id
int64 1
213k
| paper_arxiv_id
stringlengths 10
13
| path
float64 | caption
stringlengths 9
127k
⌀ | label
stringlengths 7
125
⌀ | table_text
stringlengths 33
131k
⌀ |
|---|---|---|---|---|---|
212,397
|
2203.03539v2
| null |
\caption{Self-supervised learning approach performs reasonably on recovering the topic posterior mean. \textbf{Left:} TV distance between recovered topic posterior (self-supervised learning approach) and true topic posterior for different topic models. \textbf{Right:} Major topic(s) recovery accuracy for different topic models. The 95\% confidence interval is reported in both tables.}
|
\label{tab:tv_map}
|
\begin{tabular}{ccccc}
\hline
\begin{tabular}[c]{@{}c@{}}{\tiny Major Topic(s)} \\ {\tiny Recovery Accuracy}\end{tabular} & \multicolumn{4}{c}{Document Type} \\
$\alpha$ & Pure & LDA & CTM & PAM \\ \hline
1 & \begin{tabular}[c]{@{}c@{}}1.0000\\ ± 0.0000\end{tabular} & \begin{tabular}[c]{@{}c@{}}0.9050 \\ ± 0.0406\end{tabular} & \begin{tabular}[c]{@{}c@{}}0.9175 \\ ± 0.0275\end{tabular} & \begin{tabular}[c]{@{}c@{}}0.9025 \\ ± 0.0330\end{tabular} \\ \hline
3 & \begin{tabular}[c]{@{}c@{}}1.0000\\ ± 0.0000\end{tabular} & \begin{tabular}[c]{@{}c@{}}0.8750 \\ ± 0.0458\end{tabular} & \begin{tabular}[c]{@{}c@{}}0.8900 \\ ± 0.0311\end{tabular} & \begin{tabular}[c]{@{}c@{}}0.8950 \\ ± 0.0344\end{tabular} \\ \hline
5 & \begin{tabular}[c]{@{}c@{}}1.0000\\ ± 0.0000\end{tabular} & \begin{tabular}[c]{@{}c@{}}0.9000 \\ ± 0.0416\end{tabular} & \begin{tabular}[c]{@{}c@{}}0.8975 \\ ± 0.0296\end{tabular} & \begin{tabular}[c]{@{}c@{}}0.8675 \\ ± 0.0407\end{tabular} \\ \hline
7 & \begin{tabular}[c]{@{}c@{}}1.0000\\ ± 0.0000\end{tabular} & \begin{tabular}[c]{@{}c@{}}0.9150 \\ ± 0.0387\end{tabular} & \begin{tabular}[c]{@{}c@{}}0.8675 \\ ± 0.0383\end{tabular} & \begin{tabular}[c]{@{}c@{}}0.8675 \\ ± 0.0407\end{tabular} \\ \hline
9 & \begin{tabular}[c]{@{}c@{}}0.9950\\ ± 0.0098\end{tabular} & \begin{tabular}[c]{@{}c@{}}0.9100 \\ ± 0.0397\end{tabular} & \begin{tabular}[c]{@{}c@{}}0.8850 \\ ± 0.0330\end{tabular} & \begin{tabular}[c]{@{}c@{}}0.8325 \\ ± 0.0461\end{tabular} \\ \hline
\end{tabular}
|
212,398
|
2203.03539v2
| null |
\caption{Robustness of self-supervised learning approach. \textbf{Left:} TV distance between recovered topic posterior and true topic posterior of self-supervised learning approach versus posterior inference via Markov Chain Monte Carlo assuming a specific prior for $\alpha=1$. \textbf{Right:} Major topic recovery rate of our approach versus posterior inference via Markov Chain Monte Carlo assuming a specific prior for $\alpha=1$. In both table, the 95\% confidence interval is reported.}
|
\label{table: tv_ssl_vs_other}
|
\begin{tabular}{lccc}
\hline
Major Topic(s) Recovery & \multicolumn{3}{c}{Document Type ($\alpha=1$)} \\
Method & LDA & CTM & PAM \\ \hline
LDA & \textbf{\begin{tabular}[c]{@{}c@{}}0.9150 \\ ± 0.0387\end{tabular}} & \begin{tabular}[c]{@{}c@{}}0.8850 \\ ± 0.0344\end{tabular} & \begin{tabular}[c]{@{}c@{}}0.8500 \\ ± 0.0360\end{tabular} \\ \hline
CTM & \begin{tabular}[c]{@{}c@{}}0.9000 \\ ± 0.0416\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}0.9175 \\ ± 0.0275\end{tabular}} & \begin{tabular}[c]{@{}c@{}}0.8300 \\ ± 0.0370\end{tabular} \\ \hline
PAM & \begin{tabular}[c]{@{}c@{}}0.8750 \\ ± 0.0458\end{tabular} & \begin{tabular}[c]{@{}c@{}}0.7250 \\ ± 0.0397\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}0.9050 \\ ± 0.0320\end{tabular}} \\ \hline
\textbf{SSL (ours)} & \begin{tabular}[c]{@{}c@{}}0.9050 \\ ± 0.0406\end{tabular} & \textbf{\begin{tabular}[c]{@{}c@{}}0.9175 \\ ± 0.0275\end{tabular}} & \begin{tabular}[c]{@{}c@{}}0.9025 \\ ± 0.0330\end{tabular} \\ \hline
\end{tabular}
|
212,399
|
2203.03539v2
| null | null |
\label{lem:reconst_error}
| null |
212,400
|
2203.03539v2
| null | null |
\label{lem:contrast_repr}
| null |
212,401
|
2203.03539v2
| null |
\caption{TV between our recovered topic posterior and the true topic posterior of our self-supervised learning approach versus topic inference via Markov Chain Monte Carlo assuming a specific prior for $\alpha=1,3,5,7,9$. The 95\% confidence interval is reported.}
|
\label{table: ssl_vs_other_appendix_1}
|
\begin{tabular}{lcccc}
\hline
& \multicolumn{4}{c}{Document Type ($\alpha=9$)} \\
Method & Pure & LDA & CTM & PAM \\ \hline
LDA & 0.1051 ± 0.0065 & - & 0.2559 ± 0.0162 & 0.2281 ± 0.0152 \\ \hline
CTM & 0.3129 ± 0.0079 & 0.2580 ± 0.0092 & - & 0.3556 ± 0.0084 \\ \hline
PAM & 0.4805 ± 0.0071 & 0.3776 ± 0.0093 & 0.5156 ± 0.0088 & - \\ \hline
\textbf{SSL-6M (ours)} & \textbf{0.0517 ± 0.0045} & \textbf{0.1358 ± 0.0089} & \textbf{0.1101 ± 0.0064} & \textbf{0.0971 ± 0.0053} \\ \hline
\textbf{SSL-120K (ours)} & 0.1121 ± 0.0202 & 0.2221 ± 0.0123 & 0.1449 ± 0.0077 & 0.1460 ± 0.0070 \\ \hline
\end{tabular}
|
212,402
|
2203.03539v2
| null |
\caption{Major topic recovery rate of our approach versus posterior inference via Markov Chain Monte Carlo assuming a specific prior for $\alpha=1,3,5,7,9$. We report the 95\% confidence interval.}
|
\label{tab:ssl_vs_other_appendix_2}
|
\begin{tabular}{lcccc}
\hline
& \multicolumn{3}{c}{Document Type ($\alpha=9$)} \\
Method &Pure & LDA & CTM & PAM \\ \hline
LDA & \textbf{1.0 ± 0.0} & 0.8850 ± 0.0442 & 0.7750 ± 0.0384 & 0.7350 ± 0.0432 \\ \hline
CTM & \textbf{1.0 ± 0.0} & 0.8950 ± 0.0425 & 0.8800 ± 0.0348 & 0.5925 ± 0.0392 \\ \hline
PAM & \textbf{1.0 ± 0.0} & 0.8650 ± 0.0474 & 0.5675 ± 0.0358 & \textbf{0.8600 ± 0.0428} \\ \hline
\textbf{SSL-6M (ours)} & 0.9950 ± 0.0098 & \textbf{0.9100 ± 0.0397} & \textbf{0.8850 ± 0.0330} & 0.8325 ± 0.0461 \\ \hline
\textbf{SSL-120K (ours)} & 0.9800 ± 0.0194 & 0.8600 ± 0.0481 & 0.8800 ± 0.0341 & 0.8425 ± 0.0430 \\ \hline
\end{tabular}
|
212,403
|
2203.03539v2
| null |
\caption{Major topic recovery rate of our approach versus posterior inference via Variational Inference assuming a specific prior for $\alpha=1,3,5,7,9$. We report the 95\% confidence interval.}
|
\label{tab:ssl_vs_other_appendix_3}
|
\begin{tabular}{lcccc}
\hline
& \multicolumn{3}{c}{Document Type ($\alpha=9$)} \\
Method & Pure & LDA & CTM & PAM \\ \hline
LDA & \textbf{1.0 ± 0.0} & 0.9050 ± 0.0410 & 0.7250 ± 0.0396 & 0.6975 ± 0.0410 \\ \hline
CTM & \textbf{1.0 ± 0.0} & \textbf{0.9200 ± 0.0382} & 0.8725 ± 0.0339 & 0.5425 ± 0.0368 \\ \hline
PAM & \textbf{1.0 ± 0.0} & 0.8750 ± 0.0452 & 0.5300 ± 0.0325 & \textbf{0.8425 ± 0.0452} \\ \hline
\textbf{SSL-6M (ours)} & 0.9950 ± 0.0098 & 0.9100 ± 0.0397 & \textbf{0.8850 ± 0.0330} & 0.8325 ± 0.0461 \\ \hline
\textbf{SSL-120K (ours)} & 0.9800 ± 0.0194 & 0.8600 ± 0.0481 & 0.8800 ± 0.0341 & \textbf{0.8425 ± 0.0430} \\ \hline
\end{tabular}
|
212,404
|
2203.03539v2
| null |
\caption{Topic posterior recovery loss of our self-supervised learning approach, measured in Total Variation distance, for all four types of documents and $\alpha=1,3,5,7,9$. The number of training epochs ranges from 25 to 200, and we sample 60K new training documents in every 2 epochs, which corresponds to 720K to 6M training documents.}
|
\label{table:TV_less_epochs}
|
\begin{tabular}{lllllll}
\hline
& & \multicolumn{5}{c}{Dirichlet hyperparameter $\alpha$} \\
Document type & Number of epochs & \multicolumn{1}{c}{1} & \multicolumn{1}{c}{3} & \multicolumn{1}{c}{5} & \multicolumn{1}{c}{7} & \multicolumn{1}{c}{9} \\ \hline
Pure-topic & 25 & 0.0505 & 0.1025 & 0.1756 & 0.6222 & 0.8285 \\
& 50 & 0.0310 & 0.0744 & 0.0991 & 0.1522 & 0.2873 \\
& 100 & 0.0160 & 0.0363 & 0.0511 & 0.0911 & 0.1257 \\
& 200 & 0.0148 & 0.0308 & 0.0501 & 0.0391 & 0.0517 \\ \hline
LDA & 25 & 0.1236 & 0.1583 & 0.1812 & 0.2022 & 0.2218 \\
& 50 & 0.0967 & 0.1428 & 0.1387 & 0.1953 & 0.2117 \\
& 100 & 0.0805 & 0.1018 & 0.1101 & 0.1361 & 0.1538 \\
& 200 & 0.0709 & 0.0951 & 0.1045 & 0.1222 & 0.1377 \\ \hline
CTM & 25 & 0.1089 & 0.1599 & 0.1563 & 0.1747 & 0.1971 \\
& 50 & 0.0900 & 0.1132 & 0.1425 & 0.1455 & 0.1501 \\
& 100 & 0.0623 & 0.0948 & 0.1102 & 0.1187 & 0.1243 \\
& 200 & 0.0550 & 0.0799 & 0.0970 & 0.1071 & 0.1101 \\ \hline
PAM & 25 & 0.1072 & 0.1342 & 0.1480 & 0.1688 & 0.1754 \\
& 50 & 0.0820 & 0.1036 & 0.1246 & 0.1185 & 0.1395 \\
& 100 & 0.0543 & 0.0755 & 0.0859 & 0.1087 & 0.1095 \\
& 200 & 0.0436 & 0.0659 & 0.0787 & 0.0953 & 0.0971 \\ \hline
\end{tabular}
|
212,405
|
2203.03539v2
| null |
\caption{Fully-connected neural network's performance in the LDA $\alpha=1$ scenario.}
|
\label{tab:param_fc_lda}
|
\begin{tabular}{llll}
\hline
TV & \multicolumn{3}{c}{Hidden dimension} \\
\# layers (residual) & \multicolumn{1}{c}{1024} & \multicolumn{1}{c}{2048} & \multicolumn{1}{c}{4096} \\ \hline
3 & 0.1509 & 0.1112 & 0.1095 \\ \hline
6 & 0.7902 & 0.7893 & 0.7901 \\ \hline
3 (residual) & 0.0871 & 0.0867 & 0.0862 \\ \hline
6 (residual) & 0.0822 & 0.0835 & \textbf{0.0709} \\ \hline
\end{tabular}
|
212,406
|
2203.03539v2
| null |
\caption{Fully-connected neural network's performance in the CTM $\alpha=1$ scenario with 4096 hidden dimensions.}
|
\label{tab:param_fc_ctm}
|
\begin{tabular}{lll}
\hline
TV & \multicolumn{2}{c}{Layer Type} \\
\# layers & \multicolumn{1}{c}{regular} & \multicolumn{1}{c}{residual} \\ \hline
3 & 0.1824 & 0.1665 \\ \hline
4 & 0.1724 & 0.1656 \\ \hline
6 & 0.1700 & 0.1556 \\ \hline
\end{tabular}
|
212,407
|
2203.03539v2
| null |
\caption{Attention-based neural network's performance on CTM documents for $\alpha=1,3,5$.}
|
\label{tab:param_attn_ctm}
|
\begin{tabular}{lllll}
\hline
TV & & \multicolumn{3}{c}{\# layers} \\
$\alpha$ & Attention dimension & \multicolumn{1}{c}{4} & \multicolumn{1}{c}{6} & \multicolumn{1}{c}{8} \\ \hline
1 & 768 & 0.0902 & 0.0844 & \textbf{0.0814} \\
& 1024 & 0.0851 & 0.0890 & 0.0836 \\ \hline
3 & 768 & 0.1471 & 0.1429 & 0.1398 \\
& 1024 & 0.1440 & 0.1384 & \textbf{0.1383} \\ \hline
5 & 768 & 0.1794 & 0.1767 & \textbf{0.1708} \\
& 1024 & 0.1878 & 0.1787 & 0.1782 \\ \hline
\end{tabular}
|
212,408
|
2203.03539v2
| null |
\caption{4-layer attention-base neural network's performance on CTM documents when attention layer's dimension varies, for $\alpha=3,5$.}
|
\label{tab:param_attn_ctm_dim}
|
\begin{tabular}{llll}
\hline
TV & \multicolumn{3}{c}{Attention dimension} \\
$\alpha$ & \multicolumn{1}{c}{768} & \multicolumn{1}{c}{1024} & \multicolumn{1}{c}{2048} \\ \hline
3 & 0.1471 & \textbf{0.1440} & 0.1670 \\
5 & \textbf{0.1794} & 0.1878 & 0.1922 \\ \hline
\end{tabular}
|
212,409
|
2203.03539v2
| null |
\caption{The number of attention blocks used in the neural network architecture for $t=2$ for all four types of documents and $\alpha=1,3,5,7,9$. }
|
\label{tab:syn_t2_arc}
|
\begin{tabular}{lcccc}
\hline
& \multicolumn{3}{c}{Document Type} \\
$\alpha$ & Pure & LDA & CTM & PAM \\ \hline
1 & 8 & 8 & 6 & 6 \\ \hline
3 & 8 & 8 & 6 & 6 \\ \hline
5 & 8 & 8 & 6 & 6 \\ \hline
7 & 8 & 8 & 8 & 6 \\ \hline
9 & 8 & 8 & 8 & 10 \\ \hline
\end{tabular}
|
212,410
|
2203.03539v2
| null |
\caption{Topic posterior recovery loss of our self-supervised learning approach for $t=2$, measured in Total Variation distance, for all four types of documents and $\alpha=1,3,5,7,9$. We report the 95\% confidence interval.}
|
\label{tab:syn_t2_tv}
|
\begin{tabular}{lcccc}
\hline
TV Distance & \multicolumn{3}{c}{Document Type} \\
$\alpha$ & Pure & LDA & CTM & PAM \\ \hline
1 & 0.0081 ± 0.0002 & 0.0636 ± 0.0031 & 0.0748 ± 0.0053 & 0.0748 ± 0.0035 \\ \hline
3 & 0.0291 ± 0.0003 & 0.0870 ± 0.0039 & 0.1206 ± 0.0069 & 0.1045 ± 0.0047 \\ \hline
5 & 0.0431 ± 0.0004 & 0.1075 ± 0.0050 & 0.1523 ± 0.0067 & 0.1544 ± 0.0062 \\ \hline
7 & 0.0578 ± 0.00004 & 0.1258 ± 0.0049 & 0.1627 ± 0.0058 & 0.1693 ± 0.0065 \\ \hline
9 & 0.0662 ± 0.0006 & 0.1413 ± 0.0052 & 0.1859 ± 0.0061 & 0.2770 ± 0.0152 \\ \hline
\end{tabular}
|
212,411
|
2203.03539v2
| null |
\caption{The major topic pair recovery rate of our self-supervised learning approach for $t=2$ for all four types of documents and $\alpha=1,3,5,7,9$. We report the 95\% confidence interval.}
|
\label{tab:syn_t2_major_topics}
|
\begin{tabular}{lcccc}
\hline
& \multicolumn{3}{c}{Document Type} \\
$\alpha$ & Pure & LDA & CTM & PAM \\ \hline
1 & 1.0 ± 0.0 & 0.9350 ± 0.0342 & 0.9088 ± 0.0286 & 0.8638 ± 0.0402 \\ \hline
3 & 1.0 ± 0.0 & 0.8800 ± 0.0450 & 0.8700 ± 0.0363 & 0.8075 ± 0.0476 \\ \hline
5 & 1.0 ± 0.0 & 0.8950 ± 0.0425 & 0.8050 ± 0.0473 & 0.8012 ± 0.0512 \\ \hline
7 & 1.0 ± 0.0 & 0.9100 ± 0.0397 & 0.8450 ± 0.0428 & 0.8125 ± 0.0472 \\ \hline
9 & 1.0 ± 0.0 & 0.8950 ± 0.0425 & 0.7675 ± 0.0505 & 0.6625 ± 0.0485 \\ \hline
\end{tabular}
|
212,412
|
2203.03539v2
| null |
\caption{Test Accuracy for different embedding dimension used for fitting Skip-gram word embedding model on the AG News Dataset (Using N=10 for 95\% Confidence Interval). We fixed the window-size to be default of 5 and used1 all 4000 training examples. }
|
\label{tab:word2vec_emb}
|
\begin{tabular}{llll}
\hline
Embedding dimension & Test Accuracy\\
\hline
200 & \textbf{0.8498 ± 0.0007}\\
\hline
300 & 0.8488 ± 0.0007\\
\hline
500 & 0.8484 ± 0.0007\\
\hline
700 & 0.8483 ± 0.0006\\
\hline
1000 & 0.8488 ± 0.0009\\
\hline
2000 & 0.8473 ± 0.0006\\
\hline
3000 & 0.8481 ± 0.0001\\
\hline
5000 & 0.8482 ± 0.0006\\
\hline
\end{tabular}
|
212,413
|
2203.03539v2
| null |
\caption{Test Accuracy for different representation extraction method. We fixed the rest of hyperparameters to be: 5000 embedding dimension, 150 epochs, 0.0002 learning rate, sampling 4 words in labels, weight decay of 0.01 and resample rate of 2. }
|
\label{tab:rep_comp_data}
|
\begin{tabular}{llll}
\hline
Test Accuracy & \multicolumn{3}{c}{Method} \\
\# layers (residual) & \multicolumn{1}{c}{\emph{Word2vec+last layer}} & \multicolumn{1}{c}{\emph{softmax+second2last layer}}& \multicolumn{1}{c}{\emph{softmax+lastlayer}}\\ \hline
3 & 0.8508 & \textbf{0.8714 } & 0.8395 \\ \hline
4 & 0.8450 & \textbf{0.8621 } & 0.8430 \\ \hline
5 & 0.8492 & \textbf{0.8689 } & 0.8382 \\ \hline
\end{tabular}
|
212,414
|
2203.03539v2
| null |
\caption{Test Accuracy for different embedding dimension used for fitting Skip-gram word
embedding model on the IMDB and Small DBpedia Dataset (Using N=10 for 95\% Confidence Interval). We fixed the window-size to be default of 5 and use all available training examples. We use a dimension of 500 for both datasets for final baseline based on the results above. }
|
\label{tab:word2vec_tune_other}
|
\begin{tabular}{llll}
\hline
Test Accuracy & \multicolumn{2}{c}{Datasets} \\
Embedding Dimensions & \multicolumn{1}{c}{IMDB} & \multicolumn{1}{c}{Small DBpedia} \\ \hline
100 & 0.8423$\pm$0.0014 & 0.9440$\pm$ 0.0007 \\ \hline
200 & 0.8529$\pm$0.0012 & 0.9541$\pm$0.0004 \\ \hline
500 & \textbf{0.8545$\pm$0.0011} & \textbf{0.9568$\pm$0.0004} \\ \hline
1000 &0.8544$\pm$ 0.0005 & 0.9567$\pm$ 0.0003 \\ \hline
\end{tabular}
|
212,415
|
2209.15283v1
| null |
\caption{Caption}
|
\label{tab:my_label}
|
\begin{tabular}{+c^c^c^c}
\toprule
& \makecell{Technique used for\\feature extraction} & \makecell{Feature extraction\\kept during MLP\\training \& interference?} & \makecell{Feature extraction\\trained together with\\MLP using gradient descent?} & Feature extraction sparse?\\\midrule
\textbf{TabNN} & GBDT & no & no\\
\textbf{SAINT} & Transformers & yes & yes\\
\textbf{Our approach} & RF / GBDT / DF & yes & yes\\\bottomrule
\end{tabular}
|
212,416
|
2209.15283v1
| null |
\caption{\small{Best scores for Protocol 2. For each data set, predictors performing at least as well as the best (resp.\ DL) score up to its standard deviation are highlighted in \textbf{bold} (resp.\ \underline{underlined}). The scores are based on 5 times repeated (stratified) 5-fold cross validation. For each model, HP have been chosen via the ``{optuna}'' library with 100 iterations. *score based on a simple 5-fold cross validation. See Appendix \ref{app:add_final_performances_p2} for a comparison with literature results.}}
|
\label{tab:p2_final_performance}
|
\begin{tabular}{+c|^c^c^c^c^c^c^c^c^c^c^c}\toprule
\rowstyle{\bfseries}
\multirow{1}{*}{\backslashbox{Model}{Data set}}&\multicolumn{1}{c}{{Housing}} & \multicolumn{1}{c}{{Airbnb}} & \multicolumn{1}{c}{{Diamonds}} & \multicolumn{1}{c}{{Adult}} & \multicolumn{1}{c}{{Bank}} & \multicolumn{1}{c}{{Blastchar}} &
\multicolumn{1}{c}{{Heloc}} &
\multicolumn{1}{c}{{Higgs}} & \multicolumn{1}{c}{{Covertype}} & \multicolumn{1}{c}{{Volkert}}\\[6pt]
& \multicolumn{1}{c}{\scriptsize \shortstack[c]{MSE $\downarrow$\\ \ }} & \multicolumn{1}{c}{\scriptsize \shortstack[c]{MSE $\downarrow$\\{(x10\textsuperscript{3})}}} &
\multicolumn{1}{c}{\scriptsize \shortstack[c]{MSE $\downarrow$\\{(x10\textsuperscript{-3})}}} &
\multicolumn{1}{c}{\scriptsize \shortstack{AUC $\uparrow$\\ \scriptsize{(in \%)}}} &
\multicolumn{1}{c}{\scriptsize \shortstack{AUC $\uparrow$\\ \scriptsize{(in \%)}}} &
\multicolumn{1}{c}{\scriptsize \shortstack{AUC $\uparrow$\\(in \%)}} & \multicolumn{1}{c}{\scriptsize \shortstack{AUC $\uparrow$\\ \scriptsize{(in \%)}}} &
\multicolumn{1}{c}{\scriptsize \shortstack{AUC $\uparrow$\\ \scriptsize{(in \%)}}} &
\multicolumn{1}{c}{\scriptsize \shortstack{Acc.\ $\uparrow$\\(in \%)}} & \multicolumn{1}{c}{\scriptsize \shortstack{Acc.\ $\uparrow$\\(in \%)}}\\\midrule
Random Forest & 0.263\scriptsize{\pm0.009}& 5.39\scriptsize{\pm0.13}& 9.80\scriptsize{\pm0.35}& 91.6\scriptsize{\pm0.3} & 92.8\scriptsize{\pm0.3} & \textbf{84.5\scriptsize{\pm1.2}} & 91.3\scriptsize{\pm0.6} & 80.4\scriptsize{\pm0.1} & 83.6\scriptsize{\pm0.1} & 64.2\scriptsize{\pm0.3}\\
GBDT & \textbf{0.208\scriptsize{\pm0.010}}& \textbf{4.71\scriptsize{\pm0.15}}& \textbf{7.38\scriptsize{\pm0.28}} & \textbf{92.7\scriptsize{\pm0.3}} & \textbf{93.3\scriptsize{\pm0.3}} & \textbf{84.7\scriptsize{\pm1.0}}& \textbf{92.1\scriptsize{\pm0.4}} & 82.8\scriptsize{\pm0.1} & \textbf{97.0\scriptsize{\pm0.0}} & 71.3\scriptsize{\pm0.4}\\
Deep Forest & 0.225\scriptsize{\pm0.008} & \textbf{4.68\scriptsize{\pm}0.16} & 8.23\scriptsize{\pm0.29} & 91.8\scriptsize{\pm0.3} & 92.9\scriptsize{\pm0.2}& \textbf{83.7\scriptsize{\pm1.2}} & 90.3\scriptsize{\pm0.5} & 81.2\scriptsize{\pm0.0}* & 92.4\scriptsize{\pm0.1}*& 66.3\scriptsize{\pm0.4}\\\midrule
MLP rand. init. & 0.258\scriptsize{\pm0.011}& 5.07\scriptsize{\pm}0.16 & 15.5\scriptsize{\pm12.5}& 90.5\scriptsize{\pm0.4} & 91.0\scriptsize{\pm0.3} & 81.4\scriptsize{\pm1.2} & 80.1\scriptsize{\pm0.1} & 83.2\scriptsize{\pm0.3} & \underline{96.7\scriptsize{\pm0.0}}& 72.2\scriptsize{\pm0.4}\\
MLP RF init.& 0.222\scriptsize{\pm0.009} & \textbf{\underline{4.66\scriptsize{\pm}0.16}} & \underline{7.93\scriptsize{\pm0.22}}& \underline{92.1\scriptsize{\pm0.3}} & \underline{92.4\scriptsize{\pm0.4}} & \underline{\textbf{84.4\scriptsize{\pm1.2}}} & \underline{\textbf{91.7\scriptsize{\pm0.4}}} & \underline{\textbf{83.6\scriptsize{\pm0.1}}}& \underline{96.7\scriptsize{\pm0.0}}& \underline{\textbf{74.1\scriptsize{\pm0.4}}}\\
MLP GBDT init.& \underline{\textbf{0.206\scriptsize{\pm}0.007}} & \underline{\textbf{4.70\scriptsize{\pm0.09}}} & 8.15\scriptsize{\pm0.35} & \underline{92.2\scriptsize{\pm0.3}} & \underline{92.5\scriptsize{\pm0.3}} & \textbf{\underline{84.6\scriptsize{\pm1.2}}} & 91.5\scriptsize{\pm0.6} & 83.0\scriptsize{\pm0.0} & 96.2\scriptsize{\pm0.0}& 73.5\scriptsize{\pm0.5}\\
MLP DF init.& 0.234\scriptsize{\pm0.016} & \underline{\textbf{4.81\scriptsize{\pm0.13}}}& 8.28\scriptsize{\pm0.24}& 91.9\scriptsize{\pm0.4}& 92.2\scriptsize{\pm0.3}& \underline{\textbf{84.2\scriptsize{\pm1.0}}} & 91.4\scriptsize{\pm0.6} & 83.3\scriptsize{\pm0.1}* & 94.5\scriptsize{\pm0.3}*& 71.3\scriptsize{\pm0.5}\\
SAINT & 0.258\scriptsize{\pm0.011}& \underline{\textbf{4.81\scriptsize{\pm0.15}}}& 17.7\scriptsize{\pm3.83}& 91.6\scriptsize{\pm0.3}& 92.2\scriptsize{\pm0.4}& \underline{\textbf{84.0\scriptsize{\pm0.8}}} & 90.2\scriptsize{\pm0.7} & \underline{\textbf{83.7\scriptsize{\pm0.1}}}* & 96.6\scriptsize{\pm0.1}*& 70.1\scriptsize{\pm0.4}\\\bottomrule
\end{tabular}
|
212,417
|
2209.15283v1
| null |
\caption{Scores from Table \ref{tab:p2_final_performance} with some detail on the MLP architecture used to achieve each score (width/depth). The mean squared error, the AUROC score (in \%) and the accuracy (in \%) are reported for regression, binary classification and multi-class classification problems respectively.}
|
\label{tab:p2_final_performance_MLPDims}
|
\begin{tabular}{+c|^c^c^c^c^c^c^c^c^c^c^c^c^c^c}\toprule
\rowstyle{\bfseries}
\backslashbox{Model}{Data set}&\multicolumn{3}{c}{\textbf{Housing $\downarrow$}} & \multicolumn{3}{c}{\textbf{Adult $\uparrow$}} & \multicolumn{2}{c}{\textbf{Bank $\uparrow$}} & \multicolumn{2}{c}{\textbf{Covertype $\uparrow$}} \\[6pt]
&\small\shortstack[c]{our\\results} &\small\shortstack[c]{MLP\\width}& \small\shortstack[c]{MLP\\depth} &\small\shortstack[c]{our\\results} &\small\shortstack[c]{MLP\\width}& \small\shortstack[c]{MLP\\depth}\\\midrule
MLP rand. init. & 0.254\pm0.010 & 1109 & 4 & 90.6\pm0.4 & 1175 & 4 & 91.5\pm0.2 & 91.1\pm0.3 & 91.0\pm0.4\\
MLP RF init.& 0.221\pm0.009 & 2048 & 10 & 91.9\pm0.3 & 2048 & 5 & - & 92.4\pm0.4 & -\\
MLP XGB init.& 0.205\pm0.007 & 2048 & 4 & {92.2\pm0.3} & 2048 & 6 & - & 92.6\pm0.3 & -\\
MLP DF init.& 0.234\pm0.016 & 2048 & 23 & 91.6\pm0.4 & 2048 & 9 & - & & -\\\bottomrule
\end{tabular}
|
212,418
|
2209.15283v1
| null |
\caption{\small Links to data sets.}
|
\label{tab:data_set_links}
|
\begin{tabular}{c|c}\toprule
Data set& Link \\\midrule
Housing&\href{https://inria.github.io/scikit-learn-mooc/python_scripts/datasets_california_housing.html}{Scikit-learn}\\
Airbnb&\href{http://insideairbnb.com/get-the-data/}{Inside Airbnb}\\
Diamond& \href{https://www.openml.org/search?type=data&sort=runs&id=44059}{OpenML}\\
Adult&\href{https://archive.ics.uci.edu/ml/datasets/adult}{UCI Machine Learning Repository}\\
Bank&\href{https://archive.ics.uci.edu/ml/datasets/bank+marketing}{UCI Machine Learning Repository}\\
Blastchar & \href{https://www.kaggle.com/datasets/blastchar/telco-customer-churn}{Kaggle}\\
Heloc & \href{https://community.fico.com/s/explainable-machine-learning-challenge}{FICO}\\
Higgs & \href{https://archive.ics.uci.edu/ml/datasets/HIGGS}{UCI Machine Learning Repository}\\
Covertype&\href{https://archive.ics.uci.edu/ml/datasets/covertype}{UCI Machine Learning Repository}\\
Volkert&\href{https://automl.chalearn.org/data}{AutoML}\\\bottomrule
\end{tabular}
|
212,419
|
2209.15283v1
| null |
\caption{\small Main properties of the data sets.}
|
\label{tab:dataset_description}
|
\begin{tabular}{cccccccccccc} \toprule
&Housing & Airbnb & Diamonds & Adult & Bank & Blastchar & Heloc & Higgs & Covertype & Volkert\\\midrule
Dataset size &
20\,640 &
119\,268 &
53\,940&
32\,561 &
45\,211 &
7\,043&
9\,871&
550\,000&
581\,012 &
58\,310\\
\# Num. features &
8 &
10 &
6&
6 &
7 &
3&
21&
27&
44 &
147\\
\# Cat. features &
0 &
3 &
3&
8 &
9 &
17&
2&
1&
10 &
0\\
Task &
Regr. &
Regr. &
Regr.&
Classif. &
Classif. &
Classif.&
Classif.&
Classif.&
Classif. &
Classif.\\
\# Classes &
- &
- &
- &
2 &
2 &
2&
2&
2&
7 &
10\\\bottomrule
\end{tabular}
|
212,420
|
2209.15283v1
| null |
\caption{\footnotesize{Best scores for Protocol P2. For each data set, our best overall score is highlighted in \textbf{bold} and our best Deep Learning score is \underline{underlined}. Our scores are based on 5 times repeated (stratified) 5-fold cross validation. For each of our models, HP were selected via the \texttt{optuna} library (100 iterations). Sources for literature values: \cite{borisov2021deep} ($\dagger$) and \cite{somepalli2021saint} ($\mathsection$). *score based on a single 5-fold cross validation.}}
|
\label{tab:p2_final_performance_appendix}
|
\begin{tabular}{+c|^c^c|^c^c|^c^c|^c^c|^c^c}\toprule
\multirow{4}{*}{\backslashbox{\textbf{Model}}{\textbf{Data set}}} & \multicolumn{2}{c}{\textbf{Adult ($\dagger$)}} & \multicolumn{2}{c}{\textbf{Bank ($\mathsection$)}} & \multicolumn{2}{c}{\textbf{Blastchar ($\mathsection$)}} & \multicolumn{2}{c}{\textbf{Heloc ($\dagger$)}} & \multicolumn{2}{c}{\textbf{Higgs ($\dagger$)}}\\[6pt]
& \multicolumn{2}{c}{\small \underline{AUC $\uparrow$ in \%}} & \multicolumn{2}{c}{\small \underline{AUC $\uparrow$ in \%}} & \multicolumn{2}{c}{\small \underline{AUC $\uparrow$ in \%}} & \multicolumn{2}{c}{\small \underline{AUC $\uparrow$ in \%}} & \multicolumn{2}{c}{\small \underline{AUC $\uparrow$ in \%}} \\[4pt]
&\small\shortstack[c]{perf. in\\literature} & \small\shortstack[t]{our\\results}&\small\shortstack[c]{perf. in\\literature} & \small\shortstack[t]{our\\results}&\small\shortstack[c]{perf. in\\literature} & \small\shortstack[t]{our\\results} & \small\shortstack[c]{perf. in\\literature} & \small\shortstack[t]{our\\results} & \small\shortstack[c]{perf. in\\literature} & \small\shortstack[t]{our\\results}\\\midrule
Random Forest & 91.7\pm0.2 & 91.6\pm0.3 & 89.1\pm0.3 & 92.8\pm0.3 & 80.6\pm0.7 & \textbf{84.5\pm1.2} & 90.0\pm0.2 & 91.3\pm0.6 & 79.7\pm0.0 & 80.4\pm0.1\\
GBDT & {92.8\pm0.1} & \textbf{92.7\pm0.3} & 93.0\pm0.2 & \textbf{93.3\pm0.3} & 81.8\pm0.3 & \textbf{84.7\pm1.0}&92.2\pm0.0 & \textbf{92.1\pm0.4} & 85.9\pm0.0 & 82.8\pm0.1\\
Deep Forest & - & 91.8\pm0.3 & - & 92.9\pm0.2 & - & \textbf{83.7\pm1.2} & - & 90.3\pm0.5 & - & 81.2\pm0.0*\\
\midrule
MLP rand.\ init.\ & 90.3\pm0.2 & 90.5\pm0.4 & 91.5\pm0.2 & 91.0\pm0.3 & 59.6\pm0.3 & 81.4\pm1.2 & 80.3\pm0.1 & 80.1\pm0.1 & 85.6\pm0.0 & 83.2\pm0.3\\
MLP RF init\ & - & \underline{92.1\pm0.3} & - & \underline{92.4\pm0.4} & - & \underline{\textbf{84.4\pm1.2}} & - & \underline{\textbf{91.7\pm0.4}} & - & \underline{\textbf{83.6\pm0.1}}\\
MLP GBDT init.\ & - & \underline{92.2\pm0.3} & - & \underline{92.5\pm0.3} & - & \underline{\textbf{84.6\pm1.2}} & - & 91.5\pm0.6 & - & 83.0\pm0.0\\
MLP DF init.\ & - & 91.9\pm0.4 & - & 92.2\pm0.3 & - & \underline{\textbf{84.2\pm1.0}} & - & 91.4\pm0.6 & - & 83.3\pm0.1*\\
SAINT & 91.6\pm0.4 & 91.6\pm0.3& 93.3\pm0.1 & 92.2\pm0.4 & 84.7\pm0.3 & \underline{\textbf{84.0\pm0.8}} & 90.7\pm0.2 & 90.2\pm0.7 & 88.3\pm0.0 & \underline{\textbf{83.7\pm0.1*}}\\\bottomrule
\end{tabular}
|
212,421
|
2209.15283v1
| null |
\caption{\small Comparison of number of parameters for each model.}
|
\label{tab:parameters}
|
\begin{tabular}{c|cccccccccc}
\toprule
\backslashbox{\textbf{Model}}{\textbf{Data set}}& Housing & Airbnb & Diamonds & Adult & Bank & Blastchar & Heloc & Higgs & Covertype & Volkert \\\midrule
MLP rand. init. &
2.47M &
1.86M &
363k&
1.09M &
52.4K &
13.1M&
11.3M&
11.6M&
1.14M &
9.03M\\
MLP RF init. &
33.6M &
12.6M &
8.42M&
29.4M &
8.43M &
25.2M&
16.8M&
4.26M&
21.1M &
17.1M\\
MLP GBDT init. &
8.41M &
12.6M &
12.6M&
33.6M &
8.43M &
16.8M&
25.2M&
8.46M&
4.32M &
21.3M\\
MLP DF init.&
88.1M &
34.0M &
59.3M&
42.0M &
46.2M &
34.36M&
25.8M&
43.2M&
57.6M &
34.1M\\
SAINT &
56.8M &
27.0M &
53.1M&
7.20M &
6.12M &
322M&
98.2M&
43.2M&
6.44M &
169M \\\bottomrule
\end{tabular}
|
212,422
|
2209.15283v1
| null |
\caption{\small{Comparison of the execution time in seconds for model training until the best validation lost is reached. The number of training epochs is indicated in parentheses.}}
|
\label{tab:p2_execution_times}
|
\begin{tabular}{c|cccccccccccc}
\toprule
\backslashbox{\textbf{Model}}{\textbf{Data set}}& Housing & Airbnb & Diamonds & Adult & Bank & Blastchar & Heloc & Higgs & Covertype & Volkert \\\midrule
MLP rand. init. &
5.96 \scriptsize{(32)}&
91.9 \scriptsize{(98)} &
13.3 \scriptsize{(31)}&
11.8 \scriptsize{(37)}&
21.6 \scriptsize{(62)}&
6.78 \scriptsize{(34)}&
14.3 \scriptsize{(60)}&
467 \scriptsize{(32)}&
312 \scriptsize{(69)}&
12.3 \scriptsize{(31)}\\
MLP RF init. &
35.6 \scriptsize{(29)} &
129 \scriptsize{(44)} &
24.6 \scriptsize{(25)} &
16.0 \scriptsize{(19)}&
19.6 \scriptsize{(23)}&
5.77 \scriptsize{(18)}&
4.83 \scriptsize{(15)}&
247 \scriptsize{(39)}&
2040 \scriptsize{(91)}&
24.6 \scriptsize{(25)}\\
MLP GBDT init. &
17.5 \scriptsize{(49)}&
276 \scriptsize{(95)}&
48.9 \scriptsize{(37)}&
32.2 \scriptsize{(31)}&
3.20 \scriptsize{(3)}&
1.69 \scriptsize{(7)}&
3.89 \scriptsize{(8)}&
58.8 \scriptsize{(5)}&
435 \scriptsize{(66)}&
48.9 \scriptsize{(37)}\\
MLP DF init.&
218 \scriptsize{(72)} &
355 \scriptsize{(48)}&
175 \scriptsize{(31)} &
91 \scriptsize{(54)}&
96 \scriptsize{(26)}&
22.3 \scriptsize{(52)}&
9.04 \scriptsize{(19)}&
3260 \scriptsize{(76)}&
5570 \scriptsize{(95)}&
175 \scriptsize{(31)}\\
SAINT &
81.9 \scriptsize{(37)} &
640 \scriptsize{(83)}&
394 \scriptsize{(84)}&
15.6 \scriptsize{(11)}&
52.7 \scriptsize{(32)}&
7.23 \scriptsize{(2)}&
51.0 \scriptsize{(31)}&
2310 \scriptsize{(19)}&
6580 \scriptsize{(97)}&
394 \scriptsize{(84)}\\
\bottomrule
\end{tabular}
|
212,423
|
2209.15283v1
| null |
\caption{\small{Comparison of the execution time in milliseconds of one model forward pass on the whole data set.}}
|
\label{tab:p2_execution_times_test}
|
\begin{tabular}{c|cccccccccc}
\toprule
\backslashbox{\textbf{Model}}{\textbf{Data set}}& Housing & Airbnb & Diamonds & Adult & Bank & Blastchar & Heloc & Higgs & Covertype & Volkert \\\midrule
MLP rand. init. & 13.4 & 103 & 12.2 & 98.7 & 141 & 10.4 & 36.9 & 1060 & 1070 & 160\\
MLP RF init. & 45.0 & 97.7 & 42.8 & 125 & 180 & 14.5 & 12.6 & 1020 & 1830 & 191\\
MLP GBDT init. & 19.6 & 167 & 22.1 & 147 & 182 & 11.7 & 12.5 & 378 & 1170 & 218\\
MLP DF init.& 102 & 227 & 105 & 200 & 415 & 28.5 & 13.4 & 4680 & 4060 & 330\\
SAINT & 129 & 508 & 130 & 304 & 429 & 77.9 & 114 & 2870 & 3670 & 645\\\bottomrule
\end{tabular}
|
212,429
|
2203.01629v5
| null |
\caption{We report the runtimes for the three methods used in the weakly-supervised experiment, i. e. labelVAE, adaptiveVAE and hypergeometricVAE.
We report the runtime as mean and standard deviation over all runs per experiment.
We report the runtimes for the different number of independent factors $k = \{-1, 2, 4, 6 \}$.}
|
\label{tab:exp_ws_runtimes}
|
\begin{tabular}{lcccc}
\toprule
\multirow{2}{*}{Method} & \multicolumn{4}{c}{runtime [s]} \\
& $k=-1$ & $k=2$ & $k=4$ & $k=6$ \\
\midrule
label & $21907.4 \pm 273.0$ & $20921.1 \pm 705.9$ & $21389.3 \pm 163.8$ & $21761.1 \pm 567.9$ \\
adaptive & $29071.5 \pm 133.2$ & $28609.8 \pm 439.9$ & $29479.1 \pm 487.1$ & $29966.3 \pm 303.3$ \\
hypergeometric & $21888.9 \pm 632.9$ & $21299.4 \pm 293.6$ & $21863.9 \pm 190.1$ & $22241.0 \pm 137.8$ \\
\bottomrule
\end{tabular}
|
212,424
|
2209.15283v1
| null |
\caption{Hyper-parameter search spaces used for numerical evaluations.}
|
\label{tab:search_spaces}
|
\begin{tabular}{cccc}
\toprule
\textbf{Method}&\textbf{Parameter}& \textbf{Search space} & \textbf{Function}\\\midrule
\multirow{3}{*}{Random Forests}
& max\_depth & $\{1,\dots,12\}$ & \multirow{3}{*}{see \href{https://scikit-learn.org/stable/modules/generated/sklearn.ensemble.RandomForestRegressor.html}{here}}\\
& n\_estimators & $\{1000\}$\\
& max\_features & $[0,1]$\\\midrule
\multirow{5}{*}{{GBDT}}
& max\_depth & \{1,\dots,12\} & \multirow{5}{*}{see \href{https://xgboost.readthedocs.io/en/stable/parameter.html}{here}}\\
& n\_estimators & $\{1000\}$\\
& reg\_alpha & $[10^{-8}, 1]$\\
& reg\_lambda & $[10^{-8}, 1]$\\
& learning\_rate & $[0.01, 0.3]$\\\midrule
\multirow{5}{*}{Deep Forest}&
forest\_depth & $\{1,2,3\}$ & Number of Deep Forest layers\\
& n\_forests & $\{1\}$ & Number of forests per Deep Forest layer\\
& n\_estimators & $\{1000\}$ & \multirow{3}{*}{RF parameters, see \href{https://scikit-learn.org/stable/modules/generated/sklearn.ensemble.RandomForestRegressor.html}{here}}\\%$2048/2^{\text{max\_depth}}$
& max\_depth & $\{1,\dots,12\}$ \\
& max\_features & $[0,1]$\\\midrule
\multirow{5}{*}{MLP random init.}
& learning\_rate & $[10^{-6},10^{-1}]$ & learning rate of SGD training\\
& depth & $\{1,\dots,10\}$ & number of layer\\
& width & $\{1, \dots, 2048\}$ & number of neurons per layer\\
& epochs & $\{100\}$ & number of SGD training epochs\\
& batch\_size & $\{256\}$ & batch size of SGD training\\\midrule
\multirow{10}{*}{MLP RF init.}
& max\_depth & $\{1,\dots,11\}$ & \multirow{3}{*}{Parameters of the RF initializer, see \href{https://scikit-learn.org/stable/modules/generated/sklearn.ensemble.RandomForestRegressor.html}{here}}\\
& n\_estimators & $2048/2^{\text{max\_depth}}$\\
& max\_features & $[0,1]$\\
& learning\_rate & $[10^{-6},10^{-1}]$ & learning rate of SGD training\\
& depth & $\{3,\dots,10\}$ & number of layer\\
& width & $\{2048\}$ & number of neurons per layer\\
& epochs & $\{100\}$ & number of SGD training epochs\\
& batch\_size & $\{256\}$ & batch size of SGD training\\
& strength01 & $[1,10^4]$ & \multirow{2}{*}{MLP translation parameters, see Section \ref{sec:diff_translation}}\\
& strength12 & $[0.01, 100]$\\\midrule
\multirow{12}{*}{MLP GBDT init.}
& max\_depth & \{1,\dots,11\} & \multirow{5}{*}{\shortstack[c]{Parameters of the GBDT initializer, see \href{https://xgboost.readthedocs.io/en/stable/parameter.html}{here}}}\\
& n\_estimators & $2048/(\text{n\_classes}\cdot2^{\text{max\_depth}})$ & \\
& reg\_alpha & $[10^{-8}, 1]$\\
& reg\_lambda & $[10^{-8}, 1]$\\
& learning\_rate\_GBDT & $[0.01, 0.3]$\\
& learning\_rate & $[10^{-6},10^{-1}]$ & learning rate of SGD training\\
& depth & $\{3,\dots,10\}$ & number of layer\\
& width & $\{2048\}$ & number of neurons per layer\\
& epochs & $\{100\}$ & number of SGD training epochs\\
& batch\_size & $\{256\}$ & batch size of SGD training\\
& strength01 & $[1,10^4]$ & \multirow{2}{*}{MLP translation parameters, see Section \ref{sec:diff_translation}}\\
& strength12 & $[0.01, 100]$\\\midrule
\multirow{14}{*}{MLP DF init.}
&forest\_depth & $\{1,2,3\}$ & Number of Deep Forest layers\\
& n\_forests & $\{1\}$ & Number of forests per Deep Forest layer\\
& n\_estimators & $2048/2^{\text{max\_depth}}$ & \multirow{3}{*}{RF parameters, see \href{https://scikit-learn.org/stable/modules/generated/sklearn.ensemble.RandomForestRegressor.html}{here}}\\%$2048/2^{\text{max\_depth}}$
& max\_depth & $\{1,\dots,12\}$\\
& max\_features & $[0,1]$\\
& learning\_rate & $[10^{-6},10^{-1}]$ & learning rate of SGD training\\
& depth & $\{3,\dots,10\}$ & number of layer\\
& width & $\{2048\}$ & number of neurons per layer\\
& epochs & $\{100\}$ & number of SGD training epochs\\
& batch\_size & $\{256\}$ & batch size of SGD training\\
& strength01 & $[1,10^4]$ & \multirow{4}{*}{MLP translation parameters, see Section \ref{sec:diff_translation}}\\
& strength12 & $[0.01, 100]$\\
& strength23 & $[0.01, 100]$\\
& strength\_id & $[0.01, 100]$\\
\midrule
\multirow{6}{*}{SAINT}
& epochs & $\{100\}$ & number of SGD training epochs\\
& batch\_size & $\{256\}$/$\{64\}$ & batch size of SGD training\\
& dim & $[32, 64, 128]$/$[8, 16]$ & number of neurons per layer in attention block\\
& depth & $\{1,2,3\}$ & number of layers in each attention block\\
& heads & $\{2,4,8\}$ & number of head in each attention layer\\
& dropout & $\{0,0.1,0.2,0.3,0.4,0.5,0.6,0.7,0.8\}$ & dropout used during SGD training\\\bottomrule
\end{tabular}
|
212,425
|
2209.15283v1
| null |
\caption{Hyper-parameters used for the experimental protocol P2.}
|
\label{tab:HP_ep2}
|
\begin{tabular}{cccccccc}
\toprule
\textbf{Method}&\textbf{Parameter}& \textbf{Housing} & \textbf{Airbnb} & \textbf{Adult} & \textbf{Bank} & \textbf{Covertype} & \textbf{Volkert}\\\midrule
\multirow{3}{*}{Random Forests}
& max\_depth & 12 & 12 & 11 & 12 & 12 & 12\\
& n\_estimators & 1000 & 1000 & 1000 & 1000 & 1000 & 1000\\
& max\_features & 0.437 & 0.623 & 0.596 & 0.943 & 0.811 & 0.688\\\midrule
\multirow{5}{*}{{GBDT}}
& max\_depth & 12 & 9 & 6 & 7 & 11 & 10\\
& n\_estimators & 1000 & 1000 & 1000 & 1000& 1000 & 1000\\
& reg\_alpha & 0.305 & 4.60$\times 10^{-6}$ & 2.39$\times 10^{-5}$ & 1.52$\times 10^{-4}$ & 0.728 & 4.47$\times 10^{-6}$\\
& reg\_lambda & 1.13$\times 10^{-2}$ & 1.75$\times 10^{-8}$ & 1.35$\times 10^{-6}$ & 1.07$\times 10^{-3}$ & 6.51$\times 10^{-4}$ & 1.71$\times 10^{-6}$\\
& learning\_rate & 3.82$\times 10^{-2}$ & 0.238 & 1.08$\times 10^{-2}$ & 1.34$\times 10^{-2}$ & 0.181 & 0.107\\\midrule
\multirow{5}{*}{Deep Forest}&
forest\_depth & 4 & 9 & 2 & 2 & 9 & 3\\
& n\_forests & 1 & 1 & 1 & 1 & 1 & 1\\
& n\_estimators & 1000 & 1000 & 1000 & 1000 & 1000 & 1000\\
& max\_depth & 5 & 12 & 11 & 9 & 12 & 12\\
& max\_features & 0.361 & 0.410 & 0.166 & 0.206 & 0.218 & 0.134\\\midrule
\multirow{5}{*}{MLP random init.}
& learning\_rate & 9.01$\times 10^{-4}$ & 4.21$\times 10^{-4}$ & 2.07$\times 10^{-4}$ & 1.1$\times 10^{-4}$ & 1.15$\times 10^{-4}$ & 2.29$\times 10^{-4}$\\
& depth & 4 & 4 & 4 & 4 & 4 & 6\\
& width & 1100 & 959 & 1175 & 856 & 738 & 1482\\
& epochs & 100 & 100 & 100 & 100 & 100 & 100\\
& batch\_size & 256 & 256 & 256 & 256 & 256 & 256\\\midrule
\multirow{10}{*}{MLP RF init.}
& max\_depth & 8 & 10 & 8 & 8 & 10 & 8\\
& n\_estimators & 8 & 2 & 8 & 8 & 2 & 8\\
& max\_features & 0.442 & 0.321 & 0.613 & 0.650 & 0.897 & 0.825\\
& learning\_rate & 1.04$\times 10^{-4}$ & 1.72$\times 10^{-4}$ & 1.55$\times 10^{-5}$ & 1.01$\times 10^{-4}$ & 1.04$\times 10^{-5}$ & 1.45$\times 10^{-4}$\\
& depth & 10 & 5 & 5 & 4 & 7 & 6\\
& width & 2048 & 2048 & 2048 & 2048 & 2048& 2048\\
& epochs & 100 & 100 & 100 & 100 & 100 & 100\\
& batch\_size & 256 & 256 & 256 & 256 & 256 & 256\\
& strength01 & 1090 & 668 & 537 & 71.4 & 13.7 & 1.02\\
& strength12 & 0.0749 & 1.09 & 62.7 & 34.5 & 1.05$\times 10^{-2}$ & 5.53$\times 10^{-2}$\\\midrule
\multirow{12}{*}{MLP GBDT init.}
& max\_depth & 3 & 4 & 4 & 4 & 8 & 4\\
& n\_estimators & 256 & 128 & 128 & 128 & 1 & 12\\
& reg\_alpha & 1.30$\times 10^{-7}$ & 1.10$\times 10^{-2}$ & 1.26$\times 10^{-8}$ & 0.413 & 1.33$\times 10^{-2}$ & 6.76$\times 10^{-6}$\\
& reg\_lambda & 1.57$\times 10^{-7}$ & 9.52$\times 10^{-4}$ & 7.85$\times 10^{-4}$ & 7.48$\times 10^{-3}$ & 0.643 & 1.99 $\times 10^{-7}$\\
& learning\_rate\_GBDT & 0.211 & 0.297 & 0.202 & 0.285 & 0.112 & 0.272\\
& learning\_rate & 1.11$\times 10^{-5}$ & 1.97$\times 10^{-5}$ & 4.77$\times 10^{-5}$ & 6.22$\times 10^{-4}$ & 6.19$\times 10^{-5}$ & 1.63$\times 10^{-4}$\\
& depth & 4 & 5 & 6 & 4 & 3 & 7\\
& width & 2048 & 2048 & 2048& 2048 & 2048 & 2048\\
& epochs & 100 & 100 & 100 & 100 & 100& 100\\
& batch\_size & 256 & 256 & 256 & 256 & 256 & 256\\
& strength01 & 575 & 7830 & 132 & 20.5 & 7280 & 4.08\\
& strength12 & 5.60 & 0.461 & 66.0 & 5.52 & 93.4 & 7.11$\times 10^{-2}$\\\midrule
\multirow{14}{*}{MLP DF init.}
&forest\_depth & 6 & 3 & 3 & 2 & 3 & 2\\
& n\_forests & 1 & 1 & 1 & 1 & 1 & 1\\
& n\_estimators & 16 & 2 & 64 & 32 & 2 & 8\\
& max\_depth & 7 & 10 & 5 & 6 & 10 & 8\\
& max\_features & 0.350 & 0.598 & 0.992 & 0.322 & 0.633 & 0.342\\
& learning\_rate & 1.04$\times 10^{-5}$ & 6.67$\times 10^{-5}$ & 1.54$\times 10^{-5}$ & 3.08$\times 10^{-5}$ & 1.58$\times 10^{-5}$ & 2.31$\times 10^{-4}$\\
& depth & 23 & 10 & 9 & 13 & 15 & 9\\
& width & 2048 & 2048 & 2048 & 2048 & 2048 & 2048\\
& epochs & 100 & 100 & 100 & 100 & 100 & 100\\
& batch\_size & 256 & 256 & 256 & 256 & 256 & 256\\
& strength01 & 515 & 36.6 & 41.0 & 15.5 & 51.6 & 1.41\\
& strength12 & 0.162 & 0.242 & 10.6 & 0.213 & 0.124 & 0.154\\
& strength23 & 1.94 & 10.4 & 47.8 & 1.94 & 4.26$\times 10^{-2}$ & 0.149\\
& strength\_id & 3.63$\times 10^{-2}$ & 6.34$\times 10^{-2}$ & 7.44 & 2.75$\times 10^{-2}$ & 5.09$\times 10^{-2}$ & 3.69\\
\midrule
\multirow{6}{*}{SAINT}
& epochs & 100 & 100 & 100 & 100 & 100 & 100\\
& batch\_size & 256 & 256 & 256 & 256 & 64 & 256\\
& dim & 128 & 64 & 32 & 32 & 8 & 16\\
& depth & 3 & 2 & 2 & 1 & 2 & 2\\
& heads & 2 & 8 & 2 & 8 & 4 & 8\\
& dropout & 0.2 & 0 & 0.4 & 0.8 & 0.5 & 0.8\\\bottomrule
\end{tabular}
|
212,426
|
2209.15283v1
| null |
\caption{Hyper-parameters used for the experimental protocol P2.}
|
\label{tab:HP_ep2_2}
|
\begin{tabular}{cccccccc}
\toprule
\textbf{Method}&\textbf{Parameter}& \textbf{Diamonds} & \textbf{Blastchar} & \textbf{Heloc} & \textbf{Higgs}\\\midrule
\multirow{3}{*}{Random Forests}
& max\_depth & 12 & 6 & 9 & 12\\
& n\_estimators & 1000 & 1000 & 1000 & 1000\\
& max\_features & 0.967 & 0.547 & 0.607 & 0.577\\\midrule
\multirow{5}{*}{{GBDT}}
& max\_depth & 7 & 1 & 1 & 11\\
& n\_estimators & 1000 & 1000 & 1000 & 1000\\
& reg\_alpha & 0.341 & 7.15$\times10^{-7}$ & 0.123 & 2.29$\times10^{-8}$\\
& reg\_lambda & 5.15$\times10^{-4}$ & 1.59$\times10^{-7}$ & 1.44$\times10^{-2}$ & 0.391\\
& learning\_rate & 9.17$\times10^{-2}$ & 1.48$\times10^{-2}$ & 0.282 & 2.46$\times10^{-2}$\\\midrule
\multirow{5}{*}{Deep Forest}&
forest\_depth & 4 & 7 & 10 & 3\\
& n\_forests & 1 & 1 & 1 & 1 \\
& n\_estimators & 1000 & 1000 & 1000 & 1000 \\
& max\_depth & 12 & 2 & 4 & 12\\
& max\_features & 0.454 & 0.641 & 0.196 & 0.163\\\midrule
\multirow{5}{*}{MLP random init.}
& learning\_rate & 2.35$\times10^{-4}$ & 1.05$\times10^{-4}$ & 1.14$\times10^{-6}$ & 2.26$\times10^{-5}$\\
& depth & 9 & 8 & 8 & 9\\
& width & 1011 & 1475 & 1369 & 1284\\
& epochs & 100 & 100 & 100 & 100 \\
& batch\_size & 256 & 256 & 256 & 256 \\\midrule
\multirow{10}{*}{MLP RF init.}
& max\_depth & 10 & 5 & 7 & 9\\
& n\_estimators & 2 & 64 & 16 & 4\\
& max\_features & 0.904 & 0.425 & 0.728 & 0.670\\
& learning\_rate & 6.67$\times10^{-5}$ & 5.07$\times10^{-6}$ & 7.33$\times10^{-6}$ & 2.17$\times10^{-5}$\\
& depth & 4 & 8 & 6 & 3\\
& width & 2048 & 2048 & 2048 & 2048 \\
& epochs & 100 & 100 & 100 & 100 \\
& batch\_size & 256 & 256 & 256 & 256 \\
& strength01 & 19.8 & 4500 & 331 & 1.43\\
& strength12 & 0.420 & 42.9 & 1.06 & 0.329\\\midrule
\multirow{12}{*}{MLP GBDT init.}
& max\_depth & 3 & 1 & 3 & 5\\
& n\_estimators & 256 & 1024 & 256 & 64\\
& reg\_alpha & 4.56$\times10^{-2}$ & 1.63$\times10^{-5}$ & 6.21$\times10^{-7}$ & 2.58$\times10^{-6}$\\
& reg\_lambda & 6.17$\times10^{-4}$ & 2.19$\times10^{-4}$ & 3.03$\times10^{-4}$ & 3.20$\times10^{-6}$\\
& learning\_rate\_GBDT & 0.214 & 4.72$\times10^{-2}$ & 8.42$\times10^{-2}$ & 0.290\\
& learning\_rate & 8.94$\times10^{-5}$ & 5.60$\times10^{-6}$ & 4.54$\times10^{-4}$ & 1.36$\times10^{-4}$\\
& depth & 5 & 6 & 8 & 4\\
& width & 2048 & 2048 & 2048& 2048 \\
& epochs & 100 & 100 & 100 & 100 \\
& batch\_size & 256 & 256 & 256 & 256 \\
& strength01 & 3870 & 4690 & 6550 & 4780\\
& strength12 & 56.6 & 21.0 & 31.8 & 0.423\\\midrule
\multirow{14}{*}{MLP DF init.}
&forest\_depth & 3 & 2 & 2 & 3\\
& n\_forests & 1 & 1 & 1 & 1 \\
& n\_estimators & 4 & 128 & 64 & 8\\
& max\_depth & 9 & 4 & 5 & 8\\
& max\_features & 0.695 & 0.516 & 0.280 & 0.572\\
& learning\_rate & 2.04$\times10^{-5}$ & 2.00$\times10^{-6}$ & 1.91$\times10^{-5}$ & 9.33$\times10^{-6}$\\
& depth & 16 & 10 & 8 & 12\\
& width & 2048 & 2048 & 2048 & 2048 \\
& epochs & 100 & 100 & 100 & 100 \\
& batch\_size & 256 & 256 & 256 & 256 \\
& strength01 & 21.0 & 93.0 & 97.8 & 1.12\\
& strength12 & 0.119 & 20.0 & 0.987 & 9.22$\times10^{-2}$\\
& strength23 & 5.34$\times10^{-2}$ & 0.283 & 27.1 & 0.207\\
& strength\_id & 0.358 & 0.475 & 9.70 & 0.152\\
\midrule
\multirow{6}{*}{SAINT}
& epochs & 100 & 100 & 100 & 100 \\
& batch\_size & 256 & 256 & 256 & 64 \\
& dim & 64 & 128 & 64 & 16\\
& depth & 3 & 3 & 3 & 2\\
& heads & 4 & 8 & 2 & 8\\
& dropout & 0.2 & 0.5 & 0.8 & 0.8\\\bottomrule
\end{tabular}
|
212,427
|
2203.01629v5
| null |
\caption{
To compare the three methods (LabelVAE, AdaVAE, HGVAE) in the weakly-supervised experiment, we evaluate their learned latent representations with respect to shared (S) and independent (I) generative factors.
To assess the amount of shared and independent information in the latent representation, we train linear classifiers on the respective latent dimensions only.
We report the adjusted balanced classification accuracy, such that the random classifier achieves score $0$.
%
%
}
|
\label{tab:exp_ws_downstream_task}
|
\begin{tabular}{lccccccc}
\toprule
&$s=0$&\multicolumn{2}{c}{$s=1$}&\multicolumn{2}{c}{$s=3$}&\multicolumn{2}{c}{$s=5$}\\
\cmidrule(l){2-2}\cmidrule(l){3-4}\cmidrule(l){5-6}\cmidrule(l){7-8}
& I& S & I& S & I& S & I\\
\midrule
Label&$0.14{\scriptstyle\pm 0.01}$&$0.19{\scriptstyle\pm 0.03}$&$0.16{\scriptstyle\pm 0.01}$&$\bm{0.10}{\scriptstyle\pm 0.00}$&$0.23{\scriptstyle\pm 0.01}$&$\bm{0.34}{\scriptstyle\pm 0.00}$&$0.00{\scriptstyle\pm 0.00}$\\
%
Ada&$0.12{\scriptstyle\pm 0.01}$&$0.19{\scriptstyle\pm 0.01}$&$0.15{\scriptstyle\pm 0.01}$&$\bm{0.10}{\scriptstyle\pm 0.03}$&$0.22{\scriptstyle\pm 0.02}$&$0.33{\scriptstyle\pm 0.03}$&$0.00{\scriptstyle\pm 0.00}$\\
%
HG {\tiny (Ours)}&$\bm{0.18}{\scriptstyle\pm 0.01}$&$\bm{0.22}{\scriptstyle\pm 0.05}$&$\bm{0.19}{\scriptstyle\pm 0.01}$&$0.08{\scriptstyle\pm 0.02}$&$\bm{0.28}{\scriptstyle\pm 0.01}$&$0.28{\scriptstyle\pm 0.01}$&$\bm{0.01}{\scriptstyle\pm 0.00}$\\
\bottomrule
\end{tabular}
|
212,428
|
2203.01629v5
| null |
\caption{Evaluation of the clustering experiment on the MNIST datasets.
We compare the methods on 3 different dataset versions, namely i) uniform class distribution ii) subsampling with 80\% of samples and iii) subsampling with only 60\% of samples.
We subsample half of the classes.
Accuracy (Acc), normalized mutual information (NMI), and adjusted rand index (ARI) are used as evaluation metrics. Higher is better for all metrics. Mean and standard deviations are computed across 5 runs. For fair comparison with the baselines all methods use the pretraining weights provided by \citet{jiang2016variational}.}
|
\label{tab:exp_clustering_all}
|
\begin{tabular}{llccc}
\toprule
Dataset Version & & Uniform & Categorical & Hypergeometric \\
\midrule
\multirow{3}{*}{Uniform} & Acc (\%) & $\bm{92.0 \pm 3.0}$& $87.2 \pm 5.0$ & $91.4\pm 5.0$ \\
& NMI (\%) & $84.8 \pm 2.2$ & $81.8 \pm 1.9$ & $\bm{85.6 \pm 2.0} $ \\
& ARI (\%) & $84.2 \pm 4.3$ & $78.3 \pm 4.6$ & $\bm{84.8 \pm 4.6}$\\
\midrule
\multirow{3}{*}{Subsampling 80\%} & Acc (\%) & $90.8 \pm 4.0$ & $87.4 \pm 4.7$ & $\bm{92.5 \pm 0.5}$ \\
& NMI (\%) & $84.1 \pm 2.2$& $ 81.8 \pm 2.3$& $\bm{84.6 \pm 0.8}$ \\
& ARI (\%) & $83.2 \pm 3.6$ & $ 78.2 \pm 5.0$ & $\bm{84.4 \pm 1.0}$\\
\midrule
\multirow{3}{*}{Subsampling 60\%} & Acc (\%) & $83.5 \pm 3.9$ & $ 86.5 \pm 4.9$ & $\bm{89.7 \pm 4.3}$ \\
& NMI (\%) & $80.7 \pm 1.4$ & $ 81.3 \pm 2.9$& $\bm{82.9 \pm 2.2}$ \\
& ARI (\%) & $77.6 \pm 2.6$ & $ 77.7 \pm 6.3$ & $\bm{81.5 \pm 3.9}$\\
\bottomrule
\end{tabular}
|
212,450
|
2210.16315v3
| null | null |
\label{prop:glbin}
| null |
212,451
|
2210.16315v3
| null | null |
\label{eq:lb-binning}
| null |
212,452
|
2210.16315v3
| null | null |
\label{prop:debiasing}
| null |
212,453
|
2210.16315v3
| null | null |
\label{prop:decomposition}
| null |
212,454
|
2210.16315v3
| null | null |
\label{eq:gl-var:classwise}
| null |
212,455
|
2210.16315v3
| null | null |
\label{eq:classwise6}
| null |
212,430
|
2206.07136v3
| null |
\caption{Average test accuracy and 95\% confidence interval on image tasks over 5 runs.}
|
\label{tab:CV results}
|
\begin{tabular}{|c|c|c|c|c|c|}
\hline
\multirow{3}{*}{Task}& \multirow{3}{*}{Model}&\multirow{3}{*}{$(\epsilon,\delta)$} &\multicolumn{3}{c|}{Accuracy \%} \\
\cline{4-6}
&&&\multirow{2}{*}{Abadi's clipping}&\multirow{2}{*}{AUTO-S clipping}&non-DP
\\
&&&&&($\epsilon=\infty$)\\
\hline
MNIST& 4-layer CNN&$(3,1e$-$5)$&$98.04\pm0.09$&$98.15\pm 0.07$&$99.11\pm 0.07$
\\
\hline
FashionMNIST& 4-layer CNN&$(3,1e$-$5)$&$86.04\pm0.26$&$86.36\pm 0.18$&$89.57\pm 0.13$
% \\
% \hline
% CIFAR10& 8-layer CNN&$(3,1e$-$5)$&$59.06\pm0.30$&$59.36\pm 0.25$&$78.12\pm 0.19$
\\
\hline
CIFAR10 pretrained& SimCLRv2&$(2,1e$-$5)$&$92.44\pm 0.13$&$92.70\pm 0.02$&$94.42\pm 0.01$
\\
\hline
ImageNette& ResNet9&$(8,1e$-$4)$&$60.29\pm 0.53$&$60.71\pm 0.48$&$71.11\pm 0.37$
\\
\hline
CelebA [Smiling]& ResNet9&$(8,5e$-$6)$&$90.75\pm 0.11$&$91.08\pm 0.08$&$92.61\pm0.20$
\\
\hline
CelebA [Male]& ResNet9&$(8,5e$-$6)$&$95.54\pm 0.14$&$95.70\pm 0.07$&$97.90\pm 0.04$
\\
\hline
CelebA Multi-label& ResNet9&$(3,5e$-$6)$&$86.81\pm 0.03$&$87.05\pm 0.01$&$90.30\pm 0.02$
\\
\hline
CelebA Multi-label& ResNet9&$(8,5e$-$6)$&$87.52\pm 0.15$&$87.58\pm 0.04$&$90.30\pm 0.02$
\\\hline
\end{tabular}
|
212,431
|
2206.07136v3
| null |
\caption{Test accuracy on language tasks with RoBERTa-large (24 blocks, 355 million parameters).}
|
\label{tab:sentence roberta large}
|
\begin{tabular}{|c|cccc|cccc|cccc|}
\hline
\multirow {2}{*}{Method}&\multicolumn {4}{c}{$\epsilon=3$}& \multicolumn{4}{c}{$\epsilon=8$}&\multicolumn{4}{c|}{$\epsilon=\infty$ (non-DP)}
\\\cline{2-13}
&MNLI&QQP&QNLI&SST2&MNLI&QQP&QNLI&SST2&MNLI&QQP&QNLI&SST2
\\\hline
RGP \cite{yu2021large}&-&-&-&-&86.1/86.0&86.7&90.0&93.0&-&-&-&-
\\\cline{1-13}
full \cite{li2021large}&\textbf{86.43}/86.46&86.43&90.76&93.04&87.02/\textbf{87.26}&87.47&91.10&93.81&\multirow{3}{*}{90.33/90.03}&\multirow{3}{*}{87.90}&\multirow{3}{*}{93.61}&\multirow{3}{*}{96.21}
\\
full AUTO-V&85.33/85.61&\textbf{86.61}&89.99&\textbf{93.12}&85.91/86.10&86.86&90.55&93.35&&&&
\\
full AUTO-S&86.27/\textbf{86.67}&\textbf{86.76}&\textbf{91.01}&\textbf{93.92}&\textbf{87.07}/87.16&\textbf{87.47}&\textbf{91.45}&\textbf{94.61}&&&&
\\\hline
\end{tabular}
|
212,432
|
2206.07136v3
| null |
\caption{Test performance on E2E dataset with GPT2. Additional performance measures are included in \Cref{tab:E2E GPT extended}. The best two GPT2 models for each row are marked in bold.}
|
\label{tab:E2E GPT selected}
|
\begin{tabular}{|l|c|c|c|cccccccc|}
\hline
&DP&GPT2&GPT2&\multicolumn{8}{c|}{GPT2}
\\
Metric&guarantee&large&medium&&&&&&&&
\\\cline{3-12}
&&full&full&full&full&full&LoRA&RGP&prefix&top2&retrain
\\
&&AUTO-S&AUTO-S&AUTO-S&AUTO-V&\cite{li2021large}&\cite{hu2021lora}&\cite{yu2021large}&\cite{li2021prefix}&&%\cite{li2021large}&\cite{li2021large}
\\\hline
\multirow{3}{*}{BLEU}&$\epsilon=3$&\textbf{64.180}&\textbf{63.850}&\textbf{61.340}&\textbf{61.519}&\textbf{61.519}&58.153&58.482&47.772&25.920&15.457
\\
&$\epsilon=8$&\textbf{64.640}&\textbf{64.220}&\textbf{63.600}&63.189&63.189&\textbf{63.389}&58.455&49.263&26.885&24.247
\\
&non-DP&66.840&68.500&69.463&69.463&69.463&69.682&68.328&68.845&65.752&65.731
\\\hline\hline
\multirow{3}{*}{ROGUE-L}&$\epsilon=3$&\textbf{67.857}&\textbf{67.071}&\textbf{65.872}&65.670&65.670&\textbf{65.773}&65.560&58.964&44.536&35.240
\\
&$\epsilon=8$&\textbf{68.968}&\textbf{67.533}&\textbf{67.073}&66.429&66.429&\textbf{67.525}&65.030&60.730&46.421&39.951
\\
&non-DP&70.384&71.458&71.359&71.359&71.359&71.709&68.844&70.805&68.704&68.751
\\\hline
\end{tabular}
|
212,433
|
2206.07136v3
| null |
\caption{Hyperparameters of automatic clipping and Abadi's clipping, for sentence classification in \Cref{tab:sentence roberta base} and \Cref{tab:sentence roberta large}, using either RoBERTa base or large.}
| null |
\begin{tabular}{c|cccc}
Dataset& MNLI(m/mm)&QQP&QNLI&SST2 \\
Epoch&18&18&6&3\\
Batch size&6000&6000&2000&1000\\
clipping threshold $R$&0.1&0.1&0.1&0.1\\
DP learning rate &5e-4&5e-4&5e-4&5e-4\\
non-DP learning rate &5e-5&5e-5&5e-5&5e-5\\
learning rate decay&Yes&Yes&Yes&Yes\\
AdamW weight decay&0&0&0&0\\
Max sequence length&256&256&256&256\\
\end{tabular}
|
212,434
|
2206.07136v3
| null |
\caption{SST2 accuracy with respect to learning rate.}
| null |
\begin{tabular}{c|c|c|c|c|c}
learning rate& 1e-4& 3e-4& 5e-4& 8e-4& 1e-3 \\\hline
RoBERTa-base& 93.92& 94.38& 94.49& 94.72& 93.35\\
RoBERTa-large&95.76& 96.21& 96.21& 96.33& 95.99
\end{tabular}
|
212,435
|
2206.07136v3
| null |
\caption{Hyperparameters of automatic clipping and Abadi's clipping, for the E2E generation task in \Cref{tab:E2E GPT selected}.}
| null |
\begin{tabular}{c|ccc}
Model& GPT2&GPT2 medium&GPT2 large \\
Epoch&10&10&10\\
Batch size&1024&1024&1024\\
clipping threshold $R$&0.1&0.1&0.1\\
DP learning rate &2e-3&2e-3&2e-3\\
non-DP learning rate &2e-4&1e-4&1e-4\\
learning rate decay&No&No&No\\
AdamW weight decay&0.01&0.01&0.01\\
Max sequence length&100&100&100\\
\end{tabular}
|
212,436
|
2206.07136v3
| null |
\caption{Test performance on E2E dataset with GPT2. The best two GPT2 models for each row are marked in bold.}
|
\label{tab:E2E GPT extended}
|
\begin{tabular}{|l|c|c|c|cccccccc|}
\hline
&DP&GPT2&GPT2&\multicolumn{8}{c|}{GPT2}
\\
Metric&guarantee&large&medium&&&&&&&&
\\\cline{3-12}
&&full&full&full&full&full&LoRA&RGP&prefix&top2&retrain
\\
&&AUTO-S&AUTO-S&AUTO-S&AUTO-V&\cite{li2021large}&\cite{hu2021lora}&\cite{yu2021large}&\cite{li2021prefix}&\cite{li2021large}&\cite{li2021large}
\\\hline
\multirow{3}{*}{BLEU}&$\epsilon=3$&\textbf{64.180}&\textbf{63.850}&\textbf{61.340}&\textbf{61.519}&\textbf{61.519}&58.153&58.482&47.772&25.920&15.457
\\
&$\epsilon=8$&\textbf{64.640}&\textbf{64.220}&\textbf{63.600}&63.189&63.189&\textbf{63.389}&58.455&49.263&26.885&24.247
\\
&non-DP&66.840&68.500&69.463&69.463&69.463&69.682&68.328&68.845&65.752&65.731
\\\hline\hline
\multirow{3}{*}{ROGUE-L}&$\epsilon=3$&\textbf{67.857}&\textbf{67.071}&\textbf{65.872}&65.670&65.670&\textbf{65.773}&65.560&58.964&44.536&35.240
\\
&$\epsilon=8$&\textbf{68.968}&\textbf{67.533}&\textbf{67.073}&66.429&66.429&\textbf{67.525}&65.030&60.730&46.421&39.951
\\
&non-DP&70.384&71.458&71.359&71.359&71.359&71.709&68.844&70.805&68.704&68.751
\\\hline\hline
\multirow{3}{*}{NIST}&$\epsilon=3$&\textbf{7.937}&\textbf{7.106}&\textbf{7.071}&\textbf{6.697}&\textbf{6.697}&5.463&5.775&5.249&1.510&0.376
\\
&$\epsilon=8$&\textbf{8.301}&\textbf{8.172}&\textbf{7.714}&7.444&7.444&\textbf{7.449}&6.276&5.525&1.547&1.01
\\
&non-DP&8.730&8.628&8.780&8.780&8.780&8.822&8.722&8.722&8.418&8.286
\\\hline\hline
\multirow{3}{*}{METEOR}&$\epsilon=3$&\textbf{0.403}&\textbf{0.387}&\textbf{0.387}&\textbf{0.384}&\textbf{0.384}&0.370&0.331&0.363&0.197&0.113
\\
&$\epsilon=8$&\textbf{0.420}&\textbf{0.418}&\textbf{0.404}&0.400&0.400&\textbf{0.407}&0.349&0.364&0.207&0.145
\\
&non-DP&0.460&0.449&0.461&0.461&0.461&0.463&0.456&0.445&0.443&0.429
\\\hline\hline
\multirow{3}{*}{CIDEr}&$\epsilon=3$&\textbf{2.008}&\textbf{1.754}&\textbf{1.801}&\textbf{1.761}&\textbf{1.761}&1.581&1.300&1.507&0.452&0.116
\\
&$\epsilon=8$&\textbf{2.163}&\textbf{2.081}&\textbf{1.938}&1.919&1.919&\textbf{1.948}&1.496&1.569&0.499&0.281
\\
&non-DP&2.356&2.137&2.422&2.422&2.422&2.491&2.418&2.345&2.180&2.004
\\\hline
\end{tabular}
|
212,437
|
2206.07136v3
| null |
\caption{Accuracy on CelebA dataset with settings in \Cref{app:CV settings} from one run. The green arrow indicates AUTO-S is better than Abadi's clipping under the same $\epsilon$; the red arrow indicates otherwise; the black bar indicates the same accuracy.}
| null |
\begin{tabular}{|c|c||c|c|c|c|c|}
\hline
\multirow{2}{*}{Index}&\multirow{2}{*}{Attributes}&Abadi's&AUTO-S&Abadi's&AUTO-S&non-DP\\
&&$\epsilon=3$&$\epsilon=3$&$\epsilon=8$&$\epsilon=8$&$\epsilon=\infty$\\\hline
0&5 o Clock Shadow&90.64&90.99$\color{green}{\uparrow}$&90.81&91.28$\color{green}{\uparrow}$&93.33\\
1&Arched Eyebrows&75.15&76.31$\color{green}{\uparrow}$& 76.84 &77.11$\color{green}{\uparrow}$&81.52\\
2&Attractive&75.85&76.10$\color{green}{\uparrow}$& 77.50 &77.74$\color{green}{\uparrow}$&81.15\\
3&Bags Under Eyes&80.75&81.12$\color{green}{\uparrow}$& 82.15 &82.13$\color{red}{\downarrow}$&84.81\\
4&Bald&97.84 &97.87$\color{green}{\uparrow}$&98.04 &97.98$\color{red}{\downarrow}$&98.58\\
5&Bangs&92.71 &92.68$\color{red}{\downarrow}$&93.46 &93.55$\color{green}{\uparrow}$&95.50\\
6&Big Lips&67.51&67.78$\color{green}{\uparrow}$& 68.34 &68.44$\color{green}{\uparrow}$&71.33\\
7&Big Nose&78.01 &80.23$\color{green}{\uparrow}$&76.69 &80.59$\color{green}{\uparrow}$&83.54\\
8&Black Hair&81.92&80.95$\color{red}{\downarrow}$& 83.33 &83.28$\color{red}{\downarrow}$&88.55\\
9&Blond Hair&92.25 &92.38$\color{green}{\uparrow}$&93.52 &93.09$\color{red}{\downarrow}$&95.49\\
10&Blurry&94.91 &94.82$\color{red}{\downarrow}$&95.08 &94.90$\color{red}{\downarrow}$&95.78\\
11&Brown Hair&80.13&82.50$\color{green}{\uparrow}$& 83.74 &83.89$\color{green}{\uparrow}$&87.79\\
12&Bushy Eyebrows&88.06&88.23$\color{green}{\uparrow}$& 89.72 &88.80$\color{red}{\downarrow}$&92.19\\
13&Chubby&94.72&94.54$\color{red}{\downarrow}$& 94.54 &94.50$\color{red}{\downarrow}$&95.56\\
14&Double Chin&95.19&95.49$\color{green}{\uparrow}$& 95.50 &95.51$\color{green}{\uparrow}$&96.09\\
15&Eyeglasses&97.06&97.64$\color{green}{\uparrow}$& 98.32 &98.06$\color{red}{\downarrow}$&99.39\\
16&Goatee&95.68&95.45$\color{red}{\downarrow}$& 95.84 &95.87$\color{green}{\uparrow}$&97.06\\
17&Gray Hair&96.77&96.79$\color{green}{\uparrow}$& 97.02 &97.03$\color{green}{\uparrow}$&98.06\\
18&Heavy Makeup&84.96&85.70$\color{green}{\uparrow}$& 87.58 &87.29$\color{red}{\downarrow}$&90.76\\
19&High Cheekbones&81.46&81.42$\color{red}{\downarrow}$& 82.62 &82.72$\color{green}{\uparrow}$&86.62\\
20&Male&92.05&92.17$\color{green}{\uparrow}$& 93.32 &93.17$\color{red}{\downarrow}$&97.46\\
21&Mouth Slightly Open&86.20&86.32$\color{green}{\uparrow}$& 87.84 &88.48$\color{green}{\uparrow}$&93.07\\
22&Mustache&96.05&95.96$\color{red}{\downarrow}$& 96.08 &95.99$\color{red}{\downarrow}$&96.74\\
23&Narrow Eyes&84.90&84.78$\color{red}{\downarrow}$& 85.14 &85.18$\color{green}{\uparrow}$&86.98\\
24&No Beard&91.55&91.67$\color{green}{\uparrow}$& 92.29 &92.45$\color{green}{\uparrow}$&95.18\\
25&Oval Face&71.26&71.42$\color{green}{\uparrow}$& 71.98 &71.25$\color{red}{\downarrow}$&74.62\\
26&Pale Skin&96.09&96.04$\color{red}{\downarrow}$& 96.15 &96.17$\color{green}{\uparrow}$&96.93\\
27&Pointy Nose&70.34&72.11$\color{green}{\uparrow}$& 72.23 &73.01$\color{green}{\uparrow}$&75.68\\
28&Receding Hairline&91.53&91.37$\color{red}{\downarrow}$& 91.75 &91.74$\color{red}{\downarrow}$&92.87\\
29&Rosy Cheeks&93.26&93.02$\color{red}{\downarrow}$& 93.56 &93.35$\color{red}{\downarrow}$&94.86\\
30&Sideburns&96.16&96.09$\color{red}{\downarrow}$& 96.27 &96.46$\color{green}{\uparrow}$&97.44\\
31&Smiling&86.39&87.08$\color{green}{\uparrow}$& 88.87 &88.63$\color{red}{\downarrow}$&92.25\\
32&Straight Hair&76.20&77.95$\color{green}{\uparrow}$& 78.78 &78.52$\color{red}{\downarrow}$&80.66\\
33&Wavy Hair&70.30&71.79$\color{green}{\uparrow}$& 73.58 &73.19$\color{red}{\downarrow}$&79.15\\
34&Wearing Earrings&80.53&81.52$\color{green}{\uparrow}$& 82.29 &82.20$\color{red}{\downarrow}$&87.56\\
35&Wearing Hat&96.99&96.83$\color{red}{\downarrow}$& 97.46 &97.31$\color{red}{\downarrow}$&98.68\\
36&Wearing Lipstick&88.95&88.04$\color{red}{\downarrow}$& 89.87 &90.72$\color{green}{\uparrow}$&93.49\\
37&Wearing Necklace&84.59&85.83$\color{green}{\uparrow}$& 85.93 &85.42$\color{red}{\downarrow}$&86.61\\
38&Wearing Necktie&93.91&93.91--& 94.43 &94.08$\color{red}{\downarrow}$&96.30\\
39&Young&81.35&81.21$\color{red}{\downarrow}$& 82.18 &82.52$\color{green}{\uparrow}$&87.18\\\hline
\end{tabular}
|
212,438
|
2206.07136v3
| null |
\caption{Accuracy on CelebA dataset with settings in \Cref{app:CV settings} from one run. `Single' means each attribute is learned separately as a binary classification task. `Multi' means all attributes are learned jointly as a multi-label classification task. The green arrow indicates AUTO-S is better than Abadi's clipping under the same $\epsilon$ and the same task; the red arrow indicates otherwise; the black bar indicates the same accuracy.}
| null |
\begin{tabular}{|c|c||c|c|c|c|c|}
\hline
\multirow{3}{*}{Index}&\multirow{3}{*}{Attributes}&Abadi's&AUTO-S&Abadi's&AUTO-S&non-DP\\
&&Single&Single&Multi&Multi&Multi\\
&&$\epsilon=8$&$\epsilon=8$&$\epsilon=8$&$\epsilon=8$&$\epsilon=\infty$\\\hline
0&5 o Clock Shadow&92.15&92.29$\color{green}{\uparrow}$&90.81&91.28$\color{green}{\uparrow}$&93.33\\
1&Arched Eyebrows&81.18&80.19$\color{red}{\downarrow}$& 76.84 &77.11$\color{green}{\uparrow}$&81.52\\
2&Attractive&79.31&79.79$\color{green}{\uparrow}$& 77.50 &77.74$\color{green}{\uparrow}$&81.15\\
3&Bags Under Eyes&83.52&83.48$\color{red}{\downarrow}$& 82.15 &82.13$\color{red}{\downarrow}$&84.81\\
4&Bald&97.89 &97.88$\color{red}{\downarrow}$&98.04 &97.98$\color{red}{\downarrow}$&98.58\\
5&Bangs&94.52 &94.83$\color{green}{\uparrow}$&93.46 &93.55$\color{green}{\uparrow}$&95.50\\
6&Big Lips&67.32&67.53$\color{green}{\uparrow}$& 68.34 &68.44$\color{green}{\uparrow}$&71.33\\
7&Big Nose&82.31 &82.36$\color{green}{\uparrow}$&76.69 &80.59$\color{green}{\uparrow}$&83.54\\
8&Black Hair&87.08&86.93$\color{red}{\downarrow}$& 83.33 &83.28$\color{red}{\downarrow}$&88.55\\
9&Blond Hair&94.29 &94.73$\color{green}{\uparrow}$&93.52 &93.09$\color{red}{\downarrow}$&95.49\\
10&Blurry&94.95 &95.20$\color{green}{\uparrow}$&95.08 &94.90$\color{red}{\downarrow}$&95.78\\
11&Brown Hair&87.41&87.19$\color{red}{\downarrow}$& 83.74 &83.89$\color{green}{\uparrow}$&87.79\\
12&Bushy Eyebrows&91.23&91.43$\color{green}{\uparrow}$& 89.72 &88.80$\color{red}{\downarrow}$&92.19\\
13&Chubby&94.70&94.70--& 94.54 &94.50$\color{red}{\downarrow}$&95.56\\
14&Double Chin&95.43&95.43--& 95.50 &95.51$\color{green}{\uparrow}$&96.09\\
15&Eyeglasses&98.88&99.14$\color{green}{\uparrow}$& 98.32 &98.06$\color{red}{\downarrow}$&99.39\\
16&Goatee&96.12&96.07$\color{red}{\downarrow}$& 95.84 &95.87$\color{green}{\uparrow}$&97.06\\
17&Gray Hair&97.48&97.34$\color{red}{\downarrow}$& 97.02 &97.03$\color{green}{\uparrow}$&98.06\\
18&Heavy Makeup&88.85&88.72$\color{red}{\downarrow}$& 87.58 &87.29$\color{red}{\downarrow}$&90.76\\
19&High Cheekbones&85.66&85.45$\color{red}{\downarrow}$& 82.62 &82.72$\color{green}{\uparrow}$&86.62\\
20&Male&95.42&95.70$\color{green}{\uparrow}$& 95.53 &93.17$\color{red}{\downarrow}$&97.46\\
21&Mouth Slightly Open&92.67&92.74$\color{green}{\uparrow}$& 87.84 &88.48$\color{green}{\uparrow}$&93.07\\
22&Mustache&96.13&96.13--& 96.08 &95.99$\color{red}{\downarrow}$&96.74\\
23&Narrow Eyes&85.13&85.13--& 85.14 &85.18$\color{green}{\uparrow}$&86.98\\
24&No Beard&94.26&94.58$\color{green}{\uparrow}$& 92.29 &92.45$\color{green}{\uparrow}$&95.18\\
25&Oval Face&70.77&73.05$\color{green}{\uparrow}$& 71.98 &71.25$\color{red}{\downarrow}$&74.62\\
26&Pale Skin&96.38&96.34$\color{red}{\downarrow}$& 96.15 &96.17$\color{green}{\uparrow}$&96.93\\
27&Pointy Nose&71.48&73.37$\color{green}{\uparrow}$& 72.23 &73.01$\color{green}{\uparrow}$&75.68\\
28&Receding Hairline&91.51&91.51--& 91.75 &91.74$\color{red}{\downarrow}$&92.87\\
29&Rosy Cheeks&93.26&93.35$\color{green}{\uparrow}$& 93.56 &93.35$\color{red}{\downarrow}$&94.86\\
30&Sideburns&96.46&96.34$\color{red}{\downarrow}$& 96.27 &96.46$\color{green}{\uparrow}$&97.44\\
31&Smiling&90.82&90.87$\color{green}{\uparrow}$& 88.87 &88.63$\color{red}{\downarrow}$&92.25\\
32&Straight Hair&79.01&79.01--& 78.78 &78.52$\color{red}{\downarrow}$&80.66\\
33&Wavy Hair&77.55&78.83$\color{green}{\uparrow}$& 73.58 &73.19$\color{red}{\downarrow}$&79.15\\
34&Wearing Earrings&87.33&87.50$\color{green}{\uparrow}$& 82.29 &82.20$\color{red}{\downarrow}$&87.56\\
35&Wearing Hat&98.04&98.11$\color{green}{\uparrow}$& 97.46 &97.31$\color{red}{\downarrow}$&98.68\\
36&Wearing Lipstick&92.05&90.46$\color{red}{\downarrow}$& 89.87 &90.72$\color{green}{\uparrow}$&93.49\\
37&Wearing Necklace&86.21&86.21--& 85.93 &85.42$\color{red}{\downarrow}$&86.61\\
38&Wearing Necktie&95.85&95.94$\color{green}{\uparrow}$& 94.43 &94.08$\color{red}{\downarrow}$&96.30\\
39&Young&85.19&84.12$\color{red}{\downarrow}$& 82.18 &82.52$\color{green}{\uparrow}$&87.18\\\hline
\end{tabular}
|
212,456
|
2210.16315v3
| null |
\caption{Raw values of the estimators in the vision (\autoref{fig:vision:comparison}) and NLP experiments of \autoref{sec:exp_nn}, before ($\clhat$ and $\gllbhat$) and after ($\clhat'$ and $\gllbhat'$) isotonic recalibration.}
|
\label{tab:vision-nlp:comparison}
| null |
212,457
|
2210.16315v3
| null | null | null | null |
212,458
|
2210.16315v3
| null | null | null | null |
212,439
|
2206.07136v3
| null |
\caption{Comparison between clipping functions. CV means computer vision and NLP means natural language processing. Notice that any clipping function with $R$ is not automatic and requires tuning, and that the stability constant $\gamma$ enjoys theoretical and empirical benefits.}
| null |
\begin{tabular}{c|c|c|c|c}
$C_i$&reference&clipping or not & convergence analysis& experiments \\\hline
$\min(1,\frac{R}{||\g_i||})$&\cite{abadi2016deep,li2021large}&clipping & None & CV and NLP
\\
$\min(\frac{1}{R},\frac{1}{||\g_i||})$&\cite{de2022unlocking}&clipping & None & CV only
\\
$\frac{R}{||\g_i||}$&\cite{das2021convergence}&normalization & convex and federated setting (not per-sample) & CV only
\\
$\frac{1}{||\g_i||+\gamma}$&\cite{yang2022normalized}&normalization & non-convex, relaxed Lipschitz smoothness & CV and NLP
\\
$\frac{1}{||\g_i||+\gamma}$&this work&normalization & non-convex, same smoothness as non-DP & CV and NLP
\end{tabular}
|
212,440
|
2312.11309v2
| null |
\caption{\label{tab:strongsingle}Avg.\ transferability (\%)
on \imagenet{} from Inc-v3 to
all other models when integrating individual
augmentations composed with \dstonly{}.
A vertical line separates the baselines.
from our attacks.
}
| null |
\begin{tabular}{r|rrrrrrr}
\toprule
\textbf{Attack}
&\ASDTM{} &
\dst{}&
\sdtv{}&
\multicolumn{1}{r|}{\undp{}}&
\chanshuffle{}-\dstonly{} &
\colorjitter{}-\dstonly{}&
\fancyPCA{}-\dstonly{} \\
\textbf{Avg.}
&80.6
&81.8
&\textbf{88.8}
&\multicolumn{1}{r|}{79.8}
&84.8
&85.9
&{83.5} \\
\midrule
\textbf{Attack (cont.)} &
\gsdt{} &
\randerase{}-\dstonly{}&
\cutmix{}-\dstonly{}&
\cutout{}-\dstonly{}&
\neuraltrans{}-\dstonly{}&
\sharpen{}-\dstonly{}&
\autoaug{}-\dstonly{}\\
\textbf{Avg.\ (cont.)}
&\textit{87.0}
&84.9
&54.0
&84.1
&73.5
&80.1
&82.9\\
\bottomrule
\end{tabular}
|
212,441
|
2312.11309v2
| null |
\caption{\label{tab:normal} Average transferability (\%) of black-box attacks on
\imagenet{}, from all surrogates %
to normally and adversarially trained targets. A
vertical line separates the baselines from our attacks.}
| null |
\begin{tabular}{l |rrrr|rrr}
\toprule
\textbf{Targets} & \textbf{\ASDTM{} }& \textbf{\SDTM{}} & \textbf{\sdtv{}} & \textbf{\undp{}} & \textbf{\gsdt{}}&\textbf{\bestcombo{}} & \textbf{\advancedultimatecombo{}} \\
\midrule
\textbf{Normally trained} &79.7 & 82.1 & 82.9 & 85.0 & 86.2 & \textit{89.6} & \textbf{92.6} \\
\textbf{Adversarially trained} &78.2 & 77.4 & 82.5 & 65.4 & 83.1 & \textit{85.3} & \textbf{91.8}\\
\bottomrule
\end{tabular}
|
212,442
|
2312.11309v2
| null |
\caption{\label{tab:ens2adv}Transferability (\%) on \imagenet{},
from an ensemble of normally trained surrogates (incl.\ Inc-v4,
Res-50, Res-101 and Res-152) to adversarially trained targets. A
vertical line separates the baselines from our attacks.
}
| null |
\begin{tabular}{l | r r r r | r r r}
\toprule
\textbf{Model} & \textbf{\ASDTM{}}& \textbf{\SDTM{}} &
\textbf{\sdtv{}}& \textbf{\undp{}}& \textbf{\gsdt{}}&
\textbf{\bestcombo{}}& \textbf{\advancedultimatecombo{}} \\ \midrule
\textbf{Inc-v3$_{adv}$} & 89.3 & 89.1 & 92.8 & 88.2 & 92.6 & \textit{93.2} & \textbf{96.2}\\
\textbf{Inc-v3$_{ens3}$} & 90.0 & 89.6 & 93.8 & 90.6 & 93.5 & \textit{95.4} & \textbf{96.3} \\
\textbf{Inc-v3$_{ens4}$} & 89.0 & 87.8 & 93.1 & 86.1 & 92.3 & \textit{93.4} & \textbf{96.4} \\
\textbf{IncRes-v2$_{ens}$} & 84.8 & 83.4 & 90.1 & 75.5 & 88.6 & \textit{91.1} & \textbf{94.6} \\
\bottomrule
\end{tabular}
|
212,443
|
2312.11309v2
| null |
\caption{\label{tab:defs}
Transferability (\%) from an ensemble of normally trained
surrogates (Inc-v4, Res-50, Res-101 and Res-152) to
defended \imagenet{} models.
}
| null |
\begin{tabular}{l | r r r r | r r }
\toprule
\textbf{Defense} & \textbf{\ASDTM{}} & \textbf{\dst{}} & \textbf{\sdtv{}} & \textbf{\undp{}} & \textbf{\bestcombo{}} & \textbf{\advancedultimatecombo{}} \\ \midrule
\textbf{\bitred{}} & 88.6 & 88.2 & 94.8 & 94.9 & \textbf{96.0} & \textit{95.5} \\
\textbf{\nrp{}} & 51.0 & 54.9 & \textbf{80.0} & 27.9 & \textit{65.3} & 55.8 \\
\textbf{\randsmooth{}} & 87.3 & 84.8 & 90.6 & 85.5 & \textit{88.5} & \textbf{95.6} \\
\textbf{\arandsmooth{}} & 65.4 & 62.9 & 66.5 & 61.9 & \textit{67.0} & \textbf{71.9} \\
\bottomrule
\end{tabular}
|
212,444
|
2312.11309v2
| null |
\caption{\label{tab:cifarTRS}Transferability (\%) on \cifarten{}
from a normally trained VGG surrogate to an ensemble of Res
\dnn{}s trained via \TRS{}.
}
| null |
\begin{tabular}{l|r r r r | r r r }
\toprule
\textbf{Epsilon} &\textbf{\ASDTM{}} & \textbf{\SDTM{}} & \textbf{\sdtv{}} & \textbf{\undp{}} & \textbf{\gsdt{}} & \textbf{\bestcombo{}}& \textbf{\advancedultimatecombo{}} \\ \midrule
0.02 & 23.7 & 26.6 & 22.1&24.9 & {28.2} & \textit{32.7} & \textbf{46.5} \\
0.04 & 45.8 & 46.7 & 41.9 &52.6 &{52.8} & \textit{59.4} & \textbf{80.0} \\
\bottomrule
\end{tabular}
|
212,445
|
2312.11309v2
| null |
\caption{\label{tab:googleAPI}Transferability (\%) on \imagenet{}
from an esemble of normally trained surrogates (Inc-v4, Res-50,
Res-101 and Res-152) to Google Cloud Vision.}
| null |
\begin{tabular}{r r r r | r r}
\toprule
\textbf{\ASDTM{}} & \textbf{\SDTM{}} & \textbf{\sdtv{}} & \textbf{\undp{}}& \textbf{\bestcombo{}}&\textbf{\bestcomboadv{}} \\ \midrule
76.2 & 73.4 & 72.6 & \textit{81.3} & {76.5} & \textbf{82.1} \\
\bottomrule
\end{tabular}
|
212,446
|
2312.11309v2
| null |
\caption{\label{tab:timing}The number of samples augmented and the
avg.\ time of crafting an \AE{} %
per attack. Times were measured on an Nvidia A5000 GPU,
on \imagenet{}, when attacking an
Inc-v3, and averaged for 1K samples. For best
transferability rates, \undp{} %
was run
for 100 iterations, while other attacks were run for 10.
}
| null |
\begin{tabular}{l | r r r r | r r r r}
\toprule
\textbf{Attack} & \textbf{\ASDTM{}} & \textbf{\SDTM{}} &
\textbf{\sdtv{}} & \textbf{\undp{}} & \textbf{\gsdt} &
\textbf{\bestcombo} & \textbf{\subsetattack{30}} & \textbf{\advancedultimatecombo{}} \\ \midrule
\textbf{Augmented samples} & 15 & 5 & 105 & 1 & 10 & 30 & 30 & 165 \\
\textbf{Time (s)} & 1.68 & \textbf{0.72} & 11.29 & 1.65 & \textit{1.10} &
3.62 & 5.54 & 40.27 \\
\bottomrule
\end{tabular}
|
212,447
|
2312.11309v2
| null |
\caption{\label{tab:app:imageneteps}Transferability rates (\%) on \imagenet{},
from a Inc-v3 surrogate to other normally trained models, with
perturbation norms $\epsilon\in\{\frac{8}{255}, \frac{24}{255}\}$
other than the default $\epsilon{}=\frac{16}{255}$.}
| null |
\begin{tabular}{l | l | r r r r r r r r r r r r | r}
\toprule
\textbf{$\epsilon$} & \textbf{Attack}& \textbf{Inc-v3}& \textbf{Inc-v4}&\textbf{Res-152}&\textbf{IncRes-v2}&\textbf{ Res-50}&\textbf{ Res-101}&\textbf{ \vit{}}&\textbf{ \mnas{}} & \textbf{Inc-v3$_{adv}$}& \textbf{Inc-v3$_{ens3}$}&\textbf{Inc-v3$_{ens4}$}&\textbf{ IncRes-v2$_{ens}$}&\textbf{Avg.}\\ \midrule
\multirow{5}{*}{8/255}&\ASDTM{}&98.7&75.0&60.6&69.3&68.3&62.7&33.5&67.8&57.6&52.5&51.8&36.1&57.7\\
&\SDTM{}&\textbf{99.7}&77.1&62.9&70.9&69.8&64.8&34.1&71.1&61.4&55.0&54.7&35.9&59.8\\
&\sdtv{}&99.5&80.0&68.1&75.1&74.8&68.9&39.7&74.5&68.6&64.4&64.5&47.7&66.0\\
&\undp{}&99.4&\textbf{92.2}&\textbf{82.4}&\textbf{90.6}&\textbf{85.6}&\textbf{85.1}&32.3&\textit{86.6}&48.7&43.9&40.8&20.3&64.4\\
& \gsdt{}&{99.5}&{81.3}&{71.1}&{77.6}&{77.3}&{72.0}&37.0&75.4&69.2&63.9&63.9&45.2&66.7\\
& \bestcombo{}&\textbf{99.7}&{86.0}&{75.0}&{81.8}&{81.2}&{76.3}&{40.8}&{82.5}&{72.7}&{67.2}&{64.2}&{46.2}&\textit{70.4}\\
& \advancedultimatecombo{}&\textbf{99.7}&\textit{89.8}&\textit{80.7}&\textit{87.1}&\textit{84.3}&\textit{81.5}&\textbf{52.6}&\textbf{87.8}&\textbf{80.7}&\textbf{77.3}&\textbf{75.4}&\textbf{58.4}&\textbf{77.8}\\
\midrule
\multirow{5}{*}{24/255}&\ASDTM{}&99.9&97.1&93.7&95.9&94.8&94.4&73.4&94.3&87.1&88.6&88.2&79.9&89.8\\
&\SDTM{}&\textbf{100.0}&97.8&93.6&96.8&95.1&93.9&74.9&95.9&84.6&90.5&89.3&77.2&90.0\\
&\sdtv{}&\textbf{100.0}&97.9&95.0&97.1&96.2&95.6&78.4&97.1&89.3&93.2&91.9&84.0&92.3\\
&\undp{}&99.8&98.9&{97.3}&{98.9}&{97.9}&98.0{}&78.8&\textit{97.6}&84.0&82.9&78.4&59.8&88.4
\\
& \gsdt{}&\textbf{100.0}&98.6&95.7&98.3&96.7&97.0&86.7&97.5&94.5&95.9&94.9&90.4&95.1\\
& \bestcombo{}&\textbf{100.0}&\textit{99.3}&\textit{97.4}&\textit{99.2}&\textit{97.5}&\textit{98.4}&\textit{83.8}&\textit{99.0}&\textit{92.5}&\textit{95.1}&\textit{95.0}&\textit{86.9}&\textit{94.9}\\
& \advancedultimatecombo{}&\textbf{100.0}&\textbf{100.0}&\textbf{99.8}&\textbf{100.0}&\textbf{99.6}&\textbf{99.9}&\textbf{94.1}&\textbf{99.5}&\textbf{97.9}&\textbf{99.3}&\textbf{98.4}&\textbf{95.2}&\textbf{98.5}\\
\bottomrule
\end{tabular}
|
212,448
|
2210.16315v3
| null | null |
\label{eq:gl-var}
| null |
212,449
|
2210.16315v3
| null | null |
\label{eq:gl-exp-res}
| null |
212,464
|
2212.04089v3
| null |
\caption{\textbf{Forgetting image classification tasks via negation}. Results are shown for CLIP models, reporting average accuracy (\%) on the eight target tasks we wish to forget (Cars, DTD, EuroSAT, GTSRB, MNIST, RESISC45, SUN397 and SVHN), and the control task (ImageNet). Negating task vectors reduce the accuracy of a pre-trained ViT-L/14 by 45.8 percentage points on the target tasks, with little loss on the control task. Additional details and results are shown in Appendix \ref{sec:clip-neg-extended}.}
|
\label{tab:forget_image}
|
\begin{tabular}{l@{\hskip .3in}cc|cc|cc}
\toprule
\multirow{2}{*}{Method} & \multicolumn{2}{c|}{ViT-B/32} & \multicolumn{2}{c|}{ViT-B/16} & \multicolumn{2}{c}{ViT-L/14} \\
& Target ($\downarrow$) & Control ($\uparrow$) & Target ($\downarrow$) & Control ($\uparrow$) & Target ($\downarrow$) & Control ($\uparrow$) \\\midrule
Pre-trained & 48.3 & 63.4 & 55.2 & 68.3 & 64.8 & 75.5\\\midrule
Fine-tuned & 90.2 & 48.2 & 92.5 & 58.3 & 94.0 & 72.6 \\
Gradient ascent & 2.73 & 0.25 & 1.93 & 0.68 & 3.93 & 16.3\\
Random vector & 45.7 & 61.5 & 53.1 & 66.0 & 60.9 & 72.9\\\midrule
Negative task vector & 24.0 & 60.9 & 21.3 & 65.4 & 19.0 & 72.9\\
\bottomrule
\end{tabular}
|
212,465
|
2212.04089v3
| null |
\caption{\textbf{Making language models less toxic with negative task vectors.} Results are shown for the GPT-2 Large model. Negative task vectors decrease the amount of toxic generations by 6$\times$, while resulting in a model with comparable perplexity on a control task (WikiText-103). Additional details and results are shown in Appendix \ref{sec:appendix-neg-lang}.}
|
\label{tab:toxicity}
|
\begin{tabular}{lrrr}
\toprule
Method & \% toxic generations ($\downarrow$)& Avg. toxicity score ($\downarrow$) & WikiText-103 perplexity ($\downarrow$)
\\\midrule
Pre-trained & 4.8 & 0.06 & 16.4 \\\midrule
Fine-tuned & 57 & 0.56 & 16.6 \\
Gradient ascent & 0.0 & 0.45 & $>$10$^{10}$ \\
Fine-tuned on non-toxic & 1.8 & 0.03 & 17.2 \\
Random vector & 4.8 & 0.06 & 16.4 \\\midrule
Negative task vector & 0.8 & 0.01 & 16.9 \\\bottomrule
\end{tabular}
|
212,466
|
2212.04089v3
| null |
\caption{\textbf{Improving performance on target tasks with external task vectors.} For four text classification tasks from the GLUE benchmark, adding task vectors downloaded from the Hugging Face Hub can improve accuracy of fine-tuned T5 models. Appendix \ref{sec:appendix-add-lang} shows additional details.}
|
\label{tab:glue}
|
\begin{tabular}{lccccc}
\toprule
Method & MRPC & RTE & CoLA & SST-2 & Average \\\midrule
Zero-shot & 74.8 & 52.7 & 8.29 & 92.7 & 57.1 \\
Fine-tuned & 88.5 & 77.3 & 52.3 & 94.5 & 78.1 \\
Fine-tuned + task vectors & 89.3 \tiny{(+0.8)} & 77.5 \tiny{(+0.2)}& 53.0 \tiny{(+0.7)} & 94.7 \tiny{(+0.2)} & 78.6 \tiny{(+0.5)} \\
\bottomrule
\end{tabular}
|
212,467
|
2212.04089v3
| null |
\caption{\textbf{Improving domain generalization with task analogies.} Using an auxiliary task for which labeled data is available and unlabeled data from both the auxiliary and the target datasets, task analogies improve the accuracy for multiple T5 models and two sentiment analysis target tasks \citep{zhang2015character,mcauley2013hidden}, without using any labeled data from the target tasks.}
|
\label{tab:sentiment-analog}
|
\begin{tabular}{lcccccccc}
\toprule
& & \multicolumn{3}{c}{target = Yelp} & & \multicolumn{3}{c}{target = Amazon} \\\cmidrule{3-5}\cmidrule{7-9}
Method & & T5-small & T5-base & T5-large & & T5-small & T5-base & T5-large
\\\midrule
Fine-tuned on auxiliary & & 88.6 & 92.3 & 95.0 & & 87.9 & 90.8 & 94.8 \\
Task analogies & & 89.9 & 93.0 & 95.1 & & 89.0 & 92.7 & 95.2 \\
Fine-tuned on target & & 91.1 & 93.4 & 95.5 & & 90.2 & 93.2 & 95.5 \\
\bottomrule
\end{tabular}
|
212,468
|
2212.04089v3
| null |
\caption{Forgetting via negation on image classification tasks. Results are shown for a CLIP ViT-L/14 model \citep{radford2021learning}, reporting accuracy on both the target (T) and control (C) tasks.}
|
\label{tab:forget_image_l14}
|
\begin{tabular}{lcc?cc?cc?cc?cc?cc?cc?cc}
\toprule
\multirow{2}{*}{Method} & \multicolumn{2}{c?}{{Cars}} & \multicolumn{2}{c?}{DTD} & \multicolumn{2}{c?}{EuroSAT} & \multicolumn{2}{c?}{GTSRB} & \multicolumn{2}{c?}{MNIST} & \multicolumn{2}{c?}{{RESISC45}} & \multicolumn{2}{c?}{{SUN397}} & \multicolumn{2}{c}{{SVHN}} \\
& T$\downarrow$ & C$\uparrow$ & T$\downarrow$ & C$\uparrow$ & T$\downarrow$ & C$\uparrow$ & T$\downarrow$ & C$\uparrow$ & T$\downarrow$ & C$\uparrow$ & T$\downarrow$ & C$\uparrow$ & T$\downarrow$ & C$\uparrow$ & T$\downarrow$ & C$\uparrow$ \\\midrule
Pre-trained & 77.8 & 75.5 & 55.4 & 75.5 & 60.2 & 75.5 & 50.6 & 75.5 & 76.4 & 75.5 & 71.0 & 75.5 & 68.3 & 75.5 & 58.6 & 75.5 \\
Fine-tuned & 92.8 & 73.1 & 83.7 & 72.3 & 99.2 & 70.5 & 99.3 & 73.1 & 99.8 & 72.9 & 96.9 & 73.8 & 82.4 & 72.7 & 98.0 & 72.6 \\
Neg. gradients & 0.00 & 4.82 & 2.13 & 0.10 & 9.26 & 1.07 & 1.19 & 0.07 & 9.80 & 67.0 & 2.14 & 0.07 & 0.25 & 0.00 & 6.70 & 57.2 \\%\midrule
% Task vectors & & & & & & & & & & & & & & & & & & \\
Random vector & 72.0 & 73.3 & 52.1 & 72.2 & 59.7 & 73.5 & 43.4 & 72.5 & 74.8 & 72.8 & 70.8 & 73.0 & 66.9 & 72.7 & 47.1 & 72.9\\\midrule
Neg. task vector & 32.0 & 72.4 & 26.7 & 72.2 & 7.33 & 73.3 & 6.45 & 72.2 & 2.69 & 74.9 & 19.7 & 72.9 & 50.8 & 72.6 & 6.71 & 72.7 \\
\bottomrule
\end{tabular}
|
212,469
|
2212.04089v3
| null |
\caption{Forgetting via negation on image classification tasks. Results are shown for a CLIP ViT-B/16 model \citep{radford2021learning}, reporting accuracy on both the target (T) and control (C) tasks.}
|
\label{tab:forget_image_b16}
|
\begin{tabular}{lcc?cc?cc?cc?cc?cc?cc?cc}
\toprule
\multirow{2}{*}{Method} & \multicolumn{2}{c?}{{Cars}} & \multicolumn{2}{c?}{DTD} & \multicolumn{2}{c?}{EuroSAT} & \multicolumn{2}{c?}{GTSRB} & \multicolumn{2}{c?}{MNIST} & \multicolumn{2}{c?}{{RESISC45}} & \multicolumn{2}{c?}{{SUN397}} & \multicolumn{2}{c}{{SVHN}} \\
& T$\downarrow$ & C$\uparrow$ & T$\downarrow$ & C$\uparrow$ & T$\downarrow$ & C$\uparrow$ & T$\downarrow$ & C$\uparrow$ & T$\downarrow$ & C$\uparrow$ & T$\downarrow$ & C$\uparrow$ & T$\downarrow$ & C$\uparrow$ & T$\downarrow$ & C$\uparrow$\\\midrule
Pre-trained & 64.6 & 68.3 & 44.9 & 68.3 & 53.9 & 68.3 & 43.4 & 68.3 & 51.6 & 68.3 & 65.8 & 68.3 & 65.5 & 68.3 & 52.0 & 68.3 \\
Fine-tuned & 87.0 & 61.9 & 82.3 & 57.5 & 99.1 & 56.0 & 99.0 & 54.7 & 99.7 & 55.2 & 96.4 & 62.2 & 79.0 & 61.7 & 97.7 & 56.8 \\
Neg. gradients & 0.36 & 0.11 & 2.13 & 0.09 & 9.26 & 0.14 & 0.71 & 0.10 & 0.04 & 1.20 & 2.60 & 0.10 & 0.25 & 0.00 & 0.08 & 3.69\\
Rand. task vector & 61.0 & 65.6 & 43.9 & 66.3 & 51.7 & 66.2 & 43.1 & 65.0 & 51.6 & 68.3 & 63.6 & 65.6 & 63.7 & 65.2 & 46.2 & 65.5 \\\midrule
Neg. task vector & 30.8 & 65.4 & 26.5 & 65.6 & 12.3 & 65.8 & 9.53 & 65.8 & 9.55 & 65.4 & 26.5 & 65.1 & 48.6 & 65.1 & 6.43 & 65.4 \\
\bottomrule
\end{tabular}
|
212,485
|
2303.17264v1
| null |
\caption{Disentanglement metrics on MUG.}
|
\label{tab:disentanglement_metrics_mug}
|
\begin{tabular}{lcccc}
\toprule
Method & Acc$\uparrow$ & IS$\uparrow$ & $H(y|x){\downarrow}$ & $H(y){\uparrow}$ \\
\midrule
MoCoGAN & $63.12\%$ & $4.332$ & $0.183$ & $1.721$ \\
DSVAE & $54.29\%$ & $3.608$ & $0.374$ & $1.657$ \\
R-WAE & $71.25\%$ & $5.149$ & $0.131$ & $1.771$ \\
\midrule
% \rowcolor{LightGray}
% \rowcolor{LightGray}
S3VAE & $70.51\%$ & $5.136$ & $0.135$ & $1.760$ \\
% \rowcolor{LightGray}
C-DSVAE & $\boldsymbol{81.16\%}$ & $5.341$ & $0.092$ & $\boldsymbol{1.775}$ \\
\midrule
Ours & $77.45\%$ & $\boldsymbol{5.569}$ & $\boldsymbol{0.052}$ & $1.769$ \\
\bottomrule
\hline
\end{tabular}
|
212,470
|
2212.04089v3
| null |
\caption{Forgetting via negation on image classification tasks. Results are shown for a CLIP ViT-B/32 model \citep{radford2021learning}, reporting accuracy on both the target (T) and control (C) tasks.}
|
\label{tab:forget_image_b32}
|
\begin{tabular}{lcc?cc?cc?cc?cc?cc?cc?cc}
\toprule
\multirow{2}{*}{Method} & \multicolumn{2}{c?}{{Cars}} & \multicolumn{2}{c?}{DTD} & \multicolumn{2}{c?}{EuroSAT} & \multicolumn{2}{c?}{GTSRB} & \multicolumn{2}{c?}{MNIST} & \multicolumn{2}{c?}{{RESISC45}} & \multicolumn{2}{c?}{{SUN397}} & \multicolumn{2}{c}{{SVHN}} \\
& T$\downarrow$ & C$\uparrow$ & T$\downarrow$ & C$\uparrow$ & T$\downarrow$ & C$\uparrow$ & T$\downarrow$ & C$\uparrow$ & T$\downarrow$ & C$\uparrow$ & T$\downarrow$ & C$\uparrow$ & T$\downarrow$ & C$\uparrow$ & T$\downarrow$ & C$\uparrow$ \\\midrule
Pre-trained & 59.6 & 63.4 & 44.1 & 63.4 & 45.9 & 63.4 & 32.5 & 63.4 & 48.7 & 63.4 & 60.7 & 63.4 & 63.2 & 63.4 & 31.5 & 63.4 \\
Fine-tuned & 79.2 & 55.2 & 78.7 & 49.3 & 98.6 & 47.2 & 98.5 & 39.1 & 99.6 & 42.5 & 95.0 & 53.2 & 75.1 & 54.6 & 97.2 & 44.7 \\
Neg. gradients & 0.01 & 0.11 & 2.13 & 0.10 & 9.26 & 0.10 & 1.19 & 0.07 & 0.00 & 1.22 & 2.60 & 0.10 & 0.25 & 0.01 & 6.38 & 0.29 \\
Rand. task vector & 54.1 & 60.9 & 39.9 & 61.5 & 45.8 & 63.4 & 27.9 & 60.7 & 48.3 & 63.4 & 57.1 & 60.9 & 61.3 & 60.5 & 31.2 & 60.7 \\\midrule
Neg. task vector & 36.0 & 61.1 & 27.8 & 60.2 & 13.6 & 61.3 & 8.13 & 61.4 & 16.7 & 60.7 & 31.7 & 61.0 & 50.7 & 60.5 & 7.65 & 61.0 \\
\bottomrule
\end{tabular}
|
212,471
|
2212.04089v3
| null |
\caption{The effect of semantic overlap with the control task in forgetting experiments on image classification tasks. Results are shown for a CLIP ViT-L/14 model, reporting accuracy both on the target task and control task (Ctrl, ImageNet).}
|
\label{tab:overlap-ablation}
|
\begin{tabular}{lcc?cc?cc?cc}
\toprule
\multirow{2}{*}{Method} & \multicolumn{4}{c?}{{Without filtering}} & \multicolumn{4}{c}{With filtering} \\
& Cars ($\downarrow$) & Ctrl ($\uparrow$) & SUN397 ($\downarrow$) & Ctrl ($\uparrow$) & Cars ($\downarrow$) & Ctrl ($\uparrow$) & SUN397 ($\downarrow$) & Ctrl ($\uparrow$) \\\midrule
Pre-trained & 77.8 & 75.5 & 68.3 & 75.5 & 77.8 & 75.5 & 68.3 & 76.1 \\
Fine-tuned & 92.8 & 73.1 & 82.4 & 72.7 & 92.8 & 73.3 & 82.4 & 73.1 \\\midrule
Neg. task vector & 32.0 & 72.4 & 50.8 & 72.6 & 32.0 & 72.5 & 48.1 & 72.4\\
\bottomrule
\end{tabular}
|
212,472
|
2212.04089v3
| null |
\caption{Making language models less toxic with negative task vectors. Results are shown for the GPT-2 Medium model.}
|
\label{tab:toxicity_gpt2med}
|
\begin{tabular}{lrrr}
\toprule
Method & \% toxic generations ($\downarrow$)& Avg. toxicity score ($\downarrow$) & WikiText-103 perplexity ($\downarrow$)
\\\midrule
Pre-trained & 4.3 & 0.06 & 18.5 \\
Fine-tuned & 54.5 & 0.54 & 20.2 \\
Gradient ascent & 0.0 & 0.00 & $>$10$^{10}$ \\
Random task vector & 4.2 & 0.05 & 18.5 \\\midrule
Negative task vector & 1.8 & 0.02 & 18.9 \\\bottomrule
\end{tabular}
|
212,473
|
2212.04089v3
| null |
\caption{Making language models less toxic with negative task vectors. Results are shown for the GPT-2 Small model.}
|
\label{tab:toxicity_gpt2small}
|
\begin{tabular}{lrrr}
\toprule
Method & \% toxic generations ($\downarrow$)& Avg. toxicity score ($\downarrow$) & WikiText-103 perplexity ($\downarrow$)
\\\midrule
Pre-trained & 3.7 & 0.04 & 25.2 \\
Fine-tuned & 62.9 & 0.61 & 28.1 \\
Gradient ascent & 0.0 & 0.00 & $>$10$^{10}$ \\
Random task vector & 3.2 & 0.04 & 25.3 \\\midrule
Negative task vector & 2.5 & 0.03 & 25.3 \\\bottomrule
\end{tabular}
|
212,474
|
2212.04089v3
| null |
\caption{\textbf{Learning via analogy.} By leveraging vectors from related tasks, we can improve accuracy on four new target tasks without any training data, and with little change on control settings. Results are shown for the CLIP models \citep{radford2019language}, additional details are provided in Appendix \ref{sec:appendix-kingsandqueens}.}
|
\label{tab:kingsandqueens}
|
\begin{tabular}{lcccccccc}
\toprule
\multirow{2}{*}{Method} & \multicolumn{2}{c}{Queens} & \multicolumn{2}{c}{Kings} & \multicolumn{2}{c}{Woman} & \multicolumn{2}{c}{Men}
\\
& Target & Control & Target & Control & Target & Control & Target & Control \\
\midrule
ViT-B/32 & 0.00 & 63.4 & 0.00 & 63.4 & 0.00 & 63.4 & 0.00 & 63.4\\
\quad{+ task vectors} & 42.0 & 62.4 & 30.0 & 62.4 & 69.4 & 62.5 & 58.0 & 62.6\\\midrule
ViT-B/16 & 0.00 & 68.3 & 0.00 & 68.3 & 0.00 & 68.3 & 0.00 & 68.3 \\
\quad{+ task vectors} & 66.0 & 67.5 & 94.0 & 67.4 & 87.8 & 67.5 & 62.0 & 67.6 \\\midrule
ViT-L/14 & 0.00 & 75.5 & 0.00 & 75.5 & 0.00 & 75.5 & 0.00 & 75.5 \\
\quad{+ task vectors} & 100 & 74.7 & 100 & 74.5 & 100 & 74.6 & 96.0 & 74.6\\
\bottomrule
\end{tabular}
|
212,475
|
2212.04089v3
| null |
\caption{\textbf{Learning by analogy on subpopulations.} Results are shown for multiple CLIP models, as detailed in Section \ref{sec:appendix-sketches}.}
|
\label{tab:sketches}
|
\begin{tabular}{lccccccc}
\toprule
\multirow{2}{*}{Model} & Samples & \multirow{2}{*}{Task vectors} & \multicolumn{5}{c}{Accuracy} \\
& per class & & Sketches-A & Sketches-B & ImageNet-A & ImageNet-B & Average\\\midrule
\multirow{8}{*}{ViT-B/32} & 0 & \xmark & 0.712 & 0.677 & 0.861 & 0.923& 0.793 \\
& 0 & \cmark & 0.782 & 0.758 & 0.861 & 0.926& 0.832 \\
& 1 & \xmark & 0.754 & 0.758 & 0.868 & 0.919& 0.825 \\
& 1 & \cmark & 0.782 & 0.766 & 0.866 & 0.922& 0.834 \\
& 2 & \xmark & 0.768 & 0.778 & 0.868 & 0.919& 0.833 \\
& 2 & \cmark & 0.786 & 0.800 & 0.867 & 0.922& 0.844 \\
& 4 & \xmark & 0.810 & 0.780 & 0.871 & 0.926& 0.847 \\
& 4 & \cmark & 0.802 & 0.796 & 0.871 & 0.927& 0.849 \\\midrule
\multirow{8}{*}{ViT-B/16} & 0 & \xmark & 0.716 & 0.732 & 0.885 & 0.946& 0.820\\
& 0 & \cmark & 0.794 & 0.794 & 0.889 & 0.953& 0.858\\
& 1 & \xmark & 0.758 & 0.812 & 0.894 & 0.948& 0.853\\
& 1 & \cmark & 0.796 & 0.804 & 0.897 & 0.957& 0.863\\
& 2 & \xmark & 0.792 & 0.817 & 0.897 & 0.951& 0.865\\
& 2 & \cmark & 0.804 & 0.829 & 0.899 & 0.956& 0.872\\
& 4 & \xmark & 0.815 & 0.812 & 0.904 & 0.952& 0.871\\
& 4 & \cmark & 0.831 & 0.825 & 0.904 & 0.953& 0.878\\\midrule
\multirow{8}{*}{ViT-L/14} & 0 & \xmark & 0.823 & 0.831 & 0.913 & 0.962& 0.882\\
& 0 & \cmark & 0.879 & 0.861 & 0.922 & 0.968& 0.908\\
& 1 & \xmark & 0.845 & 0.863 & 0.923 & 0.971& 0.900\\
& 1 & \cmark & 0.879 & 0.863 & 0.930 & 0.973& 0.911\\
& 2 & \xmark & 0.865 & 0.881 & 0.925 & 0.973& 0.911\\
& 2 & \cmark & 0.875 & 0.881 & 0.932 & 0.975& 0.916\\
& 4 & \xmark & 0.875 & 0.883 & 0.934 & 0.973& 0.916\\
& 4 & \cmark & 0.903 & 0.887 & 0.941 & 0.975& 0.927\\
\bottomrule
\end{tabular}
|
212,476
|
2110.09022v3
| null |
\caption{Comparison of test accuracies of each method under instance-based label noise. CE (unfixed $f$ with random init.) is a common baseline adopted in the literature which trains a random initialized DNN using CE loss on noisy dataset. }
|
\label{table:instance}
|
\begin{tabular}{c|cccccc}
\hline
\multirow{2}{*}{Method} & \multicolumn{3}{c}{\emph{Inst. CIFAR10} } & \multicolumn{3}{c}{\emph{Inst. CIFAR100} } \\
% \cline{3-8}
& $\varepsilon = 0.2$&$\varepsilon = 0.4$&$\varepsilon = 0.6$ & $\varepsilon = 0.2$&$\varepsilon = 0.4$&$\varepsilon = 0.6$\\
\hline\hline
CORES \citep{sieve2020} &89.50 &82.84 & 79.66 & 61.25 & 47.81 &37.85\\
CE (unfixed $f$ with random init.)&87.16&75.16 & 44.64 &58.72 & 41.14 & 25.29\\
CE (fixed $f$, no sampling) &88.74 & 75.71 & 25.7 & 59.38 & 46.13 &24.75\\
CE (fixed $f$, down-sampling) &\textbf{90.12} & \textbf{84.19} & \textbf{82.06} & \textbf{62.88} & \textbf{61.1} &\textbf{58.93}\\
\hline
\end{tabular}
|
212,477
|
2110.09022v3
| null |
\caption{Comparison of test accuracies with each method on CIFAR10. The model is \textit{learned from scratch} for all methods with $\lambda = 1$. Best and last epoch accuracies are reported: best/last.}
|
\label{cifar10_reg}
|
\begin{tabular}{c|ccc}
\hline
\multirow{2}{*}{Method} & \multicolumn{2}{c}{\emph{Symm. CIFAR10} } & \multicolumn{1}{c}{\emph{Asymm. CIFAR10} } \\
% \cline{3-8}
& $\varepsilon = 0.6$ & $\varepsilon = 0.8$& $\varepsilon = 0.4$ \\
\hline
CE & 61.29/32.83 & 38.46/15.05 & 67.28/56.6\\
CE + Regularizer & \textbf{69.02}/\textbf{65.13} & \textbf{61.94}/\textbf{56.78}&\textbf{73.38}/\textbf{58.51}\\
\hline
GCE \citep{zhang2018generalized} & 72.56/62.84 & 40.71/20.53 & 69.19/53.24\\
GCE + Regularizer & \textbf{72.61/68.38} & \textbf{63.63/63.05}&\textbf{69.79.61.32}\\
\hline
FW \citep{patrini2017making} & 65.95/60.13 & 40.08/26.7 & 68.62/58.01\\
FW + Regularizer & \textbf{68.73/65.90} & \textbf{60.94/59.88}&\textbf{75.64/67.66}\\
\hline
HOC \citep{zhu2021clusterability} & 62.53/46.17 & 39.61/16.90 & \textbf{85.88}/78.89\\
HOC + Regularizer & \textbf{70.07/66.94} & \textbf{60.9/34.90}&83.53/\textbf{82.56}\\
\hline
Peer Loss \citep{liu2020peer} & 77.52/\textbf{76.07} & 15.60/10.00&\textbf{84.47}/68.93\\
Peer Loss + Regularizer & \textbf{77.61}/73.26&\textbf{61.64/53.52}&81.58/\textbf{75.38}\\
\hline
\end{tabular}
|
212,478
|
2110.09022v3
| null |
\caption{Test accuracy for each method on CIFAR10N and CIFAR100N. }
|
\label{table:cifarn}
|
\begin{tabular}{c|ccccccc}
\hline
& CE & GCE &Co-Teaching+&Peer Loss&JoCoR& ELR& CE + Regularizer \\
\hline
CIFAR10N (Worst)& 77.69& 80.66& 83.26& 82.53& 83.37&83.58&\textbf{88.74} \\
CIFAR100N&55.50&56.73&57.88&57.59&59.97&58.94&\textbf{60.81}\\
\hline
\end{tabular}
|
212,479
|
2110.09022v3
| null |
\caption{Test accuracy for each method on Clothing1M dataset. All the methods use ResNet50 backbones. DS: Down-Sampling. Reg: With structural regularizer. }
|
\label{table:clothing1m}
|
\begin{tabular}{c|cccc|ccc}
\hline
& Foward-T & Co-teaching &CORES+DS&ELR+DS&CE& CE + DS& CE + DS + Reg \\
\hline
Initializer& ImageNet& ImageNet& ImageNet& ImageNet& SimCLR& SimCLR&SimCLR \\
Accuracy&70.83&69.21&73.24&72.87&70.90&72.95&\textbf{73.48}\\
\hline
\end{tabular}
|
212,480
|
2110.09022v3
| null |
\caption{Test accuracy for each method on Clothing1M dataset. All the methods use ResNet50 backbones. SimCLR is adopted for SSL -pretraining. DS: Down-Sampling. Reg: With structural regularizer. \zzw{}{this table takes too much space. Consider to make it in three rows? Row1: method, Row2 Initializer, Row3 accuracy.}}
|
\label{table:clothing1m}
|
\begin{tabular}{ccc}
\hline
Method & Initializer & Accuracy \\
\hline
Foward-T & ImageNet &70.83\\
Co-teaching & ImageNet &69.21\\
CORES & ImageNet &73.24\\
ELR & ImageNet &72.87\\
\hline
CE & SimCLR & 70.90\\
CE + DS & SimCLR & 72.95\\ % 73.52. 72.95 use SGD
CE + DS + Reg & SimCLR & 73.48\\ % tuning parameters
% \zzw{imagenet?}{} \\
% \zzw{+ warmup?} \\
% best 73.48, last 72.63. reg: rkd_dis, alpha: 0.2, beta: 0.2, gamma: 0.2, lr_plan: 1, optimizer: sgd
% best 73.42, last 72.63. reg: rkd_dis, alpha: 0.2, beta: 0.2, gamma: 1, lr_plan: 1, optimizer: sgd
% best 73.38, last 72.82. reg: rkd_dis, alpha: 0.2, beta: 0.2, gamma: 5, lr_plan: 1, optimizer: sgd
% best 73.44, last 72.64. reg: rkd_dis, alpha: 0.2, beta: 1, gamma: 0.2, lr_plan: 1, optimizer: sgd
% best 73.29, last 72.42. reg: rkd_dis, alpha: 0.2, beta: 1, gamma: 1, lr_plan: 1, optimizer: sgd
% [Imagenet] best 72.49, last 5 avg 71.81333333333333. tau: 0.5, warm: 1, reg: rkd_dis, alpha: 0.2, beta: 0.2, gamma: 5, lr_plan: 1, optimizer: sgd
\hline
\end{tabular}
|
212,481
|
2110.09022v3
| null |
\caption{Comparison of different methods for dealing with label noise.}
|
\label{method_compare}
|
\begin{tabular}{c|c|c|c}
\hline
Methods & \thead{constrain the \\logits with the label} & \thead{constrain the \\model representation} & \thead{performance on \\high noise regime}\\
\hline
Robust loss function & $\checkmark$ & $\times$ & $\times$ \\
\hline
Loss correction & $\checkmark$ & $ \times$ & $ \times$ \\
\hline
Loss re-weighting & $\checkmark$ & $ \times$ & $ \times$ \\
\hline
Our framework & $\checkmark$ & $ \checkmark$ & $ \checkmark$ \\
\hline
\end{tabular}
|
212,482
|
2110.09022v3
| null |
\caption{Comparing different SSL methods on CIFAR10 with symmetric label noise}
|
\label{table:exp_moco}
|
\begin{tabular}{c|ccccc}
\hline
\multirow{2}{*}{Method} & \multicolumn{4}{c}{\emph{Symm label noise ratio}} \\
% \cline{2-7}
& 0.2&0.4&0.6&0.8\\
\hline
\hline
CE (fixed encoder with SimCLR init) & 91.06&90.73 &90.2 &88.24 \\
CE (fixed encoder with MoCo init)&
91.55&91.12 &90.45 &88.51\\
\hline
\end{tabular}
|
212,483
|
2210.00750v2
| null |
\caption{Suboptimality gaps for different algorithms with differentiable function class \ref{def:function_class}. Here we omit the higher order term for clear comparison. With Concentrability, we can only achieve the worst case bound that does not explicit depend on the function model $f$. With the stronger uniform coverage \ref{assum:cover}, better instance-dependent characterizations become available. Here $C_\mathrm{eff}$ is in \ref{assume:con_co}, $\Sigma^\star$ in \ref{thm:PFQL}, $\Lambda^\star$ in \ref{thm:VAFQL} and $\epsilon_{\mathcal{F}}$ in \ref{assum:R+BC}.}
|
\label{tab: result for comparison}
|
\begin{tabular}{|c|c|c|c|}
\hline
Algorithm & Assumption & Suboptimality Gap $v^\star-v^{\widehat{\pi}}$\\%[5pt]
\hline
\hline
VFQL, Theorem~\ref{thm:VFQL} & Concentrability \ref{assume:con_co} & $\sqrt{C_{\mathrm{eff}}}H\cdot\sqrt{\frac{H^2d+\lambda C^2_\Theta}{K}}+\sqrt[\frac{1}{4}]{\frac{H^3d\epsilon_{\mathcal{F}}}{K}} +\sqrt{C_\mathrm{eff} H^3 \epsilon_{\mathcal{F}}}+H\epsilon_{\mathcal{F}}$ \\%[1mm]
\hline
PFQL, Theorem~\ref{thm:PFQL} & Uniform Coverage \ref{assum:cover} & $\sum_{h=1}^H 16 dH\cdot \E_{\pi^\star}\left[\sqrt{\nabla^\top_\theta f({\theta}^\star_h,\phi(s_h,a_h))\Sigma^{\star-1}_h\nabla_\theta f({\theta}^\star_h,\phi(s_h,a_h))}\right]$\\
\hline
VAFQL, Theorem~\ref{thm:VAFQL} & Uniform Coverage \ref{assum:cover}& $ 16 d\cdot\sum_{h=1}^H \E_{\pi^\star}\left[\sqrt{\nabla^\top_\theta f({\theta}^\star_h,\phi(s_h,a_h))\Lambda^{\star-1}_h\nabla_\theta f({\theta}^\star_h,\phi(s_h,a_h))}\right]$\\
\hline
\end{tabular}
|
212,484
|
2303.17264v1
| null |
\caption{Accuracy measures of factorial swap experiments.}
|
\label{tab:factorial_swap}
|
\begin{tabular}{lccccc}
\toprule
Test & action & skin & top & pants & hair \\
\midrule
hair swap & $10.51\%$ & $16.25\%$ & $16.33\%$ & $35.51\%$ & $\boldsymbol{90.59\%}$ \\
skin swap & $10.55\%$ & $\boldsymbol{73.01\%}$ & $16.29\%$ & $30.55\%$ & $17.70\%$ \\
\bottomrule
\hline
\end{tabular}
|
212,567
|
2210.05359v1
| null |
\caption{The effect of using different sizes of text-to-code models (T2C) with GPT-3 175B/1.3B as the foundation model (FM) in the zero-shot and few-shot settings.}
|
\label{tab:text-to-code}
|
\begin{tabular}{@{}lll@{}}
\toprule
LM Size (T2C + FM) & Zero-shot & Few-shot \\ \midrule
0.3B + 1.3B & 25.1 $_{1.77}$ & 43.3 $_{1.12}$ \\
0.3B + 175B & 48.3 $_{1.61}$ & 82.1 $_{0.89}$ \\
1.5B + 1.3B & 29.8 $_{1.65}$ & 46.7 $_{0.82}$ \\ \midrule
1.5B + 175B (\textit{default}) & 51.9 $_{1.53}$ & 84.2 $_{0.79}$ \\
\bottomrule
\end{tabular}
|
212,486
|
2303.17264v1
| null |
\caption{Disentanglement metrics on Sprites.}
|
\label{tab:disentanglement_metrics_sprites}
|
\begin{tabular}{lcccc}
\toprule
Method & Acc$\uparrow$ & IS$\uparrow$ & $H(y|x){\downarrow}$ & $H(y){\uparrow}$ \\
\midrule
MoCoGAN & $92.89\%$ & $8.461$ & $0.090$ & $2.192$ \\
DSVAE & $90.73\%$ & $8.384$ & $0.072$ & $2.192$ \\
R-WAE & $98.98\%$ & $8.516$ & $0.055$ & $2.197$ \\
\midrule
% \rowcolor{LightGray}
S3VAE & $99.49\%$ & $8.637$ & $0.041$ & $2.197$ \\
% \rowcolor{LightGray}
C-DSVAE & $99.99\%$ & $8.871$ & $0.014$ & $2.197$ \\
\midrule
Ours & $\boldsymbol{100\%}$ & $\boldsymbol{8.999}$ & $\boldsymbol{\expnumber{1.6}{-7}}$ & $\boldsymbol{2.197}$ \\
\bottomrule
\hline
\end{tabular}
|
212,487
|
2303.17264v1
| null |
\caption{Disentanglement metrics on MUG.}
|
\label{tab:disentanglement_metrics_mug}
|
\begin{tabular}{lcccc}
\toprule
Method & Acc$\uparrow$ & IS$\uparrow$ & $H(y|x){\downarrow}$ & $H(y){\uparrow}$ \\
\midrule
MoCoGAN & $63.12\%$ & $4.332$ & $0.183$ & $1.721$ \\
DSVAE & $54.29\%$ & $3.608$ & $0.374$ & $1.657$ \\
R-WAE & $71.25\%$ & $5.149$ & $0.131$ & $1.771$ \\
\midrule
% \rowcolor{LightGray}
% \rowcolor{LightGray}
S3VAE & $70.51\%$ & $5.136$ & $0.135$ & $1.760$ \\
% \rowcolor{LightGray}
C-DSVAE & $\boldsymbol{81.16\%}$ & $5.341$ & $0.092$ & $\boldsymbol{1.775}$ \\
\midrule
Ours & $77.45\%$ & $\boldsymbol{5.569}$ & $\boldsymbol{0.052}$ & $1.769$ \\
\bottomrule
\hline
\end{tabular}
|
212,488
|
2303.17264v1
| null |
\caption{Disentanglement metrics on TIMIT.}
|
\label{tab:disentanglement_metrics_timit}
|
\begin{tabular}{lccc|cccc|c}
\toprule
Method & FHVAE & DSVAE & R-WAE & S3VAE & C-DSVAE$^\ast$ & C-DSVAE$^\dagger$ & C-DSVAE & Ours \\ %[0.5ex]
\midrule
Static EER$\downarrow$ & $5.06\%$ & $5.65\%$ & $4.73\%$ & $5.02\%$ & $5.09\%$ & $4.31\%$ & $\boldsymbol{4.03\%}$ & $4.46\%$ \\
Dynamic EER$\uparrow$ & $22.77\%$ & $19.20\%$ & $23.41\%$ & $25.51\%$ & $24.30\%$ & $31.09\%$ & $\boldsymbol{31.81\%}$ & $26.78\%$ \\
\bottomrule
\hline
\end{tabular}
|
212,489
|
2303.17264v1
| null |
\caption{Architecture details.}
|
\label{tab:arch}
|
\begin{tabular}{l|l}
\toprule
Encoder & Decoder \\
\midrule
$64 \times64 \times 3$ image & Z \\
Conv2D$(3,32,4,2,1) \rightarrow$ BN2D$(32) \rightarrow$ LeakyReLU & LSTM$(k, h)$ \\
Conv2D$(32,64,4,2,1) \rightarrow$ BN2D$(64) \rightarrow$ LeakyReLU & Conv2DT$(h, 256, 4, 1, 0) \rightarrow$ BN2D$(256) \rightarrow$ LeakyReLU \\
Conv2D$(64,128, 4,2,1) \rightarrow$ BN2D$(128) \rightarrow$ LeakyReLU & Conv2DT$(256, 128, 4, 1, 0) \rightarrow$ BN2D$(128) \rightarrow$ LeakyReLU \\
Conv2D$(128, 256, 4,2,1) \rightarrow$ BN2D$(256) \rightarrow$ LeakyReLU & Conv2DT$(128, 64, 4, 1, 0) \rightarrow$ BN2D$(64) \rightarrow$ LeakyReLU \\
Conv2D$(256, k, 4,2,1) \rightarrow$ BN2D$(k) \rightarrow$ LeakyReLU & Conv2DT$(64, 32, 4, 1, 0) \rightarrow$ BN2D$(32) \rightarrow$ LeakyReLU \\
LSTM$(k, k)$ & Conv2DT$(32, 3, 4, 1, 0) \rightarrow$ Sigmoid \\
\bottomrule
\end{tabular}
|
212,490
|
2303.17264v1
| null |
\caption{Hyperparameter details.}
|
\label{tab:hyp}
|
\begin{tabular}{lcccc|ccccc}
\toprule
Dataset & $b$ & $k$ & $h$ & $\#$epochs & $\lambda_\mathrm{rec}$ & $\lambda_\mathrm{pred}$ & $\lambda_\mathrm{eig}$ & $k_s$ & $\epsilon$ \\
\midrule
Sprites & $32$ & $40$ & $40$ & $800$ & $15$ & $1$ & $1$ & $8$ & $0.5$ \\
MUG & $16$ & $40$ & $100$ & $1000$ & $20$ & $1$ & $1$ & $5$ & $0.5$ \\
TIMIT & $30$ & $165$ & - & $400$ & $15$ & $3$ & $1$ & $15$ & $0$ \\
\bottomrule
\end{tabular}
|
212,491
|
2303.17264v1
| null |
\caption{Accuracy measures of factorial swap experiments, see Tab.~\ref{tab:factorial_swap}.}
|
\label{tab:factorial_swap_app}
|
\begin{tabular}{>{}l>{}c>{}c>{}c>{}c>{}c}
\toprule
Test & action & skin & top & pants & hair \\
\midrule
hair swap & $11.35\% \pm 0.65\%$ & $17.40\% \pm 0.79\%$ & $17.07\% \pm 0.77\%$ & $36.29\% \pm 0.88\%$ & $\boldsymbol{90.20\% \pm 0.52\%}$ \\
skin swap & $11.35\% \pm 0.65\%$ & $\boldsymbol{72.72\% \pm 0.68\%}$ & $17.23\% \pm 0.89\%$ & $31.22\% \pm 0.84\%$ & $16.92\% \pm 0.77\%$ \\
\bottomrule
\hline
\end{tabular}
|
212,492
|
2303.17264v1
| null |
\caption{Disentanglement metrics on Sprites and MUG, see Tabs.~\ref{tab:disentanglement_metrics_sprites} \ref{tab:disentanglement_metrics_mug}.}
|
\label{tab:disentanglement_metrics_app}
|
\begin{tabular}{>{}l>{}c>{}c>{}c>{}c}
\toprule
Method & Acc$\uparrow$ & IS$\uparrow$ & $H(y|x){\downarrow}$ & $H(y){\uparrow}$ \\
\midrule
Sprites & $\boldsymbol{100\% \pm 0\%}$ & $\boldsymbol{8.999 \pm \expnumber{2.3}{-6}}$ & $\boldsymbol{\expnumber{1.6}{-7} \pm \expnumber{2.2}{-7}}$ & $\boldsymbol{2.197 \pm 0}$ \\
\midrule
MUG & $77.45\% \pm 0.62\%$ & $\boldsymbol{5.569 \pm 0.026}$ & $\boldsymbol{0.052 \pm 0.004}$ & $1.769 \pm 0$ \\
\bottomrule
\end{tabular}
|
212,493
|
2303.17264v1
| null |
\caption{Computational resources comparison.}
|
\label{tab:resources}
|
\begin{tabular}{lcc|cc|c}
\toprule
Method & DSVAE & R-WAE & S3VAE & C-DSVAE & Ours \\
\midrule
Type & unsupervised & (weakly) unsupervised & self-supervised & self-supervised & unsupervised \\
Params & 21M & 121M & 11M & 11M & 2M \\
Data & - & labels & supervisory signals & data augmentation ($\times 16$) & - \\
\bottomrule
\hline
\end{tabular}
|
212,494
|
2305.13656v1
| null |
\caption{A summary of dataset statistics. }
|
\label{tab::dataset}
|
\begin{tabular}{ccccccc}
\toprule
& \#Nodes & \#Edges & \#Attrs & Avg. degree & Density\\
\midrule
\textsc{Cora} & 2,708 & 5,278 & 1,433 & 3.90 & 0.14\% \\
\textsc{CiteSeer} & 3,327 & 4,552 & 3,703 & 2.74 & 0.08\% \\
\textsc{PubMed} & 19,717 & 44,324 & 500 & 4.50 & 0.02\% \\
\textsc{Photo} & 7,650 & 119,081 & 745 & 31.13 & 0.41\%\\
\textsc{Computers} & 13,752 & 245,861 & 767 & 35.76 & 0.26\% \\
\bottomrule
\end{tabular}
|
212,495
|
2305.13656v1
| null |
\caption{Link prediction performance comparison (mean ± std AP). Gelato consistently outperforms GNN-based methods, topological heuristics, and two-stage approaches combining attributes and topology. }
|
\label{tab::performance_ap}
|
\begin{tabular}{ccccccc}
\toprule
& & \textsc{Cora} & \textsc{CiteSeer} & \textsc{PubMed} & \textsc{Photo} & \textsc{Computers} \\
\midrule
\multirow{9}[2]{*}{GNN}
& GAE & 0.27 ± 0.02 & 0.66 ± 0.11 & 0.26 ± 0.03 & 0.28 ± 0.02 & 0.30 ± 0.02 \\
& SEAL & 1.89 ± 0.74 & 0.91 ± 0.66 & *** & 10.49 ± 0.86 & 6.84\tnote{*} \\
& HGCN & 0.82 ± 0.03 & 0.74 ± 0.10 & 0.35 ± 0.01 & 2.11 ± 0.10 & 2.30 ± 0.14 \\
& LGCN & 1.14 ± 0.04 & 0.86 ± 0.09 & 0.44 ± 0.01 & 3.53 ± 0.05 & 1.96 ± 0.03 \\
& TLC-GNN & 0.29 ± 0.09 & 0.35 ± 0.18 & OOM & 1.77 ± 0.11 & OOM \\
& Neo-GNN & 2.05 ± 0.61 & 1.61 ± 0.36 & 1.21 ± 0.14 & 10.83 ± 1.53 & 6.75\tnote{*} \\
& NBFNet & 1.36 ± 0.17 & 0.77 ± 0.22 & *** & 11.99 ± 1.60 & *** \\
& BScNets & 0.32 ± 0.08 & 0.20 ± 0.06 & 0.22 ± 0.08 & 2.47 ± 0.18 & 1.45 ± 0.10 \\
& WalkPool & 2.04 ± 0.07 & 1.39 ± 0.11 & 1.31\tnote{*} & OOM & OOM \\
\midrule
\multicolumn{1}{c}{\multirow{4}[2]{*}{\parbox{1.8cm}{\centering Topological Heuristics}}}
& CN & 1.10 ± 0.00 & 0.74 ± 0.00 & 0.36 ± 0.00 & 7.73 ± 0.00 & 5.09 ± 0.00 \\
& AA & 2.07 ± 0.00 & 1.24 ± 0.00 & 0.45 ± 0.00 & 9.67 ± 0.00 & 6.52 ± 0.00 \\
& RA & 2.02 ± 0.00 & 1.19 ± 0.00 & 0.33 ± 0.00 & 10.77 ± 0.00 & 7.71 ± 0.00 \\
& AC & 2.43 ± 0.00 & 2.65 ± 0.00 & 2.50 ± 0.00 & 16.63 ± 0.00 & 11.64 ± 0.00 \\
\midrule
\multicolumn{1}{c}{\multirow{5}[2]{*}{\parbox{1.8cm}{\centering Attributes + Topology}}}
& MLP & 0.30 ± 0.05 & 0.44 ± 0.09 & 0.14 ± 0.06 & 1.01 ± 0.26 & 0.41 ± 0.23 \\
& Cos & 0.42 ± 0.00 & 1.89 ± 0.00 & 0.07 ± 0.00 & 0.11 ± 0.00 & 0.07 ± 0.00 \\
& MLP+AC & 3.24 ± 0.03 & 1.95 ± 0.05 & 2.61 ± 0.06 & 15.99 ± 0.21 & 11.25 ± 0.13 \\
& Cos+AC & 3.60 ± 0.00 & 4.46 ± 0.00 & 0.51 ± 0.00 & 10.01 ± 0.00 & 5.20 ± 0.00 \\
& MLP+Cos+AC& 3.39 ± 0.06 & 4.15 ± 0.14 & 0.55 ± 0.03 & 10.88 ± 0.09 & 5.75 ± 0.11 \\
\midrule
\multicolumn{2}{c}{Gelato} & \textbf{3.90 ± 0.03} & \textbf{4.55 ± 0.02} & \textbf{2.88 ± 0.09} & \textbf{25.68 ± 0.53} & \textbf{18.77 ± 0.19} \\
\bottomrule
\end{tabular}
|
212,496
|
2305.13656v1
| null | null | null |
\begin{tabular}{ccccccc}
\toprule
& & \textsc{Cora} & \textsc{CiteSeer} & \textsc{PubMed} & \textsc{Photo} & \textsc{Computers} \\
\midrule
\multirow{9}[2]{*}{GNN}
& GAE & 0.27 ± 0.02 & 0.66 ± 0.11 & 0.26 ± 0.03 & 0.28 ± 0.02 & 0.30 ± 0.02 \\
& SEAL & 1.89 ± 0.74 & 0.91 ± 0.66 & *** & 10.49 ± 0.86 & 6.84\tnote{*} \\
& HGCN & 0.82 ± 0.03 & 0.74 ± 0.10 & 0.35 ± 0.01 & 2.11 ± 0.10 & 2.30 ± 0.14 \\
& LGCN & 1.14 ± 0.04 & 0.86 ± 0.09 & 0.44 ± 0.01 & 3.53 ± 0.05 & 1.96 ± 0.03 \\
& TLC-GNN & 0.29 ± 0.09 & 0.35 ± 0.18 & OOM & 1.77 ± 0.11 & OOM \\
& Neo-GNN & 2.05 ± 0.61 & 1.61 ± 0.36 & 1.21 ± 0.14 & 10.83 ± 1.53 & 6.75\tnote{*} \\
& NBFNet & 1.36 ± 0.17 & 0.77 ± 0.22 & *** & 11.99 ± 1.60 & *** \\
& BScNets & 0.32 ± 0.08 & 0.20 ± 0.06 & 0.22 ± 0.08 & 2.47 ± 0.18 & 1.45 ± 0.10 \\
& WalkPool & 2.04 ± 0.07 & 1.39 ± 0.11 & 1.31\tnote{*} & OOM & OOM \\
\midrule
\multicolumn{1}{c}{\multirow{4}[2]{*}{\parbox{1.8cm}{\centering Topological Heuristics}}}
& CN & 1.10 ± 0.00 & 0.74 ± 0.00 & 0.36 ± 0.00 & 7.73 ± 0.00 & 5.09 ± 0.00 \\
& AA & 2.07 ± 0.00 & 1.24 ± 0.00 & 0.45 ± 0.00 & 9.67 ± 0.00 & 6.52 ± 0.00 \\
& RA & 2.02 ± 0.00 & 1.19 ± 0.00 & 0.33 ± 0.00 & 10.77 ± 0.00 & 7.71 ± 0.00 \\
& AC & 2.43 ± 0.00 & 2.65 ± 0.00 & 2.50 ± 0.00 & 16.63 ± 0.00 & 11.64 ± 0.00 \\
\midrule
\multicolumn{1}{c}{\multirow{5}[2]{*}{\parbox{1.8cm}{\centering Attributes + Topology}}}
& MLP & 0.30 ± 0.05 & 0.44 ± 0.09 & 0.14 ± 0.06 & 1.01 ± 0.26 & 0.41 ± 0.23 \\
& Cos & 0.42 ± 0.00 & 1.89 ± 0.00 & 0.07 ± 0.00 & 0.11 ± 0.00 & 0.07 ± 0.00 \\
& MLP+AC & 3.24 ± 0.03 & 1.95 ± 0.05 & 2.61 ± 0.06 & 15.99 ± 0.21 & 11.25 ± 0.13 \\
& Cos+AC & 3.60 ± 0.00 & 4.46 ± 0.00 & 0.51 ± 0.00 & 10.01 ± 0.00 & 5.20 ± 0.00 \\
& MLP+Cos+AC& 3.39 ± 0.06 & 4.15 ± 0.14 & 0.55 ± 0.03 & 10.88 ± 0.09 & 5.75 ± 0.11 \\
\midrule
\multicolumn{2}{c}{Gelato} & \textbf{3.90 ± 0.03} & \textbf{4.55 ± 0.02} & \textbf{2.88 ± 0.09} & \textbf{25.68 ± 0.53} & \textbf{18.77 ± 0.19} \\
\bottomrule
\end{tabular}
|
212,497
|
2305.13656v1
| null | null | null | null |
212,498
|
2305.13656v1
| null |
\caption{Results of the ablation study based on AP scores. Each component of Gelato plays an important role in enabling state-of-the-art link prediction performance. }
|
\label{tab::ablation}
|
\begin{tabular}{lccccc}
\toprule
& \textsc{Cora} & \textsc{CiteSeer} & \textsc{PubMed} & \textsc{Photo} & \textsc{Computers} \\
\midrule
\emph{Gelato$-$MLP} & 2.43 ± 0.00 & 2.65 ± 0.00 & 2.50 ± 0.00 & 16.63 ± 0.00 & 11.64 ± 0.00 \\
\emph{Gelato$-$AC} & 1.94 ± 0.18 & 3.91 ± 0.37 & 0.83 ± 0.05 & 7.45 ± 0.44 & 4.09 ± 0.16 \\
\emph{Gelato$-$NP+UT} & 2.98 ± 0.20 & 1.96 ± 0.11 & 2.35 ± 0.24 & 14.87 ± 1.41 & 9.77 ± 2.67 \\
\emph{Gelato$-$NP} & 1.96 ± 0.01 & 1.77 ± 0.20 & 2.32 ± 0.16 & 19.63 ± 0.38 & 9.84 ± 4.42 \\
\emph{Gelato$-$UT} & 3.07 ± 0.01 & 1.95 ± 0.05 & 2.52 ± 0.09 & 23.66 ± 1.01 & 11.59 ± 0.35 \\
\emph{Gelato} & \textbf{3.90 ± 0.03} & \textbf{4.55 ± 0.02} & \textbf{2.88 ± 0.09} & \textbf{25.68 ± 0.53} & \textbf{18.77 ± 0.19} \\
\bottomrule
\end{tabular}
|
212,499
|
2305.13656v1
| null | null | null |
\begin{tabular}{lccccc}
\toprule
& \textsc{Cora} & \textsc{CiteSeer} & \textsc{PubMed} & \textsc{Photo} & \textsc{Computers} \\
\midrule
\emph{Gelato$-$MLP} & 2.43 ± 0.00 & 2.65 ± 0.00 & 2.50 ± 0.00 & 16.63 ± 0.00 & 11.64 ± 0.00 \\
\emph{Gelato$-$AC} & 1.94 ± 0.18 & 3.91 ± 0.37 & 0.83 ± 0.05 & 7.45 ± 0.44 & 4.09 ± 0.16 \\
\emph{Gelato$-$NP+UT} & 2.98 ± 0.20 & 1.96 ± 0.11 & 2.35 ± 0.24 & 14.87 ± 1.41 & 9.77 ± 2.67 \\
\emph{Gelato$-$NP} & 1.96 ± 0.01 & 1.77 ± 0.20 & 2.32 ± 0.16 & 19.63 ± 0.38 & 9.84 ± 4.42 \\
\emph{Gelato$-$UT} & 3.07 ± 0.01 & 1.95 ± 0.05 & 2.52 ± 0.09 & 23.66 ± 1.01 & 11.59 ± 0.35 \\
\emph{Gelato} & \textbf{3.90 ± 0.03} & \textbf{4.55 ± 0.02} & \textbf{2.88 ± 0.09} & \textbf{25.68 ± 0.53} & \textbf{18.77 ± 0.19} \\
\bottomrule
\end{tabular}
|
212,500
|
2305.13656v1
| null |
\caption{Selected hyperparameters of Gelato. }
|
\label{tab::hyperparameters}
|
\begin{tabular}{cccccc}
\toprule
& \textsc{Cora} & \textsc{CiteSeer} & \textsc{PubMed} & \textsc{Photo} & \textsc{Computers} \\
\midrule
$\eta$ & 0.5 & 0.75 & 0.0 & 0.0 & 0.0\\ %
$\alpha$ & 0.5 & 0.5 & 0.0 & 0.0 & 0.0\\ %
$\beta$ & 0.25 & 0.5 & 1.0 & 1.0 & 1.0\\ %
\bottomrule
\end{tabular}
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.