.468",
"0.515",
"0.0801",
"0.515"
],
[
"0.5034",
"0.5",
"0.8906",
"0.5",
"0.8906",
"0.513",
"0.5034",
"0.513"
],
[
"0.0801",
"0.5166",
"0.468",
"0.5166",
"0.468",
"0.53",
"0.0801",
"0.53"
],
[
"0.5024",
"0.5146",
"0.892",
"0.5146",
"0.892",
"0.528",
"0.5024",
"0.528"
],
[
"0.0801",
"0.532",
"0.468",
"0.532",
"0.468",
"0.5454",
"0.0801",
"0.5454"
],
[
"0.5034",
"0.529",
"0.892",
"0.529",
"0.892",
"0.5425",
"0.5034",
"0.5425"
],
[
"0.0809",
"0.549",
"0.4697",
"0.549",
"0.4697",
"0.5596",
"0.0809",
"0.5596"
],
[
"0.5024",
"0.5435",
"0.8896",
"0.5435",
"0.8896",
"0.558",
"0.5024",
"0.558"
],
[
"0.0784",
"0.5625",
"0.4697",
"0.561",
"0.4697",
"0.5747",
"0.0784",
"0.5757"
],
[
"0.5024",
"0.5596",
"0.8906",
"0.5596",
"0.8906",
"0.5728",
"0.5024",
"0.5728"
],
[
"0.0784",
"0.5757",
"0.468",
"0.577",
"0.468",
"0.5913",
"0.0784",
"0.5903"
],
[
"0.5024",
"0.5747",
"0.8906",
"0.5747",
"0.8906",
"0.5884",
"0.5024",
"0.5884"
],
[
"0.0801",
"0.5923",
"0.468",
"0.5923",
"0.468",
"0.606",
"0.0801",
"0.606"
],
[
"0.5034",
"0.5913",
"0.564",
"0.5913",
"0.564",
"0.603",
"0.5034",
"0.603"
],
[
"0.0801",
"0.608",
"0.468",
"0.608",
"0.468",
"0.622",
"0.0801",
"0.622"
],
[
"0.5215",
"0.604",
"0.8906",
"0.604",
"0.8906",
"0.617",
"0.5215",
"0.617"
],
[
"0.0801",
"0.6226",
"0.468",
"0.6226",
"0.468",
"0.636",
"0.0801",
"0.636"
],
[
"0.5024",
"0.6187",
"0.8906",
"0.619",
"0.8906",
"0.634",
"0.5024",
"0.6333"
],
[
"0.0801",
"0.638",
"0.4697",
"0.638",
"0.4697",
"0.6514",
"0.0801",
"0.6514"
],
[
"0.5024",
"0.6353",
"0.892",
"0.6353",
"0.892",
"0.6484",
"0.5024",
"0.6484"
],
[
"0.0809",
"0.655",
"0.468",
"0.655",
"0.468",
"0.6655",
"0.0809",
"0.6655"
],
[
"0.5024",
"0.6494",
"0.892",
"0.6494",
"0.892",
"0.6636",
"0.5024",
"0.6636"
],
[
"0.0784",
"0.6665",
"0.4697",
"0.6675",
"0.4697",
"0.6816",
"0.0784",
"0.681"
],
[
"0.501",
"0.664",
"0.8906",
"0.664",
"0.8906",
"0.678",
"0.501",
"0.678"
],
[
"0.0801",
"0.683",
"0.286",
"0.683",
"0.286",
"0.6963",
"0.0801",
"0.6963"
],
[
"0.5034",
"0.681",
"0.892",
"0.681",
"0.892",
"0.6943",
"0.5034",
"0.6943"
],
[
"0.5024",
"0.696",
"0.892",
"0.696",
"0.892",
"0.7104",
"0.5024",
"0.7104"
],
[
"0.0801",
"0.707",
"0.437",
"0.707",
"0.437",
"0.72",
"0.0801",
"0.72"
],
[
"0.5024",
"0.711",
"0.732",
"0.711",
"0.732",
"0.7246",
"0.5024",
"0.7246"
],
[
"0.10046",
"0.729",
"0.4673",
"0.729",
"0.4673",
"0.742",
"0.10046",
"0.742"
],
[
"0.6406",
"0.732",
"0.7524",
"0.7344",
"0.7524",
"0.752",
"0.6406",
"0.75"
],
[
"0.8677",
"0.735",
"0.8945",
"0.735",
"0.8945",
"0.752",
"0.8677",
"0.752"
],
[
"0.0801",
"0.7456",
"0.468",
"0.7456",
"0.468",
"0.759",
"0.0801",
"0.759"
],
[
"0.0801",
"0.7603",
"0.4697",
"0.7603",
"0.4697",
"0.7734",
"0.0801",
"0.7734"
],
[
"0.5024",
"0.762",
"0.892",
"0.762",
"0.892",
"0.78",
"0.5024",
"0.78"
],
[
"0.0784",
"0.773",
"0.4697",
"0.7715",
"0.4697",
"0.789",
"0.0784",
"0.7905"
],
[
"0.5024",
"0.781",
"0.892",
"0.781",
"0.892",
"0.7944",
"0.5024",
"0.7944"
],
[
"0.0768",
"0.7905",
"0.4697",
"0.787",
"0.4697",
"0.806",
"0.0768",
"0.8096"
],
[
"0.501",
"0.797",
"0.892",
"0.7944",
"0.892",
"0.8096",
"0.501",
"0.811"
],
[
"0.0801",
"0.8066",
"0.468",
"0.8066",
"0.468",
"0.821",
"0.0801",
"0.821"
],
[
"0.5024",
"0.811",
"0.8906",
"0.811",
"0.8906",
"0.8247",
"0.5024",
"0.8247"
],
[
"0.0801",
"0.8228",
"0.468",
"0.8228",
"0.468",
"0.836",
"0.0801",
"0.836"
],
[
"0.5024",
"0.827",
"0.8906",
"0.827",
"0.8906",
"0.8403",
"0.5024",
"0.8403"
],
[
"0.0801",
"0.8384",
"0.468",
"0.8384",
"0.468",
"0.8516",
"0.0801",
"0.8516"
],
[
"0.5024",
"0.8413",
"0.892",
"0.8413",
"0.892",
"0.855",
"0.5024",
"0.855"
],
[
"0.0784",
"0.8496",
"0.4714",
"0.8486",
"0.4714",
"0.866",
"0.0784",
"0.867"
],
[
"0.5034",
"0.858",
"0.8906",
"0.858",
"0.8906",
"0.8687",
"0.5034",
"0.8687"
],
[
"0.10046",
"0.8735",
"0.4697",
"0.875",
"0.4697",
"0.8896",
"0.10046",
"0.888"
],
[
"0.501",
"0.8706",
"0.8906",
"0.872",
"0.8906",
"0.886",
"0.501",
"0.8853"
],
[
"0.0784",
"0.886",
"0.281",
"0.888",
"0.281",
"0.9014",
"0.0784",
"0.8994"
],
[
"0.5024",
"0.887",
"0.8936",
"0.887",
"0.8936",
"0.901",
"0.5024",
"0.901"
]
] | [
"sources. More recent works [6, 3] directly learn a function",
"the C seen classes. Each row of Y has a unique \"1\" indi-",
"to measure the compatibiliy between documents and visual",
"cating its corresponding class label. At the test stage, the",
"features.However,compared withthe state-of-the-artzero-",
"document representations of the C unseen classes are given",
"shot learning methods, their performance seems to be dis-",
"and our task is to assign C unseen class labels to the test",
"appointing even though some advanced technologies, such",
"images.",
"as deep learning, has been applied [3].",
"3.4.Formulation",
"3. Our Approach",
"Our method is inspired by a recently proposed zero-shot",
"3.1.Overview",
"learning approach [20] which has demonstrated impressive",
"performance despite a very simple learning process. More",
"The overview of our method is depicted in Figure 1. It",
"specifically, it learns a matrix V which optimizes the fol-",
"starts with a raw document representation which is simply",
"lowing objective function.",
"a binarized histogram of words. This representation is fed",
"in xS-+S +xT +V",
"into our zero-shot learning algorithm to generate a classi-",
"fier to detect relevant images.In the process of generat-",
"(1)",
"ing this classifier, the noise suppression regularizer in our",
"method will automatically suppress the impact of less rele-",
"where S denotes the semantic attribute matrix and it can be",
"vant words (illustrated as the red words in Figure 1).",
"either a binary matrix or a real value matrix. The scalars",
"and A are weights controlling the prominence of the var-",
"3.2.Text representation",
"ious terms. The underlying idea of this algorithm can be",
"We extract our text representation based on a simple bag-",
"understood as follows. If the task is to classify X into C",
"of-words model. We start by a preprocessing step of tok-",
"categories, we can simply learn a linear classifier by fitting",
"enizing the words and removing stop words and punctua-",
"to Y, that is, minw IxT W -- Yll2. However, in this case",
"tions. Then a histogram of the remaining word occurrences",
"W cannot be transferred to unseen classes. Thus we fur-",
"is calculated and is subsequently binarized as the text rep-",
"therimposethatW=VS.Inotherwords.theclassifier",
"resentation. In other words, once a word appears in a doc-",
"of a class is generated from its attributes. With this require-",
"ument, its corresponding dimension within the text repre-",
"ment, the classifier of an unseen class can be easily obtained",
"sentation is set to\"1\".One more commonly used choice for",
"and utilized to predict the category of a test image. Simi-",
"the text representation is based on TF-IDF as in [6, 3]. How-",
"larly, we can also treat X V as the classifier operated on",
"ever, we find it produces worse performance 2 than directly",
"the attributes S. The above understanding naturally gives",
"using thebinarizedrepresentation.Thisisprobablybecause",
"rise to the regularization terms >/VS2 and XTVl2",
"the weighting calculated of TF-IDF is not suitable for our",
"which play the same role of the Frobenius norm regularizer",
"zero-shot learning although it is considered to be less noisy",
"as commonly introduced in multi-class classification or re-",
"for applications like document classification. In the bina-",
"gression.",
"rized histogram we essentialy treat each word in a document",
"Since our document representation can also be seen as an",
"equally and this inevitably introduces a lot of noisy signals.",
"attribute vector, the method in [20] can be readily applied",
"However, thanks to our noise suppressing zero-shot learn-",
"to our problem by simply setting S = Z. However, this",
"ing algorithm,we can substantially down-weightthe less",
"naive solution ignores an important fact that the document",
"relevant words and achieve good performance even with a",
"representation is much more noisy than the human specified",
"noisy document representation.",
"attribute vectors. To handle this issue, we introduce a noise",
"suppression mechanism into Eq. (1). More specifically, we",
"3.3. Learning to match text and visual features",
"first decompose V into two terms:",
"We first formally define our problem and introduce the",
"V=WWz,",
"(2)",
"notation used in the following sections. At the training",
"stage, both image features and document descriptions for",
"where Wx E Rmxd and Wz E Rmxd. These two matri-",
"C seen categories are available. Let X e RdxN denote the",
"ces will play different roles in our method. Wz is used to",
"image features of N training examples and Z e (0, 1dxc",
"suppress the noisy components of Z and transform Z into",
"the aforementioned document representations for C seen",
"a m × C intermediate representation. Wx is used to gen-",
"classes, where d and d are the dimensionality of the doc-",
"erate the image classifier from the noise-suppressed inter-",
"ument representation and the image features respectively.",
"mediate representation. Thus, two different regularization",
"We also define Y e [0, 1)NxC as the indicator matrix for",
"terms are imposed to suit these two different roles. The first",
"2Using TD-IDF is about 7% and 5% inferior to binarized representa-",
"term is the l2.1-norm of Wz which achieves the noise sup-",
"tions on AwA and CUB, respectively.",
"pression effect. The second term is the Frobenius norm of"
] | {
"arxiv_id": "1604.01146",
"language": "en",
"page_idx": 2
} |
||
1604.01146-en-3 | 1604.01146-en-3.jpg | <FigureHere>
Figure 1. Overview of our zero-shot learning approach. The text representations are processed by the noise suppression mechanism to
generate a classifier to detect relevant images and the noisy components of text representations are suppressed to gain better performance.
$\mathbf{W}_{\mathbf{x}}^{\top}\mathbf{W}_{\mathbf{z}}\mathbf{Z}$which is similar to the $\lambda\|\mathbf{VS}\|_F^{2}$ term in Eq. (1).
The formulation of our method is expressed as follows:
Similar to [20], once V, in our case $\mathbf{V}=\mathbf{W_x}^\top\mathbf{W_z}$, is learned, we can infer the class label of a test image x using the following rule:
$$\min_{\mathbf{W_{x},W_{z}}}\:L(\mathbf{W_{x},W_{z}})+\lambda_{1}\|\mathbf{W_{x}^{\top}W_{z}Z}\|_{F}^{2}+\lambda_{2}\|\mathbf{W_{z}^{\top}}\|_{2,1}$$
(4)
$$c^*=\max_c\mathbf{x}^\top\mathbf{W}_\mathbf{x}^\top\mathbf{W}_\mathbf{z}\mathbf{z}_c,$$
(3)
$$L(\mathbf{W_x},\mathbf{W_z})=\|\mathbf{X^\top W_x^\top W_zZ-Y}\|_F^2.$$
where z$_c$ is the document representation for the $c$-th candi-
date test class.
## $3.5.0ptimization$
Eq.(3) is convex for $\mathbf{W_{x}}$ and $\mathbf{W_{z}}$ individually but not convex for both of them. Therefore we can solve it using an alternating method, that is, we first fix $\mathbf{W_x}$ and solve for $\mathbf{W_{z};then~fix~W_{z}~and~solve~for~W_{x}.}$
$(I)Fix\textbf{W}_{\mathbf{x}}and\textit{solve for W}_{\mathbf{z}}{:}$
This sub-problem is a regression problem with $l_{2,1}$-norm regularization. Nie et $al.[15]$ proposes an iterative framework to efficiently solve it. It has been shown that the original problem is equivalent to sequentially solving the following problem until convergence
The $l_{2,1}$-norm is defined as $\|\mathbf{W}_{\mathbf{z}}^{T}\|_{2,1}=\sum_{i=1}^{\hat{d}}\|\mathbf{w}_{\mathbf{z}}^{i}\|_{2}$, where $\mathbf{w_z}^i$ denotes the $i$-th column of W$_\mathbf{z}.$ It is known that the $l_{2,1}$-norm will encourage the column vectors of $\mathbf{W_z}$ to have few large values, which means that the impact of noisy dimensions of Z will be substantially suppressed or even completely eliminated. In fact, if $\lambda_2$ becomes sufficient large, it achieves the effect of feature selection on the document representation. However, by cross-validating $\lambda_1$ and $\lambda_2$, our method does not lead to an exactly sparse solution as it seems that the algorithm prefers to keep the majority of the dimensions in $\tilde{\mathbf{Z}}$ for zero-shot learning. This is probably due to the joint regularization effect of $\|\mathbf{W_x}^\top\mathbf{W_zZ}\|_F^2$ or the fact that dimensions corresponding to lower values of $\|\mathbf{w}_{\mathbf{z}}^{i}\|_{2}$ are still useful for zero-shot learning. Therefore we consider the use of the $l_2,1$-norm here as a noise suppression mechanism rather than a feature selection mechanism. We drop out the other regularization terms in Eq. (1) since we find them have little impact on performance.
(5)
$$\begin{aligned}\min_{\mathbf{W_{z}},\mathbf{D}}\:L(\mathbf{W_{x}},\mathbf{W_{z}})+\lambda_{1}\|\mathbf{W_{x}^{\top}W_{z}Z}\|_{F}^{2}+\\\lambda_{2}Tr(\mathbf{W_{z}D}^{t}\mathbf{W_{z}^{\top}}),\end{aligned}$$
where $\mathbf{D}^t$ is a diagonal matrix whose $i$-th diagonal ele- | [
[
"0.10706",
"0.1092",
"0.3472",
"0.1092",
"0.3472",
"0.1193",
"0.10706",
"0.1193"
],
[
"0.5884",
"0.1092",
"0.822",
"0.1092",
"0.822",
"0.1193",
"0.5884",
"0.1193"
],
[
"0.10864",
"0.1193",
"0.367",
"0.1193",
"0.367",
"0.13",
"0.10864",
"0.13"
],
[
"0.5825",
"0.1193",
"0.841",
"0.1193",
"0.841",
"0.13",
"0.5825",
"0.13"
],
[
"0.10455",
"0.13",
"0.335",
"0.13",
"0.335",
"0.1401",
"0.10455",
"0.1401"
],
[
"0.5767",
"0.1288",
"0.8145",
"0.1276",
"0.8145",
"0.1414",
"0.5767",
"0.1427"
],
[
"0.4722",
"0.1383",
"0.5",
"0.1383",
"0.5",
"0.1559",
"0.4722",
"0.1559"
],
[
"0.1054",
"0.1604",
"0.3594",
"0.1604",
"0.3594",
"0.1704",
"0.1054",
"0.1704"
],
[
"0.4631",
"0.1642",
"0.5073",
"0.1642",
"0.5073",
"0.1787",
"0.4631",
"0.1787"
],
[
"0.581",
"0.1604",
"0.834",
"0.1604",
"0.834",
"0.1704",
"0.581",
"0.1704"
],
[
"0.10706",
"0.1704",
"0.3635",
"0.1704",
"0.3635",
"0.1812",
"0.10706",
"0.1812"
],
[
"0.581",
"0.1704",
"0.8384",
"0.1704",
"0.8384",
"0.1812",
"0.581",
"0.1812"
],
[
"0.10297",
"0.1812",
"0.2974",
"0.1799",
"0.2974",
"0.1913",
"0.10297",
"0.1925"
],
[
"0.4412",
"0.1818",
"0.5317",
"0.1818",
"0.5317",
"0.1932",
"0.4412",
"0.1932"
],
[
"0.5786",
"0.1799",
"0.7695",
"0.1799",
"0.7695",
"0.1932",
"0.5786",
"0.1932"
],
[
"0.2197",
"0.1957",
"0.2303",
"0.1957",
"0.2303",
"0.2084",
"0.2197",
"0.2084"
],
[
"0.7036",
"0.1957",
"0.7134",
"0.1957",
"0.7134",
"0.2084",
"0.7036",
"0.2084"
],
[
"0.1054",
"0.2122",
"0.3757",
"0.2108",
"0.3757",
"0.2247",
"0.1054",
"0.226"
],
[
"0.58",
"0.2122",
"0.8506",
"0.2122",
"0.8506",
"0.226",
"0.58",
"0.226"
],
[
"0.10864",
"0.2247",
"0.3733",
"0.2247",
"0.3733",
"0.2349",
"0.10864",
"0.2349"
],
[
"0.5825",
"0.2247",
"0.847",
"0.2247",
"0.847",
"0.2349",
"0.5825",
"0.2349"
],
[
"0.1095",
"0.2361",
"0.1969",
"0.2361",
"0.1969",
"0.2437",
"0.1095",
"0.2437"
],
[
"0.585",
"0.2361",
"0.6724",
"0.2361",
"0.6724",
"0.2437",
"0.585",
"0.2437"
],
[
"0.6006",
"0.2798",
"0.6284",
"0.2798",
"0.6284",
"0.3018",
"0.6006",
"0.3018"
],
[
"0.694",
"0.2922",
"0.759",
"0.2922",
"0.759",
"0.307",
"0.694",
"0.307"
],
[
"0.6855",
"0.307",
"0.7656",
"0.3093",
"0.7656",
"0.324",
"0.6855",
"0.3213"
],
[
"0.2173",
"0.3833",
"0.2573",
"0.3833",
"0.2573",
"0.3977",
"0.2173",
"0.3977"
],
[
"0.0801",
"0.4507",
"0.892",
"0.4507",
"0.892",
"0.464",
"0.0801",
"0.464"
],
[
"0.0801",
"0.464",
"0.888",
"0.464",
"0.888",
"0.4778",
"0.0801",
"0.4778"
],
[
"0.0809",
"0.5005",
"0.4673",
"0.5024",
"0.4673",
"0.5176",
"0.0809",
"0.515"
],
[
"0.523",
"0.5024",
"0.892",
"0.5024",
"0.892",
"0.5166",
"0.523",
"0.5166"
],
[
"0.0809",
"0.518",
"0.4429",
"0.518",
"0.4429",
"0.532",
"0.0809",
"0.532"
],
[
"0.501",
"0.5176",
"0.892",
"0.518",
"0.892",
"0.5327",
"0.501",
"0.532"
],
[
"0.5024",
"0.5327",
"0.6284",
"0.5327",
"0.6284",
"0.547",
"0.5024",
"0.547"
],
[
"0.0948",
"0.547",
"0.4617",
"0.5474",
"0.4617",
"0.565",
"0.0948",
"0.5645"
],
[
"0.08905",
"0.558",
"0.1323",
"0.558",
"0.1323",
"0.5684",
"0.08905",
"0.5684"
],
[
"0.613",
"0.563",
"0.7817",
"0.563",
"0.7817",
"0.5776",
"0.613",
"0.5776"
],
[
"0.869",
"0.563",
"0.8936",
"0.563",
"0.8936",
"0.577",
"0.869",
"0.577"
],
[
"0.4468",
"0.5713",
"0.4697",
"0.5713",
"0.4697",
"0.586",
"0.4468",
"0.586"
],
[
"0.1405",
"0.5903",
"0.4004",
"0.5903",
"0.4004",
"0.608",
"0.1405",
"0.608"
],
[
"0.5034",
"0.595",
"0.8906",
"0.595",
"0.8906",
"0.61",
"0.5034",
"0.61"
],
[
"0.5034",
"0.613",
"0.5986",
"0.613",
"0.5986",
"0.625",
"0.5034",
"0.625"
],
[
"0.09314",
"0.6206",
"0.4722",
"0.6235",
"0.4722",
"0.6484",
"0.09314",
"0.646"
],
[
"0.5034",
"0.638",
"0.638",
"0.638",
"0.638",
"0.653",
"0.5034",
"0.653"
],
[
"0.0801",
"0.6445",
"0.468",
"0.643",
"0.468",
"0.6577",
"0.0801",
"0.659"
],
[
"0.0809",
"0.66",
"0.4697",
"0.66",
"0.4697",
"0.674",
"0.0809",
"0.674"
],
[
"0.5215",
"0.662",
"0.892",
"0.662",
"0.892",
"0.6753",
"0.5215",
"0.6753"
],
[
"0.0809",
"0.677",
"0.4673",
"0.677",
"0.4673",
"0.6875",
"0.0809",
"0.6875"
],
[
"0.5024",
"0.678",
"0.892",
"0.678",
"0.892",
"0.6914",
"0.5024",
"0.6914"
],
[
"0.0801",
"0.69",
"0.468",
"0.69",
"0.468",
"0.704",
"0.0801",
"0.704"
],
[
"0.5024",
"0.6934",
"0.892",
"0.6934",
"0.892",
"0.707",
"0.5024",
"0.707"
],
[
"0.0801",
"0.7056",
"0.4697",
"0.7056",
"0.4697",
"0.7197",
"0.0801",
"0.7197"
],
[
"0.5034",
"0.7085",
"0.744",
"0.7085",
"0.744",
"0.7217",
"0.5034",
"0.7217"
],
[
"0.0784",
"0.72",
"0.468",
"0.7197",
"0.468",
"0.7344",
"0.0784",
"0.735"
],
[
"0.505",
"0.7227",
"0.709",
"0.7227",
"0.709",
"0.7363",
"0.505",
"0.7363"
],
[
"0.0809",
"0.7363",
"0.468",
"0.7363",
"0.468",
"0.75",
"0.0809",
"0.75"
],
[
"0.523",
"0.7407",
"0.892",
"0.7407",
"0.892",
"0.754",
"0.523",
"0.754"
],
[
"0.0809",
"0.752",
"0.468",
"0.752",
"0.468",
"0.765",
"0.0809",
"0.765"
],
[
"0.5024",
"0.756",
"0.8906",
"0.756",
"0.8906",
"0.7695",
"0.5024",
"0.7695"
],
[
"0.0801",
"0.7666",
"0.4673",
"0.7666",
"0.4673",
"0.78",
"0.0801",
"0.78"
],
[
"0.5034",
"0.7715",
"0.8896",
"0.7715",
"0.8896",
"0.782",
"0.5034",
"0.782"
],
[
"0.0801",
"0.781",
"0.4673",
"0.781",
"0.4673",
"0.7944",
"0.0801",
"0.7944"
],
[
"0.5024",
"0.785",
"0.892",
"0.785",
"0.892",
"0.799",
"0.5024",
"0.799"
],
[
"0.0801",
"0.797",
"0.4673",
"0.797",
"0.4673",
"0.81",
"0.0801",
"0.81"
],
[
"0.5024",
"0.8003",
"0.705",
"0.8003",
"0.705",
"0.8145",
"0.5024",
"0.8145"
],
[
"0.0801",
"0.811",
"0.4697",
"0.811",
"0.4697",
"0.8247",
"0.0801",
"0.8247"
],
[
"0.0825",
"0.827",
"0.468",
"0.827",
"0.468",
"0.8403",
"0.0825",
"0.8403"
],
[
"0.5596",
"0.8276",
"0.835",
"0.827",
"0.835",
"0.8447",
"0.5596",
"0.845"
],
[
"0.869",
"0.83",
"0.892",
"0.83",
"0.892",
"0.8433",
"0.869",
"0.8433"
],
[
"0.0801",
"0.8413",
"0.468",
"0.8413",
"0.468",
"0.855",
"0.0801",
"0.855"
],
[
"0.562",
"0.839",
"0.597",
"0.839",
"0.597",
"0.851",
"0.562",
"0.851"
],
[
"0.0809",
"0.858",
"0.468",
"0.858",
"0.468",
"0.8687",
"0.0809",
"0.8687"
],
[
"0.705",
"0.856",
"0.8325",
"0.856",
"0.8325",
"0.8706",
"0.705",
"0.8706"
],
[
"0.0801",
"0.872",
"0.4697",
"0.872",
"0.4697",
"0.886",
"0.0801",
"0.886"
],
[
"0.0801",
"0.887",
"0.3718",
"0.887",
"0.3718",
"0.901",
"0.0801",
"0.901"
],
[
"0.5024",
"0.8853",
"0.8896",
"0.886",
"0.8896",
"0.901",
"0.5024",
"0.8994"
]
] | [
"An antelope is a member ofa number of even-toed",
"nantelope isamemberofanumberofeven-toed",
"ungulate species indigenous to various regions in Africa",
"ungulate species indigenous to various regions in Africa",
"and Eurasia.Agroupof antelope iscalled a herd",
"and Eurasia. A group of antelope is called a herd.",
"W.",
"The beaver is a primarily nocturnal, large,semiaquatic",
"Noise",
"Thebeaver is aprimarilynocturnal, large,semiaquatic",
"rodent. Beavers are known for building dams. They are",
"rodent. Beavers are known for building dams. They are",
"the second-largest rodent in the world...",
"Suppression",
"the second-largest rodent in the world..",
"...",
"..",
"Cattles are the most common type of large domesticated",
"Cattles are the most common type of large domesticated",
"ungulates.They are a prominent modern member of the",
"ungulates. Theyare a prominent modern member of the",
"subfamilyBovinae",
"subfamilyBovinae",
"W.",
"Classifier",
"Generating",
"PTM",
"Figure 1. Overview of our zero-shot learning approach. The text representations are processed by the noise suppression mechanism to",
"generate a classifier to detect relevant images and the noisy components of text representations are suppressed to gain better performance.",
"WTW,Zwhich is similar to the^//VS2term in Eq.(1)",
"Similar to [20], once V, in our case V = WxT Wz, is",
"The formulation of our method is expressed as follows:",
"learned, we can infer the class label of a test image x using",
"the following rule:",
"minL(Wx,Wz)+入1//WzZ/+入2/wIl2,1,",
"w..w.",
"c* = max x W. Wzze,",
"(4)",
"(3)",
"L(wx, w.) = IIxw,w,z- Y?",
"where ze is the document representation for the c-th candi-",
"date test class.",
"The 12,1-norm is defined as IW 2.1 = Z /wi2,",
"3.5.Optimization",
"where w, denotes the i-th column of Wz. It is known that",
"the l2.1-norm will encourage the column vectors of Wz to",
"Eg. (3) is convex for Wxand Wz individually but not",
"havefewlargevalues,whichmeansthattheimpactofnoisy",
"convex for both of them. Therefore we can solve it using",
"dimensions of Z will be substantially suppressed or even",
"an alternating method, that is, we first fix Wx and solve for",
"completely eliminated. In fact, if 2 becomes sufficient",
"Wz; then fix Wz and solve for Wx",
"large, it achieves the effect of feature selection on the doc-",
"(l) Fix Wx and solve for Wz:",
"ument representation. However, by cross-validating A and",
"This sub-problem is a regression problem with l2.1-norm",
"A2, our method does not lead to an exactly sparse solution",
"regularization. Nie et al. [15] proposes an iterative frame-",
"as it seems that the algorithm prefers to keep the majority",
"work to efficiently solve it.It has been shown that the origi-",
"of the dimensions in Z for zero-shot learning. This is prob-",
"nal problem is equivalent to sequentially solving the follow-",
"ably due to the joint regularization effect of //W WzZ?",
"ing problem until convergence",
"or the fact that dimensions corresponding to lower values of",
"I/wz ll2 are still useful for zero-shot learning. Therefore we",
"min L(Wx, Wz) + > l/w,W.zi/+",
"(5)",
"consider the use of the l2.1-norm here as a noise suppression",
"Wz,D",
"mechanismratherthanafeatureselectionmechanism.We",
"A2Tr(WzDtW)",
"drop out the other regularization terms in Eq. (1) since we",
"find them have little impact on performance.",
"where Dt is a diagonal matrix whose i-th diagonal ele-"
] | {
"arxiv_id": "1604.01146",
"language": "en",
"page_idx": 3
} |
||
1604.01146-en-4 | 1604.01146-en-4.jpg | ment is $1/(2\|(\mathbf{w}_{z}^{i})^{(t-1)}\|_{2})^{3}$ at the $t$-th iteration, where $(\mathbf{w}_z^i)^{(t-1)^{\prime}\text{is the }i\text{-th column of the optimal W}_z}$ solved at the $(t-1)$-th iteration. The problem in Eq.(5)further reduces to a Sylvester equation of Wz
(6)
$$\begin{aligned}&\mathbf{A}\mathbf{W}_{\mathbf{z}}+\mathbf{W}_{\mathbf{z}}\mathbf{B}=\mathbf{C},\\&\mathbf{A}=\lambda_{2}(\mathbf{W}_{\mathbf{x}}\mathbf{X}\mathbf{X}^{\top}\mathbf{W}_{\mathbf{x}}^{\top}+\lambda_{1}\mathbf{W}_{\mathbf{x}}\mathbf{W}_{\mathbf{x}}^{\top})^{-1},\\&\mathbf{B}=\mathbf{Z}\mathbf{Z}^{\top}(\mathbf{D})^{-1},\\&\mathbf{C}=\frac{1}{\lambda_{2}}\mathbf{A}\mathbf{W}_{\mathbf{x}}\mathbf{X}\mathbf{Y}\mathbf{Z}^{\top}(\mathbf{D})^{-1}.\end{aligned}$$
The Sylvester equation has a unique solution if and only if A and $-\mathbf{B}$ do not share any eigenvalues. Many stateof-the-art toolboxes are able to solve it efficiently. In our setting, since both A and B are positive definite, A has only positive eigenvalues and $-\mathbf{B}$ has only negative eigenvalues. Therefore Eq.(6) has a unique solution. In summary, the sub-problem of fxing $\mathbf{W_x}$ to solve $\mathbf{W_z}$ can be solved via the algorithm listed in Algorithm 1.
(2) Fix $\mathbf{W} _{\mathbf{z} }$ and solve for $\mathbf{W} _{\mathbf{x} }\nobreak {: } $
This sub-problem is a conventional least squares minimization problem which has the following closed-form solution
<table>
<tbody>
<tr>
<td>$\textbf{L}$ $fi$ $\hat{\mathbf{e}}$</td>
<td>ize W$_{\mathbf{x}}^{0}$ with Gaussian distribution. $=1\cdots\tau$do Solve (5) iteratively for $\mathbf{W_{z}^{t}}$ with $\mathbf{W_{x}^{t-1}}$ according $z$ $x$ ${\mathrm{Algorithm~1.}}$ ${\mathrm{Solve~}(7)}{\mathrm{for}}{\mathbf{W}_{\mathbf{x}}^{t}}{\mathrm{with}}{\mathbf{W}_{\mathbf{z}}^{t}}$ $x$ $z^{\cdot}$ $Converges\textbf{ then}$ $\cdot$ $\bar{\text{Break}} .$ $\mathbf{d}$ $\mathbf{T}\mathbf{X}\boldsymbol{T}$ $\mathbf{W}$</td>
</tr>
</tbody>
</table>
## $4. \textbf{Experiments}$
We divide our experiments into two parts. In the first part we evaluate the proposed method and compare it against both of the methods utilizing online textual sources and human-specified semantic attributes. In the second part we analyse in-depth the noise suppression effect of the proposed method and provide insight into what kind of infor- $\hat{\text{mation in a document is useful for zero-shot learning.}}$
# 4.1. Experimental setting
Datasets: We test our approach on two widely used benchmarks for attribute learning and zero-shot learning: Animals with Attributes [12] (AwA) and Caltech-UCSD $birds- 200- 2011[ 25] ( CUB- 200- 2011) .$ AwAconsistsof 30,475 images of 50 mammals classes with 85 attributes including color, skin texture, body size, body part, affordance, food source, habitat, and behaviour. CUB-200-2011 contains 11,788 images of 200 categories of bird subspecies with 312 fine-grained attributes such as color/shape/texture of body parts. We follow the train/test split according to [12] and [25], where 10 and 50 testing classes are treated as unseen for AwA and CUB-200-2011, respectively.
Textual document sources: We extract the text representation according to scheme introduced in Section 3.2. The raw textual sources are collected from Wikipedia articles describing each of the categories. When constructing the vocabulary, we use the articles of seen classes only. The dimensionality of the text representation is 3506 for AwA and 6815 for CUB-200-2011, respectively.
Image features: To make fair a comparison, two types of image features, the low-level features in [19] and the fully connected layer activations from the“imagenet-vggverydeep-19"[22] CNN are used in our experiments.
Implementation details: The Sylvester equation in Eq.(6) is solved by a MATLAB built-in function, which takes only around 5 seconds on an Intel Core i7 CPU at 3.40GHz. The number of rows of matrices $\mathbf{W_x}$ and $\mathbf{W_z}$ is equal to the number of seen classes. We choose the
$$\mathbf{W_x^\top}=(\mathbf{X}\mathbf{X}^\top+\lambda_1\mathbf{I})^{-1}\mathbf{X}\mathbf{Y}\mathbf{Z}^\top\mathbf{W_z^\top}(\mathbf{W_z}\mathbf{Z}\mathbf{Z}^\top\mathbf{W_z^\top})^{-1}.$$
By alternating between the above two matrices, the overall alternating optimization algorithm for Eq. (3) is listed in Algorithm 2.
$^{3}$In practice, we relax $1/(2||\mathbf{w_{z}}^{i}||_{2})$ to $1/(2\sqrt\mathbf{w_{z}^{i\top}\mathbf{w_{z}^{i}}+\sigma}),\sigma\to0$,as the $i$-th diagonal element to avoid the case of zero columns, and the $l_{2,1}$ as the $i$-th diagonal element to avoid the case of zero columns, and the $l_{2,1}$norm is therefore approximated by $\sum_{i=1}^{\hat{d}}\sqrt{\mathbf{w}_{\mathbf{z}}^{i}}^{\top}\mathbf{w}_{\mathbf{z}}^{i}+\sigma.$ It has beenprovedin that that this approximation guarantees the convergence and the result approaches to that of $l_{2,1}$-norm as $\sigma \to 0$ . | [
[
"0.0825",
"0.09344",
"0.3186",
"0.09344",
"0.3186",
"0.1042",
"0.0825",
"0.1042"
],
[
"0.5024",
"0.0903",
"0.8594",
"0.09155",
"0.8594",
"0.1061",
"0.5024",
"0.1048"
],
[
"0.08496",
"0.1092",
"0.452",
"0.1092",
"0.452",
"0.1225",
"0.08496",
"0.1225"
],
[
"0.509",
"0.1092",
"0.8784",
"0.1092",
"0.8784",
"0.1225",
"0.509",
"0.1225"
],
[
"0.1365",
"0.1256",
"0.4084",
"0.1256",
"0.4084",
"0.1364",
"0.1365",
"0.1364"
],
[
"0.558",
"0.1256",
"0.7812",
"0.1256",
"0.7812",
"0.1364",
"0.558",
"0.1364"
],
[
"0.08417",
"0.137",
"0.3635",
"0.1351",
"0.3635",
"0.1528",
"0.08417",
"0.1547"
],
[
"0.5073",
"0.137",
"0.7817",
"0.1383",
"0.7817",
"0.1528",
"0.5073",
"0.1515"
],
[
"0.0866",
"0.154",
"0.2115",
"0.154",
"0.2115",
"0.1674",
"0.0866",
"0.1674"
],
[
"0.509",
"0.154",
"0.634",
"0.154",
"0.634",
"0.1674",
"0.509",
"0.1674"
],
[
"0.1095",
"0.1686",
"0.4395",
"0.1666",
"0.4395",
"0.1812",
"0.1095",
"0.1831"
],
[
"0.5312",
"0.1654",
"0.876",
"0.1666",
"0.876",
"0.1843",
"0.5312",
"0.1831"
],
[
"0.11273",
"0.1843",
"0.452",
"0.1843",
"0.452",
"0.1976",
"0.11273",
"0.1976"
],
[
"0.535",
"0.1849",
"0.636",
"0.1849",
"0.636",
"0.1963",
"0.535",
"0.1963"
],
[
"0.11273",
"0.1989",
"0.452",
"0.1989",
"0.452",
"0.2134",
"0.11273",
"0.2134"
],
[
"0.5337",
"0.1963",
"0.7295",
"0.1989",
"0.7295",
"0.2134",
"0.5337",
"0.2108"
],
[
"0.1095",
"0.2122",
"0.2402",
"0.214",
"0.2402",
"0.2291",
"0.1095",
"0.2267"
],
[
"0.5337",
"0.2122",
"0.6562",
"0.214",
"0.6562",
"0.2291",
"0.5337",
"0.2267"
],
[
"0.1095",
"0.2279",
"0.2345",
"0.2291",
"0.2345",
"0.2437",
"0.1095",
"0.2424"
],
[
"0.5522",
"0.2291",
"0.6177",
"0.2291",
"0.6177",
"0.2405",
"0.5522",
"0.2405"
],
[
"0.1316",
"0.2437",
"0.1936",
"0.2437",
"0.1936",
"0.2551",
"0.1316",
"0.2551"
],
[
"0.5337",
"0.2437",
"0.5645",
"0.2437",
"0.5645",
"0.2556",
"0.5337",
"0.2556"
],
[
"0.1111",
"0.2583",
"0.1405",
"0.2583",
"0.1405",
"0.2708",
"0.1111",
"0.2708"
],
[
"0.509",
"0.2583",
"0.54",
"0.2583",
"0.54",
"0.2734",
"0.509",
"0.2734"
],
[
"0.08826",
"0.2734",
"0.1168",
"0.2734",
"0.1168",
"0.286",
"0.08826",
"0.286"
],
[
"0.509",
"0.2734",
"0.6396",
"0.2734",
"0.6396",
"0.2886",
"0.509",
"0.2886"
],
[
"0.0866",
"0.2886",
"0.1847",
"0.2886",
"0.1847",
"0.303",
"0.0866",
"0.303"
],
[
"0.5034",
"0.315",
"0.6313",
"0.315",
"0.6313",
"0.33",
"0.5034",
"0.33"
],
[
"0.0801",
"0.3308",
"0.4697",
"0.3308",
"0.4697",
"0.3452",
"0.0801",
"0.3452"
],
[
"0.5195",
"0.339",
"0.892",
"0.3403",
"0.892",
"0.3547",
"0.5195",
"0.3542"
],
[
"0.0825",
"0.3467",
"0.4697",
"0.3467",
"0.4697",
"0.361",
"0.0825",
"0.361"
],
[
"0.0801",
"0.361",
"0.468",
"0.3623",
"0.468",
"0.377",
"0.0801",
"0.3757"
],
[
"0.5024",
"0.356",
"0.892",
"0.356",
"0.892",
"0.3694",
"0.5024",
"0.3694"
],
[
"0.0801",
"0.377",
"0.3186",
"0.377",
"0.3186",
"0.3901",
"0.0801",
"0.3901"
],
[
"0.5034",
"0.3726",
"0.8906",
"0.3726",
"0.8906",
"0.3833",
"0.5034",
"0.3833"
],
[
"0.5024",
"0.3865",
"0.892",
"0.3865",
"0.892",
"0.3997",
"0.5024",
"0.3997"
],
[
"0.5024",
"0.4016",
"0.8906",
"0.4016",
"0.8906",
"0.4153",
"0.5024",
"0.4153"
],
[
"0.1365",
"0.4192",
"0.2769",
"0.4192",
"0.2769",
"0.4343",
"0.1365",
"0.4343"
],
[
"0.4468",
"0.4204",
"0.4697",
"0.4204",
"0.4697",
"0.4343",
"0.4468",
"0.4343"
],
[
"0.501",
"0.4167",
"0.892",
"0.4153",
"0.892",
"0.43",
"0.501",
"0.4312"
],
[
"0.501",
"0.4312",
"0.8545",
"0.432",
"0.8545",
"0.4458",
"0.501",
"0.4443"
],
[
"0.134",
"0.4375",
"0.4028",
"0.4382",
"0.4028",
"0.4558",
"0.134",
"0.455"
],
[
"0.1365",
"0.4602",
"0.2524",
"0.4602",
"0.2524",
"0.4749",
"0.1365",
"0.4749"
],
[
"0.5024",
"0.455",
"0.697",
"0.457",
"0.697",
"0.4717",
"0.5024",
"0.4697"
],
[
"0.1765",
"0.4817",
"0.1896",
"0.4817",
"0.1896",
"0.4912",
"0.1765",
"0.4912"
],
[
"0.5195",
"0.476",
"0.8936",
"0.4766",
"0.8936",
"0.4944",
"0.5195",
"0.4937"
],
[
"0.1365",
"0.488",
"0.1683",
"0.488",
"0.1683",
"0.5005",
"0.1365",
"0.5005"
],
[
"0.1912",
"0.486",
"0.33",
"0.486",
"0.33",
"0.5005",
"0.1912",
"0.5005"
],
[
"0.1748",
"0.5",
"0.1912",
"0.5",
"0.1912",
"0.509",
"0.1748",
"0.509"
],
[
"0.5034",
"0.4956",
"0.8896",
"0.4956",
"0.8896",
"0.506",
"0.5034",
"0.506"
],
[
"0.5024",
"0.509",
"0.8896",
"0.509",
"0.8896",
"0.523",
"0.5024",
"0.523"
],
[
"0.0801",
"0.521",
"0.468",
"0.521",
"0.468",
"0.535",
"0.0801",
"0.535"
],
[
"0.5024",
"0.5234",
"0.892",
"0.5234",
"0.892",
"0.537",
"0.5024",
"0.537"
],
[
"0.0784",
"0.535",
"0.468",
"0.536",
"0.468",
"0.5503",
"0.0784",
"0.55"
],
[
"0.5024",
"0.539",
"0.892",
"0.539",
"0.892",
"0.553",
"0.5024",
"0.553"
],
[
"0.0801",
"0.552",
"0.4697",
"0.552",
"0.4697",
"0.565",
"0.0801",
"0.565"
],
[
"0.5024",
"0.5537",
"0.892",
"0.5537",
"0.892",
"0.5674",
"0.5024",
"0.5674"
],
[
"0.0801",
"0.5664",
"0.4673",
"0.5664",
"0.4673",
"0.58",
"0.0801",
"0.58"
],
[
"0.501",
"0.5693",
"0.8906",
"0.5684",
"0.8906",
"0.582",
"0.501",
"0.5835"
],
[
"0.0801",
"0.582",
"0.4673",
"0.582",
"0.4673",
"0.595",
"0.0801",
"0.595"
],
[
"0.5024",
"0.5854",
"0.892",
"0.5854",
"0.892",
"0.5986",
"0.5024",
"0.5986"
],
[
"0.0801",
"0.595",
"0.4697",
"0.5967",
"0.4697",
"0.6113",
"0.0801",
"0.61"
],
[
"0.5034",
"0.6016",
"0.8906",
"0.6016",
"0.8906",
"0.6123",
"0.5034",
"0.6123"
],
[
"0.0801",
"0.6123",
"0.4697",
"0.6123",
"0.4697",
"0.6255",
"0.0801",
"0.6255"
],
[
"0.5024",
"0.6157",
"0.892",
"0.6157",
"0.892",
"0.629",
"0.5024",
"0.629"
],
[
"0.0801",
"0.627",
"0.3145",
"0.627",
"0.3145",
"0.64",
"0.0801",
"0.64"
],
[
"0.5024",
"0.629",
"0.892",
"0.63",
"0.892",
"0.6445",
"0.5024",
"0.643"
],
[
"0.0825",
"0.64",
"0.286",
"0.6416",
"0.286",
"0.656",
"0.0825",
"0.655"
],
[
"0.5024",
"0.643",
"0.835",
"0.6445",
"0.835",
"0.659",
"0.5024",
"0.6577"
],
[
"0.10046",
"0.657",
"0.468",
"0.657",
"0.468",
"0.6704",
"0.10046",
"0.6704"
],
[
"0.5215",
"0.659",
"0.8906",
"0.659",
"0.8906",
"0.6724",
"0.5215",
"0.6724"
],
[
"0.0801",
"0.6724",
"0.468",
"0.6724",
"0.468",
"0.686",
"0.0801",
"0.686"
],
[
"0.5034",
"0.677",
"0.8896",
"0.677",
"0.8896",
"0.6875",
"0.5034",
"0.6875"
],
[
"0.0801",
"0.688",
"0.1217",
"0.688",
"0.1217",
"0.6997",
"0.0801",
"0.6997"
],
[
"0.5034",
"0.69",
"0.892",
"0.69",
"0.892",
"0.704",
"0.5034",
"0.704"
],
[
"0.5024",
"0.7056",
"0.892",
"0.7056",
"0.892",
"0.7197",
"0.5024",
"0.7197"
],
[
"0.5024",
"0.72",
"0.892",
"0.72",
"0.892",
"0.7344",
"0.5024",
"0.7344"
],
[
"0.0948",
"0.728",
"0.4478",
"0.728",
"0.4478",
"0.742",
"0.0948",
"0.742"
],
[
"0.505",
"0.7373",
"0.8906",
"0.7373",
"0.8906",
"0.7476",
"0.505",
"0.7476"
],
[
"0.4468",
"0.7456",
"0.468",
"0.7456",
"0.468",
"0.759",
"0.4468",
"0.759"
],
[
"0.501",
"0.75",
"0.7817",
"0.7505",
"0.7817",
"0.765",
"0.501",
"0.7646"
],
[
"0.10046",
"0.7715",
"0.468",
"0.7715",
"0.468",
"0.785",
"0.10046",
"0.785"
],
[
"0.523",
"0.7666",
"0.8936",
"0.7666",
"0.8936",
"0.78",
"0.523",
"0.78"
],
[
"0.0784",
"0.785",
"0.468",
"0.786",
"0.468",
"0.8003",
"0.0784",
"0.8"
],
[
"0.5034",
"0.7827",
"0.8906",
"0.7827",
"0.8906",
"0.7935",
"0.5034",
"0.7935"
],
[
"0.5024",
"0.7954",
"0.8906",
"0.797",
"0.8906",
"0.811",
"0.5024",
"0.81"
],
[
"0.0809",
"0.803",
"0.165",
"0.803",
"0.165",
"0.8145",
"0.0809",
"0.8145"
],
[
"0.5034",
"0.811",
"0.8506",
"0.811",
"0.8506",
"0.8247",
"0.5034",
"0.8247"
],
[
"0.1013",
"0.8306",
"0.468",
"0.8306",
"0.468",
"0.8486",
"0.1013",
"0.8486"
],
[
"0.5215",
"0.827",
"0.892",
"0.827",
"0.892",
"0.8403",
"0.5215",
"0.8403"
],
[
"0.0801",
"0.8467",
"0.468",
"0.8467",
"0.468",
"0.86",
"0.0801",
"0.86"
],
[
"0.5034",
"0.8413",
"0.892",
"0.8413",
"0.892",
"0.855",
"0.5034",
"0.855"
],
[
"0.076",
"0.8594",
"0.4739",
"0.8594",
"0.4739",
"0.88",
"0.076",
"0.88"
],
[
"0.5024",
"0.857",
"0.8936",
"0.857",
"0.8936",
"0.8706",
"0.5024",
"0.8706"
],
[
"0.0784",
"0.8755",
"0.4697",
"0.8755",
"0.4697",
"0.8896",
"0.0784",
"0.8896"
],
[
"0.5024",
"0.872",
"0.892",
"0.872",
"0.892",
"0.8853",
"0.5024",
"0.8853"
],
[
"0.0784",
"0.888",
"0.3406",
"0.888",
"0.3406",
"0.9014",
"0.0784",
"0.9014"
],
[
"0.501",
"0.886",
"0.892",
"0.8853",
"0.892",
"0.8994",
"0.501",
"0.901"
]
] | [
"Algorithm1Fix Wxand solveW,",
"Algorithm 2Alternating algorithm for solving Eq. (3)",
"Input: Wx; X of seen classes; Z of seen classes; Ai",
"Input: X of seen classes; Z of seen classes; Ai and A2;",
"and >2;maximum number of iterations T",
"maximum number of iterations T.",
"Initialize DO as identity matrix I e IRdxd.",
"Initialize WQ with Gaussian distribution.",
"for t = 1...T do",
"for t = 1...↑ do",
"Solve Sylvester equation (6)forWt withDt-1",
": Solve (5) iteratively for Wt with Wt-1 according",
"Update the diagonal matrix Dt with its i-diagonal",
"toAlgorithm 1.",
"element as 1/(2/l(w;)(t)|/2), where (wz)(t) is the",
"Solve(7)forWtwithWt",
"i-thcolumnofWt",
"if Converges then",
"if Converges then",
"Break.",
".Break",
"end",
"end",
"end",
"end",
"Output: Wx, Wz",
"Output: Wz.",
"4. Experiments",
"ment is 1/(2/l(wi)(t-1)ll2)3 at the t-th iteration, where",
"We divide our experiments into two parts. In the first part",
"(wi)(t-1) is the i-th column of the optimal Wz solved at",
"the (t - 1)-th iteration. The problem in Eq. (5) further re-",
"we evaluate the proposed method and compare it against",
"duces to a Sylvester equation of W,",
"both of the methods utilizing online textual sources and",
"human-specified semantic attributes. In the second part we",
"analyse in-depth the noise suppression effect of the pro-",
"AWz + W,B = C,",
"(6)",
"posed method and provide insight into what kind of infor-",
"mation in a document is useful for zero-shot learning.",
"A =^2(WxXXTWI +>WW)-",
"B =ZzT(D)-1",
"4.1.Experimental setting",
"1",
"Datasets: We test our approach on two widely used",
"C=",
"AWxXYZT(D)-1",
"12",
"benchmarks for attribute learning and zero-shot learning:",
"Animals with Attributes [12] (AwA) and Caltech-UCSD",
"The Sylvester equation has a unique solution if and only",
"birds-200-2011 [25] (CUB-200-2011). AwA consists of",
"if A and -B do not share any eigenvalues. Many state-",
"30,475 images of 50 mammals classes with 85 attributes",
"of-the-art toolboxes are able to solve it efficiently. In our",
"including color, skin texture, body size, body part, affor-",
"setting, since both A and B are positive definite, A has only",
"dance, food source, habitat, and behaviour. CUB-200-2011",
"positive eigenvalues and --B has only negative eigenvalues.",
"contains 11,788 images of 200 categories of bird subspecies",
"Therefore Eg. (6) has a unique solution. In summary, the",
"with312fine-grained attributes suchas color/shape/texture",
"sub-problem of fixing Wx to solve Wz can be solved via",
"of body parts. We follow the train/test split according to",
"the algorithm listed in Algorithm 1.",
"[12] and [25], where 10 and 50 testing classes are treated as",
"(2) Fix Wz and solve for Wx:",
"unseen for AwA and CUB-200-2011, respectively.",
"This sub-problem is a conventional least squares mini-",
"Textual document sources: We extract the text repre-",
"mization problem which has the following closed-form so-",
"sentation according to scheme introduced in Section 3.2.",
"lution",
"The raw textual sources are collected from Wikipedia ar-",
"ticles describing each of the categories. When constructing",
"the vocabulary, we use the articles of seen classes only. The",
"WT = (XXT +A,I)-1XYZTW,(W,ZZTWT)-1",
"dimensionalityofthetextrepresentationis3506forAwA",
"(7)",
"and 6815 for CUB-200-2011, respectively.",
"By alternating between the above two matrices, the over-",
"Image features: To make fair a comparison, two types",
"all alternating optimization algorithm for Eq. (3) is listed in",
"of image features, the low-level features in [19l and the",
"fully connected layer activations from the \"imagenet-vgg-",
"Algorithm 2.",
"verydeep-19\" [22] CNN are used in our experiments.",
"3In practice, we relax 1/(2w2ll2) to1/(2/w2T wz + ),g → 0,",
"Implementation details:The Sylvester equation in",
"as the i-th diagonal element to avoid the case of zero columns, and the l2,1",
"Eq. (6) is solved by a MATLAB built-in function, which",
"norm is therefore approximated by -wTw +o. It has been",
"takes only around 5 seconds on an Intel Core i7 CPU at",
"proved in [15] that this approximation guarantees the convergence and the",
"3.40GHz. The number of rows of matrices Wx and Wz",
"result approaches to that of l2,1-norm as g → 0 .",
"is equal to the number of seen classes.We choose the"
] | {
"arxiv_id": "1604.01146",
"language": "en",
"page_idx": 4
} |
||
2107.09044-en-0 | 2107.09044-en-0.jpg | ## Just Train Twice: Improving Group Robustness without Training Group Information
Evan Zheran Liu$^{*1}$ Behzad Haghgoo$^{*1}$ Annie S.Chen$^{*1}$ Aditi Raghunathan$^{1}$ Pang Wei Koh$^{1}$
Shiori Sagawa$^{1}$ Percy Liang$^{1}\textbf{Chelsea Finn}^{1}$
## $\mathbf{Abstract}$
$tions.$ For example, in the task of classifying whether an online comment is toxic, the training data is often biased so that mentions of particular demographics (e.g., certain races or religions) are correlated with toxicity. Models trained via ERM then associate these demographics with toxicity and thus perform poorly on groups of examples in which the correlation does not hold, such as non-toxic comments mentioning a particular demographic (Borkan et al., 2019). Similar performance disparities due to spurious correlations occur in many other applications, including other language tasks, facial recognition, and medical imaging (Gururangan et al., 2018; McCoy et al., 2019; Badgeley et al., 2019; Sagawa et al., 2020a; Oakden-Rayner et al., 2020).
Standard training via empirical risk minimization (ERM) can produce models that achieve high accuracy on average but low accuracy on certain groups, especially in the presence of spurious correlations between the input and label. Prior approaches that achieve high worst-group accuracy, like group distributionally robust optimization (group DRO) require expensive group annotations for each training point, whereas approaches that do not use such group annotations typically achieve unsatisfactory worst-group accuracy. In this paper, we propose a simple two-stage approach, JTT, that first trains a standard ERM model for several epochs, and then trains a second model that upweights the training examples that the first model misclassified. Intuitively, this upweights examples from groups on which standard ERM models perform poorly, leading to improved worst-group performance. Averaged over four image classification and natural language processing tasks with spurious correlations, JTT closes 75% of the gap in worst-group accuracy between standard ERM and group DRO, while only requiring group annotations on a small validation set in order to tune hyperparameters.
Following prior work, we formalize this setting by considering a set of pre-defined groups (e.g., corresponding to different demographics) and seeking models that have low worst-group error (Sagawa et al., 2020a). Previous approaches typically require annotations of the group membership of each training example (Sagawa et al., 2020a; Goel et al., 2020; Zhang et al., 2020). While these approaches have been successful at improving worst-group performance, the required training group annotations are often expensive to obtain; for example, in the toxicity classification task mentioned above, each comment has to be annotated with all the demographic identities that are mentioned.
In this paper, we propose a simple algorithm, JTT (Just Train Twice), for improving the worst-group error without training group annotations, instead only requiring group annotations on a much smaller validation set to tune hyperparameters. JTT is composed of two stages: we first identify training examples that are misclassified by a standard ERM model, and then we train the final model by upweighting the examples identified in the first stage. Intuitively, this procedure exploits the observation that sufficiently-regularized ERM models tend to incur high worst-group training error (and subsequently high worst-group test error). This makes
## $1. \textbf{Introduction}$
The standard approach of empirical risk minimization (ERM)—training machine learning models to minimize average training loss—can produce models that achieve low test error on average but still incur high error on certain groups of examples (Hovy & Søgaard, 2015; Blodgett et al., 2016; Tatman, 2017; Hashimoto et al., 2018; Duchi et al., 2019). These performance disparities across groups can be especially pronounced in the presence of spurious correla-
selecting misclassified examples an effective heuristic for identifying examples from groups that ERM models fail on, such as minority groups. Since the final classifier upweights such examples, it performs better on such groups and achieves better minority group performance.
$^{*}$Equal contribution$^{\text{I}}$Stanford University. Correspondence to:
Evan $\dot{\text{Zheran Liu}} <$evanliu@cs.stanford.edu>.
Proceedings of the 38 th International Conference on Machine $Learning$,PMLR 139, 2021. Copyright 2021 by the author(s). | [
[
"0.2484",
"0.11426",
"0.726",
"0.11426",
"0.726",
"0.1288",
"0.2484",
"0.1288"
],
[
"0.3015",
"0.1364",
"0.6724",
"0.1364",
"0.6724",
"0.154",
"0.3015",
"0.154"
],
[
"0.1422",
"0.1989",
"0.8286",
"0.1989",
"0.8286",
"0.2166",
"0.1422",
"0.2166"
],
[
"0.3228",
"0.2166",
"0.6465",
"0.2166",
"0.6465",
"0.2311",
"0.3228",
"0.2311"
],
[
"0.2443",
"0.2588",
"0.3203",
"0.2588",
"0.3203",
"0.2734",
"0.2443",
"0.2734"
],
[
"0.5",
"0.26",
"0.8857",
"0.26",
"0.8857",
"0.2734",
"0.5",
"0.2734"
],
[
"0.0196",
"0.2683",
"0.0678",
"0.2683",
"0.0678",
"0.3757",
"0.0196",
"0.3757"
],
[
"0.11926",
"0.279",
"0.4412",
"0.279",
"0.4412",
"0.2935",
"0.11926",
"0.2935"
],
[
"0.4983",
"0.2747",
"0.8857",
"0.2747",
"0.8857",
"0.2886",
"0.4983",
"0.2886"
],
[
"0.501",
"0.2903",
"0.8857",
"0.2903",
"0.8857",
"0.3037",
"0.501",
"0.3037"
],
[
"0.1209",
"0.2966",
"0.4412",
"0.2966",
"0.4412",
"0.31",
"0.1209",
"0.31"
],
[
"0.1217",
"0.3113",
"0.4412",
"0.3113",
"0.4412",
"0.3245",
"0.1217",
"0.3245"
],
[
"0.5",
"0.3062",
"0.886",
"0.3062",
"0.886",
"0.3193",
"0.5",
"0.3193"
],
[
"0.1209",
"0.327",
"0.4412",
"0.327",
"0.4412",
"0.3403",
"0.1209",
"0.3403"
],
[
"0.5",
"0.3213",
"0.8857",
"0.3213",
"0.8857",
"0.3352",
"0.5",
"0.3352"
],
[
"0.1217",
"0.3416",
"0.4412",
"0.3416",
"0.4412",
"0.3547",
"0.1217",
"0.3547"
],
[
"0.501",
"0.336",
"0.886",
"0.336",
"0.886",
"0.3499",
"0.501",
"0.3499"
],
[
"0.1209",
"0.3567",
"0.4412",
"0.3567",
"0.4412",
"0.3706",
"0.1209",
"0.3706"
],
[
"0.501",
"0.3516",
"0.8857",
"0.3516",
"0.8857",
"0.3655",
"0.501",
"0.3655"
],
[
"0.02696",
"0.3643",
"0.05554",
"0.3643",
"0.05554",
"0.3965",
"0.02696",
"0.3965"
],
[
"0.1209",
"0.3718",
"0.4429",
"0.3718",
"0.4429",
"0.3865",
"0.1209",
"0.3865"
],
[
"0.5",
"0.3662",
"0.886",
"0.3655",
"0.886",
"0.3801",
"0.5",
"0.3806"
],
[
"0.1209",
"0.385",
"0.4412",
"0.3865",
"0.4412",
"0.401",
"0.1209",
"0.3997"
],
[
"0.501",
"0.3818",
"0.8857",
"0.3818",
"0.8857",
"0.3958",
"0.501",
"0.3958"
],
[
"0.1217",
"0.4016",
"0.4412",
"0.4016",
"0.4412",
"0.4153",
"0.1217",
"0.4153"
],
[
"0.501",
"0.3965",
"0.886",
"0.3965",
"0.886",
"0.4104",
"0.501",
"0.4104"
],
[
"0.02533",
"0.406",
"0.058",
"0.406",
"0.058",
"0.4507",
"0.02533",
"0.4507"
],
[
"0.1209",
"0.4153",
"0.4412",
"0.4167",
"0.4412",
"0.4312",
"0.1209",
"0.43"
],
[
"0.5024",
"0.4136",
"0.8857",
"0.4136",
"0.8857",
"0.4236",
"0.5024",
"0.4236"
],
[
"0.1217",
"0.432",
"0.4412",
"0.432",
"0.4412",
"0.4458",
"0.1217",
"0.4458"
],
[
"0.501",
"0.4268",
"0.886",
"0.4268",
"0.886",
"0.44",
"0.501",
"0.44"
],
[
"0.02777",
"0.4414",
"0.05963",
"0.4414",
"0.05963",
"0.4841",
"0.02777",
"0.4841"
],
[
"0.501",
"0.4414",
"0.834",
"0.4414",
"0.834",
"0.455",
"0.501",
"0.455"
],
[
"0.11926",
"0.4475",
"0.4395",
"0.4475",
"0.4395",
"0.461",
"0.11926",
"0.461"
],
[
"0.1217",
"0.4622",
"0.4412",
"0.4622",
"0.4412",
"0.476",
"0.1217",
"0.476"
],
[
"0.5",
"0.464",
"0.886",
"0.464",
"0.886",
"0.4778",
"0.5",
"0.4778"
],
[
"0.11926",
"0.4766",
"0.4395",
"0.4766",
"0.4395",
"0.4905",
"0.11926",
"0.4905"
],
[
"0.5",
"0.4797",
"0.886",
"0.4797",
"0.886",
"0.4944",
"0.5",
"0.4944"
],
[
"0.02287",
"0.4912",
"0.05963",
"0.4912",
"0.05963",
"0.712",
"0.02287",
"0.712"
],
[
"0.1209",
"0.4924",
"0.4412",
"0.4924",
"0.4412",
"0.506",
"0.1209",
"0.506"
],
[
"0.4983",
"0.4944",
"0.886",
"0.4937",
"0.886",
"0.5083",
"0.4983",
"0.509"
],
[
"0.11926",
"0.506",
"0.4412",
"0.507",
"0.4412",
"0.5215",
"0.11926",
"0.521"
],
[
"0.501",
"0.5103",
"0.8857",
"0.5103",
"0.8857",
"0.5234",
"0.501",
"0.5234"
],
[
"0.1209",
"0.523",
"0.4412",
"0.523",
"0.4412",
"0.536",
"0.1209",
"0.536"
],
[
"0.5",
"0.5244",
"0.8857",
"0.5244",
"0.8857",
"0.539",
"0.5",
"0.539"
],
[
"0.1217",
"0.5386",
"0.4412",
"0.5386",
"0.4412",
"0.552",
"0.1217",
"0.552"
],
[
"0.501",
"0.5405",
"0.886",
"0.5405",
"0.886",
"0.5537",
"0.501",
"0.5537"
],
[
"0.1217",
"0.5537",
"0.4395",
"0.5537",
"0.4395",
"0.5645",
"0.1217",
"0.5645"
],
[
"0.5",
"0.5537",
"0.8857",
"0.5547",
"0.8857",
"0.5693",
"0.5",
"0.5684"
],
[
"0.1209",
"0.5674",
"0.4412",
"0.5674",
"0.4412",
"0.5806",
"0.1209",
"0.5806"
],
[
"0.5",
"0.5693",
"0.886",
"0.571",
"0.886",
"0.584",
"0.5",
"0.5835"
],
[
"0.11926",
"0.582",
"0.4412",
"0.5806",
"0.4412",
"0.595",
"0.11926",
"0.5967"
],
[
"0.501",
"0.5854",
"0.8857",
"0.5854",
"0.8857",
"0.5986",
"0.501",
"0.5986"
],
[
"0.1209",
"0.5986",
"0.4395",
"0.5986",
"0.4395",
"0.6094",
"0.1209",
"0.6094"
],
[
"0.501",
"0.601",
"0.886",
"0.601",
"0.886",
"0.6143",
"0.501",
"0.6143"
],
[
"0.1217",
"0.613",
"0.4412",
"0.613",
"0.4412",
"0.627",
"0.1217",
"0.627"
],
[
"0.501",
"0.6157",
"0.886",
"0.6157",
"0.886",
"0.629",
"0.501",
"0.629"
],
[
"0.1209",
"0.6274",
"0.4429",
"0.627",
"0.4429",
"0.6416",
"0.1209",
"0.6426"
],
[
"0.501",
"0.631",
"0.824",
"0.631",
"0.824",
"0.6445",
"0.501",
"0.6445"
],
[
"0.1209",
"0.643",
"0.3105",
"0.643",
"0.3105",
"0.657",
"0.1209",
"0.657"
],
[
"0.501",
"0.653",
"0.8857",
"0.653",
"0.8857",
"0.6675",
"0.501",
"0.6675"
],
[
"0.501",
"0.6685",
"0.886",
"0.6685",
"0.886",
"0.6816",
"0.501",
"0.6816"
],
[
"0.08905",
"0.6846",
"0.2157",
"0.686",
"0.2157",
"0.7007",
"0.08905",
"0.6987"
],
[
"0.5",
"0.6846",
"0.8857",
"0.6846",
"0.8857",
"0.6978",
"0.5",
"0.6978"
],
[
"0.5",
"0.6987",
"0.886",
"0.6987",
"0.886",
"0.712",
"0.5",
"0.712"
],
[
"0.08905",
"0.711",
"0.4739",
"0.711",
"0.4739",
"0.726",
"0.08905",
"0.726"
],
[
"0.5",
"0.714",
"0.8857",
"0.714",
"0.8857",
"0.728",
"0.5",
"0.728"
],
[
"0.08826",
"0.7266",
"0.4756",
"0.7266",
"0.4756",
"0.7407",
"0.08826",
"0.7407"
],
[
"0.501",
"0.728",
"0.8857",
"0.728",
"0.8857",
"0.741",
"0.501",
"0.741"
],
[
"0.08826",
"0.742",
"0.4739",
"0.742",
"0.4739",
"0.756",
"0.08826",
"0.756"
],
[
"0.501",
"0.7437",
"0.8857",
"0.7437",
"0.8857",
"0.757",
"0.501",
"0.757"
],
[
"0.0866",
"0.757",
"0.4739",
"0.756",
"0.4739",
"0.771",
"0.0866",
"0.7715"
],
[
"0.5",
"0.759",
"0.886",
"0.759",
"0.886",
"0.773",
"0.5",
"0.773"
],
[
"0.08905",
"0.773",
"0.4756",
"0.773",
"0.4756",
"0.786",
"0.08905",
"0.786"
],
[
"0.5",
"0.7734",
"0.886",
"0.7734",
"0.886",
"0.787",
"0.5",
"0.787"
],
[
"0.08905",
"0.787",
"0.4756",
"0.787",
"0.4756",
"0.8003",
"0.08905",
"0.8003"
],
[
"0.501",
"0.789",
"0.886",
"0.789",
"0.886",
"0.803",
"0.501",
"0.803"
],
[
"0.08905",
"0.802",
"0.4739",
"0.802",
"0.4739",
"0.815",
"0.08905",
"0.815"
],
[
"0.5",
"0.8047",
"0.886",
"0.8047",
"0.886",
"0.8184",
"0.5",
"0.8184"
],
[
"0.08826",
"0.8174",
"0.4739",
"0.8174",
"0.4739",
"0.8306",
"0.08826",
"0.8306"
],
[
"0.501",
"0.8193",
"0.8857",
"0.8193",
"0.8857",
"0.8335",
"0.501",
"0.8335"
],
[
"0.5",
"0.834",
"0.884",
"0.834",
"0.884",
"0.8486",
"0.5",
"0.8486"
],
[
"0.10455",
"0.8403",
"0.4756",
"0.8403",
"0.4756",
"0.855",
"0.10455",
"0.855"
],
[
"0.08826",
"0.854",
"0.367",
"0.854",
"0.367",
"0.867",
"0.08826",
"0.867"
],
[
"0.501",
"0.8496",
"0.8857",
"0.8496",
"0.8857",
"0.863",
"0.501",
"0.863"
],
[
"0.5",
"0.8657",
"0.886",
"0.8657",
"0.886",
"0.879",
"0.5",
"0.879"
],
[
"0.08905",
"0.8774",
"0.4756",
"0.8774",
"0.4756",
"0.8926",
"0.08905",
"0.8926"
],
[
"0.4983",
"0.88",
"0.819",
"0.88",
"0.819",
"0.8945",
"0.4983",
"0.8945"
],
[
"0.08826",
"0.8926",
"0.455",
"0.8926",
"0.455",
"0.906",
"0.08826",
"0.906"
]
] | [
"Just Train Twice: Improving Group Robustness",
"without Training Group Information",
"Evan Zheran Liu*1 Behzad Haghgoo*I Annie S. Chen*1 Aditi RaghunathanI Pang Wei Koh !",
"Shiori Sagawa' Percy Liang Chelsea Finn!",
"Abstract",
"tions. For example, in the task of classifying whether an",
"120z des 4",
"Standard training via empirical risk minimization",
"online comment is toxic, the training data is often biased so",
"that mentions of particular demographics (e.g., certain races",
"(ERM) can produce models that achieve high",
"accuracy on average but low accuracy on cer-",
"or religions) are correlated with toxicity. Models trained",
"tain groups, especially in the presence of spu-",
"via ERM then associate these demographics with toxicity",
"rious correlations between the input and label.",
"and thus perform poorly on groups of examples in which",
"Prior approaches that achieve high worst-group",
"the correlation does not hold, such as non-toxic comments",
"2",
"accuracy, like group distributionally robust opti-",
"mentioning a particular demographic (Borkan et al., 2019).",
"mization (group DRO)require expensive group",
"Similar performance disparities due to spurious correlations",
"annotations for each training point, whereas ap-",
"occur in many other applications, including other language",
"[DT",
"proaches that do not use such group annotations",
"tasks,facial recognition,and medical imaging (Gururan-",
"typically achieve unsatisfactory worst-group accu-",
"gan et al., 2018; McCoy et al., 2019; Badgeley et al., 2019;",
"I'so",
"Sagawa et al., 2020a; Oakden-Rayner et al., 2020).",
"racy. In this paper, we propose a simple two-stage",
"approach, JTT, that first trains a standard ERM",
"Following prior work, we formalize this setting by con-",
"model for several epochs, and then trains a second",
"sidering a set of pre-defined groups (e.g., corresponding",
"0600",
"model that upweights the training examples that",
"to different demographics) and seeking models that have",
"the first model misclassified. Intuitively, this up",
"low worst-group error (Sagawa et al., 2020a).Previous",
"weights examples from groups on which standard",
"approaches typically require annotations of the group mem-",
"ERM models perform poorly, leading to improved",
"bership of each training example (Sagawa et al., 2020a; Goel",
"worst-groupperformance.Averagedoverfourim",
"et al., 2020; Zhang et al., 2020). While these approaches",
"age classification and natural language processing",
"have been successful at improving worst-group performance,",
"tasks with spurious correlations, JTT closes 75%",
"the required training group annotations are often expensive",
"ofthegapinworst-groupaccuracybetweenstan",
"to obtain; for example, in the toxicity classification task",
"dard ERM and group DRO, while only requiring",
"mentioned above, each comment has to be annotated with",
"group annotations on a small validation set in or-",
"all the demographic identities that are mentioned.",
"der to tune hyperparameters.",
"In this paper, we propose a simple algorithm, JTT (Just",
"Train Twice), for improving the worst-group error without",
"1. Introduction",
"training group annotations, instead only requiring group",
"annotations on a much smaller validation set to tune hyper-",
"The standard approach of empirical risk minimization",
"parameters. JTT is composed of two stages: we first identify",
"(ERM)training machine learning models to minimize av-",
"training examples that are misclassified by a standard ERM",
"erage training loss-can produce models that achieve low",
"model, and then we train the final model by upweighting the",
"test error on average but still incur high error on certain",
"examples identified in the first stage. Intuitively, this pro-",
"groups of examples (Hovy & Sogaard, 2015; Blodgett et al.,",
"cedure exploits the observation that sufficiently-regularized",
"2016; Tatman, 2017; Hashimoto et al., 2018; Duchi et al.,",
"ERM models tend to incur high worst-group training error",
"2019). These performance disparities across groups can be",
"(and subsequently high worst-group test error). This makes",
"especially pronounced in the presence of spurious correla-",
"selecting misclassified examples an effective heuristic for",
"identifying examples from groups that ERM models fail",
"\"Equal contribution IStanford University. Correspondence to:",
"Evan Zheran Liu <evanliu@es.stanford.edu>.",
"on, such as minority groups. Since the final classifier up",
"weights such examples, it performs better on such groups",
"Proceedings of the 38th International Conference on Machine",
"and achieves better minority group performance.",
"Learning, PMLR 139, 2021. Copyright 2021 by the author(s)."
] | {
"arxiv_id": "2107.09044",
"language": "en",
"page_idx": 0
} |
||
2107.09044-en-1 | 2107.09044-en-1.jpg | Just Train Twice: Improving Group Robustness without Training Group Information
gio et al., 2013; Szegedy et al., 2014) or domain generalization (Blanchard et al., 2011; Muandet et al., 2013). Approaches for group robustness fall into the two main categories we discuss below.
Robustness using group information. Several approaches leverage group information during training, either to combat spurious correlations or handle shifts in group proportions between train and test distributions. For example, Mohri et al. (2019); Sagawa et al. (2020a); Zhang et al. (2020) minimize the worst-group loss during training; Goel et al. (2020) synthetically expand the minority groups via generative modeling; Shimodaira (2000); Byrd & Lipton (2019); Sagawa et al. (2020b) reweight or subsample to artificially balance the majority and minority groups; Cao et al. (2019; 2020) impose heavy Lipschitz regularization around minority points. These approaches substantially reduce worst-group error, but obtaining group annotations for the entire training set can be extremely expensive.
Another line of work studies worst-group performance in the context of fairness (Hardt et al., 2016; Woodworth et al., 2017;Pleiss et al., 2017; Agarwal et al., 2018; Khani et al., 2019). While these works also aim to improve the worstgroup loss, they explicitly focus on equalizing the loss across all groups.
We evaluate JTT on two image classification datasets with spurious correlations, Waterbirds (Wah et al., 2011; Sagawa et al., 2020a) and CelebA (Liu et al., 2015) and two natural language processing datasets, MultiNLI (Williams et al., 2018) and CivilComments-WILDS (Borkan et al., 2019; Koh et al., 2021). We use the versions of Waterbirds, CelebA, and MultiNLI from Sagawa et al. (2020a), where in Waterbirds, the label waterbird or landbird spuriously correlates with water in the background; in CelebA, the label blond or non-blond spuriously correlates with binary gender; and in MultiNLI, the label spuriously correlates with the presence of negation words. In CivilComments-WILDS, where the input is online comments, the label $toxic,non-$ toxic spuriously correlates with the mention of particular demographics, as discussed above. Our method outperforms ERM on all four datasets, with an average worst-group accuracy improvement of 16.2%, while maintaining competitive average accuracy (only $4.2\%$ worse on average). Furthermore, despite having no group annotations during training, JTT closes 75% of the gap between ERM and group DRO, which uses complete group information on the training data. We then empirically analyze JTT. First, we analyze the examples identified by JTT and show that JTT upweights groups on which standard ERM models perform poorly, e.g., minority groups that do not have the spurious correlation (such as waterbirds on land in the Waterbirds dataset). Second, we show that having validation group annotations is essential for hyperparameter tuning for JTT and other related algorithms.
Finally, we compare JTT with the distributionally robust optimization (DRO) algorithm that minimizes the conditional value at risk (CVaR). CVaR DRO aims to train models that are robust to a wide range of potential distribution shifts by minimizing the worst-case loss over all subsets of the training set of a certain size (Duchi et al., 2019). This objective does not require training group annotations, and it can be optimized by dynamically upweighting training examples with the highest losses in each minibatch (Levy et al., 2020). Though CVaR DRO and JTT share conceptual similarities—they both upweight high loss training points and do not require training group information—the difference is that JTT upweights a static set of examples, while CVaR DRO dynamically re-computes which examples to update. Empirically, we find that JTT empirically substantially outperforms CVaR DRO on worst-group accuracy in the datasets we tested.
## $2. \textbf{Related Work}$
In this paper, we focus on group robustness (i.e., training models that obtain good performance on each of a set of predefined groups in the dataset), though other notions of robustness are also studied, such as adversarial examples (Big-
Robustness without group information. We focus on the setting where group annotations are expensive and unavailable on the training data, and potentially only available on a much smaller validation set. Many approaches for this setting fall under the general DRO framework, where models are trained to minimize the worst-case loss across all distributions in a ball around the empirical distribution (BenTal et al., 2013; Lam & Zhou, 2015; Duchi et al., 2016; Namkoong& Duchi, 2017; Oren et al., 2019). Pezeshki et al. (2020) modify the dynamics of stochastic gradient descent to avoid learning spurious correlations. Sohoni et al. (2020) automatically identify groups based by clustering the data points. Kim et al. (2019) propose an auditing scheme that searches for high-loss groups defined by a function within a pre-specified complexity class and postprocess the model to minimize discrepancies identified by the auditor. Khani et al. (2019) minimize the variance in the loss across all data points to encourage lower discrepancy in the losses across all possible groups. Another approach is to directly learn how to reweight the training examples either using small amounts of metadata (Shu et al., 2019) or automatically via meta-learning (Ren et al., 2018).
Most closely related to JTT are several approaches that also train a pair of models, where the performance of the first
model is used to help train the second model ( Yaghoobzadeh et al., 2019; Utama et al., 2020; Nam et al., 2020). We compare JTT with one such approach, called Learning from | [
[
"0.2197",
"0.05682",
"0.7563",
"0.05682",
"0.7563",
"0.0707",
"0.2197",
"0.0707"
],
[
"0.0866",
"0.084",
"0.4756",
"0.0852",
"0.4756",
"0.09973",
"0.0866",
"0.0985"
],
[
"0.5",
"0.0852",
"0.8857",
"0.0852",
"0.8857",
"0.09973",
"0.5",
"0.09973"
],
[
"0.08905",
"0.1016",
"0.4756",
"0.1016",
"0.4756",
"0.11554",
"0.08905",
"0.11554"
],
[
"0.4983",
"0.101",
"0.886",
"0.101",
"0.886",
"0.11426",
"0.4983",
"0.11426"
],
[
"0.08826",
"0.11615",
"0.4756",
"0.11615",
"0.4756",
"0.13",
"0.08826",
"0.13"
],
[
"0.4983",
"0.11743",
"0.886",
"0.11743",
"0.886",
"0.1307",
"0.4983",
"0.1307"
],
[
"0.08826",
"0.132",
"0.4739",
"0.132",
"0.4739",
"0.1459",
"0.08826",
"0.1459"
],
[
"0.4983",
"0.132",
"0.6914",
"0.132",
"0.6914",
"0.1459",
"0.4983",
"0.1459"
],
[
"0.08826",
"0.1465",
"0.4756",
"0.1465",
"0.4756",
"0.1604",
"0.08826",
"0.1604"
],
[
"0.08826",
"0.1604",
"0.4756",
"0.161",
"0.4756",
"0.1755",
"0.08826",
"0.1749"
],
[
"0.5",
"0.1642",
"0.886",
"0.1642",
"0.886",
"0.1787",
"0.5",
"0.1787"
],
[
"0.08905",
"0.1768",
"0.4739",
"0.1768",
"0.4739",
"0.1901",
"0.08905",
"0.1901"
],
[
"0.5",
"0.1799",
"0.886",
"0.1799",
"0.886",
"0.1932",
"0.5",
"0.1932"
],
[
"0.08905",
"0.1932",
"0.4739",
"0.1932",
"0.4739",
"0.2039",
"0.08905",
"0.2039"
],
[
"0.4983",
"0.1932",
"0.886",
"0.1945",
"0.886",
"0.209",
"0.4983",
"0.2084"
],
[
"0.08905",
"0.2084",
"0.4722",
"0.2084",
"0.4722",
"0.2184",
"0.08905",
"0.2184"
],
[
"0.501",
"0.2102",
"0.886",
"0.2102",
"0.886",
"0.2235",
"0.501",
"0.2235"
],
[
"0.08826",
"0.2229",
"0.4756",
"0.2229",
"0.4756",
"0.2361",
"0.08826",
"0.2361"
],
[
"0.501",
"0.2247",
"0.886",
"0.2247",
"0.886",
"0.238",
"0.501",
"0.238"
],
[
"0.08905",
"0.2374",
"0.4739",
"0.2374",
"0.4739",
"0.2507",
"0.08905",
"0.2507"
],
[
"0.5",
"0.2405",
"0.886",
"0.2405",
"0.886",
"0.254",
"0.5",
"0.254"
],
[
"0.08905",
"0.252",
"0.4739",
"0.252",
"0.4739",
"0.2651",
"0.08905",
"0.2651"
],
[
"0.5",
"0.2551",
"0.886",
"0.2551",
"0.886",
"0.2683",
"0.5",
"0.2683"
],
[
"0.08826",
"0.2676",
"0.4756",
"0.2676",
"0.4756",
"0.281",
"0.08826",
"0.281"
],
[
"0.5",
"0.2683",
"0.886",
"0.2695",
"0.886",
"0.2842",
"0.5",
"0.2827"
],
[
"0.08826",
"0.2827",
"0.4739",
"0.2827",
"0.4739",
"0.2966",
"0.08826",
"0.2966"
],
[
"0.5",
"0.2854",
"0.886",
"0.2854",
"0.886",
"0.2986",
"0.5",
"0.2986"
],
[
"0.08826",
"0.2974",
"0.4756",
"0.2974",
"0.4756",
"0.3113",
"0.08826",
"0.3113"
],
[
"0.5",
"0.3005",
"0.886",
"0.3005",
"0.886",
"0.3145",
"0.5",
"0.3145"
],
[
"0.08826",
"0.3132",
"0.4756",
"0.3132",
"0.4756",
"0.327",
"0.08826",
"0.327"
],
[
"0.501",
"0.315",
"0.886",
"0.315",
"0.886",
"0.3289",
"0.501",
"0.3289"
],
[
"0.08826",
"0.3276",
"0.4756",
"0.3276",
"0.4756",
"0.3416",
"0.08826",
"0.3416"
],
[
"0.5",
"0.3308",
"0.886",
"0.3308",
"0.886",
"0.3447",
"0.5",
"0.3447"
],
[
"0.08905",
"0.3435",
"0.4756",
"0.3435",
"0.4756",
"0.3567",
"0.08905",
"0.3567"
],
[
"0.501",
"0.3467",
"0.886",
"0.3467",
"0.886",
"0.3599",
"0.501",
"0.3599"
],
[
"0.08826",
"0.358",
"0.4739",
"0.358",
"0.4739",
"0.3718",
"0.08826",
"0.3718"
],
[
"0.501",
"0.361",
"0.8286",
"0.361",
"0.8286",
"0.375",
"0.501",
"0.375"
],
[
"0.08826",
"0.3726",
"0.4756",
"0.3726",
"0.4756",
"0.387",
"0.08826",
"0.387"
],
[
"0.0866",
"0.3882",
"0.4756",
"0.3882",
"0.4756",
"0.4016",
"0.0866",
"0.4016"
],
[
"0.5",
"0.3838",
"0.886",
"0.3838",
"0.886",
"0.3977",
"0.5",
"0.3977"
],
[
"0.5",
"0.3984",
"0.888",
"0.3984",
"0.888",
"0.4124",
"0.5",
"0.4124"
],
[
"0.08826",
"0.4104",
"0.4739",
"0.4104",
"0.4739",
"0.4236",
"0.08826",
"0.4236"
],
[
"0.5",
"0.4136",
"0.888",
"0.414",
"0.888",
"0.428",
"0.5",
"0.4268"
],
[
"0.08826",
"0.4255",
"0.4739",
"0.4255",
"0.4739",
"0.4395",
"0.08826",
"0.4395"
],
[
"0.501",
"0.428",
"0.886",
"0.428",
"0.886",
"0.4414",
"0.501",
"0.4414"
],
[
"0.08826",
"0.4414",
"0.4756",
"0.4414",
"0.4756",
"0.455",
"0.08826",
"0.455"
],
[
"0.501",
"0.4443",
"0.886",
"0.4443",
"0.886",
"0.4583",
"0.501",
"0.4583"
],
[
"0.08826",
"0.457",
"0.4756",
"0.457",
"0.4756",
"0.4702",
"0.08826",
"0.4702"
],
[
"0.4983",
"0.459",
"0.6167",
"0.459",
"0.6167",
"0.473",
"0.4983",
"0.473"
],
[
"0.08826",
"0.4717",
"0.4756",
"0.4717",
"0.4756",
"0.4849",
"0.08826",
"0.4849"
],
[
"0.08905",
"0.486",
"0.4756",
"0.486",
"0.4756",
"0.5",
"0.08905",
"0.5"
],
[
"0.501",
"0.4905",
"0.886",
"0.4905",
"0.886",
"0.505",
"0.501",
"0.505"
],
[
"0.08826",
"0.502",
"0.4756",
"0.502",
"0.4756",
"0.515",
"0.08826",
"0.515"
],
[
"0.5",
"0.507",
"0.886",
"0.507",
"0.886",
"0.521",
"0.5",
"0.521"
],
[
"0.08826",
"0.5166",
"0.2141",
"0.5166",
"0.2141",
"0.53",
"0.08826",
"0.53"
],
[
"0.5",
"0.523",
"0.886",
"0.523",
"0.886",
"0.536",
"0.5",
"0.536"
],
[
"0.08826",
"0.537",
"0.4739",
"0.5386",
"0.4739",
"0.553",
"0.08826",
"0.552"
],
[
"0.5",
"0.537",
"0.886",
"0.537",
"0.886",
"0.5503",
"0.5",
"0.5503"
],
[
"0.08826",
"0.5537",
"0.4756",
"0.5537",
"0.4756",
"0.5674",
"0.08826",
"0.5674"
],
[
"0.501",
"0.552",
"0.8857",
"0.552",
"0.8857",
"0.565",
"0.501",
"0.565"
],
[
"0.0907",
"0.571",
"0.4739",
"0.571",
"0.4739",
"0.5806",
"0.0907",
"0.5806"
],
[
"0.5",
"0.5674",
"0.8857",
"0.5674",
"0.8857",
"0.5806",
"0.5",
"0.5806"
],
[
"0.08905",
"0.586",
"0.4722",
"0.586",
"0.4722",
"0.5967",
"0.08905",
"0.5967"
],
[
"0.5024",
"0.584",
"0.8857",
"0.584",
"0.8857",
"0.5947",
"0.5024",
"0.5947"
],
[
"0.08905",
"0.601",
"0.4722",
"0.601",
"0.4722",
"0.6113",
"0.08905",
"0.6113"
],
[
"0.5",
"0.5977",
"0.886",
"0.5977",
"0.886",
"0.6113",
"0.5",
"0.6113"
],
[
"0.08826",
"0.6143",
"0.4756",
"0.6143",
"0.4756",
"0.6274",
"0.08826",
"0.6274"
],
[
"0.4983",
"0.6123",
"0.886",
"0.6123",
"0.886",
"0.6255",
"0.4983",
"0.6255"
],
[
"0.08826",
"0.63",
"0.4756",
"0.63",
"0.4756",
"0.6445",
"0.08826",
"0.6445"
],
[
"0.5",
"0.6274",
"0.886",
"0.6274",
"0.886",
"0.6416",
"0.5",
"0.6416"
],
[
"0.08826",
"0.646",
"0.4756",
"0.646",
"0.4756",
"0.659",
"0.08826",
"0.659"
],
[
"0.5",
"0.643",
"0.886",
"0.643",
"0.886",
"0.657",
"0.5",
"0.657"
],
[
"0.08826",
"0.66",
"0.4739",
"0.66",
"0.4739",
"0.674",
"0.08826",
"0.674"
],
[
"0.4983",
"0.6577",
"0.8857",
"0.6577",
"0.8857",
"0.672",
"0.4983",
"0.672"
],
[
"0.08826",
"0.675",
"0.4756",
"0.675",
"0.4756",
"0.688",
"0.08826",
"0.688"
],
[
"0.5",
"0.674",
"0.886",
"0.674",
"0.886",
"0.6875",
"0.5",
"0.6875"
],
[
"0.08826",
"0.69",
"0.4756",
"0.69",
"0.4756",
"0.704",
"0.08826",
"0.704"
],
[
"0.5",
"0.688",
"0.8857",
"0.688",
"0.8857",
"0.702",
"0.5",
"0.702"
],
[
"0.08826",
"0.705",
"0.478",
"0.705",
"0.478",
"0.7197",
"0.08826",
"0.7197"
],
[
"0.5",
"0.704",
"0.8857",
"0.7026",
"0.8857",
"0.7163",
"0.5",
"0.7173"
],
[
"0.08826",
"0.72",
"0.4739",
"0.72",
"0.4739",
"0.7344",
"0.08826",
"0.7344"
],
[
"0.4983",
"0.7173",
"0.884",
"0.7173",
"0.884",
"0.731",
"0.4983",
"0.731"
],
[
"0.08826",
"0.7344",
"0.4722",
"0.7344",
"0.4722",
"0.7476",
"0.08826",
"0.7476"
],
[
"0.501",
"0.735",
"0.884",
"0.735",
"0.884",
"0.7456",
"0.501",
"0.7456"
],
[
"0.08826",
"0.7505",
"0.4739",
"0.7505",
"0.4739",
"0.7646",
"0.08826",
"0.7646"
],
[
"0.501",
"0.7485",
"0.8857",
"0.7485",
"0.8857",
"0.762",
"0.501",
"0.762"
],
[
"0.0866",
"0.7646",
"0.4739",
"0.765",
"0.4739",
"0.78",
"0.0866",
"0.779"
],
[
"0.501",
"0.7646",
"0.8857",
"0.7646",
"0.8857",
"0.778",
"0.501",
"0.778"
],
[
"0.08826",
"0.781",
"0.2361",
"0.781",
"0.2361",
"0.7944",
"0.08826",
"0.7944"
],
[
"0.5",
"0.779",
"0.8857",
"0.779",
"0.8857",
"0.7925",
"0.5",
"0.7925"
],
[
"0.501",
"0.7944",
"0.886",
"0.7944",
"0.886",
"0.808",
"0.501",
"0.808"
],
[
"0.08826",
"0.815",
"0.2263",
"0.815",
"0.2263",
"0.829",
"0.08826",
"0.829"
],
[
"0.5",
"0.8096",
"0.714",
"0.8096",
"0.714",
"0.8228",
"0.5",
"0.8228"
],
[
"0.4983",
"0.83",
"0.886",
"0.8306",
"0.886",
"0.845",
"0.4983",
"0.8447"
],
[
"0.08905",
"0.8413",
"0.4739",
"0.8413",
"0.4739",
"0.855",
"0.08905",
"0.855"
],
[
"0.4968",
"0.8467",
"0.8857",
"0.8467",
"0.8857",
"0.86",
"0.4968",
"0.86"
],
[
"0.08826",
"0.857",
"0.4756",
"0.857",
"0.4756",
"0.8706",
"0.08826",
"0.8706"
],
[
"0.5",
"0.863",
"0.8857",
"0.863",
"0.8857",
"0.8735",
"0.5",
"0.8735"
],
[
"0.08826",
"0.872",
"0.4763",
"0.872",
"0.4763",
"0.8853",
"0.08826",
"0.8853"
],
[
"0.5",
"0.877",
"0.886",
"0.877",
"0.886",
"0.89",
"0.5",
"0.89"
],
[
"0.08826",
"0.887",
"0.4739",
"0.887",
"0.4739",
"0.901",
"0.08826",
"0.901"
],
[
"0.5",
"0.8926",
"0.886",
"0.8926",
"0.886",
"0.906",
"0.5",
"0.906"
]
] | [
"Just Train Twice: Improving Group Robustness without Training Group Information",
"We evaluate JTT on two image classification datasets with",
"gio et al., 2013; Szegedy et al., 2014) or domain gener",
"spurious correlations, Waterbirds (Wah et al., 201l; Sagawa",
"alization (Blanchard et al., 2011; Muandet et al., 2013).",
"et al., 2020a) and CelebA (Liu et al., 2015) and two natural",
"Approaches for group robustness fall into the two main",
"language processing datasets, MultiNLI (Williams et al.,",
"categories we discuss below.",
"2018) and CivilComments-WILDS (Borkan et al., 2019;",
"Koh et al., 2021).We use the versions of Waterbirds,",
"Robustness using group information. Several ap",
"CelebA, and MultiNLI from Sagawa et al. (2020a), where in",
"proaches leverage group information during training, either",
"Waterbirds,thelabelwaterbird orlandbird spuriouslycor-",
"to combat spurious correlations or handle shifts in group",
"relateswithwaterinthebackground:inCelebA.thelabel",
"proportions between train and test distributions. For exam-",
"blond or non-blond spuriously correlates with binary gen-",
"ple, Mohri et al. (2019); Sagawa et al. (2020a); Zhang et al.",
"der; and in MultiNLI, the label spuriously correlates with",
"(2020) minimize the worst-group loss during training; Goel",
"the presence of negation words. In CivilComments-WILDS",
"et al. (2020) synthetically expand the minority groups via",
"where the input is online comments, the label toxic, non-",
"generative modeling; Shimodaira (2000); Byrd & Lipton",
"toxic spuriously correlates with the mention of particular",
"(2019); Sagawa et al. (2020b) reweight or subsample to",
"demographics, as discussed above. Our method outperforms",
"artificially balance the majority and minority groups; Cao",
"ERM on all four datasets, with an average worst-group accu-",
"et al. (2019; 2020) impose heavy Lipschitz regularization",
"racy improvement of 16.2%, while maintaining competitive",
"around minority points. These approaches substantially re-",
"average accuracy (only 4.2% worse on average). Further-",
"duce worst-group error, but obtaining group annotations for",
"more, despite having no group annotations during training,",
"the entire training set can be extremely expensive.",
"JTT closes 75% of the gap between ERM and group DRO,",
"which uses complete group information on the training data.",
"Another line of work studies worst-group performance in",
"the context of fairness (Hardt et al., 2016; Woodworth et al.,",
"We then empirically analyze JTT. First, we analyze the",
"2017; Pleiss et al., 2017; Agarwal et al.,2018; Khani et al.,",
"examples identified by JTT and show that JTT upweights",
"2019). While these works also aim to improve the worst-",
"groups on which standard ERM models perform poorly,",
"group loss, they explicitly focus on equalizing the loss",
"e.g., minority groups that do not have the spurious corela-",
"across all groups.",
"tion (such as waterbirds on land in the Waterbirds dataset).",
"Second, we show that having validation group annotations",
"Robustness without group information.We focus on",
"is essential for hyperparameter tuning for JTT and other",
"the setting where group annotations are expensive and un-",
"related algorithms.",
"available on the training data, and potentially only available",
"Finally, we compare JTT with the distributionally robust op-",
"on a much smaller validation set. Many approaches for this",
"timization (DRO) algorithm that minimizes the conditional",
"setting fall under the general DRO framework, where mod",
"valueatrisk(CVaR).CVaRDROaimstotrainmodelsthat",
"els are trained to minimize the worst-case loss across all",
"are robust to a wide range of potential distribution shifts",
"distributions ina ball around theempirical distribution (Ben-",
"by minimizing the worst-case loss over all subsets of the",
"Tal et al., 2013; Lam & Zhou, 2015; Duchi et al., 2016;",
"training set of a certain size (Duchi et al., 2019). This",
"Namkoong & Duchi,2017; Oren et al., 2019).Pezeshkiet al.",
"objective does not require training group annotations, and",
"(2020) modify the dynamics of stochastic gradient descent",
"it can be optimized by dynamically upweighting training",
"to avoid learning spurious correlations. Sohoni et al. (2020)",
"examples with the highest losses in each minibatch (Levy",
"automatically identify groups based by clustering the data",
"et al., 2020). Though CVaR DRO and JTT share conceptual",
"points. Kim et al. (2019) propose an auditing scheme that",
"similarities-they both upweight high loss training points",
"searches for high-loss groups defined by a function within a",
"and do not require training group information-the differ-",
"pre-specified complexity class and postprocess the model",
"ence is that JTT upweights a static set of examples, while",
"to minimize discrepancies identified by the auditor.Khani",
"CVaR DRO dynamically re-computes which examples to",
"et al.(2019)minimizethe variance intheloss acrossall data",
"update. Empirically, we find that JTT empirically substan-",
"points to encourage lower discrepancy in the losses across",
"tially outperforms CVaR DRO on worst-group accuracy in",
"all possible groups. Another approach is to directly learn",
"the datasets we tested.",
"how to reweight the training examples either using small",
"amounts of metadata (Shu et al., 2019) or automatically via",
"2. Related Work",
"meta-learning (Ren et al., 2018).",
"Most closely related to JTT are several approaches that also",
"In this paper, we focus on group robustness (i.e., training",
"train a pair of models, where the performance of the first",
"models that obtain good performance on each of a set of",
"modelisusedtohelptrainthe secondmodel (Yaghoobzadeh",
"predefined groups in the dataset), though other notions of ro-",
"et al., 2019; Utama et al., 2020; Nam et al., 2020). We com-",
"bustness are also studied, such as adversarial examples (Big",
"pare JTT with one such approach, called Learning from"
] | {
"arxiv_id": "2107.09044",
"language": "en",
"page_idx": 1
} |
||
2107.09044-en-3 | 2107.09044-en-3.jpg | Just Train Twice: Improving Group Robustness without Training Group Information
3.2.4. GROUP DISTRIBUTIONALLY ROBUST
OPTIMIZATION (GROUP DRO)
and then minimize the expected loss over the worst-case
distribution in this uncertainty set (Duchi et al., 2019).
Group DRO uses training group annotations to directly minimize the worst-group error on the training set. Our primary focus in this paper is the setting where we do not have access to training group annotations. However, we use group DRO as an oracle method that upper bounds the performance we can expect without any training group annotations. Assume we have access to group annotations on the training data such that the $n$ training points are $\{(x_1,y_1,g_1),\ldots(x_n,y_n,g_n)\}.$ For some loss function $\ell(x,y;\theta)$,the group DRO objective can then be written as:
In this paper, we study a classic instance of this type of worst-case loss known as the conditional value at risk (CVaR) at level $\alpha\in(0,1]$, which corresponds to an uncertainty set that contains all $\alpha$-sized subpopulations of the training distribution (Rockafellar & Uryasev, 2000). The idea is that the worst loss over $\alpha$-sized subpopulations upper bounds the worst-group loss over the (unknown) groups in $\mathcal{G}$ when $\alpha$ is close to the size of the smallest group in $\mathcal{G}.$
In practice, we treat $\alpha$ as a hyperparameter. Concretely, for some loss function $\ell(x,y;\theta)$, the CVaR objective can be written as
(6)
$$J_{\text{group-DRO}}(\theta)=\max_{g\in\mathcal{G}}\frac{1}{n_g}\sum_{i|g_i=g}\ell(x_i,y_i;\theta)$$
$$J_{\mathrm{CVaR}}(\theta,\alpha)=\sup\limits_{q\in\Delta^{n}}\bigg\{\sum\limits_{i=1}^{n}q_{i}\ell(x_{i},y_{i};\theta)\:\mathrm{s.t.}\:\|q\|_{\infty}\leq\frac{1}{\alpha n}\bigg\},$$
(3)
where $n_g$ is the number of training points with group $g_i=g.$
## 4. JTT: Just Train Twice
where $\Delta^n$ is the probability simplex in $\mathbb{R}^n.$
Note that the CVaR objective is equivalent to the average loss incurred by the $\alpha$-fraction of training points that have the highest loss.
We now present JTT, a simple two-stage approach that does not require group annotations at training time. In the first stage, we train an identification model and select examples with high training loss. Then, in the second stage, we train a final model while upweighting the selected examples.
3.2.3. LEARNING FROM FAILURE (LFF)
LfF attempts to automatically upweight examples from challenging groups, such as those where the spurious correlation does not hold. It does this by learning two models $f_{B}(y\mid x;\theta_{B})$ and $f_D(y\mid x;\theta_D)$,parameterized by $\theta_B$ and $\theta_D.$
The first model $f_B$ is trained with ERM using generalized
cross-entropy (GCE) loss (Zhang & Sabuncu, 2018):
Stage 1 (identification). The key empirical observation that JTT builds on is that sufficiently low complexity ERM models tend to fit groups with easy-to-learn spurious correlations (e.g., landbirds on land and waterbirds on water in the Waterbirds dataset), but not groups that do not exhibit the same correlation (e.g., waterbirds on land) (Sagawa et al., 2020a). We therefore use the simple heuristic of first training an identifcation model $\hat{f}_{\mathrm{id}}$ via ERM and then identifying an error set E of training examples that $\hat{f}_{\mathrm{id}}$ misclassifies:
$$\ell_{\mathrm{GCE}}(x_i,y_i;\theta_B,q)=\frac{1-f_B(y_i\mid x_i;\theta_B)^q}{q},$$
(4)
$$E=\{(x_i,y_i)\text{s.t.}\hat{f}_{\mathrm{id}}(x_i)\neq y_i\}.$$
(7)
Stage 2 (upweighting). Next, we train a final model $\hat{f}_{\mathrm{final}}$ by upweighting the points in the error set $E$ identified in step one:
where $q\in[0,1)$ is a hyperparameter. Compared to standard cross-entropy loss, the gradient of GCE loss upweights examples where $f_B(y_i\mid x_i;\theta_B)$ is large, which intentionally biases $f_B$ to perform better on easier examples and poorly on examples group challenging groups.
The second model is also trained with ERM, using crossentropy loss, where each example $(x_i,y_i)$ is reweighted by a factor of:
$$J_{\text{up-ERM}}(\theta,E)=\left(\lambda_{\text{up}}\sum_{(x,y)\in E}\ell(x,y;\theta)+\sum_{(x,y)\not\in E}\ell(x,y;\theta)\right),$$
(8)
(5)
$$\mathcal{W}(x_i,y_i)=\frac{\log f_B(y_i\mid x_i)}{\log f_B(y_i\mid x_i)+\log f_D(y_i\mid x_i)}.$$
where $\lambda_{\mathrm{up}}\in\mathbb{R}_{+}$ is a hyperparameter. The hope is that if the examples in the error sets come from challenging groups, such as those where the spurious correlation does not hold, then upweighting them will lead to better worstgroup performance.
The hope is that early in training, $\log f_B(y_i\mid x_i)$ will be smaller than $\log f_D(y_i\mid x_i)$ on the easier examples, which leads to smaller weights on the easier examples and larger weights on the challenging examples.
Practical implementation. Overall, training JTT is summarized in Algorithm 1. In practice, to restrict the capacity | [
[
"0.2181",
"0.05682",
"0.755",
"0.05682",
"0.755",
"0.0707",
"0.2181",
"0.0707"
],
[
"0.08905",
"0.0859",
"0.4756",
"0.0859",
"0.4756",
"0.09973",
"0.08905",
"0.09973"
],
[
"0.5",
"0.0859",
"0.8",
"0.0859",
"0.8",
"0.09973",
"0.5",
"0.09973"
],
[
"0.08826",
"0.101",
"0.4453",
"0.101",
"0.4453",
"0.11554",
"0.08826",
"0.11554"
],
[
"0.546",
"0.101",
"0.7563",
"0.101",
"0.7563",
"0.11426",
"0.546",
"0.11426"
],
[
"0.08826",
"0.1237",
"0.4756",
"0.1237",
"0.4756",
"0.137",
"0.08826",
"0.137"
],
[
"0.5",
"0.125",
"0.886",
"0.125",
"0.886",
"0.1395",
"0.5",
"0.1395"
],
[
"0.08826",
"0.1395",
"0.4756",
"0.1395",
"0.4756",
"0.1528",
"0.08826",
"0.1528"
],
[
"0.4983",
"0.1401",
"0.886",
"0.1401",
"0.886",
"0.154",
"0.4983",
"0.154"
],
[
"0.08826",
"0.154",
"0.4756",
"0.154",
"0.4756",
"0.1686",
"0.08826",
"0.1686"
],
[
"0.4983",
"0.1547",
"0.886",
"0.1547",
"0.886",
"0.1692",
"0.4983",
"0.1692"
],
[
"0.08826",
"0.1692",
"0.4739",
"0.1692",
"0.4739",
"0.1831",
"0.08826",
"0.1831"
],
[
"0.4983",
"0.1704",
"0.8857",
"0.1704",
"0.8857",
"0.1843",
"0.4983",
"0.1843"
],
[
"0.08826",
"0.1831",
"0.4722",
"0.1831",
"0.4722",
"0.1963",
"0.08826",
"0.1963"
],
[
"0.5",
"0.1849",
"0.8857",
"0.1849",
"0.8857",
"0.1989",
"0.5",
"0.1989"
],
[
"0.08905",
"0.1995",
"0.4722",
"0.1995",
"0.4722",
"0.2134",
"0.08905",
"0.2134"
],
[
"0.4983",
"0.2008",
"0.8857",
"0.2008",
"0.8857",
"0.214",
"0.4983",
"0.214"
],
[
"0.08826",
"0.214",
"0.4722",
"0.214",
"0.4722",
"0.2279",
"0.08826",
"0.2279"
],
[
"0.5",
"0.2166",
"0.886",
"0.2166",
"0.886",
"0.2299",
"0.5",
"0.2299"
],
[
"0.08826",
"0.2299",
"0.4534",
"0.2299",
"0.4534",
"0.2437",
"0.08826",
"0.2437"
],
[
"0.5",
"0.2311",
"0.8857",
"0.2311",
"0.8857",
"0.2443",
"0.5",
"0.2443"
],
[
"0.5",
"0.2456",
"0.8857",
"0.2443",
"0.8857",
"0.2588",
"0.5",
"0.26"
],
[
"0.08826",
"0.252",
"0.4739",
"0.252",
"0.4739",
"0.2651",
"0.08826",
"0.2651"
],
[
"0.5",
"0.26",
"0.8857",
"0.2588",
"0.8857",
"0.2734",
"0.5",
"0.2747"
],
[
"0.08826",
"0.2676",
"0.4739",
"0.2676",
"0.4739",
"0.2822",
"0.08826",
"0.2822"
],
[
"0.0866",
"0.2827",
"0.1528",
"0.2827",
"0.1528",
"0.2935",
"0.0866",
"0.2935"
],
[
"0.7036",
"0.286",
"0.7183",
"0.286",
"0.7183",
"0.2954",
"0.7036",
"0.2954"
],
[
"0.7295",
"0.289",
"0.746",
"0.2935",
"0.7427",
"0.3005",
"0.7256",
"0.2966"
],
[
"0.749",
"0.2903",
"0.831",
"0.2922",
"0.831",
"0.308",
"0.749",
"0.3062"
],
[
"0.863",
"0.2922",
"0.886",
"0.2922",
"0.886",
"0.3062",
"0.863",
"0.3062"
],
[
"0.6626",
"0.2942",
"0.761",
"0.307",
"0.7573",
"0.327",
"0.6587",
"0.3145"
],
[
"0.556",
"0.2922",
"0.699",
"0.2922",
"0.699",
"0.307",
"0.556",
"0.307"
],
[
"0.08905",
"0.3213",
"0.1854",
"0.3213",
"0.1854",
"0.336",
"0.08905",
"0.336"
],
[
"0.1871",
"0.3245",
"0.2181",
"0.3245",
"0.2181",
"0.3333",
"0.1871",
"0.3333"
],
[
"0.2607",
"0.3213",
"0.4207",
"0.3208",
"0.4207",
"0.3352",
"0.2607",
"0.336"
],
[
"0.4355",
"0.332",
"0.4592",
"0.332",
"0.4592",
"0.3435",
"0.4355",
"0.3435"
],
[
"0.5",
"0.33",
"0.888",
"0.3308",
"0.888",
"0.3452",
"0.5",
"0.3447"
],
[
"0.183",
"0.3372",
"0.2181",
"0.3333",
"0.2214",
"0.3435",
"0.1854",
"0.348"
],
[
"0.452",
"0.3499",
"0.4756",
"0.3499",
"0.4756",
"0.363",
"0.452",
"0.363"
],
[
"0.501",
"0.3655",
"0.7075",
"0.3655",
"0.7075",
"0.3801",
"0.501",
"0.3801"
],
[
"0.08496",
"0.377",
"0.3708",
"0.3757",
"0.3708",
"0.3933",
"0.08496",
"0.3945"
],
[
"0.5",
"0.3894",
"0.8857",
"0.3901",
"0.8857",
"0.4048",
"0.5",
"0.404"
],
[
"0.08826",
"0.4016",
"0.4739",
"0.4016",
"0.4739",
"0.4153",
"0.08826",
"0.4153"
],
[
"0.5",
"0.4072",
"0.886",
"0.4072",
"0.886",
"0.4204",
"0.5",
"0.4204"
],
[
"0.0866",
"0.4167",
"0.4756",
"0.4172",
"0.4756",
"0.432",
"0.0866",
"0.4312"
],
[
"0.5",
"0.4216",
"0.886",
"0.4216",
"0.886",
"0.435",
"0.5",
"0.435"
],
[
"0.08905",
"0.4343",
"0.1969",
"0.4343",
"0.1969",
"0.4443",
"0.08905",
"0.4443"
],
[
"0.5024",
"0.4382",
"0.8857",
"0.4382",
"0.8857",
"0.449",
"0.5024",
"0.449"
],
[
"0.5",
"0.452",
"0.863",
"0.452",
"0.863",
"0.4653",
"0.5",
"0.4653"
],
[
"0.08905",
"0.4622",
"0.3652",
"0.4622",
"0.3652",
"0.476",
"0.08905",
"0.476"
],
[
"0.08905",
"0.486",
"0.4722",
"0.486",
"0.4722",
"0.5",
"0.08905",
"0.5"
],
[
"0.501",
"0.4817",
"0.886",
"0.4817",
"0.886",
"0.4956",
"0.501",
"0.4956"
],
[
"0.08905",
"0.502",
"0.4739",
"0.502",
"0.4739",
"0.515",
"0.08905",
"0.515"
],
[
"0.5",
"0.4976",
"0.886",
"0.4976",
"0.886",
"0.511",
"0.5",
"0.511"
],
[
"0.5",
"0.512",
"0.886",
"0.512",
"0.886",
"0.526",
"0.5",
"0.526"
],
[
"0.0907",
"0.518",
"0.4722",
"0.518",
"0.4722",
"0.529",
"0.0907",
"0.529"
],
[
"0.0907",
"0.531",
"0.4722",
"0.531",
"0.4722",
"0.5444",
"0.0907",
"0.5444"
],
[
"0.5",
"0.528",
"0.8857",
"0.528",
"0.8857",
"0.5415",
"0.5",
"0.5415"
],
[
"0.08826",
"0.5474",
"0.1136",
"0.5474",
"0.1136",
"0.56",
"0.08826",
"0.56"
],
[
"0.501",
"0.5425",
"0.8857",
"0.5425",
"0.8857",
"0.556",
"0.501",
"0.556"
],
[
"0.501",
"0.558",
"0.886",
"0.558",
"0.886",
"0.5713",
"0.501",
"0.5713"
],
[
"0.08826",
"0.5684",
"0.4739",
"0.5684",
"0.4739",
"0.582",
"0.08826",
"0.582"
],
[
"0.501",
"0.5728",
"0.886",
"0.5728",
"0.886",
"0.586",
"0.501",
"0.586"
],
[
"0.0866",
"0.5835",
"0.433",
"0.5835",
"0.433",
"0.5967",
"0.0866",
"0.5967"
],
[
"0.5",
"0.5884",
"0.886",
"0.5884",
"0.886",
"0.6016",
"0.5",
"0.6016"
],
[
"0.5",
"0.603",
"0.876",
"0.6016",
"0.876",
"0.616",
"0.5",
"0.617"
],
[
"0.134",
"0.6187",
"0.4207",
"0.605",
"0.4224",
"0.63",
"0.1356",
"0.643"
],
[
"0.452",
"0.6226",
"0.4739",
"0.6226",
"0.4739",
"0.6353",
"0.452",
"0.6353"
],
[
"0.5884",
"0.627",
"0.7983",
"0.627",
"0.7983",
"0.6416",
"0.5884",
"0.6416"
],
[
"0.3423",
"0.6333",
"0.3594",
"0.6333",
"0.3594",
"0.648",
"0.3423",
"0.648"
],
[
"0.863",
"0.629",
"0.886",
"0.629",
"0.886",
"0.6426",
"0.863",
"0.6426"
],
[
"0.0866",
"0.6577",
"0.4739",
"0.659",
"0.4739",
"0.674",
"0.0866",
"0.6724"
],
[
"0.4983",
"0.66",
"0.8857",
"0.662",
"0.8857",
"0.6797",
"0.4983",
"0.678"
],
[
"0.08826",
"0.675",
"0.4739",
"0.674",
"0.4739",
"0.688",
"0.08826",
"0.6895"
],
[
"0.4983",
"0.6787",
"0.8857",
"0.678",
"0.8857",
"0.6924",
"0.4983",
"0.6934"
],
[
"0.08826",
"0.6895",
"0.4722",
"0.688",
"0.4722",
"0.7026",
"0.08826",
"0.704"
],
[
"0.501",
"0.696",
"0.562",
"0.696",
"0.562",
"0.707",
"0.501",
"0.707"
],
[
"0.0866",
"0.704",
"0.4722",
"0.705",
"0.4722",
"0.7197",
"0.0866",
"0.7183"
],
[
"0.0866",
"0.7197",
"0.3506",
"0.72",
"0.3506",
"0.735",
"0.0866",
"0.7344"
],
[
"0.5024",
"0.729",
"0.597",
"0.729",
"0.597",
"0.7437",
"0.5024",
"0.7437"
],
[
"0.6567",
"0.726",
"0.88",
"0.726",
"0.88",
"0.747",
"0.6567",
"0.747"
],
[
"0.08826",
"0.742",
"0.4739",
"0.742",
"0.4739",
"0.756",
"0.08826",
"0.756"
],
[
"0.6504",
"0.747",
"0.7017",
"0.747",
"0.7017",
"0.761",
"0.6504",
"0.761"
],
[
"0.773",
"0.7476",
"0.8228",
"0.7456",
"0.824",
"0.7603",
"0.774",
"0.762"
],
[
"0.08826",
"0.7583",
"0.4739",
"0.7583",
"0.4739",
"0.773",
"0.08826",
"0.773"
],
[
"0.863",
"0.761",
"0.886",
"0.761",
"0.886",
"0.7744",
"0.863",
"0.7744"
],
[
"0.08905",
"0.7734",
"0.1626",
"0.7734",
"0.1626",
"0.785",
"0.08905",
"0.785"
],
[
"0.5",
"0.784",
"0.8857",
"0.784",
"0.8857",
"0.799",
"0.5",
"0.799"
],
[
"0.259",
"0.797",
"0.3635",
"0.799",
"0.3635",
"0.8164",
"0.259",
"0.8145"
],
[
"0.5",
"0.8",
"0.8857",
"0.8003",
"0.8857",
"0.815",
"0.5",
"0.8145"
],
[
"0.1168",
"0.806",
"0.2035",
"0.8096",
"0.2018",
"0.8247",
"0.1152",
"0.8213"
],
[
"0.1871",
"0.8125",
"0.4248",
"0.815",
"0.4248",
"0.834",
"0.1871",
"0.8306"
],
[
"0.452",
"0.8096",
"0.4739",
"0.8096",
"0.4739",
"0.8213",
"0.452",
"0.8213"
],
[
"0.5",
"0.8164",
"0.8857",
"0.815",
"0.8857",
"0.829",
"0.5",
"0.83"
],
[
"0.5",
"0.832",
"0.886",
"0.832",
"0.886",
"0.845",
"0.5",
"0.845"
],
[
"0.08826",
"0.8447",
"0.4756",
"0.8447",
"0.4756",
"0.8623",
"0.08826",
"0.8623"
],
[
"0.4983",
"0.8467",
"0.6323",
"0.845",
"0.6323",
"0.86",
"0.4983",
"0.8613"
],
[
"0.0866",
"0.8613",
"0.4739",
"0.86",
"0.4739",
"0.875",
"0.0866",
"0.8755"
],
[
"0.08826",
"0.877",
"0.4739",
"0.877",
"0.4739",
"0.89",
"0.08826",
"0.89"
],
[
"0.501",
"0.8755",
"0.886",
"0.8755",
"0.886",
"0.8896",
"0.501",
"0.8896"
],
[
"0.08826",
"0.8926",
"0.335",
"0.8926",
"0.335",
"0.906",
"0.08826",
"0.906"
],
[
"0.5",
"0.8926",
"0.886",
"0.8926",
"0.886",
"0.906",
"0.5",
"0.906"
]
] | [
"Just Train Twice: Improving Group Robustness without Training Group Information",
"and then minimize the expected loss over the worst-case",
"3.2.4. GROUP DISTRIBUTIONALLY ROBUST",
"distribution in this uncertainty set (Duchi et al., 2019).",
"OPTIMIZATION (GROUPDRO)",
"In this paper, we study a classic instance of this type of",
"Group DRO uses training group annotations to directly min-",
"worst-case loss known as the conditional value at risk",
"imize the worst-group error on the training set. Our pri-",
"(CVaR) at level α E (0, 1], which corresponds to an un-",
"mary focus in this paper is the setting where we do not",
"certainty set that contains all α-sized subpopulations of the",
"have access to training group annotations. However, we",
"training distribution (Rockafellar & Uryasev, 2000).The",
"use group DRO as an oracle method that upper bounds",
"idea is that the worst loss over Q-sized subpopulations upper",
"the performance we can expect without any training group",
"bounds the worst-group loss over the (unknown) groups in",
"annotations. Assume we have access to group annota-",
"G when Q is close to the size of the smallest group in g.",
"tions on the training data such that the n training points",
"are ((r1, y1,g1),.. (n, Yn, gn)). For some loss function",
"In practice, we treat α as a hyperparameter. Concretely, for",
"(c, y; 0), the group DRO objective can then be written as:",
"some loss function l(α, y; 0), the CVaR objective can be",
"written as",
"1",
"",
"l(ci,yi;0)",
"(6)",
"meo ng ilgi=g",
"Jgroup-DRo(0) = max",
"JcVaR(0, α) =",
"sup",
"qil(ci,Ys; O) s.t. Ilglloo",
"an",
"where ng is the number of training points with group gi = g.",
"en",
"(3)",
"4. JTT: Just Train Twice",
"where \" is the probability simplex in Rn.",
"We now present JTT, a simple two-stage approach that does",
"Note that the CVaR objective is equivalent to the average",
"not require group annotations at training time. In the first",
"loss incurred by the α-fraction of training points that have",
"stage, we train an identification model and select examples",
"the highest loss.",
"withhightrainingloss.Then,inthesecond stage,wetrain",
"a final model while upweighting the selected examples.",
"3.2.3. LEARNING FROM FAILURE (LFF)",
"LfF attempts to automatically upweight examples from chal.",
"Stage 1 (identification).The key empirical observation",
"lenging groups, such as those where the spurious correla",
"that JTT builds on is that sufficiently low complexity ERM",
"models tend to fit groups with easy-to-learn spurious correla-",
"tion doesnothold.Itdoes this bylearning twomodels",
"f(yr; @B) and fp(y ; D), parameterized by B and",
"tions (e.g., landbirds on land and waterbirds on water in the",
"OD.",
"Waterbirds dataset), but not groups that do not exhibit the",
"same correlation (e.g., waterbirds on land) (Sagawa et al.,",
"The first model fB is trained with ERM using generalized",
"2020a). We therefore use the simple heuristic of first train-",
"cross-entropy (GCE) loss (Zhang & Sabuncu, 2018):",
"ing an identification model fid via ERM and then identifying",
"an error set E of training examples that fid misclassifies:",
"lGce(ai, ;B,9) = 1-fB(: ;OB)",
"(4)",
"E ={(ci, Yi) s.t. fid(ci) 手 yil.",
"q",
"(7)",
"where g E [0, 1) is a hyperparameter. Compared to standard",
"Stage 2 (upweighting). Next, we train a final model frinal",
"cross-entropy loss, the gradient of GCE loss upweights ex-",
"by upweighting the points in the error set E identified in",
"amples where fB(y|i;B)is large, which intentionally",
"step one:",
"biases f to perform better on easier examples and poorly",
"on examples group challenging groups.",
"Jup-ERM(0, E)",
"Z l(a,y;0)+ Z e(a,y;0)",
"The second model is also trained with ERM, using cross-",
"(r,y)eE",
"(a,y)&E",
"entropy loss, where each example (ci, yi) is reweighted by",
"(8)",
"a factor of:",
"where Aup E IR+ is a hyperparameter. The hope is that",
"log fb(yi [ ri)",
"if the examples in the error sets come from challenging",
"W(ri, yi) =",
"=logfe(yi|ri)+log fp(y:|ai)",
"(5)",
"groups, such as those where the spurious correlation does",
"not hold, then upweighting them will lead to better worst-",
"The hope is that early in training, log fe(yi I ci) will be",
"group performance.",
"smaller than log fp(yi ri) on the easier examples, which",
"leads to smaller weights on the easier examples and larger",
"Practical implementation.Overall, training JTT is sum-",
"weights on the challenging examples.",
"marized in Algorithm 1. In practice, to restrict the capacity"
] | {
"arxiv_id": "2107.09044",
"language": "en",
"page_idx": 3
} |
||
2107.09044-en-4 | 2107.09044-en-4.jpg | Just Train Twice: Improving Group Robustness without Training Group Information
<table>
<tbody>
<tr>
<td>$\overline{\mathbf{A}}$ A 1</td>
<td>$\overline{\text{training}}$</td>
</tr>
</tbody>
</table>
of the identification model, we only train it for $T$ steps, where $T$ is a hyperparameter (line 1). This prevents it from potentially overfitting the training data and yielding an empty error set. To implement the upweighted objective (8), we simply upsample the examples from the error set by $\lambda_\mathrm{up}$ (line 3) and train the final model on the upsampled data (line 4). Specifically, in each epoch of training, we sample each example from the error set $\lambda_\mathrm{up}$ times and all other examples only once.
We tune the algorithm hyperparameters (the number of training epochs $T$ for the identification model $\hat{f}_\mathrm{id}$, and the upweight factor $\lambda_\mathrm{up})$ and both identification and final model hyperparameters (e.g., the learning rate and $\ell_2$ regularization strength) based on the worst-group error of the final model $\hat{f}_{\mathrm{final}}$ on the validation set. In our experiments, we share the same hyperparameters and architecture between the identification and final models, outside of the early stopping $T$ of the identification model, and we sometimes find it helpful to learn them with different optimizers. Note that setting the upweight factor $\lambda_\mathrm{up}$ to 1 recovers ERM, so JTT should perform at least as well as ERM, given a sufficiently large validation set. We describe full training details in Appendix A.
# 5. Experiments
In our experiments, we first demonstrate that JTT substantially improves worst-group performance compared to standard ERM models (Section 5.2). We also show that it recovers a significant fraction of the performance gains yielded by group DRO, which, as discussed in Section 3, is an oracle that relies on group annotations on training examples. We then present empirical analysis of JTT, including the analysis of the error set (Section 5.3), exploration on the role of the validation set (Section 5.4), and comparison with CVaR DRO (Section 5.5).
# 5.1. Setup
We study four datasets in which prior work has observed poor worst-group performance due to spurious correlations
(Figure 2). Full details about these datasets are in Ap-
pendix B.
$\cdot$ Waterbirds (Wah et al., 2011; Sagawa et al., 2020a): The task is to classify images of birds as “waterbird” or “landbird”, and the label is spuriously correlated with the image background, which is either“land”or “water.”
$\cdot$ CelebA (Liu et al., 2015): We consider the task from Sagawa et al. (2020a) of classifying the hair color of celebrities as“blond”or“not blond."The label is spuriously correlated with gender, which is either “male” or “female.”
$\cdot$ MultiNLI (Williams et al., 2018): Given a pair of sen-
tences, the task is to classify whether the second sentence is entailed by, neutral with, or contradicts the first sentence. We use the spurious attribute from Sagawa et al. (2020a), which is the presence of negation words in the second sentence; due to the artifacts from the data collection process, contradiction examples often include negation words.
$\cdot$ CivilComments-WILDS (Borkan et al., 2019; Koh et al., 2021): The task is to classify whether an online comment is toxic or non-toxic, and the label is spuriously correlated with mentions of certain demographic identities (male, female, White, Black, LGBTQ, Mus- $\lim$, Christian, and other religion). We use the evaluation metric from Koh et al. (2021), which defines 16 overlapping groups $(a,toxic)$ and $(a,non-toxic)$ for each of the above 8 demographic identities $a$, and report the worst-group performance over these groups.
Points of comparison. We aim to answer two main questions: (1) How does JTT compare with other approaches that also do not use training group information? (2) How does JTT compare with approaches that $do$ use training group information?
To answer the first question, we compare JTT with ERM, CVaR DRO, and a recently proposed approach called Learning from Failure (LfF) (Nam et al.,2020). To answer the second question, we compare JTT with group DRO (Sagawa et al., 2020a), an oracle that uses training group annotations. For details about these approaches, see Section 3. Note that on CivilComments, group DRO cannot be directly applied on the 16 defined groups, since it is not designed for overlapping groups. Instead, our group DRO minimizes worst-group loss over 4 groups $(y,a)$,where the spurious attribute $a$ is a binary indicator of whether any demographic identity is mentioned and the label $y$ is toxic or non-toxic. We tune the hyperparameters of all approaches based on worst-group performance on a small validation set with group annotations. | [
[
"0.2197",
"0.05682",
"0.7563",
"0.05682",
"0.7563",
"0.0707",
"0.2197",
"0.0707"
],
[
"0.08905",
"0.0852",
"0.2573",
"0.0859",
"0.2573",
"0.09973",
"0.08905",
"0.0985"
],
[
"0.501",
"0.0859",
"0.886",
"0.0859",
"0.886",
"0.09973",
"0.501",
"0.09973"
],
[
"0.10455",
"0.1016",
"0.464",
"0.1029",
"0.464",
"0.11743",
"0.10455",
"0.11615"
],
[
"0.5",
"0.101",
"0.568",
"0.101",
"0.568",
"0.11554",
"0.5",
"0.11554"
],
[
"0.10455",
"0.11743",
"0.2769",
"0.11615",
"0.2769",
"0.1307",
"0.10455",
"0.132"
],
[
"0.10455",
"0.132",
"0.4453",
"0.132",
"0.4453",
"0.1465",
"0.10455",
"0.1465"
],
[
"0.517",
"0.132",
"0.886",
"0.132",
"0.886",
"0.1459",
"0.517",
"0.1459"
],
[
"0.10297",
"0.1477",
"0.423",
"0.1477",
"0.423",
"0.161",
"0.10297",
"0.161"
],
[
"0.5293",
"0.1465",
"0.886",
"0.1459",
"0.886",
"0.1604",
"0.5293",
"0.161"
],
[
"0.10455",
"0.1635",
"0.3245",
"0.1635",
"0.3245",
"0.178",
"0.10455",
"0.178"
],
[
"0.5317",
"0.161",
"0.886",
"0.161",
"0.886",
"0.1749",
"0.5317",
"0.1749"
],
[
"0.1054",
"0.178",
"0.3865",
"0.178",
"0.3865",
"0.1925",
"0.1054",
"0.1925"
],
[
"0.5337",
"0.178",
"0.886",
"0.178",
"0.886",
"0.1913",
"0.5337",
"0.1913"
],
[
"0.10455",
"0.1932",
"0.4739",
"0.1932",
"0.4739",
"0.2084",
"0.10455",
"0.2084"
],
[
"0.5312",
"0.1945",
"0.584",
"0.1945",
"0.584",
"0.2058",
"0.5312",
"0.2058"
],
[
"0.1054",
"0.209",
"0.455",
"0.209",
"0.455",
"0.2229",
"0.1054",
"0.2229"
],
[
"0.5195",
"0.2166",
"0.8857",
"0.2166",
"0.8857",
"0.2299",
"0.5195",
"0.2299"
],
[
"0.10455",
"0.226",
"0.4697",
"0.226",
"0.4697",
"0.2393",
"0.10455",
"0.2393"
],
[
"0.5317",
"0.2317",
"0.8857",
"0.2317",
"0.8857",
"0.2456",
"0.5317",
"0.2456"
],
[
"0.5317",
"0.2468",
"0.8857",
"0.2468",
"0.8857",
"0.2615",
"0.5317",
"0.2615"
],
[
"0.5317",
"0.262",
"0.886",
"0.262",
"0.886",
"0.276",
"0.5317",
"0.276"
],
[
"0.5337",
"0.2766",
"0.613",
"0.2766",
"0.613",
"0.2886",
"0.5337",
"0.2886"
],
[
"0.0866",
"0.2842",
"0.4739",
"0.2854",
"0.4739",
"0.2998",
"0.0866",
"0.2986"
],
[
"0.08826",
"0.3005",
"0.4739",
"0.3005",
"0.4739",
"0.3145",
"0.08826",
"0.3145"
],
[
"0.5186",
"0.3005",
"0.886",
"0.3005",
"0.886",
"0.3145",
"0.5186",
"0.3145"
],
[
"0.08905",
"0.315",
"0.4739",
"0.315",
"0.4739",
"0.3289",
"0.08905",
"0.3289"
],
[
"0.5317",
"0.3164",
"0.886",
"0.3164",
"0.886",
"0.33",
"0.5317",
"0.33"
],
[
"0.08826",
"0.3308",
"0.4739",
"0.3308",
"0.4739",
"0.3447",
"0.08826",
"0.3447"
],
[
"0.5337",
"0.3333",
"0.8857",
"0.3333",
"0.8857",
"0.3435",
"0.5337",
"0.3435"
],
[
"0.08826",
"0.3447",
"0.4739",
"0.3452",
"0.4739",
"0.3599",
"0.08826",
"0.3591"
],
[
"0.5317",
"0.3467",
"0.886",
"0.3467",
"0.886",
"0.3599",
"0.5317",
"0.3599"
],
[
"0.08826",
"0.361",
"0.4739",
"0.361",
"0.4739",
"0.375",
"0.08826",
"0.375"
],
[
"0.5337",
"0.3599",
"0.8857",
"0.3599",
"0.8857",
"0.3738",
"0.5337",
"0.3738"
],
[
"0.08826",
"0.377",
"0.4739",
"0.377",
"0.4739",
"0.3901",
"0.08826",
"0.3901"
],
[
"0.5337",
"0.377",
"0.8857",
"0.377",
"0.8857",
"0.3901",
"0.5337",
"0.3901"
],
[
"0.08826",
"0.3914",
"0.4739",
"0.3914",
"0.4739",
"0.4048",
"0.08826",
"0.4048"
],
[
"0.5317",
"0.3926",
"0.886",
"0.3926",
"0.886",
"0.406",
"0.5317",
"0.406"
],
[
"0.08826",
"0.4077",
"0.1569",
"0.4077",
"0.1569",
"0.4192",
"0.08826",
"0.4192"
],
[
"0.5317",
"0.4072",
"0.6914",
"0.4072",
"0.6914",
"0.4204",
"0.5317",
"0.4204"
],
[
"0.0866",
"0.4268",
"0.4739",
"0.4287",
"0.4739",
"0.443",
"0.0866",
"0.4414"
],
[
"0.5186",
"0.4287",
"0.886",
"0.4287",
"0.886",
"0.443",
"0.5186",
"0.443"
],
[
"0.08905",
"0.443",
"0.4722",
"0.443",
"0.4722",
"0.4583",
"0.08905",
"0.4583"
],
[
"0.5312",
"0.4458",
"0.886",
"0.4443",
"0.886",
"0.4583",
"0.5312",
"0.459"
],
[
"0.08826",
"0.459",
"0.4739",
"0.459",
"0.4739",
"0.4734",
"0.08826",
"0.4734"
],
[
"0.535",
"0.4622",
"0.884",
"0.4622",
"0.884",
"0.473",
"0.535",
"0.473"
],
[
"0.08905",
"0.4734",
"0.4739",
"0.4734",
"0.4739",
"0.4873",
"0.08905",
"0.4873"
],
[
"0.5317",
"0.476",
"0.8857",
"0.476",
"0.8857",
"0.4893",
"0.5317",
"0.4893"
],
[
"0.08826",
"0.4893",
"0.4739",
"0.4893",
"0.4739",
"0.5024",
"0.08826",
"0.5024"
],
[
"0.5337",
"0.4905",
"0.8857",
"0.4905",
"0.8857",
"0.504",
"0.5337",
"0.504"
],
[
"0.08826",
"0.504",
"0.4739",
"0.504",
"0.4739",
"0.518",
"0.08826",
"0.518"
],
[
"0.5317",
"0.506",
"0.886",
"0.506",
"0.886",
"0.5195",
"0.5317",
"0.5195"
],
[
"0.08905",
"0.5195",
"0.4739",
"0.5195",
"0.4739",
"0.5327",
"0.08905",
"0.5327"
],
[
"0.5317",
"0.521",
"0.886",
"0.521",
"0.886",
"0.534",
"0.5317",
"0.534"
],
[
"0.08905",
"0.535",
"0.4722",
"0.535",
"0.4722",
"0.549",
"0.08905",
"0.549"
],
[
"0.5337",
"0.537",
"0.886",
"0.537",
"0.886",
"0.5503",
"0.5337",
"0.5503"
],
[
"0.08826",
"0.55",
"0.4739",
"0.549",
"0.4739",
"0.5625",
"0.08826",
"0.563"
],
[
"0.5337",
"0.552",
"0.886",
"0.552",
"0.886",
"0.565",
"0.5337",
"0.565"
],
[
"0.0907",
"0.5664",
"0.4722",
"0.5664",
"0.4722",
"0.577",
"0.0907",
"0.577"
],
[
"0.5317",
"0.5664",
"0.8774",
"0.5664",
"0.8774",
"0.58",
"0.5317",
"0.58"
],
[
"0.08826",
"0.58",
"0.4739",
"0.58",
"0.4739",
"0.5933",
"0.08826",
"0.5933"
],
[
"0.08905",
"0.5933",
"0.4722",
"0.5933",
"0.4722",
"0.608",
"0.08905",
"0.608"
],
[
"0.4983",
"0.595",
"0.886",
"0.5977",
"0.886",
"0.6123",
"0.4983",
"0.61"
],
[
"0.08905",
"0.61",
"0.4756",
"0.61",
"0.4756",
"0.6235",
"0.08905",
"0.6235"
],
[
"0.5",
"0.613",
"0.886",
"0.613",
"0.886",
"0.627",
"0.5",
"0.627"
],
[
"0.08826",
"0.627",
"0.1732",
"0.627",
"0.1732",
"0.638",
"0.08826",
"0.638"
],
[
"0.5",
"0.627",
"0.886",
"0.6274",
"0.886",
"0.6426",
"0.5",
"0.6416"
],
[
"0.5",
"0.6426",
"0.8857",
"0.643",
"0.8857",
"0.6577",
"0.5",
"0.657"
],
[
"0.08826",
"0.659",
"0.2173",
"0.659",
"0.2173",
"0.674",
"0.08826",
"0.674"
],
[
"0.4983",
"0.659",
"0.5884",
"0.657",
"0.5884",
"0.672",
"0.4983",
"0.674"
],
[
"0.08905",
"0.685",
"0.4739",
"0.685",
"0.4739",
"0.6987",
"0.08905",
"0.6987"
],
[
"0.5",
"0.6797",
"0.886",
"0.6797",
"0.886",
"0.6943",
"0.5",
"0.6943"
],
[
"0.5",
"0.6934",
"0.886",
"0.6943",
"0.886",
"0.709",
"0.5",
"0.7085"
],
[
"0.08826",
"0.6997",
"0.4722",
"0.6997",
"0.4722",
"0.7134",
"0.08826",
"0.7134"
],
[
"0.08826",
"0.7153",
"0.4756",
"0.7153",
"0.4756",
"0.729",
"0.08826",
"0.729"
],
[
"0.501",
"0.7104",
"0.8857",
"0.7104",
"0.8857",
"0.7236",
"0.501",
"0.7236"
],
[
"0.08905",
"0.731",
"0.4756",
"0.731",
"0.4756",
"0.744",
"0.08905",
"0.744"
],
[
"0.5",
"0.726",
"0.886",
"0.726",
"0.886",
"0.7393",
"0.5",
"0.7393"
],
[
"0.08826",
"0.747",
"0.4756",
"0.747",
"0.4756",
"0.7603",
"0.08826",
"0.7603"
],
[
"0.501",
"0.741",
"0.886",
"0.741",
"0.886",
"0.755",
"0.501",
"0.755"
],
[
"0.08826",
"0.761",
"0.4739",
"0.761",
"0.4739",
"0.7744",
"0.08826",
"0.7744"
],
[
"0.501",
"0.756",
"0.8857",
"0.756",
"0.8857",
"0.771",
"0.501",
"0.771"
],
[
"0.08826",
"0.7764",
"0.4739",
"0.7764",
"0.4739",
"0.7905",
"0.08826",
"0.7905"
],
[
"0.5",
"0.7715",
"0.886",
"0.7715",
"0.886",
"0.785",
"0.5",
"0.785"
],
[
"0.08826",
"0.7915",
"0.4739",
"0.7915",
"0.4739",
"0.8047",
"0.08826",
"0.8047"
],
[
"0.5",
"0.786",
"0.886",
"0.785",
"0.886",
"0.8",
"0.5",
"0.8003"
],
[
"0.08826",
"0.8066",
"0.4739",
"0.8066",
"0.4739",
"0.821",
"0.08826",
"0.821"
],
[
"0.501",
"0.802",
"0.8857",
"0.802",
"0.8857",
"0.815",
"0.501",
"0.815"
],
[
"0.08826",
"0.8213",
"0.2607",
"0.8213",
"0.2607",
"0.8354",
"0.08826",
"0.8354"
],
[
"0.501",
"0.8174",
"0.8857",
"0.8174",
"0.8857",
"0.8306",
"0.501",
"0.8306"
],
[
"0.5024",
"0.8335",
"0.884",
"0.8335",
"0.884",
"0.8433",
"0.5024",
"0.8433"
],
[
"0.501",
"0.8467",
"0.886",
"0.8467",
"0.886",
"0.86",
"0.501",
"0.86"
],
[
"0.08826",
"0.853",
"0.1602",
"0.853",
"0.1602",
"0.867",
"0.08826",
"0.867"
],
[
"0.4983",
"0.86",
"0.886",
"0.8613",
"0.886",
"0.8755",
"0.4983",
"0.875"
],
[
"0.08826",
"0.8755",
"0.4739",
"0.8755",
"0.4739",
"0.8896",
"0.08826",
"0.8896"
],
[
"0.5",
"0.877",
"0.886",
"0.877",
"0.886",
"0.89",
"0.5",
"0.89"
],
[
"0.08826",
"0.8926",
"0.4756",
"0.8926",
"0.4756",
"0.906",
"0.08826",
"0.906"
],
[
"0.5",
"0.8926",
"0.625",
"0.8926",
"0.625",
"0.906",
"0.5",
"0.906"
]
] | [
"Just Train Twice: Improving Group Robustness without Training Group Information",
"Algorithm1JTTtraining",
"(Figure 2). Full details about these datasets are in Ap-",
"Input: Training set D and hyperparameters T and Aup.",
"pendix B.",
"Stage one: identification",
"1. Train fid on D via ERM for T steps (Equation 2).",
". Waterbirds (Wah et al., 2011; Sagawa et al., 2020a):",
"2. Construct the error set E of training examples",
"The task is to classify images of birds as “waterbird",
"misclassified by fid (Equation 7).",
"or \"landbird\", and the label is spuriously correlated",
"Stage two: upweighting identified points",
"with the image background, which is either“landor",
"3. Construct upsampled dataset Dup containing examples",
"\"\"water.\"",
"in the error set Aup times and all other examples once.",
". CelebA (Liu et al., 2015): We consider the task from",
"4. Train final model final on Dup via ERM (Equation 2),",
"Sagawa et al. (2020a) of classifying the hair color of",
"celebrities as \"blond\" or “not blond.\" The label is spu-",
"riously correlated with gender, which is either “male\"",
"or\"female.\"",
"of the identification model, we only train it for T steps.",
"where T is a hyperparameter (line 1). This prevents it",
" MultiNLI (Williams et al., 2018): Given a pair of sen-",
"from potentially overfitting the training data and yielding an",
"tences, the task is to classify whether the second sen-",
"empty error set. To implement the upweighted objective (8),",
"tenceis entailed by,neutral with,or contradicts thefirst",
"we simply upsample the examples from the error set by Aup",
"sentence. We use the spurious attribute from Sagawa",
"(line 3) and train the final model on the upsampled data (line",
"et al. (2020a), which is the presence of negation words",
"4). Specifically, in each epoch of training, we sample each",
"in the second sentence; due to the artifacts from the",
"example from the error set Aup times and all other examples",
"data collection process, contradiction examples often",
"only once.",
"include negation words.",
"We tune the algorithm hyperparameters (the number of train-",
". CivilComments-WILDS (Borkan et al., 2019; Koh",
"ing epochs T for the identification model fid, and the up",
"et al.,2021):Thetask is to classify whether an online",
"weight factor Xup) and both identification and final model",
"commentistoxicornon-toxic,andthelabelis spuri",
"hyperparameters (e.g., the learning rate and l2 regulariza-",
"ously correlated with mentions of certain demographic",
"tion strength) based on the worst-group error of the final",
"identities (male, female, White, Black, LGBTQ, Mus-",
"model final on the validation set. In our experiments, we",
"lim, Christian, and other religion). We use the eval-",
"share the same hyperparameters and architecture between",
"uation metric from Koh et al. (2021), which defines",
"the identification and final models, outside of the early stop",
"16 overlapping groups (a, toxic) and (a, non-toxic) for",
"ping T of the identification model, and we sometimes find",
"each of the above 8 demographic identities a, and re-",
"ithelpful to learn them with differentoptimizers.Note that",
"port the worst-group performance over these groups.",
"setting the upweight factor Aup to 1 recovers ERM, so JTT",
"should perform at least as well as ERM, given a sufficiently",
"Points of comparison.We aim to answer two main ques-",
"large validation set. We describe full training details in",
"tions: (1) How does JTT compare with other approaches that",
"Appendix A.",
"also do not use training group information? (2) How does",
"JTT compare with approaches that do use training group",
"5. Experiments",
"information?",
"In our experiments, we first demonstrate that JTT substan-",
"To answer the first question, we compare JTT with ERM,",
"CVaR DRO, and a recently proposed approach called Learn-",
"tially improves worst-group performance compared to stan",
"dard ERM models (Section 5.2). We also show that it recov-",
"ing from Failure (LfF) (Nam et al., 2020). To answer the",
"ers a significant fraction of the performance gains yielded",
"second question, we compare JTT with group DRO (Sagawa",
"by group DRO, which, as discussed in Section 3, is an or-",
"et al., 2020a), an oracle that uses training group annotations.",
"acle that relies on group annotations on training examples.",
"For details about these approaches, see Section 3. Note",
"We then present empirical analysis of JTT, including the",
"that on CivilComments, group DRO cannot be directly ap-",
"analysis of the error set (Section 5.3), exploration on the",
"plied on the 16 defined groups, since it is not designed for",
"role of the validation set (Section 5.4), and comparison with",
"overlapping groups. Instead, our group DRO minimizes",
"CVaR DRO (Section 5.5).",
"worst-group loss over 4 groups (y, a), where the spurious",
"attributeaisabinaryindicatorofwhetheranydemographic",
"identity is mentioned and the label y is toxic or non-toxic.",
"5.1. Setup",
"We tune the hyperparameters of all approaches based on",
"We study four datasets in which prior work has observed",
"worst-group performance on a small validation set with",
"poor worst-group performance due to spurious correlations",
"group annotations."
] | {
"arxiv_id": "2107.09044",
"language": "en",
"page_idx": 4
} |
||
2403.12599-en-0 | 2403.12599-en-0.jpg | ## Preventing Eviction-Caused Homelessness through ML-Informed Distribution of Rental Assistance
$\textbf{Catalina Vajiac}^{*\:1},\textbf{Arun Frey}^{*2},\textbf{Joachim Baumann}^{*3,4},\textbf{Abigail Smith}^{*5}$,
$\textbf{Kasun Amarasinghe}^1,\textbf{Alice Lai}^1,\textbf{Kit Rodolfa}^2,\textbf{Rayid Ghani}^1$
$^{1}$Carnegie Mellon University,
$^{2}$Stanford University, $^3$University of Zurich,
$^{4}$Zurich University of Applied Sciences, $^{5}$NORC at the Úniversity of Chicago
cvajiac@ cs.cmu.edu, arunfrey@law.stanford.edu, baumann@iff.uzh.ch, als1@u.northwestern.edu, $\{$kamarasi, alicelai
ghani$\}@$andrew.cmu.edu, krodolfa@ law.stanford.edu
## $\mathbf{Abstract}$
Rental assistance programs provide individuals with financial assistance to prevent housing instabilities caused by evictions and avert homelessness. Since these programs operate under resource constraints, they must decide who to prioritize. Typically, funding is distributed by a reactive or firstcome-first serve allocation process that does not systematically consider risk of future homelessness. We partnered with Allegheny County, PA to explore a proactive allocation approach that prioritizes individuals facing eviction based on their risk of future homelessness. Our ML system that uses state and county administrative data to accurately identify individuals in need of support outperforms simpler prioritization approaches by at least 20% while being fair and equitable across race and gender. Furthermore, our approach would identify 28% of individuals who are overlooked by the current process and end up homeless. Beyond improvements to the rental assistance program in Allegheny County, this study can inform the development of evidence-based decision support tools in similar contexts, including lessons about data needs, model design, evaluation, and field validation.
## 1 Introduction
Homelessness remains a pervasive and pressing issue across the United States. In January 2022, more than 500k individuals experienced homelessness on a single night (de Sousa et al. 2022). Rising eviction rates and the lack of affordable housing contribute to this problem: as the gap between housing costs and income levels continues to grow, an increasing number of households struggle to pay rent, eventually facing eviction and, in some cases, homelessness.
To curb rates of homelessness, policymakers are increasingly seeking to reduce the rate of entry into homelessness (Culhane, Metraux, and Byrne 2011). One popular prevention strategy is rental assistance programs, which provide temporary financial assistance to individuals facing eviction
$^*$These authors contributed equally.
Copyright $\odot2024$, Association for the Advancement of Artificial
Intelligence (www.aaai.org). All rights reserved.
to keep them stably housed. Experimental evidence suggests that such assistance is effective at reducing subsequent homelessness: using quasi-random variation in funding availability of a rental assistance program in Chicago, it was shown that individuals who called when funding was available were 76% less likely to become homeless in the subsequent 6 months compared to those who called when funding was unavailable $\overset{.}{\operatorname*{(Evans,Sullivan,~and~Wallskog}}$ 2016). A similar program in NYC randomized the provision of financial assistance to households and reduced average days in shelter from 32 nights among the control group to 10 nights among the treatment group in the 2 years following funding (Rolston, Geyer, and Locke 2013).
For rental assistance programs to be a viable homelessness prevention strategy, they need to be effective (i.e. prevent individuals from falling into homelessness), efficient (i.e. target individuals at risk of falling into homelessness) (Burt 2007), and equitable. Currently, however, these programs prioritize individuals based on simple heuristics (e.g. first come, first served) instead of their likelihood of future homelessness. In addition, they place the burden of applying for funding on the individuals facing eviction and overlook those who need help but do not apply. The administrative process can create long delays, so that even eligible tenants often remain on a waitlist for multiple months and can end up evicted before they receive assistance.
In this paper, we describe our collaboration with the Allegheny County Department of Human Services (ACDHS) to improve the allocation of rental assistance by prioritizing individuals with the highest likelihood of falling into homelessness. We combine rich county and state administrative data with court records to predict homelessness among all individuals currently facing eviction, regardless of whether they contacted the county for fnancial support. Using data from January 2012 through August 2023, we develop a series of machine learning (ML) models that predict a tenant's need for homelessness services in the next 12 months, allowing ACDHS to proactively prioritize rental assistance to the most vulnerable tenants. In particular, we make the fol- | [
[
"0.0972",
"0.1237",
"0.9043",
"0.1237",
"0.9043",
"0.1383",
"0.0972",
"0.1383"
],
[
"0.4084",
"0.1427",
"0.589",
"0.1433",
"0.589",
"0.161",
"0.4084",
"0.1604"
],
[
"0.2035",
"0.1755",
"0.8",
"0.1755",
"0.8",
"0.1932",
"0.2035",
"0.1932"
],
[
"0.2418",
"0.1945",
"0.7524",
"0.1932",
"0.7524",
"0.2108",
"0.2418",
"0.2122"
],
[
"0.4019",
"0.2134",
"0.597",
"0.2153",
"0.597",
"0.2329",
"0.4019",
"0.2311"
],
[
"0.4272",
"0.2291",
"0.569",
"0.2311",
"0.569",
"0.2456",
"0.4272",
"0.2437"
],
[
"0.4248",
"0.2437",
"0.572",
"0.2443",
"0.572",
"0.2588",
"0.4248",
"0.2583"
],
[
"0.02287",
"0.2568",
"0.06046",
"0.2568",
"0.06046",
"0.3738",
"0.02287",
"0.3738"
],
[
"0.3652",
"0.2551",
"0.6323",
"0.2568",
"0.6323",
"0.2747",
"0.3652",
"0.2727"
],
[
"0.3774",
"0.2715",
"0.621",
"0.2715",
"0.621",
"0.2854",
"0.3774",
"0.2854"
],
[
"0.1136",
"0.2874",
"0.8857",
"0.2874",
"0.8857",
"0.3005",
"0.1136",
"0.3005"
],
[
"0.3228",
"0.3005",
"0.6772",
"0.3005",
"0.6772",
"0.3145",
"0.3228",
"0.3145"
],
[
"0.2507",
"0.3452",
"0.3162",
"0.348",
"0.3145",
"0.363",
"0.25",
"0.361"
],
[
"0.5186",
"0.3467",
"0.911",
"0.348",
"0.911",
"0.3623",
"0.5186",
"0.361"
],
[
"0.02696",
"0.363",
"0.05392",
"0.363",
"0.05392",
"0.3945",
"0.02696",
"0.3945"
],
[
"0.5195",
"0.3623",
"0.91",
"0.3623",
"0.91",
"0.3757",
"0.5195",
"0.3757"
],
[
"0.10297",
"0.3706",
"0.4617",
"0.3706",
"0.4617",
"0.3838",
"0.10297",
"0.3838"
],
[
"0.5195",
"0.3757",
"0.91",
"0.3757",
"0.91",
"0.3894",
"0.5195",
"0.3894"
],
[
"0.10297",
"0.3833",
"0.4617",
"0.3833",
"0.4617",
"0.3965",
"0.10297",
"0.3965"
],
[
"0.5195",
"0.3901",
"0.9126",
"0.3901",
"0.9126",
"0.404",
"0.5195",
"0.404"
],
[
"0.01878",
"0.3997",
"0.05963",
"0.3984",
"0.0621",
"0.486",
"0.02124",
"0.4873"
],
[
"0.10455",
"0.3977",
"0.4617",
"0.3977",
"0.4617",
"0.4077",
"0.10455",
"0.4077"
],
[
"0.10297",
"0.4092",
"0.4631",
"0.4092",
"0.4631",
"0.4224",
"0.10297",
"0.4224"
],
[
"0.5195",
"0.404",
"0.9126",
"0.404",
"0.9126",
"0.4172",
"0.5195",
"0.4172"
],
[
"0.04166",
"0.411",
"0.05228",
"0.411",
"0.05228",
"0.4153",
"0.04166",
"0.4153"
],
[
"0.10297",
"0.4204",
"0.4631",
"0.4204",
"0.4631",
"0.4343",
"0.10297",
"0.4343"
],
[
"0.5195",
"0.4172",
"0.9126",
"0.4172",
"0.9126",
"0.4312",
"0.5195",
"0.4312"
],
[
"0.10455",
"0.435",
"0.46",
"0.435",
"0.46",
"0.4458",
"0.10455",
"0.4458"
],
[
"0.5215",
"0.4312",
"0.9126",
"0.4312",
"0.9126",
"0.4443",
"0.5215",
"0.4443"
],
[
"0.10455",
"0.4475",
"0.4617",
"0.4475",
"0.4617",
"0.4583",
"0.10455",
"0.4583"
],
[
"0.5195",
"0.4443",
"0.9126",
"0.4443",
"0.9126",
"0.4583",
"0.5195",
"0.4583"
],
[
"0.10455",
"0.459",
"0.4617",
"0.459",
"0.4617",
"0.473",
"0.10455",
"0.473"
],
[
"0.5186",
"0.457",
"0.9126",
"0.4583",
"0.9126",
"0.473",
"0.5186",
"0.4717"
],
[
"0.10455",
"0.473",
"0.4617",
"0.473",
"0.4617",
"0.483",
"0.10455",
"0.483"
],
[
"0.5215",
"0.473",
"0.914",
"0.473",
"0.914",
"0.486",
"0.5215",
"0.486"
],
[
"0.10297",
"0.4841",
"0.4631",
"0.4841",
"0.4631",
"0.4976",
"0.10297",
"0.4976"
],
[
"0.5195",
"0.486",
"0.914",
"0.486",
"0.914",
"0.5",
"0.5195",
"0.5"
],
[
"0.02287",
"0.4956",
"0.06046",
"0.4956",
"0.06046",
"0.5947",
"0.02287",
"0.5947"
],
[
"0.1054",
"0.4976",
"0.46",
"0.4976",
"0.46",
"0.5083",
"0.1054",
"0.5083"
],
[
"0.5215",
"0.5005",
"0.9126",
"0.5005",
"0.9126",
"0.5146",
"0.5215",
"0.5146"
],
[
"0.10297",
"0.509",
"0.4617",
"0.509",
"0.4617",
"0.523",
"0.10297",
"0.523"
],
[
"0.5195",
"0.5146",
"0.826",
"0.5146",
"0.826",
"0.528",
"0.5195",
"0.528"
],
[
"0.1013",
"0.5215",
"0.4631",
"0.5215",
"0.4631",
"0.535",
"0.1013",
"0.535"
],
[
"0.10297",
"0.534",
"0.4631",
"0.534",
"0.4631",
"0.5474",
"0.10297",
"0.5474"
],
[
"0.536",
"0.529",
"0.91",
"0.529",
"0.91",
"0.5425",
"0.536",
"0.5425"
],
[
"0.1013",
"0.5454",
"0.46",
"0.5454",
"0.46",
"0.5596",
"0.1013",
"0.5596"
],
[
"0.5195",
"0.5425",
"0.911",
"0.5425",
"0.911",
"0.556",
"0.5195",
"0.556"
],
[
"0.1013",
"0.56",
"0.4631",
"0.56",
"0.4631",
"0.5737",
"0.1013",
"0.5737"
],
[
"0.5195",
"0.556",
"0.9126",
"0.556",
"0.9126",
"0.5693",
"0.5195",
"0.5693"
],
[
"0.10455",
"0.5737",
"0.46",
"0.5737",
"0.46",
"0.584",
"0.10455",
"0.584"
],
[
"0.5215",
"0.571",
"0.9126",
"0.571",
"0.9126",
"0.584",
"0.5215",
"0.584"
],
[
"0.0196",
"0.577",
"0.058",
"0.577",
"0.06046",
"0.712",
"0.02287",
"0.712"
],
[
"0.10297",
"0.584",
"0.4617",
"0.584",
"0.4617",
"0.5977",
"0.10297",
"0.5977"
],
[
"0.5215",
"0.584",
"0.9126",
"0.584",
"0.9126",
"0.5977",
"0.5215",
"0.5977"
],
[
"0.10297",
"0.5967",
"0.4631",
"0.5967",
"0.4631",
"0.61",
"0.10297",
"0.61"
],
[
"0.5215",
"0.5986",
"0.9126",
"0.5986",
"0.9126",
"0.6123",
"0.5215",
"0.6123"
],
[
"0.1013",
"0.608",
"0.415",
"0.608",
"0.415",
"0.622",
"0.1013",
"0.622"
],
[
"0.5195",
"0.6123",
"0.914",
"0.6123",
"0.914",
"0.6255",
"0.5195",
"0.6255"
],
[
"0.5195",
"0.6255",
"0.914",
"0.6255",
"0.914",
"0.6396",
"0.5195",
"0.6396"
],
[
"0.2402",
"0.643",
"0.3513",
"0.643",
"0.3513",
"0.6577",
"0.2402",
"0.6577"
],
[
"0.5195",
"0.64",
"0.914",
"0.64",
"0.914",
"0.654",
"0.5195",
"0.654"
],
[
"0.5195",
"0.654",
"0.914",
"0.654",
"0.914",
"0.6675",
"0.5195",
"0.6675"
],
[
"0.2173",
"0.646",
"0.2255",
"0.646",
"0.2255",
"0.655",
"0.2173",
"0.655"
],
[
"0.0866",
"0.662",
"0.4795",
"0.662",
"0.4795",
"0.6753",
"0.0866",
"0.6753"
],
[
"0.5195",
"0.6675",
"0.914",
"0.6675",
"0.914",
"0.681",
"0.5195",
"0.681"
],
[
"0.0866",
"0.677",
"0.478",
"0.677",
"0.478",
"0.69",
"0.0866",
"0.69"
],
[
"0.5215",
"0.683",
"0.9126",
"0.683",
"0.9126",
"0.6934",
"0.5215",
"0.6934"
],
[
"0.0866",
"0.69",
"0.4795",
"0.69",
"0.4795",
"0.704",
"0.0866",
"0.704"
],
[
"0.523",
"0.6963",
"0.789",
"0.6963",
"0.789",
"0.707",
"0.523",
"0.707"
],
[
"0.0866",
"0.705",
"0.478",
"0.704",
"0.478",
"0.7153",
"0.0866",
"0.7163"
],
[
"0.535",
"0.7104",
"0.911",
"0.7104",
"0.911",
"0.7236",
"0.535",
"0.7236"
],
[
"0.0866",
"0.7183",
"0.4763",
"0.7183",
"0.4763",
"0.732",
"0.0866",
"0.732"
],
[
"0.5195",
"0.7236",
"0.9126",
"0.7236",
"0.9126",
"0.7373",
"0.5195",
"0.7373"
],
[
"0.0866",
"0.7344",
"0.4763",
"0.7344",
"0.4763",
"0.744",
"0.0866",
"0.744"
],
[
"0.5186",
"0.7373",
"0.9126",
"0.738",
"0.9126",
"0.752",
"0.5186",
"0.7505"
],
[
"0.08417",
"0.744",
"0.478",
"0.7456",
"0.478",
"0.7603",
"0.08417",
"0.759"
],
[
"0.5195",
"0.752",
"0.911",
"0.752",
"0.911",
"0.765",
"0.5195",
"0.765"
],
[
"0.08496",
"0.7603",
"0.3652",
"0.7603",
"0.3652",
"0.7734",
"0.08496",
"0.7734"
],
[
"0.5195",
"0.765",
"0.9126",
"0.765",
"0.9126",
"0.779",
"0.5195",
"0.779"
],
[
"0.10297",
"0.773",
"0.4739",
"0.773",
"0.4739",
"0.786",
"0.10297",
"0.786"
],
[
"0.5215",
"0.781",
"0.9126",
"0.781",
"0.9126",
"0.7915",
"0.5215",
"0.7915"
],
[
"0.08496",
"0.787",
"0.478",
"0.787",
"0.478",
"0.8003",
"0.08496",
"0.8003"
],
[
"0.5195",
"0.7925",
"0.9126",
"0.7925",
"0.9126",
"0.806",
"0.5195",
"0.806"
],
[
"0.08496",
"0.802",
"0.478",
"0.802",
"0.478",
"0.815",
"0.08496",
"0.815"
],
[
"0.5195",
"0.8066",
"0.9126",
"0.8066",
"0.9126",
"0.821",
"0.5195",
"0.821"
],
[
"0.08496",
"0.815",
"0.478",
"0.815",
"0.478",
"0.829",
"0.08496",
"0.829"
],
[
"0.5195",
"0.821",
"0.911",
"0.821",
"0.911",
"0.834",
"0.5195",
"0.834"
],
[
"0.08496",
"0.83",
"0.4795",
"0.83",
"0.4795",
"0.8433",
"0.08496",
"0.8433"
],
[
"0.5195",
"0.834",
"0.9126",
"0.834",
"0.9126",
"0.8477",
"0.5195",
"0.8477"
],
[
"0.10864",
"0.8516",
"0.3162",
"0.8516",
"0.3162",
"0.8657",
"0.10864",
"0.8657"
],
[
"0.5215",
"0.8496",
"0.911",
"0.8496",
"0.911",
"0.86",
"0.5215",
"0.86"
],
[
"0.08496",
"0.863",
"0.4795",
"0.863",
"0.4795",
"0.877",
"0.08496",
"0.877"
],
[
"0.5195",
"0.8623",
"0.914",
"0.8623",
"0.914",
"0.8755",
"0.5195",
"0.8755"
],
[
"0.08496",
"0.877",
"0.3718",
"0.877",
"0.3718",
"0.89",
"0.08496",
"0.89"
],
[
"0.5186",
"0.8755",
"0.9126",
"0.8755",
"0.9126",
"0.8896",
"0.5186",
"0.8896"
]
] | [
"Preventing Eviction-Caused Homelessness through ML-Informed Distribution of",
"Rental Assistance",
"Catalina Vajiac* 1, Arun Frey*2, Joachim Baumann*3.4, Abigail Smith*5,",
"Kasun Amarasinghe', Alice Lai', Kit Rodolfa?, Rayid Ghani'",
"ICarnegie Mellon University,",
"2Stanford University,",
"3University of Zurich,",
"",
"4Zurich University of Applied Sciences,",
"5NORC at the University of Chicago",
"cvajiac @cs.cmu.edu, arunfrey @law.stanford.edu, baumann@ifi.uzh.ch, alsl @u.northwestern.edu, [kamarasi, alicelai,",
"ghani) @andrew.cmu.edu,krodolfa@law.stanford.edu",
"Abstract",
"to keep them stably housed. Experimental evidence sug-",
"61",
"gests that such assistance is effective at reducing subse-",
"Rental assistance programs provide individuals with finan-",
"quent homelessness: using quasi-random variation in fund-",
"cial assistance to prevent housing instabilities caused by evic-",
"ing availability of a rental assistance program in Chicago,",
"['s]",
"tions and averthomelessness.Since these programs operate",
"under resource constraints, they must decide who to priori-",
"it was shown that individuals who called when funding was",
"",
"tize. Typically, funding is distributed by a reactive or first-",
"available were 76% less likely to become homeless in the",
"come-first serve allocation process that does not systemati-",
"subsequent 6 months compared to those who called when",
"callyconsiderrisk of futurehomelessness.Wepartnered with",
"funding was unavailable (Evans, Sullivan, and Wallskog",
"Allegheny County, PA to explore a proactive allocation ap-",
"2016). A similar program in NYC randomized the provision",
"proach that prioritizes individuals facing eviction based on",
"of financial assistance to households and reduced average",
"their risk of future homelessness. Our ML system that uses",
"days in shelter from 32 nights among the control group to",
"466521",
"stateand county administrativedatato accuratelyidentifyin-",
"10 nights among the treatment group in the 2 years follow-",
"dividuals in need of support outperforms simpler prioritiza-",
"ing funding (Rolston, Geyer, and Locke 2013)",
"tion approaches by at least 20% while being fair and equitable",
"across race and gender. Furthermore, our approach would",
"For rental assistance programs to be a viable homeless-",
"identify 28% of individuals who are overlooked by the cur-",
"ness prevention strategy, they need to be effective (i.e. pre-",
"rent process and end up homeless. Beyond improvements to",
"vent individuals from falling into homelessness), efficient",
"therental assistanceprogram in AlleghenyCounty,this study",
"(i.e. target individuals at risk of falling into homelessness)",
"C",
"can inform the development of evidence-based decision sup-",
"(Burt 2007), and equitable. Currently, however, these pro-",
"port tools in similar contexts, including lessons about data",
"grams prioritize individuals based on simple heuristics (e.g.",
"needs, model design, evaluation, and field validation.",
"first come,first served)instead of their likelihood of future",
"homelessness. In addition, they place the burden of applying",
"Introduction",
"for funding on the individuals facing eviction and overlook",
"those who need help but do not apply. The administrative",
"1",
"Homelessness remains a pervasive and pressing issue across",
"process can create long delays, so that even eligible tenants",
"the United States. In January 2022, more than 500k individ-",
"oftenremainonawaitlistformultiplemonthsandcanend",
"uals experienced homelessness on a single night (de Sousa",
"upevictedbeforetheyreceiveassistance.",
"etal.2022).Rising evictionratesand thelackof affordable",
"In this paper, we describe our collaboration with the Al-",
"housing contribute to this problem: as the gap between hous-",
"legheny County Department of Human Services (ACDHS)",
"ingcostsandincomelevelscontinuestogrow,anincreasing",
"to improve the allocation ofrental assistancebyprioritizing",
"number of households struggle to pay rent, eventually facing",
"individuals with the highest likelihood of falling into home-",
"eviction and, in some cases, homelessness.",
"lessness. We combine rich county and state administrative",
"To curb rates of homelessness, policymakers are increas",
"data with courtrecords topredicthomelessness among all",
"ingly seeking to reduce the rate of entry into homelessness",
"individuals currently facing eviction, regardless of whether",
"(Culhane, Metraux, and Byrne 2011). One popular preven-",
"they contacted the county for financial support. Using data",
"tion strategy is rental assistance programs, which provide",
"from January 2012 through August 2023, we develop a se-",
"temporary financial assistance to individuals facing eviction",
"ries of machine learning (ML) models that predict a tenant's",
"*These authors contributed equally.",
"needforhomelessness services inthenext 12months,al-",
"Copyright 2024, Association for the Advancement of Artificial",
"lowing ACDHS to proactively prioritize rental assistance to",
"Intelligence (www.aaai.org).All rights reserved.",
"the most vulnerable tenants. In particular, we make the fol-"
] | {
"arxiv_id": "2403.12599",
"language": "en",
"page_idx": 0
} |
||
2403.12599-en-1 | 2403.12599-en-1.jpg | lowing contributions:
$\cdot$ Enabling a proactive approach: We consider a signifi-
cantly higher proportion of residents, namely all tenants in the county facing eviction, instead of only those who call the county for help. Our models identify 28% of people who are overlooked by the current process and end up homeless. By shifting the burden away from those impacted, this proactive approach also reduces administrative effort and is more likely to provide preventative rental assistance to tenants before their eviction.
$\cdot$ Implementing need-based prioritization: Our models identify individuals in need of future homelessness support services with at least 20% improvement over simpler baselines, and are 10x better than random selection while also being fair and equitable.
$\cdot$ Field validation: We conduct a shadow mode deployment to validate our models on new data, mitigating the risk of leakage in the future. Additionally, we are planning a randomized control trial to compare our proposed solution to the current process and to evaluate the effectiveness of rental assistance in preventing entry into homelessness among targeted individuals.
$\cdot$ Lessons learned: We reflect on pitfalls and successes that may inform AI researchers seeking to ethically design predictive decision support tools in other contexts.
Reproducibility. Our code is available at https://github
com/dssg/acdhs housing -public.
## 2 $\textbf{Ethical Considerations}$
Since we are informing the allocation of scarce and critical resources to vulnerable people who could become homeless, we bring into focus the ethical considerations that informed and were embedded in every phase of the scoping, design, and development of our approach.
Equitable outcomes through the allocation of resources $\bar{\text{The use of AI in real-world contexts can perpetuate systemic}}$ biases such as racial disparities (Chen, Joshi, and Ghassemi 2020). To mitigate this risk, we carefully designed the scope of our work and our formulation and analyzed our model results to guard against bias against certain demographic groups, i.e. race and gender (see Section 5.3). Our field trial will also test the equity in the impact of the downstream decisions made using our system.
Transparency and interpretability. Our goal is to develop a system that supports and informs social workers in making better decisions that lead to improved, more equitable outcomes. To this end, we train models that are more explicit about which features they learn from (see Section 4.3) and analyze which features end up most predictive of future homelessness (see Section 5.4). Our solution is designed to augment the current process, supporting rather than excluding social workers from the decisionmaking loop and providing them with additional information beyond predictions to help them make better-informed decisions. We also consulted impacted community members in the design and validation process.
Privacy concerns of using sensitive data. The collection of personal information always comes with privacy concerns. In our case, particularly because ACDHS connects administrative data across different facets of residents' lives, we want to be sure that the use of this data offers significant enough improvements to the community, particularly people potentially facing homelessness. In Sections 5 and 6, we extensively evaluate our models with historical data, and a field trial to ensure that this work improves outcomes. We are currently soliciting feedback from groups potentially impacted by the system through a community engagement process to better assess tradeoffs before conducting a validation trial.
# 3 Current Approaches
We first consider how ACDHS currently prioritizes residents
for rental assistance and highlight relevant prior research.
# 3.1 ACDHS’ Current Process
The current process for obtaining rental assistance in Allegheny County is illustrative of how such programs are usually managed throughout the United States. Tenants with an eviction notice can contact the Allegheny County Link helpline to request rental assistance, or they may be referred to the rental assistance program by a mediation program or a housing assistance organization. Applicants are prescreened according to available funding and county-level eligibility requirements, including income and the amount of rent owed. Tenants are then placed on a waitlist and will be considered on a first-come-first-served basis. When the tenant reaches the front of the list, a social worker examines their funding request. If they are deemed eligible and can provide the required documentation, they will receive a payment covering the rent owed.
There are many issues with this reactive process, and the county is seeking our help in improving it. First, it puts heavy logistical strain on tenants facing eviction, who have to know to contact the helpline or other housing stability organizations to apply for assistance, and must be able to prove their eligibility with documentation. This requirement excludes all individuals at risk of homelessness due to eviction who do not proactively seek help. For those who apply, the substantial delay between application and receipt of funding means that many tenants have already been evicted by the time they are considered for assistance, or their rental debt has substantially increased. Since eligibility requirements change over time and are not always easily accessible, the current decision system is opaque to those seeking assistance. Finally, rental assistance is not distributed according to who has the most need, but instead to those who know to call the helpline, get through the waitlist in time, and meet the eligibility requirements before being evicted.
# 3.2 Related Work
Previous work has aimed to predict future homelessness to identify those most in need and more efficiently allocate homelessness prevention resources (Fowler et al. 2019). In New York City, re-evaluation of an existing homelessness | [
[
"0.0866",
"0.06946",
"0.2279",
"0.06946",
"0.2279",
"0.0833",
"0.0866",
"0.0833"
],
[
"0.5215",
"0.06946",
"0.9126",
"0.06946",
"0.9126",
"0.0833",
"0.5215",
"0.0833"
],
[
"0.09314",
"0.089",
"0.4756",
"0.089",
"0.4756",
"0.1042",
"0.09314",
"0.1042"
],
[
"0.5195",
"0.0833",
"0.911",
"0.0833",
"0.911",
"0.0966",
"0.5195",
"0.0966"
],
[
"0.5195",
"0.09784",
"0.9126",
"0.09784",
"0.9126",
"0.1111",
"0.5195",
"0.1111"
],
[
"0.10706",
"0.1042",
"0.478",
"0.1042",
"0.478",
"0.11743",
"0.10706",
"0.11743"
],
[
"0.10706",
"0.1187",
"0.478",
"0.1187",
"0.478",
"0.132",
"0.10706",
"0.132"
],
[
"0.5215",
"0.113",
"0.911",
"0.113",
"0.911",
"0.1237",
"0.5215",
"0.1237"
],
[
"0.5215",
"0.1256",
"0.9126",
"0.1256",
"0.9126",
"0.1395",
"0.5215",
"0.1395"
],
[
"0.10706",
"0.132",
"0.4756",
"0.132",
"0.4756",
"0.1459",
"0.10706",
"0.1459"
],
[
"0.5215",
"0.1414",
"0.911",
"0.1414",
"0.911",
"0.1515",
"0.5215",
"0.1515"
],
[
"0.10864",
"0.1477",
"0.4763",
"0.1477",
"0.4763",
"0.1578",
"0.10864",
"0.1578"
],
[
"0.10706",
"0.159",
"0.4795",
"0.159",
"0.4795",
"0.1736",
"0.10706",
"0.1736"
],
[
"0.5215",
"0.154",
"0.9067",
"0.1528",
"0.9067",
"0.1642",
"0.5215",
"0.1654"
],
[
"0.5186",
"0.1666",
"0.9126",
"0.1654",
"0.9126",
"0.1787",
"0.5186",
"0.1799"
],
[
"0.10706",
"0.1736",
"0.4763",
"0.1736",
"0.4763",
"0.1875",
"0.10706",
"0.1875"
],
[
"0.5195",
"0.1799",
"0.91",
"0.1799",
"0.91",
"0.1932",
"0.5195",
"0.1932"
],
[
"0.10864",
"0.1881",
"0.478",
"0.1881",
"0.478",
"0.1989",
"0.10864",
"0.1989"
],
[
"0.5195",
"0.1945",
"0.9126",
"0.1945",
"0.9126",
"0.2084",
"0.5195",
"0.2084"
],
[
"0.10706",
"0.2008",
"0.4224",
"0.2008",
"0.4224",
"0.214",
"0.10706",
"0.214"
],
[
"0.5215",
"0.209",
"0.914",
"0.209",
"0.914",
"0.2229",
"0.5215",
"0.2229"
],
[
"0.0948",
"0.2172",
"0.4763",
"0.2172",
"0.4763",
"0.2311",
"0.0948",
"0.2311"
],
[
"0.5195",
"0.2229",
"0.902",
"0.2229",
"0.902",
"0.2361",
"0.5195",
"0.2361"
],
[
"0.1054",
"0.2311",
"0.4756",
"0.2317",
"0.4756",
"0.2456",
"0.1054",
"0.2443"
],
[
"0.1095",
"0.2468",
"0.478",
"0.2468",
"0.478",
"0.2568",
"0.1095",
"0.2568"
],
[
"0.6094",
"0.25",
"0.819",
"0.2507",
"0.819",
"0.2683",
"0.6094",
"0.2676"
],
[
"0.10706",
"0.2588",
"0.4795",
"0.2588",
"0.4795",
"0.2727",
"0.10706",
"0.2727"
],
[
"0.1054",
"0.2727",
"0.2957",
"0.2727",
"0.2957",
"0.286",
"0.1054",
"0.286"
],
[
"0.523",
"0.2747",
"0.9126",
"0.2747",
"0.9126",
"0.2886",
"0.523",
"0.2886"
],
[
"0.09314",
"0.2903",
"0.4756",
"0.2903",
"0.4756",
"0.3037",
"0.09314",
"0.3037"
],
[
"0.5195",
"0.2886",
"0.8945",
"0.2886",
"0.8945",
"0.3018",
"0.5195",
"0.3018"
],
[
"0.10706",
"0.3062",
"0.4763",
"0.3062",
"0.4763",
"0.3164",
"0.10706",
"0.3164"
],
[
"0.10706",
"0.3176",
"0.4756",
"0.3176",
"0.4756",
"0.3308",
"0.10706",
"0.3308"
],
[
"0.5195",
"0.3176",
"0.7573",
"0.3176",
"0.7573",
"0.332",
"0.5195",
"0.332"
],
[
"0.10706",
"0.332",
"0.4795",
"0.332",
"0.4795",
"0.3452",
"0.10706",
"0.3452"
],
[
"0.523",
"0.339",
"0.91",
"0.339",
"0.91",
"0.3528",
"0.523",
"0.3528"
],
[
"0.10706",
"0.348",
"0.4763",
"0.348",
"0.4763",
"0.358",
"0.10706",
"0.358"
],
[
"0.10706",
"0.3599",
"0.4795",
"0.3599",
"0.4795",
"0.3738",
"0.10706",
"0.3738"
],
[
"0.523",
"0.3547",
"0.9087",
"0.3547",
"0.9087",
"0.3655",
"0.523",
"0.3655"
],
[
"0.5215",
"0.3674",
"0.9126",
"0.3674",
"0.9126",
"0.3806",
"0.5215",
"0.3806"
],
[
"0.1054",
"0.3738",
"0.3816",
"0.3738",
"0.3816",
"0.387",
"0.1054",
"0.387"
],
[
"0.523",
"0.3833",
"0.911",
"0.3833",
"0.911",
"0.3933",
"0.523",
"0.3933"
],
[
"0.09314",
"0.3901",
"0.478",
"0.3901",
"0.478",
"0.404",
"0.09314",
"0.404"
],
[
"0.5215",
"0.3958",
"0.9126",
"0.3958",
"0.9126",
"0.4092",
"0.5215",
"0.4092"
],
[
"0.10864",
"0.406",
"0.4756",
"0.406",
"0.4756",
"0.4167",
"0.10864",
"0.4167"
],
[
"0.5195",
"0.4092",
"0.9126",
"0.4092",
"0.9126",
"0.4224",
"0.5195",
"0.4224"
],
[
"0.10706",
"0.4185",
"0.468",
"0.4185",
"0.468",
"0.432",
"0.10706",
"0.432"
],
[
"0.5215",
"0.4248",
"0.9087",
"0.4248",
"0.9087",
"0.435",
"0.5215",
"0.435"
],
[
"0.0866",
"0.4395",
"0.4756",
"0.4395",
"0.4756",
"0.4539",
"0.0866",
"0.4539"
],
[
"0.523",
"0.4382",
"0.9087",
"0.4382",
"0.9087",
"0.449",
"0.523",
"0.449"
],
[
"0.08496",
"0.4539",
"0.3015",
"0.4539",
"0.3015",
"0.4673",
"0.08496",
"0.4673"
],
[
"0.523",
"0.452",
"0.911",
"0.452",
"0.911",
"0.4622",
"0.523",
"0.4622"
],
[
"0.5195",
"0.464",
"0.9126",
"0.464",
"0.9126",
"0.4778",
"0.5195",
"0.4778"
],
[
"0.1691",
"0.481",
"0.3938",
"0.481",
"0.3938",
"0.4956",
"0.1691",
"0.4956"
],
[
"0.5215",
"0.4792",
"0.9087",
"0.4792",
"0.9087",
"0.4893",
"0.5215",
"0.4893"
],
[
"0.5215",
"0.4937",
"0.911",
"0.4937",
"0.911",
"0.504",
"0.5215",
"0.504"
],
[
"0.0866",
"0.5",
"0.478",
"0.5",
"0.478",
"0.513",
"0.0866",
"0.513"
],
[
"0.5195",
"0.506",
"0.9126",
"0.506",
"0.9126",
"0.5195",
"0.5195",
"0.5195"
],
[
"0.08496",
"0.5146",
"0.4739",
"0.5146",
"0.4739",
"0.528",
"0.08496",
"0.528"
],
[
"0.5215",
"0.5195",
"0.911",
"0.5195",
"0.911",
"0.5327",
"0.5215",
"0.5327"
],
[
"0.0866",
"0.53",
"0.4763",
"0.53",
"0.4763",
"0.5405",
"0.0866",
"0.5405"
],
[
"0.5215",
"0.534",
"0.7114",
"0.534",
"0.7114",
"0.5474",
"0.5215",
"0.5474"
],
[
"0.08496",
"0.5425",
"0.478",
"0.5425",
"0.478",
"0.556",
"0.08496",
"0.556"
],
[
"0.536",
"0.5474",
"0.9126",
"0.547",
"0.9126",
"0.561",
"0.536",
"0.5625"
],
[
"0.0866",
"0.5547",
"0.3137",
"0.5547",
"0.3137",
"0.5684",
"0.0866",
"0.5684"
],
[
"0.523",
"0.563",
"0.911",
"0.563",
"0.911",
"0.5737",
"0.523",
"0.5737"
],
[
"0.0866",
"0.577",
"0.4763",
"0.577",
"0.4763",
"0.5903",
"0.0866",
"0.5903"
],
[
"0.5195",
"0.5757",
"0.9126",
"0.5757",
"0.9126",
"0.589",
"0.5195",
"0.589"
],
[
"0.08826",
"0.5923",
"0.4763",
"0.5923",
"0.4763",
"0.603",
"0.08826",
"0.603"
],
[
"0.5215",
"0.5913",
"0.91",
"0.5913",
"0.91",
"0.6016",
"0.5215",
"0.6016"
],
[
"0.08826",
"0.606",
"0.478",
"0.606",
"0.478",
"0.616",
"0.08826",
"0.616"
],
[
"0.5215",
"0.604",
"0.9126",
"0.604",
"0.9126",
"0.617",
"0.5215",
"0.617"
],
[
"0.0866",
"0.617",
"0.4763",
"0.617",
"0.4763",
"0.631",
"0.0866",
"0.631"
],
[
"0.5215",
"0.617",
"0.911",
"0.617",
"0.911",
"0.631",
"0.5215",
"0.631"
],
[
"0.0866",
"0.6333",
"0.478",
"0.6333",
"0.478",
"0.6465",
"0.0866",
"0.6465"
],
[
"0.5215",
"0.6333",
"0.911",
"0.6333",
"0.911",
"0.643",
"0.5215",
"0.643"
],
[
"0.08496",
"0.6465",
"0.478",
"0.6465",
"0.478",
"0.66",
"0.08496",
"0.66"
],
[
"0.5215",
"0.646",
"0.9126",
"0.646",
"0.9126",
"0.659",
"0.5215",
"0.659"
],
[
"0.0866",
"0.66",
"0.478",
"0.66",
"0.478",
"0.674",
"0.0866",
"0.674"
],
[
"0.5215",
"0.659",
"0.911",
"0.659",
"0.911",
"0.6724",
"0.5215",
"0.6724"
],
[
"0.0866",
"0.675",
"0.4756",
"0.675",
"0.4756",
"0.688",
"0.0866",
"0.688"
],
[
"0.5215",
"0.675",
"0.91",
"0.675",
"0.91",
"0.685",
"0.5215",
"0.685"
],
[
"0.08496",
"0.688",
"0.2932",
"0.688",
"0.2932",
"0.702",
"0.08496",
"0.702"
],
[
"0.523",
"0.688",
"0.911",
"0.688",
"0.911",
"0.6987",
"0.523",
"0.6987"
],
[
"0.5195",
"0.7007",
"0.911",
"0.7007",
"0.911",
"0.714",
"0.5195",
"0.714"
],
[
"0.0866",
"0.709",
"0.4756",
"0.709",
"0.4756",
"0.7227",
"0.0866",
"0.7227"
],
[
"0.5215",
"0.7163",
"0.91",
"0.7163",
"0.91",
"0.7266",
"0.5215",
"0.7266"
],
[
"0.0866",
"0.7236",
"0.478",
"0.7236",
"0.478",
"0.7373",
"0.0866",
"0.7373"
],
[
"0.5195",
"0.728",
"0.9087",
"0.729",
"0.9087",
"0.7407",
"0.5195",
"0.7393"
],
[
"0.08496",
"0.7373",
"0.4763",
"0.7363",
"0.4763",
"0.75",
"0.08496",
"0.7505"
],
[
"0.5215",
"0.7437",
"0.911",
"0.7437",
"0.911",
"0.754",
"0.5215",
"0.754"
],
[
"0.0866",
"0.7505",
"0.4795",
"0.7505",
"0.4795",
"0.765",
"0.0866",
"0.765"
],
[
"0.5195",
"0.756",
"0.914",
"0.756",
"0.914",
"0.7695",
"0.5195",
"0.7695"
],
[
"0.0866",
"0.765",
"0.478",
"0.765",
"0.478",
"0.779",
"0.0866",
"0.779"
],
[
"0.5195",
"0.7695",
"0.9126",
"0.7695",
"0.9126",
"0.7827",
"0.5195",
"0.7827"
],
[
"0.08905",
"0.78",
"0.4756",
"0.78",
"0.4756",
"0.7905",
"0.08905",
"0.7905"
],
[
"0.5186",
"0.784",
"0.841",
"0.784",
"0.841",
"0.7974",
"0.5186",
"0.7974"
],
[
"0.0866",
"0.7925",
"0.4763",
"0.7925",
"0.4763",
"0.806",
"0.0866",
"0.806"
],
[
"0.08496",
"0.8047",
"0.478",
"0.806",
"0.478",
"0.8193",
"0.08496",
"0.8184"
],
[
"0.5195",
"0.811",
"0.669",
"0.811",
"0.669",
"0.8257",
"0.5195",
"0.8257"
],
[
"0.08905",
"0.8213",
"0.4763",
"0.8213",
"0.4763",
"0.832",
"0.08905",
"0.832"
],
[
"0.0866",
"0.834",
"0.478",
"0.834",
"0.478",
"0.8477",
"0.0866",
"0.8477"
],
[
"0.523",
"0.8354",
"0.911",
"0.8354",
"0.911",
"0.845",
"0.523",
"0.845"
],
[
"0.08496",
"0.8477",
"0.478",
"0.8477",
"0.478",
"0.8613",
"0.08496",
"0.8613"
],
[
"0.5215",
"0.8477",
"0.9126",
"0.8477",
"0.9126",
"0.8613",
"0.5215",
"0.8613"
],
[
"0.08496",
"0.8623",
"0.4795",
"0.8623",
"0.4795",
"0.8755",
"0.08496",
"0.8755"
],
[
"0.523",
"0.863",
"0.9126",
"0.863",
"0.9126",
"0.8735",
"0.523",
"0.8735"
],
[
"0.08496",
"0.8755",
"0.3096",
"0.8755",
"0.3096",
"0.8896",
"0.08496",
"0.8896"
],
[
"0.5195",
"0.8755",
"0.914",
"0.8755",
"0.914",
"0.8896",
"0.5195",
"0.8896"
]
] | [
"lowing contributions:",
"Privacy concerns of using sensitive data.The collection",
". Enabling a proactive approach: We consider a signifi-",
"of personal information always comes with privacy con-",
"cerns. In our case, particularly because ACDHS connects",
"cantly higher proportion of residents, namely all tenants",
"in the county facing eviction, instead of only those who",
"administrativedataacross differentfacetsof residents'lives.",
"we want to be sure that the use of this data offers significant",
"call the county for help. Our models identify 28% of peo-",
"enoughimprovementstothecommunity,particularlypeople",
"ple who are overlooked bythe currentprocess and end",
"up homeless. By shifting the burden away from those",
"potentiallyfacinghomelessness.InSections5and6,weex",
"tensively evaluate our models with historical data, and a field",
"impacted, this proactive approach also reduces adminis-",
"trial to ensure that this work improves outcomes. We are cur-",
"trativeeffort and ismorelikelytoprovidepreventative",
"rently soliciting feedback from groups potentially impacted",
"rental assistance to tenants before their eviction.",
"by the system through a community engagement process to",
".Implementing need-based prioritization: Our models",
"better assess tradeoffs before conducting a validation trial.",
"identify individuals in need of future homelessness sup-",
"portserviceswithatleast20%improvementoversimpler",
"3Current Approaches",
"baselines, and are 10x better than random selection while",
"also being fair and equitable.",
"We first consider how ACDHS currently prioritizes residents",
": Field validation: We conduct a shadow mode deploy-",
"for rental assistance and highlight relevant prior research.",
"mentto validateourmodels onnewdata,mitigatingthe",
"risk of leakage in the future. Additionally,we are plan-",
"3.1ACDHS' Current Process",
"ning a randomized control trial to compare our proposed",
"The current process for obtaining rental assistance in Al-",
"solution to the current process and to evaluate the ef-",
"fectiveness of rental assistance in preventing entry into",
"leghenyCountyisillustrativeofhowsuchprogramsareusu",
"ally managed throughout the United States. Tenants with",
"homelessness among targeted individuals",
"aneviction notice cancontacttheAlleghenyCountyLink",
". Lessons learned: We reflect on pitfalls and successes",
"helpline to request rental assistance, or they may be referred",
"thatmay informAl researchers seekingtoethically de-",
"to the rental assistance program by a mediation program",
"sign predictive decision support tools in other contexts.",
"or a housing assistance organization. Applicants are pre.",
"Reproducibility. Our code is available at https://github",
"screenedaccordingtoavailablefundingandcounty-levelel",
"com/dssg/acdhs_housing-public.",
"igibilityrequirements,includingincomeandtheamountof",
"rent owed. Tenants are then placed on a waitlist and will be",
"2Ethical Considerations",
"considered on a first-come-first-served basis.When the ten-",
"ant reaches the front of the list,a social worker examines",
"Since we are informing the allocation of scarce and critical",
"their funding request. If they are deemed eligible and can",
"resources to vulnerable people who could become homeless",
"provide the required documentation, they will receive a pay-",
"webring into focus the ethical considerations that informed",
"ment covering the rent owed.",
"and were embedded in every phase of the scoping, design,",
"There are many issues with this reactive process, and the",
"and development of our approach.",
"county is seeking ourhelp inimproving it.First,itputs",
"Equitable outcomes through the allocation of resources",
"heavy logistical strain on tenants facing eviction, who have",
"TheuseofAIinreal-world contextscanperpetuatesystemic",
"toknowtocontactthehelplineorotherhousingstabilityor-",
"biases such as racial disparities (Chen, Joshi,and Ghassemi",
"ganizations to apply for assistance, and must be able to prove",
"2020). To mitigate this risk, we carefully designed the scope",
"their eligibility with documentation. This requirement ex-",
"of our work and our formulation and analyzed our model",
"cludesallindividualsatriskofhomelessnessduetoeviction",
"results to guard against bias against certain demographic",
"who do not proactively seek help. For those who apply, the",
"groups, i.e. race and gender (see Section 5.3). Our field trial",
"substantial delay between application and receipt of fund-",
"will also test the equity in the impact of the downstream de-",
"ingmeansthat manytenantshave alreadybeen evicted by",
"cisions made using our system.",
"thetime they are consideredfor assistanceor theirrental",
"debt has substantially increased. Since eligibility require-",
"Transparency and interpretability.Our goal is to de",
"mentschangeovertimeand arenotalways easilyaccessible.",
"velop a system that supports and informs social workers",
"thecurrent decision systemisopaquetothose seeking assis",
"in making better decisions that lead to improved, more eq",
"tance.Finally,rental assistance is not distributed according",
"uitable outcomes. To this end, we train models that are",
"to who has the most need, but instead to those who know to",
"more explicit about which features they learn from (see",
"call the helpline, get through the waitlist in time, and meet",
"Section4.3)andanalyzewhichfeaturesendupmostpre",
"the eligibility requirements before being evicted.",
"dictive of future homelessness (see Section 5.4). Our so-",
"lution is designed to augment the current process, support-",
"3.2 Related Work",
"ingratherthan excludingsocial workersfromthedecision-",
"making loop and providing them with additional information",
"Previousworkhasaimedtopredictfuturehomelessnessto",
"beyond predictions to help them make better-informed deci-",
"identify those most in need and more efficiently allocate",
"sions. We also consulted impacted community members in",
"homelessnesspreventionresources(Fowleretal.2019).In",
"the design and validation process.",
"New York City, re-evaluation of an existing homelessness"
] | {
"arxiv_id": "2403.12599",
"language": "en",
"page_idx": 1
} |
||
2403.12599-en-2 | 2403.12599-en-2.jpg | <table>
<tbody>
<tr>
<td>$\cdot Evictioncohortinclusioncr$</td>
<td>NITE criteria:</td>
</tr>
<tr>
<td>has eviction filing in with</td>
<td>Tilinnir analvsis 1arp</td>
</tr>
<tr>
<td>has lost an eviction case wi</td>
<td>Hatein</td>
</tr>
<tr>
<td>. Homelessness cohort exclu</td>
<td>criteria:</td>
</tr>
<tr>
<td>started or stopped utilizatic</td>
<td>stnr</td>
</tr>
<tr>
<td>currentlv isino homelessne</td>
<td> </td>
</tr>
<tr>
<td>enrolled in rehousing progra $\cdot$Homelessness label: utilizes</td>
<td>renousin Toroqr DVEO IT alle : OT analysi ness label: utilizes homeless services in 2</td>
</tr>
<tr>
<td>date ot analysis</td>
<td>date ot analysis</td>
</tr>
<tr>
<td>Emonths</td>
<td>$12months$</td>
</tr>
</tbody>
</table>
Figure 1: Inclusion/exclusion criteria for homelessness and eviction. Indicators for homelessness in ACDHS data include clients' interactions with homelessness services, such as staying in a shelter. We also include clients enrolled in rehousing programs that have not moved in as of the prediction date, as being homeless is a prerequisite for enrollment.
prevention service resulted in a model based on 15 risk factors that would have outperformed human judgment in selecting at-risk individuals (Shinn et al.2013).In another work focusing on those who were previously homeless, administrative data was used to better predict future homelessness within two years to best match people with programs, though this also disadvantaged other service recipients (Kube, Das, and Fowler 2019). Because of these tradeoffs, full automation of resource allocation is not recommended: human decision-makers should still be in control.
Work by the California Policy Lab predicted the risk of homelessness among all residents in Los Angeles County (Wachter et al. 2019). Several factors were found to be indicative of future homelessness, including repeated and recent interactions with government agencies or county services and prior mental health crises. Their predictive model achieves $29\%$ precision for all residents.
Research Gaps. To our knowledge, no previous work has predicted future homelessness among those who are facing eviction. he California Policy Lab's work likely suffers from label bias due to the low percentage of homeless individuals utilizing county services (see Section 7), while our focus on informing a concrete action, rental assistance, allows our system to fit into the existing support infrastructure.
# 4 Predictive Approach
We develop ML models to facilitate need-based prioritization for proactive distribution of rental assistance. One limitation we encounter with this type of data is that it is conditioned on past rental assistance allocations and offers no insights into the counterfactual outcomes. As a result, our approach resembles an augmentation of the existing process. This is pertinent not only for the technical implementation but also for the interpretation of our results (e.g., in Section 5.2, we concentrate on the outcomes for vulnerable individuals who are not served under the current practice) and the conceptualization of the overall project, which is centered on social impact. and equity. In Section 6, we outline the transition from this stage to a fully deployed system.
## 4.1 Problem Formulation
We formulate our task as a binary classification problem: identifying tenants that will interact with homelessness services within 12 months of the date of prediction. At each prediction date, we consider individuals who have an active eviction filing against them within the past four months and are not currently homeless.$^{\mathrm{1}}$From this cohort, our models identify the 100 individuals with the highest chance of interacting with homelessness services in the next 12 months. We select 100 individuals since this corresponds to the monthly intervention capacity of ACDHS. Specific inclusion and exclusion criteria are detailed in Figure 1.
## 4.2 Data and Feature Engineering
ACDHS collects and combines administrative data from various county and state-level programs, including information
More details on the demographic composition of this group are
provided in Supplementary Materials B.
on individuals who have previous eviction filings, homelessness spells, interactions with mental, behavioral, and physical health institutions, address changes, or who have been enrolled in a variety of other ACDHS and state programs.
For each client at each analysis date, we generate $\sim7000$ features based on the data sources described in Supplementary Materials A. The features can be classified as follows:
Demographics and Event features. For each tenant, we include the total number of interactions with each ACDHS and state program, total number of evictions, and total number of physical and mental health visits. We also include categorical features, such as the particular type of physical or mental health visit and demographic features (race, gender, and age).
Temporal aggregation features. For each data source, we generate several temporal features to capture the dynamic nature of the process and assess how frequently an individual interacts with state and local services within the last 3 & < 6 months, and 1, 2, 3, 4, and 5 years. For example, using the eviction data, we generate the following features per individual: number of days since most recent eviction; number of evictions in the specified time period; sum, min, max, average rent owed in eviction cases; min, max, average interarrival times between evictions.
## 43 Model Training and Validation
We use various supervised classification methods to predict entry into homelessness: logistic regression (LR), decision trees (DT), random forests (RF), Adaboost, Light Gradient Boosting Model (LGBM), and XG boost using ScikitLearn (Pedregosa et al. 2011) and the hyperparameter grid specified in Supplementary Materials C. To most closely re- $\hat{\text{produce the context in which our models will be deployed,}}$ we use temporal validation (Hoptroff 1993), generating different training and evaluation matrices for each analysis date. For example, when evaluating the efficacy of predicting homelessness as of January 2019, we train models on feature label pairs using data up to January 2019, and evaluate them based on how many people become homeless between January 2019 and January 2020 (see Supplementary | [
[
"0.0866",
"0.06946",
"0.4756",
"0.06946",
"0.4756",
"0.0833",
"0.0866",
"0.0833"
],
[
"0.5195",
"0.0707",
"0.7134",
"0.0707",
"0.7134",
"0.07764",
"0.5195",
"0.07764"
],
[
"0.08826",
"0.0852",
"0.4756",
"0.0852",
"0.4756",
"0.09534",
"0.08826",
"0.09534"
],
[
"0.5337",
"0.0802",
"0.899",
"0.0802",
"0.899",
"0.0903",
"0.5337",
"0.0903"
],
[
"0.5337",
"0.0903",
"0.819",
"0.0903",
"0.819",
"0.101",
"0.5337",
"0.101"
],
[
"0.0866",
"0.09784",
"0.4795",
"0.09784",
"0.4795",
"0.1111",
"0.0866",
"0.1111"
],
[
"0.517",
"0.0985",
"0.755",
"0.09973",
"0.755",
"0.113",
"0.517",
"0.11237"
],
[
"0.08826",
"0.113",
"0.4763",
"0.113",
"0.4763",
"0.1237",
"0.08826",
"0.1237"
],
[
"0.5337",
"0.11237",
"0.863",
"0.11237",
"0.863",
"0.1225",
"0.5337",
"0.1225"
],
[
"0.0866",
"0.1256",
"0.478",
"0.1256",
"0.478",
"0.1395",
"0.0866",
"0.1395"
],
[
"0.5337",
"0.1237",
"0.869",
"0.1237",
"0.869",
"0.1338",
"0.5337",
"0.1338"
],
[
"0.0866",
"0.1395",
"0.478",
"0.1395",
"0.478",
"0.1528",
"0.0866",
"0.1528"
],
[
"0.5337",
"0.1338",
"0.9126",
"0.1338",
"0.9126",
"0.1445",
"0.5337",
"0.1445"
],
[
"0.5215",
"0.1445",
"0.8145",
"0.1445",
"0.8145",
"0.1547",
"0.5215",
"0.1547"
],
[
"0.0866",
"0.1528",
"0.4756",
"0.1528",
"0.4756",
"0.1666",
"0.0866",
"0.1666"
],
[
"0.08826",
"0.1666",
"0.4763",
"0.1666",
"0.4763",
"0.1799",
"0.08826",
"0.1799"
],
[
"0.6055",
"0.1666",
"0.6396",
"0.1666",
"0.6396",
"0.1736",
"0.6055",
"0.1736"
],
[
"0.7656",
"0.1622",
"0.7754",
"0.1704",
"0.7686",
"0.1755",
"0.7573",
"0.1686"
],
[
"0.0866",
"0.1812",
"0.478",
"0.1812",
"0.478",
"0.1945",
"0.0866",
"0.1945"
],
[
"0.0866",
"0.1957",
"0.468",
"0.1957",
"0.468",
"0.2058",
"0.0866",
"0.2058"
],
[
"0.5215",
"0.202",
"0.9126",
"0.202",
"0.9126",
"0.2153",
"0.5215",
"0.2153"
],
[
"0.10297",
"0.2084",
"0.478",
"0.2084",
"0.478",
"0.2216",
"0.10297",
"0.2216"
],
[
"0.523",
"0.2172",
"0.9087",
"0.2172",
"0.9087",
"0.2279",
"0.523",
"0.2279"
],
[
"0.08826",
"0.2235",
"0.478",
"0.2235",
"0.478",
"0.2343",
"0.08826",
"0.2343"
],
[
"0.523",
"0.2311",
"0.911",
"0.2311",
"0.911",
"0.2412",
"0.523",
"0.2412"
],
[
"0.08826",
"0.2374",
"0.478",
"0.2374",
"0.478",
"0.2474",
"0.08826",
"0.2474"
],
[
"0.5195",
"0.2437",
"0.9126",
"0.2437",
"0.9126",
"0.2568",
"0.5195",
"0.2568"
],
[
"0.08496",
"0.2507",
"0.4795",
"0.2507",
"0.4795",
"0.2644",
"0.08496",
"0.2644"
],
[
"0.5195",
"0.2583",
"0.91",
"0.2583",
"0.91",
"0.2715",
"0.5195",
"0.2715"
],
[
"0.08496",
"0.2644",
"0.4763",
"0.2644",
"0.4763",
"0.2778",
"0.08496",
"0.2778"
],
[
"0.5195",
"0.2715",
"0.9126",
"0.2715",
"0.9126",
"0.2854",
"0.5195",
"0.2854"
],
[
"0.08496",
"0.2778",
"0.478",
"0.2778",
"0.478",
"0.2917",
"0.08496",
"0.2917"
],
[
"0.08496",
"0.2935",
"0.3506",
"0.2935",
"0.3506",
"0.3037",
"0.08496",
"0.3037"
],
[
"0.0866",
"0.3113",
"0.478",
"0.3113",
"0.478",
"0.3245",
"0.0866",
"0.3245"
],
[
"0.5195",
"0.3132",
"0.9126",
"0.3132",
"0.9126",
"0.3276",
"0.5195",
"0.3276"
],
[
"0.0866",
"0.3257",
"0.478",
"0.3257",
"0.478",
"0.339",
"0.0866",
"0.339"
],
[
"0.5195",
"0.3276",
"0.91",
"0.3276",
"0.91",
"0.3416",
"0.5195",
"0.3416"
],
[
"0.0866",
"0.339",
"0.478",
"0.339",
"0.478",
"0.3528",
"0.0866",
"0.3528"
],
[
"0.5195",
"0.3416",
"0.9126",
"0.3416",
"0.9126",
"0.3547",
"0.5195",
"0.3547"
],
[
"0.08496",
"0.3528",
"0.4763",
"0.3528",
"0.4763",
"0.3662",
"0.08496",
"0.3662"
],
[
"0.5186",
"0.3542",
"0.8945",
"0.3547",
"0.8945",
"0.3694",
"0.5186",
"0.3687"
],
[
"0.0866",
"0.3687",
"0.4763",
"0.3687",
"0.4763",
"0.379",
"0.0866",
"0.379"
],
[
"0.536",
"0.3687",
"0.9126",
"0.3687",
"0.9126",
"0.3818",
"0.536",
"0.3818"
],
[
"0.08496",
"0.3806",
"0.4795",
"0.3806",
"0.4795",
"0.3945",
"0.08496",
"0.3945"
],
[
"0.5195",
"0.3833",
"0.911",
"0.3833",
"0.911",
"0.3965",
"0.5195",
"0.3965"
],
[
"0.08496",
"0.3945",
"0.4265",
"0.3945",
"0.4265",
"0.4077",
"0.08496",
"0.4077"
],
[
"0.5215",
"0.3984",
"0.9043",
"0.3984",
"0.9043",
"0.4092",
"0.5215",
"0.4092"
],
[
"0.1765",
"0.4185",
"0.3857",
"0.4192",
"0.3857",
"0.4375",
"0.1765",
"0.4363"
],
[
"0.5215",
"0.4192",
"0.9126",
"0.4192",
"0.9126",
"0.433",
"0.5215",
"0.433"
],
[
"0.08496",
"0.4375",
"0.478",
"0.4382",
"0.478",
"0.4526",
"0.08496",
"0.452"
],
[
"0.5215",
"0.4343",
"0.911",
"0.4343",
"0.911",
"0.4443",
"0.5215",
"0.4443"
],
[
"0.5195",
"0.4463",
"0.911",
"0.4458",
"0.911",
"0.4602",
"0.5195",
"0.461"
],
[
"0.08826",
"0.455",
"0.4756",
"0.455",
"0.4756",
"0.4653",
"0.08826",
"0.4653"
],
[
"0.08496",
"0.4666",
"0.4756",
"0.4666",
"0.4756",
"0.4797",
"0.08496",
"0.4797"
],
[
"0.5215",
"0.4622",
"0.9087",
"0.4622",
"0.9087",
"0.473",
"0.5215",
"0.473"
],
[
"0.5195",
"0.4749",
"0.9126",
"0.4749",
"0.9126",
"0.488",
"0.5195",
"0.488"
],
[
"0.0866",
"0.481",
"0.478",
"0.481",
"0.478",
"0.4944",
"0.0866",
"0.4944"
],
[
"0.08417",
"0.4937",
"0.4756",
"0.4944",
"0.4756",
"0.5083",
"0.08417",
"0.507"
],
[
"0.5195",
"0.488",
"0.9126",
"0.488",
"0.9126",
"0.5024",
"0.5195",
"0.5024"
],
[
"0.0866",
"0.5083",
"0.4763",
"0.5083",
"0.4763",
"0.5215",
"0.0866",
"0.5215"
],
[
"0.5195",
"0.5024",
"0.5825",
"0.5024",
"0.5825",
"0.5176",
"0.5195",
"0.5176"
],
[
"0.08496",
"0.521",
"0.4795",
"0.5215",
"0.4795",
"0.536",
"0.08496",
"0.535"
],
[
"0.5215",
"0.5244",
"0.9126",
"0.5244",
"0.9126",
"0.5386",
"0.5215",
"0.5386"
],
[
"0.08496",
"0.536",
"0.4756",
"0.536",
"0.4756",
"0.55",
"0.08496",
"0.55"
],
[
"0.5215",
"0.539",
"0.9126",
"0.539",
"0.9126",
"0.553",
"0.5215",
"0.553"
],
[
"0.08496",
"0.55",
"0.4763",
"0.55",
"0.4763",
"0.563",
"0.08496",
"0.563"
],
[
"0.5215",
"0.5537",
"0.91",
"0.5537",
"0.91",
"0.5645",
"0.5215",
"0.5645"
],
[
"0.08496",
"0.5625",
"0.4795",
"0.563",
"0.4795",
"0.5776",
"0.08496",
"0.577"
],
[
"0.5215",
"0.565",
"0.9126",
"0.565",
"0.9126",
"0.579",
"0.5215",
"0.579"
],
[
"0.0866",
"0.579",
"0.4739",
"0.579",
"0.4739",
"0.589",
"0.0866",
"0.589"
],
[
"0.5215",
"0.579",
"0.911",
"0.579",
"0.911",
"0.5923",
"0.5215",
"0.5923"
],
[
"0.08496",
"0.5913",
"0.478",
"0.5913",
"0.478",
"0.605",
"0.08496",
"0.605"
],
[
"0.5215",
"0.5933",
"0.9126",
"0.5947",
"0.9126",
"0.608",
"0.5215",
"0.607"
],
[
"0.08496",
"0.606",
"0.4534",
"0.606",
"0.4534",
"0.619",
"0.08496",
"0.619"
],
[
"0.5195",
"0.608",
"0.911",
"0.608",
"0.911",
"0.622",
"0.5195",
"0.622"
],
[
"0.5186",
"0.6206",
"0.9126",
"0.622",
"0.9126",
"0.636",
"0.5186",
"0.6353"
],
[
"0.08496",
"0.6274",
"0.2942",
"0.629",
"0.2942",
"0.643",
"0.08496",
"0.6426"
],
[
"0.5186",
"0.6353",
"0.911",
"0.636",
"0.911",
"0.6494",
"0.5186",
"0.6484"
],
[
"0.08826",
"0.6465",
"0.478",
"0.6465",
"0.478",
"0.661",
"0.08826",
"0.661"
],
[
"0.517",
"0.6484",
"0.7295",
"0.6484",
"0.7295",
"0.662",
"0.517",
"0.662"
],
[
"0.0866",
"0.662",
"0.4739",
"0.662",
"0.4739",
"0.6724",
"0.0866",
"0.6724"
],
[
"0.08496",
"0.675",
"0.4795",
"0.675",
"0.4795",
"0.688",
"0.08496",
"0.688"
],
[
"0.5215",
"0.6753",
"0.7983",
"0.6753",
"0.7983",
"0.69",
"0.5215",
"0.69"
],
[
"0.0866",
"0.688",
"0.478",
"0.688",
"0.478",
"0.702",
"0.0866",
"0.702"
],
[
"0.5195",
"0.6934",
"0.9126",
"0.6943",
"0.9126",
"0.709",
"0.5195",
"0.7085"
],
[
"0.0866",
"0.704",
"0.4763",
"0.704",
"0.4763",
"0.714",
"0.0866",
"0.714"
],
[
"0.5195",
"0.7104",
"0.9126",
"0.7104",
"0.9126",
"0.7236",
"0.5195",
"0.7236"
],
[
"0.0866",
"0.7163",
"0.478",
"0.7163",
"0.478",
"0.73",
"0.0866",
"0.73"
],
[
"0.08417",
"0.729",
"0.4763",
"0.731",
"0.4763",
"0.744",
"0.08417",
"0.742"
],
[
"0.5195",
"0.7236",
"0.91",
"0.7236",
"0.91",
"0.7373",
"0.5195",
"0.7373"
],
[
"0.5195",
"0.7373",
"0.91",
"0.7373",
"0.91",
"0.7505",
"0.5195",
"0.7505"
],
[
"0.08496",
"0.7437",
"0.478",
"0.7437",
"0.478",
"0.757",
"0.08496",
"0.757"
],
[
"0.5215",
"0.7505",
"0.9126",
"0.7505",
"0.9126",
"0.765",
"0.5215",
"0.765"
],
[
"0.0866",
"0.759",
"0.4763",
"0.759",
"0.4763",
"0.7695",
"0.0866",
"0.7695"
],
[
"0.5195",
"0.765",
"0.91",
"0.765",
"0.91",
"0.779",
"0.5195",
"0.779"
],
[
"0.08496",
"0.7715",
"0.4763",
"0.7715",
"0.4763",
"0.785",
"0.08496",
"0.785"
],
[
"0.5195",
"0.779",
"0.911",
"0.779",
"0.911",
"0.7925",
"0.5195",
"0.7925"
],
[
"0.08496",
"0.785",
"0.3447",
"0.785",
"0.3447",
"0.799",
"0.08496",
"0.799"
],
[
"0.5215",
"0.7925",
"0.9126",
"0.7925",
"0.9126",
"0.806",
"0.5215",
"0.806"
],
[
"0.0866",
"0.8096",
"0.3586",
"0.8096",
"0.3586",
"0.8237",
"0.0866",
"0.8237"
],
[
"0.5215",
"0.806",
"0.9126",
"0.806",
"0.9126",
"0.8193",
"0.5215",
"0.8193"
],
[
"0.5215",
"0.821",
"0.911",
"0.821",
"0.911",
"0.834",
"0.5215",
"0.834"
],
[
"0.0866",
"0.827",
"0.4763",
"0.827",
"0.4763",
"0.8403",
"0.0866",
"0.8403"
],
[
"0.5195",
"0.834",
"0.914",
"0.834",
"0.914",
"0.8477",
"0.5195",
"0.8477"
],
[
"0.08496",
"0.8413",
"0.4795",
"0.8413",
"0.4795",
"0.855",
"0.08496",
"0.855"
],
[
"0.5195",
"0.8477",
"0.911",
"0.8477",
"0.911",
"0.8613",
"0.5195",
"0.8613"
],
[
"0.10706",
"0.8613",
"0.4795",
"0.8623",
"0.4795",
"0.877",
"0.10706",
"0.8755"
],
[
"0.5186",
"0.8623",
"0.91",
"0.8623",
"0.91",
"0.8755",
"0.5186",
"0.8755"
],
[
"0.08496",
"0.8755",
"0.3245",
"0.8755",
"0.3245",
"0.8896",
"0.08496",
"0.8896"
],
[
"0.5186",
"0.8755",
"0.911",
"0.8755",
"0.911",
"0.8896",
"0.5186",
"0.8896"
]
] | [
"prevention service resulted in a model based on 15 risk fac-",
".Evictioncohortinclusioncriteria:",
"tors that would have outperformed human judgment in se-",
"has evictionfiling in withoutadispositionbefore dateof analysis",
"has lost an eviction case witha dispositiondatein",
"lecting at-risk individuals (Shinn et al. 2013). In another",
".Homelessnesscohortexclusion criteria:",
"workfocusingonthosewhowerepreviouslyhomeless,ad-",
"startedorstoppedutilizationofhomelessnessservicesin,",
"ministrative data was used to better predict future home-",
"currently using homelessnessservices during dateof analysis",
"lessness within two years to best match people with pro-",
"enrolled inrehousingprogram inbutmoved inafterdateof analysis",
"Homelessnesslabel:utilizeshomelessservicesin2",
"grams, though this also disadvantaged other service recip",
"ients (Kube, Das, and Fowler 2019). Because of these trade-",
"analysis",
"2",
"offs, full automation of resource allocation is not recom-",
"mended:humandecision-makers should stillbeincontrol",
"Figure l: Inclusion/exclusion criteria for homelessness and",
"Work by the California Policy Lab predicted the risk",
"eviction. Indicators for homelessness in ACDHS data in",
"ofhomelessnessamongallresidentsinLosAngeles",
"cludeclientsinteractionswithhomelessnessservices,such",
"County(Wachteretal.2019).Severalfactors werefoundto",
"as staying in a shelter. We also include clients enrolled in",
"be indicative of future homelessness, including repeated and",
"rehousing programs that have not moved in as of the predic",
"recent interactions with government agencies or county ser-",
"tion date, as being homeless is a prerequisite for enrollment.",
"vices and prior mental health crises. Their predictive model",
"achieves29%precisionforallresidents.",
"Research Gaps.To our knowledge, no previous work has",
"on individuals who have previous eviction filings, homeless-",
"predicted future homelessness among those who are facing",
"ness spells, interactions with mental, behavioral, and phys-",
"eviction. he California Policy Lab's work likely suffers from",
"ical health institutions, address changes, or who have been",
"label bias due to the low percentage of homeless individu-",
"enrolled in a variety of other ACDHS and state programs",
"alsutilizing county services(seeSection7),whileourfocus",
"For each client at each analysis date, we generate ~7000",
"on informing a concrete action, rental assistance, allows our",
"features based on the data sources described in Supplemen-",
"system to fit into the existing support infrastructure.",
"tary Materials A.The features can be classified as follows:",
"4Predictive Approach",
"Demographics and Event features.For each tenant, we",
"We develop ML models to facilitate need-based prioritiza-",
"includethetotal numberof interactions witheachACDHS",
"and state program, total number of evictions, and total num-",
"tionforproactivedistributionofrentalassistance.Onelimi-",
"tation we encounter with this type of data is that it is condi-",
"berofphysical and mental healthvisits.Wealso includecat+",
"egorical features, such as the particular type of physical or",
"tioned on past rental assistance allocations and offers no in-",
"sights into the counterfactual outcomes. As a result, our ap-",
"mental health visit and demographic features (race, gender,",
"proach resembles an augmentation of the existing process",
"and age).",
"This is pertinent not only for the technical implementation",
"Temporal aggregation features.For each data source, we",
"but also for the interpretation of our results (e.g., in Sec-",
"generate several temporal features to capture the dynamic",
"tion 5.2, we concentrate on the outcomes for vulnerable in-",
"natureof theprocess andassesshowfrequentlyan individ-",
"dividuals who are not served under the current practice) and",
"ual interacts with state and local services within the last 3",
"the conceptualization of the overall project, which is cen",
"& 6 months, and 1, 2, 3, 4, and 5 years. For example, us-",
"tered on social impact. and equity. In Section 6, we outline",
"ing the eviction data, we generate the following features per",
"the transition from this stage to a fully deployed system.",
"individual: number of days since most recent eviction; num-",
"ber of evictions in the specified time period; sum, min, max,",
"4.1 Problem Formulation",
"average rent owed in eviction cases; min, max, average inter-",
"We formulate our task as a binary classification problem:",
"arrival times between evictions.",
"identifyingtenantsthatwillinteractwithhomelessnessser",
"vices within 12 months of the date of prediction. At each",
"4.3Model Training and Validation",
"prediction date, we consider individuals who have an active",
"We use various supervised classification methods to predict",
"evictionfilingagainstthemwithinthepastfourmonthsand",
"entry into homelessness: logistic regression (LR), decision",
"are not currently homeless.1 From this cohort, our models",
"identify the 100 individuals with the highest chance of inter-",
"trees (DT), random forests (RF), Adaboost, Light Gradi-",
"ent Boosting Model (LGBM), and XG boost using Scikit",
"acting with homelessness services in the next 12 months. We",
"Learn (Pedregosa et al. 2011) and the hyperparameter grid",
"select1ooindividualssincethiscorrespondstothemonthly",
"specified in Supplementary Materials C. To most closely re-",
"intervention capacity of ACDHS. Specific inclusion and ex-",
"produce the context in which our models will be deployed,",
"clusion criteria are detailed in Figure l.",
"we use temporal validation (Hoptroff 1993), generating dif-",
"4.2Data and Feature Engineering",
"ferent training and evaluation matrices for each analysis",
"date. For example, when evaluating the efficacy of predict-",
"ACDHS collects and combines administrative data from var-",
"ing homelessness as of January 2019, we train models on",
"ious county and state-level programs, including information",
"feature label pairs using data up to January 2019, and eval-",
"'More details on the demographic composition of this group are",
"uate them based on how many people become homeless be-",
"provided in Supplementary Materials B.",
"tween January 2019 and January 2020 (see Supplementary"
] | {
"arxiv_id": "2403.12599",
"language": "en",
"page_idx": 2
} |
||
2403.12599-en-3 | 2403.12599-en-3.jpg | <FigureHere>
Figure 2: LR and RF outperform all baselines: Precision@100 over time shows that out of all baselines, Bl: Previously Homeless performs best, but LR and RF perform better for all splits outside of the moratorium.
Materials D). Temporal validation requires that no information beyond a certain date is used to evaluate an algorithm to avoid data leakage, i.e., predicting the future using data from the future, which can be challenging when combining real-world data from disparate sources that are updated at different frequencies (see Section 8 for a full discussion).
## 4.4 Baseline Models
We compare the ML models to several simple baselines that either attempt to approximate ACDHS'current system of allocating rental assistance, or that provide simple improvements on the current status quo. Baselines represent simple heuristics that do not require implementing an ML model (i.e., ranking individuals based on a single attribute) and are, therefore, much easier to deploy and understand.
B1. Previous homelessness. Prior homelessness is a strong
indicator of future homelessness (Glendening and Shinn 2017). With this in mind, ACDHS could prioritize clients by the last date they interacted with any homelessness service. Here, a more recent date would imply a tenant is at higher risk of future homelessness.
B2. Baserate. If ACDHS were to randomly select individu-
als to give rental assistance, the precision of the approach would be equal to the proportion of individuals in our cohort who become homeless in the next year, around $2\%$ during our period of analysis.
B3. Earliest $\hat{\mathbf{OFP.As~an~approximation~of~the~current~pro-}}$
cess' frst-come-first-serve waitlist, we look for tenants who have been waiting for the longest since an Order $for\textit{Possession ( OFP) has been granted, which allows the}$ landlord to evict the tenant.
Other baselines were omitted from these results due to poor performance (see Supplementary Materials E for a full list).
## 5 $\textbf{Key Findings}$
We trained and validated over 5000 model variants to select the 100 individuals with the highest score, i.e. the highest risk of falling into homelessness within the next 12 months. The model selection metrics are measured at top-$k$,where $k=100.$
## 5.1 Predictive Models Are More Efficient and Effective than Heuristic Baselines
We evaluate the performance of predictive models based on their ability to improve the efficiency (i.e., ensuring assistance goes to those who are most in need) and effectiveness (i.e., maximizing the reach to people who would otherwise become homeless) of the allocation of rental assistance resources while satisfying the equity constraints. Efficiency and effectiveness are measured by precision@ 100 and recall@100 respectively. For each model, we calculate these metrics across all temporal validation splits and explore a broad hyperparameter space for each model type. Table 1 summarizes the metrics over time. We primarily focus on the average values of both metrics as we aim to select models that will generalize well to the future. Based on guidance from our partners at ACDHS, we exclude validation
splits that coincide with the COVID-19 eviction moratorium period when measuring average precision$^2.$ Figure 2 shows precision@ 100 over time for the best hyperparameter con- ${\tilde{\text{figuration for each model type}}}.$
We observe that RF and LR perform best and outperform simple heuristic baselines with respect to precision@ 100, showing a $\sim20\%$ improvement over the best heuristic, $Bl$, and performing 10x better than our approximation of the current allocation process, B3.
<table>
<tbody>
<tr>
<th rowspan="2">Model type</th>
<th colspan="3">Precision@ 100</th>
<th colspan="3">Recall@ 100</th>
</tr>
<tr>
<th>Avg</th>
<th>Min</th>
<th>Max</th>
<th>Avg</th>
<th>Min</th>
<th>Max</th>
</tr>
<tr>
<td>$RF$</td>
<td>0.20</td>
<td>0.14</td>
<td>0.25</td>
<td>0.22</td>
<td>0.16</td>
<td>0.34</td>
</tr>
<tr>
<td>LGBM</td>
<td>0.18</td>
<td>0.12</td>
<td>0.23</td>
<td>0.19</td>
<td>0.14</td>
<td>0.33</td>
</tr>
<tr>
<td>$\mathbf{LR}$</td>
<td>0.20</td>
<td>0.14</td>
<td>0.29</td>
<td>0.22</td>
<td>0.17</td>
<td>0.33</td>
</tr>
<tr>
<td>$DT$</td>
<td>0.16</td>
<td>0.08</td>
<td>0.21</td>
<td>0.18</td>
<td>0.09</td>
<td>0.30</td>
</tr>
<tr>
<td>XGBoost</td>
<td>0.18</td>
<td>0.13</td>
<td>0.23</td>
<td>0.20</td>
<td>0.16</td>
<td>0.29</td>
</tr>
<tr>
<td>$BI:Prev.$ $HL$ ,1</td>
<td>0.15</td>
<td>0.08</td>
<td>0.21</td>
<td>0.17</td>
<td>0.09</td>
<td>0.30</td>
</tr>
<tr>
<td>$B2:Baserate$</td>
<td>0.02</td>
<td>0.01</td>
<td>0.04</td>
<td>0.02</td>
<td>0.02</td>
<td>0.05</td>
</tr>
<tr>
<td>$B3\nobreak {: } Early$ $OFP$</td>
<td>0.03</td>
<td>0.00</td>
<td>0.05</td>
<td>0.03</td>
<td>0.00</td>
<td>0.07</td>
</tr>
</tbody>
</table>
Table 1: RF and LR outperform: Precision and recall over time for the model types considered show that our models outperform even the best baseline, $Bl: Previously\textit{Homeless. }$
## 5.2 Predictive Models Identify People Who Are Overlooked by the Current Process
The risk in relying on tenants to proactively apply for rental assistance is that the people most in need will not receive it and will end up homeless. In each cohort, about 75 people become homeless within a year of their eviction filing. Notably, a majority of these people (50 on average) do not apply for help. This missed group—those who interacted with the homelessness system within 12 months of the prediction date, did not apply for help, and did not receive rental
$^{2}$During this time, there are very few people facing eviction due to the moratorium on evictions during COVID-19 so our cohort becomes very small. See Supplementary Materials F. | [
[
"0.572",
"0.06757",
"0.6045",
"0.06757",
"0.6045",
"0.0802",
"0.572",
"0.0802"
],
[
"0.6396",
"0.0682",
"0.714",
"0.0682",
"0.714",
"0.0802",
"0.6396",
"0.0802"
],
[
"0.7812",
"0.0726",
"0.8906",
"0.0726",
"0.8906",
"0.0833",
"0.7812",
"0.0833"
],
[
"0.5767",
"0.07764",
"0.6055",
"0.07764",
"0.6055",
"0.089",
"0.5767",
"0.089"
],
[
"0.636",
"0.0789",
"0.721",
"0.0789",
"0.721",
"0.0903",
"0.636",
"0.0903"
],
[
"0.539",
"0.0802",
"0.5596",
"0.0802",
"0.5596",
"0.0903",
"0.539",
"0.0903"
],
[
"0.08496",
"0.06946",
"0.4763",
"0.06946",
"0.4763",
"0.0833",
"0.08496",
"0.0833"
],
[
"0.08496",
"0.0833",
"0.478",
"0.0833",
"0.478",
"0.0966",
"0.08496",
"0.0966"
],
[
"0.574",
"0.0903",
"0.6235",
"0.0903",
"0.6235",
"0.101",
"0.574",
"0.101"
],
[
"0.636",
"0.0903",
"0.7305",
"0.0903",
"0.7305",
"0.1016",
"0.636",
"0.1016"
],
[
"0.08496",
"0.0985",
"0.4795",
"0.0985",
"0.4795",
"0.11237",
"0.08496",
"0.11237"
],
[
"0.5186",
"0.0985",
"0.5376",
"0.0966",
"0.5415",
"0.1225",
"0.523",
"0.125"
],
[
"0.5864",
"0.1016",
"0.636",
"0.1016",
"0.636",
"0.113",
"0.5864",
"0.113"
],
[
"0.0866",
"0.11237",
"0.4795",
"0.11237",
"0.4795",
"0.1256",
"0.0866",
"0.1256"
],
[
"0.589",
"0.113",
"0.601",
"0.113",
"0.601",
"0.1218",
"0.589",
"0.1218"
],
[
"0.5146",
"0.1225",
"0.55",
"0.1061",
"0.562",
"0.1218",
"0.527",
"0.1383"
],
[
"0.5146",
"0.1237",
"0.5664",
"0.1237",
"0.5664",
"0.1831",
"0.5146",
"0.1831"
],
[
"0.08496",
"0.1395",
"0.46",
"0.1395",
"0.46",
"0.1528",
"0.08496",
"0.1528"
],
[
"0.08496",
"0.1256",
"0.4795",
"0.1256",
"0.4795",
"0.1395",
"0.08496",
"0.1395"
],
[
"0.0866",
"0.1654",
"0.2532",
"0.1654",
"0.2532",
"0.1799",
"0.0866",
"0.1799"
],
[
"0.08496",
"0.1849",
"0.4795",
"0.1849",
"0.4795",
"0.1989",
"0.08496",
"0.1989"
],
[
"0.0866",
"0.1995",
"0.4795",
"0.1995",
"0.4795",
"0.2134",
"0.0866",
"0.2134"
],
[
"0.562",
"0.1995",
"0.5933",
"0.1995",
"0.5933",
"0.2108",
"0.562",
"0.2108"
],
[
"0.621",
"0.1995",
"0.652",
"0.1995",
"0.652",
"0.2108",
"0.621",
"0.2108"
],
[
"0.6772",
"0.1995",
"0.7114",
"0.1995",
"0.7114",
"0.2108",
"0.6772",
"0.2108"
],
[
"0.738",
"0.1995",
"0.7686",
"0.1995",
"0.7686",
"0.2108",
"0.738",
"0.2108"
],
[
"0.796",
"0.1995",
"0.8286",
"0.1995",
"0.8286",
"0.2108",
"0.796",
"0.2108"
],
[
"0.8545",
"0.1995",
"0.8857",
"0.1995",
"0.8857",
"0.2108",
"0.8545",
"0.2108"
],
[
"0.08496",
"0.2134",
"0.478",
"0.2134",
"0.478",
"0.2267",
"0.08496",
"0.2267"
],
[
"0.5215",
"0.2216",
"0.911",
"0.2216",
"0.911",
"0.2361",
"0.5215",
"0.2361"
],
[
"0.0866",
"0.2291",
"0.4763",
"0.2291",
"0.4763",
"0.2393",
"0.0866",
"0.2393"
],
[
"0.08496",
"0.2412",
"0.4795",
"0.2412",
"0.4795",
"0.2551",
"0.08496",
"0.2551"
],
[
"0.5215",
"0.2374",
"0.91",
"0.2374",
"0.91",
"0.2474",
"0.5215",
"0.2474"
],
[
"0.08496",
"0.2551",
"0.4795",
"0.2551",
"0.4795",
"0.2683",
"0.08496",
"0.2683"
],
[
"0.5215",
"0.252",
"0.911",
"0.252",
"0.911",
"0.262",
"0.5215",
"0.262"
],
[
"0.08496",
"0.2695",
"0.406",
"0.2695",
"0.406",
"0.2827",
"0.08496",
"0.2827"
],
[
"0.5195",
"0.2644",
"0.7983",
"0.2644",
"0.7983",
"0.2778",
"0.5195",
"0.2778"
],
[
"0.08496",
"0.286",
"0.4795",
"0.2886",
"0.4795",
"0.303",
"0.08496",
"0.3005"
],
[
"0.10864",
"0.3037",
"0.478",
"0.3037",
"0.478",
"0.3145",
"0.10864",
"0.3145"
],
[
"0.5195",
"0.307",
"0.914",
"0.307",
"0.914",
"0.3208",
"0.5195",
"0.3208"
],
[
"0.10706",
"0.315",
"0.4795",
"0.315",
"0.4795",
"0.3289",
"0.10706",
"0.3289"
],
[
"0.5195",
"0.3213",
"0.914",
"0.3213",
"0.914",
"0.3352",
"0.5195",
"0.3352"
],
[
"0.10706",
"0.33",
"0.4795",
"0.33",
"0.4795",
"0.3435",
"0.10706",
"0.3435"
],
[
"0.5195",
"0.3352",
"0.9126",
"0.3352",
"0.9126",
"0.3484",
"0.5195",
"0.3484"
],
[
"0.10706",
"0.3447",
"0.4795",
"0.3447",
"0.4795",
"0.358",
"0.10706",
"0.358"
],
[
"0.5195",
"0.3484",
"0.7236",
"0.3484",
"0.7236",
"0.3623",
"0.5195",
"0.3623"
],
[
"0.10706",
"0.3591",
"0.353",
"0.3591",
"0.353",
"0.3694",
"0.10706",
"0.3694"
],
[
"0.536",
"0.3623",
"0.9126",
"0.3623",
"0.9126",
"0.3757",
"0.536",
"0.3757"
],
[
"0.0866",
"0.3718",
"0.4763",
"0.3718",
"0.4763",
"0.385",
"0.0866",
"0.385"
],
[
"0.5215",
"0.377",
"0.911",
"0.377",
"0.911",
"0.3901",
"0.5215",
"0.3901"
],
[
"0.10706",
"0.385",
"0.4795",
"0.385",
"0.4795",
"0.3984",
"0.10706",
"0.3984"
],
[
"0.5215",
"0.3901",
"0.914",
"0.3901",
"0.914",
"0.404",
"0.5215",
"0.404"
],
[
"0.10864",
"0.401",
"0.4763",
"0.401",
"0.4763",
"0.411",
"0.10864",
"0.411"
],
[
"0.5195",
"0.4048",
"0.914",
"0.4048",
"0.914",
"0.4192",
"0.5195",
"0.4192"
],
[
"0.10706",
"0.4136",
"0.478",
"0.4136",
"0.478",
"0.4268",
"0.10706",
"0.4268"
],
[
"0.5195",
"0.4185",
"0.721",
"0.4185",
"0.721",
"0.432",
"0.5195",
"0.432"
],
[
"0.10706",
"0.4268",
"0.3015",
"0.4268",
"0.3015",
"0.44",
"0.10706",
"0.44"
],
[
"0.0866",
"0.44",
"0.478",
"0.44",
"0.478",
"0.455",
"0.0866",
"0.455"
],
[
"0.653",
"0.4507",
"0.7573",
"0.4507",
"0.7573",
"0.4653",
"0.653",
"0.4653"
],
[
"0.8022",
"0.4507",
"0.886",
"0.4507",
"0.886",
"0.4653",
"0.8022",
"0.4653"
],
[
"0.528",
"0.457",
"0.609",
"0.459",
"0.609",
"0.4734",
"0.528",
"0.4717"
],
[
"0.649",
"0.4653",
"0.761",
"0.4653",
"0.761",
"0.4797",
"0.649",
"0.4797"
],
[
"0.10864",
"0.4558",
"0.478",
"0.4558",
"0.478",
"0.4666",
"0.10864",
"0.4666"
],
[
"0.10706",
"0.4685",
"0.4795",
"0.4685",
"0.4795",
"0.4817",
"0.10706",
"0.4817"
],
[
"0.7915",
"0.4653",
"0.9043",
"0.4653",
"0.9043",
"0.4797",
"0.7915",
"0.4797"
],
[
"0.1054",
"0.481",
"0.4795",
"0.4817",
"0.4795",
"0.4968",
"0.1054",
"0.4956"
],
[
"0.5293",
"0.4849",
"0.5557",
"0.4849",
"0.5557",
"0.4988",
"0.5293",
"0.4988"
],
[
"0.648",
"0.4849",
"0.7607",
"0.4849",
"0.7607",
"0.4988",
"0.648",
"0.4988"
],
[
"0.79",
"0.4849",
"0.903",
"0.4849",
"0.903",
"0.4988",
"0.79",
"0.4988"
],
[
"0.10706",
"0.4968",
"0.2893",
"0.4968",
"0.2893",
"0.5103",
"0.10706",
"0.5103"
],
[
"0.527",
"0.4988",
"0.581",
"0.4988",
"0.581",
"0.513",
"0.527",
"0.513"
],
[
"0.6465",
"0.4988",
"0.761",
"0.4988",
"0.761",
"0.513",
"0.6465",
"0.513"
],
[
"0.79",
"0.4988",
"0.903",
"0.4988",
"0.903",
"0.513",
"0.79",
"0.513"
],
[
"0.528",
"0.509",
"0.5596",
"0.512",
"0.5557",
"0.53",
"0.524",
"0.5264"
],
[
"0.6465",
"0.512",
"0.7607",
"0.512",
"0.7607",
"0.5264",
"0.6465",
"0.5264"
],
[
"0.79",
"0.513",
"0.903",
"0.513",
"0.903",
"0.5264",
"0.79",
"0.5264"
],
[
"0.528",
"0.5264",
"0.556",
"0.5264",
"0.556",
"0.5415",
"0.528",
"0.5415"
],
[
"0.08496",
"0.5146",
"0.4795",
"0.515",
"0.4795",
"0.53",
"0.08496",
"0.529"
],
[
"0.6465",
"0.5264",
"0.7607",
"0.5264",
"0.7607",
"0.5415",
"0.6465",
"0.5415"
],
[
"0.789",
"0.5264",
"0.903",
"0.5264",
"0.903",
"0.5405",
"0.789",
"0.5405"
],
[
"0.528",
"0.5405",
"0.596",
"0.5405",
"0.596",
"0.5547",
"0.528",
"0.5547"
],
[
"0.08496",
"0.53",
"0.478",
"0.53",
"0.478",
"0.5435",
"0.08496",
"0.5435"
],
[
"0.6465",
"0.5405",
"0.732",
"0.5405",
"0.732",
"0.5547",
"0.6465",
"0.5547"
],
[
"0.7256",
"0.5415",
"0.759",
"0.5415",
"0.759",
"0.5537",
"0.7256",
"0.5537"
],
[
"0.79",
"0.5405",
"0.903",
"0.5405",
"0.903",
"0.5547",
"0.79",
"0.5547"
],
[
"0.2091",
"0.5566",
"0.3545",
"0.5596",
"0.3545",
"0.5737",
"0.2091",
"0.5713"
],
[
"0.528",
"0.56",
"0.619",
"0.56",
"0.619",
"0.5747",
"0.528",
"0.5747"
],
[
"0.648",
"0.561",
"0.6855",
"0.561",
"0.6855",
"0.5747",
"0.648",
"0.5747"
],
[
"0.6787",
"0.561",
"0.722",
"0.561",
"0.722",
"0.5728",
"0.6787",
"0.5728"
],
[
"0.7183",
"0.561",
"0.759",
"0.561",
"0.759",
"0.5728",
"0.7183",
"0.5728"
],
[
"0.79",
"0.56",
"0.9043",
"0.56",
"0.9043",
"0.5737",
"0.79",
"0.5737"
],
[
"0.6465",
"0.5737",
"0.761",
"0.5737",
"0.761",
"0.5884",
"0.6465",
"0.5884"
],
[
"0.79",
"0.5747",
"0.903",
"0.5747",
"0.903",
"0.5884",
"0.79",
"0.5884"
],
[
"0.5312",
"0.5757",
"0.6177",
"0.5757",
"0.6177",
"0.587",
"0.5312",
"0.587"
],
[
"0.0866",
"0.5776",
"0.4795",
"0.5776",
"0.4795",
"0.5913",
"0.0866",
"0.5913"
],
[
"0.5312",
"0.5884",
"0.6323",
"0.5884",
"0.6323",
"0.5996",
"0.5312",
"0.5996"
],
[
"0.6445",
"0.5884",
"0.761",
"0.5884",
"0.761",
"0.603",
"0.6445",
"0.603"
],
[
"0.79",
"0.5884",
"0.903",
"0.5884",
"0.903",
"0.6016",
"0.79",
"0.6016"
],
[
"0.0866",
"0.608",
"0.4756",
"0.608",
"0.4756",
"0.6187",
"0.0866",
"0.6187"
],
[
"0.08496",
"0.5923",
"0.4795",
"0.5923",
"0.4795",
"0.606",
"0.08496",
"0.606"
],
[
"0.0866",
"0.6206",
"0.4795",
"0.6206",
"0.4795",
"0.634",
"0.0866",
"0.634"
],
[
"0.5186",
"0.616",
"0.9126",
"0.617",
"0.9126",
"0.632",
"0.5186",
"0.631"
],
[
"0.08496",
"0.632",
"0.1478",
"0.634",
"0.1478",
"0.6484",
"0.08496",
"0.6465"
],
[
"0.5215",
"0.634",
"0.911",
"0.634",
"0.911",
"0.6445",
"0.5215",
"0.6445"
],
[
"0.5195",
"0.6465",
"0.9126",
"0.6465",
"0.9126",
"0.66",
"0.5195",
"0.66"
],
[
"0.0866",
"0.661",
"0.1217",
"0.661",
"0.1217",
"0.674",
"0.0866",
"0.674"
],
[
"0.12335",
"0.66",
"0.4453",
"0.66",
"0.4453",
"0.675",
"0.12335",
"0.675"
],
[
"0.125",
"0.675",
"0.3857",
"0.6753",
"0.3857",
"0.69",
"0.125",
"0.6895"
],
[
"0.5215",
"0.69",
"0.55",
"0.69",
"0.55",
"0.7026",
"0.5215",
"0.7026"
],
[
"0.558",
"0.6895",
"0.892",
"0.69",
"0.892",
"0.705",
"0.558",
"0.704"
],
[
"0.0866",
"0.696",
"0.4795",
"0.696",
"0.4795",
"0.709",
"0.0866",
"0.709"
],
[
"0.08496",
"0.7104",
"0.4763",
"0.7104",
"0.4763",
"0.7236",
"0.08496",
"0.7236"
],
[
"0.5605",
"0.707",
"0.831",
"0.707",
"0.831",
"0.72",
"0.5605",
"0.72"
],
[
"0.08496",
"0.7236",
"0.4795",
"0.7236",
"0.4795",
"0.7373",
"0.08496",
"0.7373"
],
[
"0.5195",
"0.726",
"0.9126",
"0.726",
"0.9126",
"0.7407",
"0.5195",
"0.7407"
],
[
"0.0866",
"0.7373",
"0.4795",
"0.7373",
"0.4795",
"0.7505",
"0.0866",
"0.7505"
],
[
"0.5195",
"0.7407",
"0.9126",
"0.7407",
"0.9126",
"0.754",
"0.5195",
"0.754"
],
[
"0.08496",
"0.752",
"0.4763",
"0.752",
"0.4763",
"0.765",
"0.08496",
"0.765"
],
[
"0.5215",
"0.756",
"0.9126",
"0.756",
"0.9126",
"0.7666",
"0.5215",
"0.7666"
],
[
"0.08496",
"0.765",
"0.478",
"0.765",
"0.478",
"0.779",
"0.08496",
"0.779"
],
[
"0.5195",
"0.7676",
"0.91",
"0.7676",
"0.91",
"0.781",
"0.5195",
"0.781"
],
[
"0.08496",
"0.779",
"0.4763",
"0.779",
"0.4763",
"0.7925",
"0.08496",
"0.7925"
],
[
"0.5195",
"0.782",
"0.911",
"0.782",
"0.911",
"0.7954",
"0.5195",
"0.7954"
],
[
"0.08496",
"0.7925",
"0.4795",
"0.7925",
"0.4795",
"0.806",
"0.08496",
"0.806"
],
[
"0.5195",
"0.797",
"0.9126",
"0.797",
"0.9126",
"0.81",
"0.5195",
"0.81"
],
[
"0.08496",
"0.806",
"0.4795",
"0.806",
"0.4795",
"0.8193",
"0.08496",
"0.8193"
],
[
"0.5195",
"0.81",
"0.91",
"0.81",
"0.91",
"0.8237",
"0.5195",
"0.8237"
],
[
"0.08496",
"0.821",
"0.478",
"0.821",
"0.478",
"0.834",
"0.08496",
"0.834"
],
[
"0.5195",
"0.8237",
"0.9126",
"0.8237",
"0.9126",
"0.8384",
"0.5195",
"0.8384"
],
[
"0.08496",
"0.834",
"0.478",
"0.834",
"0.478",
"0.8477",
"0.08496",
"0.8477"
],
[
"0.08496",
"0.8477",
"0.4795",
"0.8477",
"0.4795",
"0.8613",
"0.08496",
"0.8613"
],
[
"0.539",
"0.8486",
"0.9126",
"0.8496",
"0.9126",
"0.8643",
"0.539",
"0.863"
],
[
"0.08496",
"0.8623",
"0.478",
"0.8623",
"0.478",
"0.8755",
"0.08496",
"0.8755"
],
[
"0.5195",
"0.863",
"0.914",
"0.863",
"0.914",
"0.877",
"0.5195",
"0.877"
],
[
"0.08496",
"0.8755",
"0.4795",
"0.8755",
"0.4795",
"0.8896",
"0.08496",
"0.8896"
],
[
"0.5186",
"0.8755",
"0.834",
"0.8755",
"0.834",
"0.8896",
"0.5186",
"0.8896"
]
] | [
"-RF",
"-- B1: Last HL",
"Eviction.moratorium",
"LR",
".--B2:Baserate",
"0.3.",
"Materials D). Temporal validation requires that no informa-",
"tion beyond a certain date is used to evaluate an algorithm",
"LGBM",
"--B3:EarlyOFP",
"to avoid data leakage, i.e., predicting the future using data",
"00L",
"XGBoost",
"from the future, which can be challenging when combining",
"DT",
"?02",
"1",
"different frequencies (see Section 8 for a full discussion).",
"real-world data from disparate sources that are updated at",
"4.4Baseline Models",
"We compare the ML models to several simple baselines that",
"either attempt to approximate ACDHS' current system of",
"2016",
"2017",
"2018",
"2019",
"2020",
"2021",
"allocating rental assistance, or that provide simple improve-",
"Figure 2: LR and RF outperform all baselines: Preci-",
"ments onthe currentstatus quo.Baselinesrepresent simple",
"heuristics that do not require implementing an ML model",
"sion@lO0overtime showsthat out ofall baselines,Bl:Pre-",
"(i.e., ranking individuals based on a single attribute) and are,",
"viouslyHomelessperformsbest,butLRandRFperformbet-",
"therefore, much easier to deploy and understand.",
"ter for all splits outside of the moratorium.",
"B1. Previous homelessness. Prior homelessness is a strong",
"indicatoroffuturehomelessness(Glendeningand Shinn",
"splits that coincide with the COVID-19 eviction moratorium",
"2017). With this in mind, ACDHS could prioritize clients",
"period when measuring average precision?. Figure 2 shows",
"by the last date they interacted with any homelessness",
"precision@ 100 over time for the best hyperparameter con-",
"service. Here, a more recent date would imply a tenant is",
"figuration for each model type.",
"at higher risk of futurehomelessness.",
"We observe that RF and LR perform best and outperform",
"B2. Baserate. If ACDHS were to randomly select individu-",
"simple heuristic baselines with respect to precision@100,",
"als to give rental assistance, the precision of the approach",
"showing a ~20% improvement over the best heuristic, Bl,",
"wouldbeequaltotheproportionofindividualsinourco-",
"and performing 10x better than our approximation of the",
"hort who become homeless in the next year, around 2%",
"current allocation process, B3.",
"during our period of analysis.",
"B3. Earliest OFP. As an approximation of the current pro-",
"Precision@100",
"Recall@100",
"Model type",
"Avg Min Max",
"cess'first-come-first-serve waitlist,welookfortenants",
"who have been waiting for the longest since an Order",
"Avg Min Max",
"for Possession (OFP) has been granted, which allows the",
"RF",
"0.20 0.14 0.25",
"0.22 0.16 0.34",
"landlord to evict the tenant.",
"LGBM",
"0.18 0.12 0.23",
"0.19 0.14 0.33",
"LR",
"0.20 0.14 0.29",
"0.22 0.17 0.33",
"DT",
"Other baselines were omitted from these results due to poor",
"0.16 0.080.21",
"0.180.090.30",
"XGBoost",
"performance (see Supplementary Materials E for a full list).",
"0.180.13",
"0.23",
"0.20 0.16 0.29",
"5 KeyFindings",
"Bl:Prev. HL",
"0.15",
"0.08",
"0.21",
"0.17 0.09 0.30",
"0.02 0.01 0.04",
"0.02 0.02 0.05",
"B2:Baserate",
"We trained and validated over 5000 model variants to select",
"B3:Early OFP",
"0.030.00 0.05",
"0.030.000.07",
"riskoffallingintohomelessnesswithinthenext12months",
"the 100 individuals with the highest score, i.e. the highest",
"The model selection metrics are measured at top-k, where",
"Table l: RF and LR outperform: Precision and recall over",
"k =100.",
"timeforthe model types considered show that our models",
"outperform even the best baseline, Bl:Previously Homeless.",
"5.1",
"Predictive Models Are More Efficient and",
"Effective than Heuristic Baselines",
"5.2",
"Predictive Models Identify People Who Are",
"We evaluate the performance of predictive models based on",
"their ability to improve the efficiency (i.e., ensuring assis-",
"Overlooked by the Current Process",
"tance goes to those who are most in need) and effectiveness",
"The risk in relying on tenants to proactively apply for rental",
"(i.e., maximizing the reach to people who would otherwise",
"assistance is that the people most in need will not receive it",
"become homeless) of the allocation of rental assistance re-",
"and willend uphomeless.Ineachcohort,about75people",
"sources while satisfying the equity constraints. Efficiency",
"become homeless within a year of their eviction filing. No-",
"and effectiveness are measured by precision@100 and re-",
"tably, a majority of these people (50 on average) do not ap-",
"call@100 respectively. For each model, we calculate these",
"ply for help. This missed group-those who interacted with",
"metrics across all temporal validation splits and explore a",
"the homelessness system within 12 months of the predic-",
"broad hyperparameter space for each model type. Table 1",
"tion date, did not apply for help, and did not receive rental",
"summarizes the metrics over time. We primarily focus on",
"the average values of both metrics as we aim to select mod-",
"2During this time,there are veryfew peoplefacing eviction due",
"els that will generalize well to the future. Based on guid-",
"to the moratorium on evictions during COVID-19 so our cohort",
"ance from our partners at ACDHS, we exclude validation",
"becomes very small. See Supplementary Materials F."
] | {
"arxiv_id": "2403.12599",
"language": "en",
"page_idx": 3
} |
||
2403.12599-en-4 | 2403.12599-en-4.jpg | assistance\_illustrates the limitations of the current reactive
allocation process.
In Figure 3, we examine what percentage of the missed group is found by each model, comparing one of our best ML models (RF) to the baselines $Bl:PreviouslyHome-$ less and B2: Baserate. B2 finds 4% of this group on average, meaning that proactive outreach to a list of 100 people from this random approach would find about 4 people every month who would be overlooked by the current rental assistance practice and would end up homeless. Both $BI$ and RF reach a substantially larger group of residents, 23% for $Bl$ and 28% for RF on average, potentially providing rental assistance to an additional 10-20 people from the missed group in each cohort every month.
<FigureHere>
Figure 3: Percentage found of missed group, the individuals who become homeless having not contacted Allegheny County Link or received rental assistance.
5.3 Predictive Models Can Promote Fairness and
Equity
Efficiency and effectiveness metrics do not reveal whether allocating resources according to these models would be biased against vulnerable groups (Barocas, Hardt, and Narayanan 2023). Since we are informing the allocation of a scarce resource, we use equality of opportunity (Hardt, Price, and Srebro 2016) as the fairness principle, which is captured using the true positive rate (TPR) of vulnerable subgroups in the top 100. We consider the following groups: Race. Black individuals are at higher risk of falling into homelessness in Allegheny County. Out of those that will become homeless in our validation cohorts, on average, 60% were Black, even though Black individuals make up only 14% of the county's population (U.S. Census Bureau 2022). To mitigate these disproportionate impacts, ML models should have a higher TPR compared to white individuals (i.e., serving a higher proportion of Black individuals who are actually at risk of homelessness). We calculate $\begin{array}{l}{\frac{P(D=1|Y=1,A=blac\dot{k})}{P(D=1|Y=1,A=white)}}&\mathrm{where~D,~Y,~and~A~represent~the~de-}\\\text{cision, outcome, and attribute of interest (e.g., race) respec-}\end{array}$ tively, finding average recall ratios of 1.34 and 1.14 respectively across temporal splits, indicating that our models meet our fairness considerations for race.
Gender. For gender, we consider men and women since data on non-binary and transgender individuals is not well-documented in ACDHS data. We calculate
$\frac{P(D=1|Y=1,A=female)}{P(D=1|Y=1,A=male)}$ for our best models and see that RF and LR models are slightly under-serving women, with average TPR ratios of 0.9 and 0.87 respectively. This issue needs to be considered further in the resource allocation process.
# 5.4 Prior Homelessness and Mental Health Crises Contribute to Homelessness Risk
We find that previous use of homelessness services and interactions with mental and behavioral health services are consistently predictive of future homelessness spells across all validation temporal splits. Most predictive features include:
Previous homelessness service utilization. The number, duration, and recency of past homelessness spells are highly predictive of future homelessness. Past referrals to homelessness services, emergency shelter utilization, and public housing utilization are identified as highly predictive. Interestingly, the most predictive feature was the number of days since the last homelessness spell (baseline $BI).$
We also compared the characteristics of the top 100 individuals versus the rest of the tenants. We found that for the validation split starting on Sept. 1, 2021, the top 100 individuals were 34x more likely to have been in an emergency shelter in their lifetime, 31x more likely to have been homeless, and likely to have spent 28x as many days in homelessness compared to other tenants.
Mental and behavioral health service interaction. Mental and behavioral health events and related service utilization are highly predictive of future homelessness. In the best performing RF, the number of days since the last mental health or behavioral health crisis event, the duration of mental health service utilization, the number of times one used mental health services, and the number of mental health crisis events were highly predictive features.
Compared to the rest of the cohort, those in the top 100 were 100x more likely to have had a mental health crisis event in the last three years and 34x more likely to have had one in their lifetime, likely to have had 28x as many days utilizing mental health services, and likely to have had 24x more behavioral health events in their lifetime.
5.5 First-Time Homelessness is Harder to Predict
About half of the people in a cohort who become homeless have not previously been homeless, but we find that our models rely on features about previous homelessness spells, resulting in lower recall for people experiencing first-time homelessness. Our models perform substantially better on people with a history of homelessness: RF's average recall (excluding the eviction moratorium) is 55% for people who have experienced previous homelessness but only 4% for people experiencing frst-time homelessness. We are exploring 1) building separate models, and 2) adding additional data sources to better predict future homelessness based on whether or not the person has previously been homeless
Interestingly, the group who receives rental assistance through the existing practice appears to mostly be people without a history of homelessness: for the validation split | [
[
"0.5186",
"0.05936",
"0.914",
"0.06757",
"0.914",
"0.089",
"0.5186",
"0.0808"
],
[
"0.08496",
"0.06946",
"0.4795",
"0.06946",
"0.4795",
"0.0833",
"0.08496",
"0.0833"
],
[
"0.523",
"0.0745",
"0.6626",
"0.07574",
"0.6626",
"0.0871",
"0.523",
"0.0859"
],
[
"0.08496",
"0.084",
"0.21",
"0.084",
"0.21",
"0.09784",
"0.08496",
"0.09784"
],
[
"0.5186",
"0.0859",
"0.91",
"0.0859",
"0.91",
"0.09973",
"0.5186",
"0.09973"
],
[
"0.10297",
"0.09784",
"0.4805",
"0.09784",
"0.4805",
"0.1111",
"0.10297",
"0.1111"
],
[
"0.5195",
"0.101",
"0.9126",
"0.101",
"0.9126",
"0.11426",
"0.5195",
"0.11426"
],
[
"0.08496",
"0.11237",
"0.4795",
"0.1111",
"0.4795",
"0.125",
"0.08496",
"0.1256"
],
[
"0.5186",
"0.11426",
"0.902",
"0.11426",
"0.902",
"0.1276",
"0.5186",
"0.1276"
],
[
"0.08826",
"0.127",
"0.4756",
"0.127",
"0.4756",
"0.137",
"0.08826",
"0.137"
],
[
"0.08496",
"0.1395",
"0.4763",
"0.1395",
"0.4763",
"0.1528",
"0.08496",
"0.1528"
],
[
"0.5215",
"0.1433",
"0.5537",
"0.1433",
"0.5537",
"0.1559",
"0.5215",
"0.1559"
],
[
"0.556",
"0.1414",
"0.911",
"0.1414",
"0.911",
"0.1547",
"0.556",
"0.1547"
],
[
"0.08417",
"0.154",
"0.4795",
"0.1528",
"0.4795",
"0.1666",
"0.08417",
"0.1674"
],
[
"0.5605",
"0.159",
"0.8145",
"0.159",
"0.8145",
"0.1724",
"0.5605",
"0.1724"
],
[
"0.0866",
"0.1674",
"0.4763",
"0.1674",
"0.4763",
"0.1812",
"0.0866",
"0.1812"
],
[
"0.0866",
"0.1812",
"0.4763",
"0.1812",
"0.4763",
"0.1945",
"0.0866",
"0.1945"
],
[
"0.5215",
"0.1787",
"0.911",
"0.1787",
"0.911",
"0.1925",
"0.5215",
"0.1925"
],
[
"0.0866",
"0.1963",
"0.478",
"0.1963",
"0.478",
"0.207",
"0.0866",
"0.207"
],
[
"0.5195",
"0.1925",
"0.91",
"0.1932",
"0.91",
"0.207",
"0.5195",
"0.2058"
],
[
"0.08826",
"0.209",
"0.478",
"0.209",
"0.478",
"0.2229",
"0.08826",
"0.2229"
],
[
"0.5215",
"0.207",
"0.9126",
"0.207",
"0.9126",
"0.2203",
"0.5215",
"0.2203"
],
[
"0.08496",
"0.2203",
"0.478",
"0.2216",
"0.478",
"0.2361",
"0.08496",
"0.2349"
],
[
"0.5215",
"0.2216",
"0.911",
"0.2216",
"0.911",
"0.2349",
"0.5215",
"0.2349"
],
[
"0.08496",
"0.2361",
"0.4795",
"0.2361",
"0.4795",
"0.25",
"0.08496",
"0.25"
],
[
"0.5215",
"0.2443",
"0.911",
"0.2443",
"0.911",
"0.2583",
"0.5215",
"0.2583"
],
[
"0.08496",
"0.2507",
"0.3137",
"0.2507",
"0.3137",
"0.2644",
"0.08496",
"0.2644"
],
[
"0.523",
"0.26",
"0.911",
"0.26",
"0.911",
"0.2708",
"0.523",
"0.2708"
],
[
"0.0768",
"0.2683",
"0.1356",
"0.2695",
"0.1299",
"0.4092",
"0.0719",
"0.4077"
],
[
"0.5215",
"0.2727",
"0.911",
"0.2727",
"0.911",
"0.286",
"0.5215",
"0.286"
],
[
"0.1584",
"0.281",
"0.246",
"0.281",
"0.246",
"0.2922",
"0.1584",
"0.2922"
],
[
"0.3423",
"0.279",
"0.452",
"0.2798",
"0.452",
"0.2942",
"0.3423",
"0.2935"
],
[
"0.1602",
"0.2935",
"0.2532",
"0.2935",
"0.2532",
"0.305",
"0.1602",
"0.305"
],
[
"0.524",
"0.2886",
"0.911",
"0.2886",
"0.911",
"0.2986",
"0.524",
"0.2986"
],
[
"0.1584",
"0.3062",
"0.2255",
"0.3062",
"0.2255",
"0.3176",
"0.1584",
"0.3176"
],
[
"0.5195",
"0.3005",
"0.911",
"0.3005",
"0.911",
"0.3145",
"0.5195",
"0.3145"
],
[
"0.1478",
"0.3193",
"0.2336",
"0.3193",
"0.2336",
"0.3308",
"0.1478",
"0.3308"
],
[
"0.5215",
"0.3145",
"0.9126",
"0.3145",
"0.9126",
"0.3276",
"0.5215",
"0.3276"
],
[
"0.5215",
"0.3289",
"0.8286",
"0.3289",
"0.8286",
"0.3423",
"0.5215",
"0.3423"
],
[
"0.536",
"0.3423",
"0.9126",
"0.3423",
"0.9126",
"0.356",
"0.536",
"0.356"
],
[
"0.10455",
"0.3567",
"0.1177",
"0.3528",
"0.1217",
"0.3599",
"0.10864",
"0.3643"
],
[
"0.523",
"0.358",
"0.911",
"0.358",
"0.911",
"0.3687",
"0.523",
"0.3687"
],
[
"0.5195",
"0.3706",
"0.911",
"0.3706",
"0.911",
"0.3838",
"0.5195",
"0.3838"
],
[
"0.5195",
"0.3838",
"0.9126",
"0.3838",
"0.9126",
"0.3977",
"0.5195",
"0.3977"
],
[
"0.1259",
"0.3945",
"0.1569",
"0.3945",
"0.1569",
"0.4072",
"0.1259",
"0.4072"
],
[
"0.1847",
"0.3945",
"0.2157",
"0.3945",
"0.2157",
"0.406",
"0.1847",
"0.406"
],
[
"0.2418",
"0.3945",
"0.2737",
"0.3945",
"0.2737",
"0.406",
"0.2418",
"0.406"
],
[
"0.2998",
"0.3945",
"0.3325",
"0.3945",
"0.3325",
"0.406",
"0.2998",
"0.406"
],
[
"0.357",
"0.3945",
"0.392",
"0.3945",
"0.392",
"0.406",
"0.357",
"0.406"
],
[
"0.4167",
"0.3945",
"0.4478",
"0.3945",
"0.4478",
"0.4072",
"0.4167",
"0.4072"
],
[
"0.5195",
"0.3977",
"0.9126",
"0.3984",
"0.9126",
"0.4124",
"0.5195",
"0.411"
],
[
"0.08496",
"0.414",
"0.4763",
"0.4153",
"0.4763",
"0.43",
"0.08496",
"0.4287"
],
[
"0.5195",
"0.4124",
"0.9126",
"0.4136",
"0.9126",
"0.4248",
"0.5195",
"0.4236"
],
[
"0.0866",
"0.43",
"0.478",
"0.43",
"0.478",
"0.443",
"0.0866",
"0.443"
],
[
"0.5195",
"0.4268",
"0.6953",
"0.4268",
"0.6953",
"0.44",
"0.5195",
"0.44"
],
[
"0.0866",
"0.443",
"0.3628",
"0.443",
"0.3628",
"0.457",
"0.0866",
"0.457"
],
[
"0.5215",
"0.4495",
"0.911",
"0.4495",
"0.911",
"0.4634",
"0.5215",
"0.4634"
],
[
"0.523",
"0.4653",
"0.91",
"0.4653",
"0.91",
"0.476",
"0.523",
"0.476"
],
[
"0.5215",
"0.4792",
"0.911",
"0.4792",
"0.911",
"0.4893",
"0.5215",
"0.4893"
],
[
"0.08826",
"0.486",
"0.1217",
"0.486",
"0.1217",
"0.4988",
"0.08826",
"0.4988"
],
[
"0.1152",
"0.486",
"0.4722",
"0.486",
"0.4722",
"0.5",
"0.1152",
"0.5"
],
[
"0.5195",
"0.4912",
"0.9126",
"0.4912",
"0.9126",
"0.505",
"0.5195",
"0.505"
],
[
"0.1259",
"0.5",
"0.1814",
"0.502",
"0.1805",
"0.5176",
"0.125",
"0.515"
],
[
"0.5195",
"0.504",
"0.91",
"0.505",
"0.91",
"0.5195",
"0.5195",
"0.518"
],
[
"0.08826",
"0.521",
"0.478",
"0.521",
"0.478",
"0.534",
"0.08826",
"0.534"
],
[
"0.5215",
"0.521",
"0.911",
"0.521",
"0.911",
"0.531",
"0.5215",
"0.531"
],
[
"0.0866",
"0.534",
"0.4763",
"0.534",
"0.4763",
"0.5474",
"0.0866",
"0.5474"
],
[
"0.523",
"0.535",
"0.911",
"0.535",
"0.911",
"0.5454",
"0.523",
"0.5454"
],
[
"0.0866",
"0.55",
"0.4795",
"0.55",
"0.4795",
"0.563",
"0.0866",
"0.563"
],
[
"0.5215",
"0.547",
"0.7935",
"0.547",
"0.7935",
"0.56",
"0.5215",
"0.56"
],
[
"0.0866",
"0.563",
"0.4763",
"0.563",
"0.4763",
"0.5737",
"0.0866",
"0.5737"
],
[
"0.5376",
"0.563",
"0.911",
"0.563",
"0.911",
"0.5737",
"0.5376",
"0.5737"
],
[
"0.0866",
"0.577",
"0.4763",
"0.577",
"0.4763",
"0.5903",
"0.0866",
"0.5903"
],
[
"0.5215",
"0.5757",
"0.911",
"0.5757",
"0.911",
"0.589",
"0.5215",
"0.589"
],
[
"0.08496",
"0.589",
"0.4763",
"0.589",
"0.4763",
"0.603",
"0.08496",
"0.603"
],
[
"0.5215",
"0.589",
"0.9126",
"0.589",
"0.9126",
"0.603",
"0.5215",
"0.603"
],
[
"0.08496",
"0.604",
"0.4795",
"0.604",
"0.4795",
"0.617",
"0.08496",
"0.617"
],
[
"0.5215",
"0.605",
"0.911",
"0.605",
"0.911",
"0.6157",
"0.5215",
"0.6157"
],
[
"0.08496",
"0.6187",
"0.478",
"0.6187",
"0.478",
"0.632",
"0.08496",
"0.632"
],
[
"0.5215",
"0.6187",
"0.911",
"0.6187",
"0.911",
"0.629",
"0.5215",
"0.629"
],
[
"0.5195",
"0.631",
"0.826",
"0.631",
"0.826",
"0.6445",
"0.5195",
"0.6445"
],
[
"0.0866",
"0.6396",
"0.478",
"0.6396",
"0.478",
"0.653",
"0.0866",
"0.653"
],
[
"0.0866",
"0.655",
"0.478",
"0.655",
"0.478",
"0.6655",
"0.0866",
"0.6655"
],
[
"0.5215",
"0.659",
"0.9087",
"0.659",
"0.9087",
"0.6724",
"0.5215",
"0.6724"
],
[
"0.08496",
"0.6675",
"0.4763",
"0.6685",
"0.4763",
"0.6816",
"0.08496",
"0.681"
],
[
"0.08496",
"0.6816",
"0.478",
"0.6816",
"0.478",
"0.696",
"0.08496",
"0.696"
],
[
"0.5215",
"0.681",
"0.9126",
"0.681",
"0.9126",
"0.6943",
"0.5215",
"0.6943"
],
[
"0.08496",
"0.696",
"0.478",
"0.696",
"0.478",
"0.709",
"0.08496",
"0.709"
],
[
"0.523",
"0.6963",
"0.911",
"0.6963",
"0.911",
"0.707",
"0.523",
"0.707"
],
[
"0.08496",
"0.709",
"0.478",
"0.709",
"0.478",
"0.7236",
"0.08496",
"0.7236"
],
[
"0.5215",
"0.7104",
"0.911",
"0.7104",
"0.911",
"0.72",
"0.5215",
"0.72"
],
[
"0.08496",
"0.7236",
"0.478",
"0.7236",
"0.478",
"0.7373",
"0.08496",
"0.7373"
],
[
"0.5215",
"0.7227",
"0.9126",
"0.7227",
"0.9126",
"0.7363",
"0.5215",
"0.7363"
],
[
"0.08496",
"0.7373",
"0.478",
"0.7373",
"0.478",
"0.7505",
"0.08496",
"0.7505"
],
[
"0.5195",
"0.7363",
"0.914",
"0.7363",
"0.914",
"0.75",
"0.5195",
"0.75"
],
[
"0.0866",
"0.7505",
"0.478",
"0.7505",
"0.478",
"0.7646",
"0.0866",
"0.7646"
],
[
"0.5186",
"0.75",
"0.914",
"0.7485",
"0.914",
"0.763",
"0.5186",
"0.7646"
],
[
"0.0866",
"0.7603",
"0.4795",
"0.7646",
"0.4795",
"0.785",
"0.0866",
"0.781"
],
[
"0.5215",
"0.7646",
"0.9126",
"0.7646",
"0.9126",
"0.778",
"0.5215",
"0.778"
],
[
"0.09235",
"0.7734",
"0.1299",
"0.776",
"0.1299",
"0.7827",
"0.09235",
"0.781"
],
[
"0.1283",
"0.778",
"0.1381",
"0.778",
"0.1381",
"0.782",
"0.1283",
"0.782"
],
[
"0.1887",
"0.7764",
"0.2173",
"0.7764",
"0.2173",
"0.781",
"0.1887",
"0.781"
],
[
"0.0866",
"0.7827",
"0.4756",
"0.7827",
"0.4756",
"0.7974",
"0.0866",
"0.7974"
],
[
"0.5195",
"0.778",
"0.911",
"0.778",
"0.911",
"0.7915",
"0.5195",
"0.7915"
],
[
"0.5186",
"0.7915",
"0.911",
"0.7905",
"0.911",
"0.8047",
"0.5186",
"0.806"
],
[
"0.0866",
"0.799",
"0.4763",
"0.799",
"0.4763",
"0.8125",
"0.0866",
"0.8125"
],
[
"0.5215",
"0.806",
"0.9126",
"0.806",
"0.9126",
"0.8193",
"0.5215",
"0.8193"
],
[
"0.0866",
"0.8125",
"0.4795",
"0.8125",
"0.4795",
"0.8257",
"0.0866",
"0.8257"
],
[
"0.5186",
"0.8193",
"0.911",
"0.8184",
"0.911",
"0.832",
"0.5186",
"0.8335"
],
[
"0.08496",
"0.8257",
"0.3228",
"0.8257",
"0.3228",
"0.839",
"0.08496",
"0.839"
],
[
"0.5215",
"0.8354",
"0.888",
"0.8354",
"0.888",
"0.845",
"0.5215",
"0.845"
],
[
"0.0866",
"0.8467",
"0.4795",
"0.8467",
"0.4795",
"0.8613",
"0.0866",
"0.8613"
],
[
"0.536",
"0.8477",
"0.9126",
"0.8477",
"0.9126",
"0.8613",
"0.536",
"0.8613"
],
[
"0.0866",
"0.863",
"0.478",
"0.863",
"0.478",
"0.8735",
"0.0866",
"0.8735"
],
[
"0.5195",
"0.8623",
"0.9126",
"0.8623",
"0.9126",
"0.8755",
"0.5195",
"0.8755"
],
[
"0.08496",
"0.8755",
"0.4795",
"0.8755",
"0.4795",
"0.8896",
"0.08496",
"0.8896"
],
[
"0.5195",
"0.8755",
"0.9126",
"0.8755",
"0.9126",
"0.8896",
"0.5195",
"0.8896"
]
] | [
"P(D=1Y-1.A-female) for our best models and see that RF",
"assistance-illustrates the limitations of the current reactive",
"P(D=1|Y=1,A=male)",
"allocation process.",
"and LR models are slightly under-serving women, with aver-",
"In Figure 3, we examine what percentage of the missed",
"age TPR ratios of 0.9 and 0.87 respectively. This issue needs",
"groupisfound by each model,comparing oneof ourbest",
"to be considered further in the resource allocation process.",
"ML models (RF)to the baselines Bl:PreviouslyHome-",
"less and B2: Baserate. B2 finds 4% of this group on aver-",
"5.4",
"Prior Homelessness and Mental Health Crises",
"age,meaningthatproactiveoutreachtoalistof1o0people",
"Contribute to Homelessness Risk",
"from this random approach would find about 4 people every",
"month who would be overlooked by the current rental as-",
"We find that previous use of homelessness services and inter-",
"sistancepractice andwould end uphomeless.BothBl and",
"actions withmental and behavioral health services are con-",
"RF reach a substantially larger group of residents, 23% for",
"sistently predictive of future homelessness spells across all",
"B1 and 28% for RF on average, potentially providing rental",
"validation temporal splits. Most predictive features include:",
" assistance to an additional 10-20 people from the missed",
"Previous homelessness service utilization.The number,",
"group in each cohort every month.",
"duration,andrecencyofpasthomelessness spellsarehighly",
"woopapuno",
"predictive of future homelessness. Past referrals to home-",
"Randomforest",
"Eviction moratorium",
"Currentprocess",
"lessnessservices,emergencyshelterutilization,andpublic",
"B1: Last HL",
"housing utilization are identified as highly predictive. Inter-",
"B2:Baserate",
"estingly, the most predictive feature was the number of days",
"since the last homelessness spell (baseline BI).",
"We also compared the characteristics of the top 100 indi-",
"20",
"viduals versus the rest of the tenants. We found that for the",
"validation split starting on Sept. 1, 2021, the top 100 individ-",
"uals were 34x more likely to have been in an emergency shel-",
"2016",
"2017",
"2018",
"2019",
"2020",
"2021",
"ter in their lifetime,3lx more likely to have been homeless,",
"Figure 3: Percentage found of missed group, the individu-",
"and likelyto havespent28xasmanydays inhomelessness",
"als who become homeless having not contacted Allegheny",
"compared to other tenants.",
"County Link or received rental assistance.",
"Mental and behavioral health service interaction.Men-",
"tal and behavioral health events and related serviceutiliza-",
"tionarehighlypredictiveoffuturehomelessness.In the best",
"5.3",
"Predictive Models Can Promote Fairness and",
"performing RF, the number of days since the last mental",
"Equity",
"health or behavioral health crisis event, the duration of men-",
"Efficiency and effectiveness metrics do not reveal whether",
"tal health service utilization, the number of times one used",
"allocating resources according to these models would be",
"mentalhealthservices,andthenumberofmentalhealthcri-",
"biased against vulnerable groups (Barocas, Hardt, and",
"sis events were highly predictive features.",
"Narayanan 2023).Since we are informing the allocation of",
"Comparedto therestofthecohort,thoseinthetop1oo",
"a scarce resource, we use equality of opportunity (Hardt",
"were 10Ox more likely to have had a mental health crisis",
"Price, and Srebro 2016) as the fairness principle, which is",
"event in the last three years and 34x more likely to have had",
"captured using the true positive rate (TPR) of vulnerable",
"one in theirlifetime,likely to have had 28x as many days",
" subgroups in the top 100. We consider the following groups:",
"utilizingmentalhealthservices.andlikelytohavehad24x",
"more behavioral health events in their lifetime.",
"Race.Black individuals are at higher risk of falling into",
"homelessness in Allegheny County.Out of those that will",
"5.5First-Time Homelessness is Harder to Predict",
"become homeless in our validation cohorts,on average,",
"60% were Black, even though Black individuals make up",
"About half of the people in a cohort who become home-",
"only 14% of the county's population (U.S. Census Bu-",
"lesshavenotpreviouslybeenhomeless,butwefindthatour",
"reau 2022). To mitigate these disproportionate impacts, ML",
"modelsrely onfeaturesaboutprevioushomelessness spells.",
"models should have a higher TPR compared to white indi-",
"resulting in lower recall for people experiencing first-time",
"viduals (i.e., serving a higher proportion of Black individu-",
"homelessness. Our models perform substantially better on",
"als who are actually at risk of homelessness). We calculate",
"people with a history of homelessness: RF's average recall",
"P(D-IV-LA-black) where D, Y, and A represent the de-",
"(excluding the eviction moratorium) is 55% for people who",
"P(D=",
"L",
"nht.",
"cision, outcome, and attribute of interest (e.g., race) respec-",
"have experienced previous homelessness but only 4% for",
"people experiencing first-time homelessness. We are explor-",
"tively, finding average recall ratios of 1.34 and 1.14 respec-",
"ing 1) building separate models, and 2) adding additional",
"tively across temporal splits, indicating that our models meet",
"data sources to better predict future homelessness based on",
"our fairness considerations for race.",
"whetherornotthepersonhaspreviouslybeenhomeless.",
"Gender. For gender, we consider men and women",
"Interestingly, the group who receives rental assistance",
"since data on non-binary and transgender individuals",
"through the existing practice appears to mostly be people",
"is not well-documented in ACDHS data. We calculate",
"without a history of homelessness: for the validation split"
] | {
"arxiv_id": "2403.12599",
"language": "en",
"page_idx": 4
} |