Datasets:

nbalepur

/

DynaMiTE

like 0

The Transformer is an extremely powerful and prominent deep learning architecture. In this work, we challenge the commonly held belief in deep learning that going deeper is better, and show an alternative design approach that is building wider attention Transformers. We demonstrate that wide single layer Transformer models can compete with or outperform deeper ones in a variety of Natural Language Processing (NLP) tasks when both are trained from scratch. The impact of changing the model aspect ratio on Transformers is then studied systematically. This ratio balances the number of layers and the number of attention heads per layer while keeping the total number of attention heads and all other hyperparameters constant. On average, across 4 NLP tasks and 10 attention types, single layer wide models perform 0.3% better than their deep counterparts. We show an in-depth evaluation and demonstrate how wide models require a far smaller memory footprint and can run faster on commodity hardware, in addition, these wider models are also more interpretable. For example, a single layer Transformer on the IMDb byte level text classification has 3.1x faster inference latency on a CPU than its equally accurate deeper counterpart, and is half the size. Our results suggest that the critical direction for building better Transformers for NLP is their width, and that their depth is less relevant.

Wide Attention Is The Way Forward For Transformers

Given the wide and ever growing range of different efficient Transformer attention mechanisms, it is important to identify which attention is most effective when given a task. In this work, we are also interested in combining different attention types to build heterogeneous Transformers. We first propose a DARTS-like Neural Architecture Search (NAS) method to find the best attention for a given task, in this setup, all heads use the same attention (homogeneous models). Our results suggest that NAS is highly effective on this task, and it identifies the best attention mechanisms for IMDb byte level text classification and Listops. We then extend our framework to search for and build Transformers with multiple different attention types, and call them heterogeneous Transformers. We show that whilst these heterogeneous Transformers are better than the average homogeneous models, they cannot outperform the best. We explore the reasons why heterogeneous attention makes sense, and why it ultimately fails.

DARTFormer: Finding The Best Type Of Attention

Infant pose monitoring during sleep has multiple applications in both healthcare and home settings. In a healthcare setting, pose detection can be used for region of interest detection and movement detection for noncontact based monitoring systems. In a home setting, pose detection can be used to detect sleep positions which has shown to have a strong influence on multiple health factors. However, pose monitoring during sleep is challenging due to heavy occlusions from blanket coverings and low lighting. To address this, we present a novel dataset, Simultaneously-collected multimodal Mannequin Lying pose (SMaL) dataset, for under the cover infant pose estimation. We collect depth and pressure imagery of an infant mannequin in different poses under various cover conditions. We successfully infer full body pose under the cover by training state-of-art pose estimation methods and leveraging existing multimodal adult pose datasets for transfer learning. We demonstrate a hierarchical pretraining strategy for transformer-based models to significantly improve performance on our dataset. Our best performing model was able to detect joints under the cover within 25mm 86% of the time with an overall mean error of 16.9mm. Data, code and models publicly available at https://github.com/DanielKyr/SMaL

Under the Cover Infant Pose Estimation using Multimodal Data

Abstractive text summarization has recently become a popular approach, but data hallucination remains a serious problem, including with quantitative data. We propose a set of probing tests to evaluate the efficacy of abstract summarization models' modeling of quantitative values found in the input text. Our results show that in most cases, the encoders of recent SOTA-performing models struggle to provide embeddings that adequately represent quantitative values in the input compared to baselines, and in particular, they outperform random representations in some, but surprisingly not all, cases. Under our assumptions, this suggests that the encoder's performance contributes to the quantity hallucination problem. One model type in particular, DistilBART-CDM, was observed to underperform randomly initialized representations for several experiments, and performance versus BERT suggests that standard pretraining and fine-tuning approaches for the summarization task may play a role in underperformance for some encoders.

Probing of Quantitative Values in Abstractive Summarization Models

Background and aim: Hip fracture can be devastating. The proximal femoral strength can be computed by subject-specific finite element (FE) analysis (FEA) using quantitative CT images. The aim of this paper is to design a deep learning-based model for hip fracture prediction with multi-view information fusion. Method: We developed a multi-view variational autoencoder (MMVAE) for feature representation learning and designed the product of expert model (PoE) for multi-view information fusion.We performed genome-wide association studies (GWAS) to select the most relevant genetic features with proximal femoral strengths and integrated genetic features with DXA-derived imaging features and clinical variables for proximal femoral strength prediction. Results: The designed model achieved the mean absolute percentage error of 0.2050,0.0739 and 0.0852 for linear fall, nonlinear fall and nonlinear stance fracture load prediction, respectively. For linear fall and nonlinear stance fracture load prediction, integrating genetic and DXA-derived imaging features were beneficial; while for nonlinear fall fracture load prediction, integrating genetic features, DXA-derived imaging features as well as clinical variables, the model achieved the best performance. Conclusion: The proposed model is capable of predicting proximal femoral strengths using genetic features, DXA-derived imaging features as well as clinical variables. Compared to performing FEA using QCT images to calculate proximal femoral strengths, the presented method is time-efficient and cost effective, and radiation dosage is limited. From the technique perspective, the final models can be applied to other multi-view information integration tasks.

Multi-view information fusion using multi-view variational autoencoders to predict proximal femoral strength

In this paper, we study the infinite-depth limit of finite-width residual neural networks with random Gaussian weights. With proper scaling, we show that by fixing the width and taking the depth to infinity, the vector of pre-activations converges in distribution to a zero-drift diffusion process. Unlike the infinite-width limit where the pre-activation converge weakly to a Gaussian random variable, we show that the infinite-depth limit yields different distributions depending on the choice of the activation function. We document two cases where these distributions have closed-form (different) expressions. We further show an intriguing phase-transition phenomenon of the post-activation norms when the width increases from 3 to 4. Lastly, we study the sequential limit infinite-depth-then-infinite-width, and show some key differences with the more commonly studied infinite-width-then-infinite-depth limit.

On the infinite-depth limit of finite-width neural networks

In this paper, we introduce a convolutional network which we call MultiPodNet consisting of a combination of two or more convolutional networks which process the input image in parallel to achieve the same goal. Output feature maps of parallel convolutional networks are fused at the fully connected layer of the network. We experimentally observed that three parallel pod networks (TripodNet) produce the best results in commonly used object recognition datasets. Baseline pod networks can be of any type. In this paper, we use ResNets as baseline networks and their inputs are augmented image patches. The number of parameters of the TripodNet is about three times that of a single ResNet. We train the TripodNet using the standard backpropagation type algorithms. In each individual ResNet, parameters are initialized with different random numbers during training. The TripodNet achieved state-of-the-art performance on CIFAR-10 and ImageNet datasets. For example, it improved the accuracy of a single ResNet from 91.66% to 92.47% under the same training process on the CIFAR-10 dataset.

Multipod Convolutional Network

This paper uses information-theoretic tools to analyze the generalization error in unsupervised domain adaptation (UDA). We present novel upper bounds for two notions of generalization errors. The first notion measures the gap between the population risk in the target domain and that in the source domain, and the second measures the gap between the population risk in the target domain and the empirical risk in the source domain. While our bounds for the first kind of error are in line with the traditional analysis and give similar insights, our bounds on the second kind of error are algorithm-dependent, which also provide insights into algorithm designs. Specifically, we present two simple techniques for improving generalization in UDA and validate them experimentally.

Information-Theoretic Analysis of Unsupervised Domain Adaptation

The extremes of lighting (e.g. too much or too little light) usually cause many troubles for machine and human vision. Many recent works have mainly focused on under-exposure cases where images are often captured in low-light conditions (e.g. nighttime) and achieved promising results for enhancing the quality of images. However, they are inferior to handling images under over-exposure. To mitigate this limitation, we propose a novel unsupervised enhancement framework which is robust against various lighting conditions while does not require any well-exposed images to serve as the ground-truths. Our main concept is to construct pseudo-ground-truth images synthesized from multiple source images that simulate all potential exposure scenarios to train the enhancement network. Our extensive experiments show that the proposed approach consistently outperforms the current state-of-the-art unsupervised counterparts in several public datasets in terms of both quantitative metrics and qualitative results. Our code is available at https://github.com/VinAIResearch/PSENet-Image-Enhancement.

PSENet: Progressive Self-Enhancement Network for Unsupervised Extreme-Light Image Enhancement

Continual learning aims to alleviate catastrophic forgetting when handling consecutive tasks under non-stationary distributions. Gradient-based meta-learning algorithms have shown the capability to implicitly solve the transfer-interference trade-off problem between different examples. However, they still suffer from the catastrophic forgetting problem in the setting of continual learning, since the past data of previous tasks are no longer available. In this work, we propose a novel efficient meta-learning algorithm for solving the online continual learning problem, where the regularization terms and learning rates are adapted to the Taylor approximation of the parameter's importance to mitigate forgetting. The proposed method expresses the gradient of the meta-loss in closed-form and thus avoid computing second-order derivative which is computationally inhibitable. We also use Proximal Gradient Descent to further improve computational efficiency and accuracy. Experiments on diverse benchmarks show that our method achieves better or on-par performance and much higher efficiency compared to the state-of-the-art approaches.

Efficient Meta-Learning for Continual Learning with Taylor Expansion Approximation

Camera physiological measurement is a fast growing field of computer vision. Remote photoplethysmography (rPPG) uses video cameras (imagers) to measure the peripheral blood volume pulse (BVP). Simply, this enables heart rate measurement via webcams, smartphone cameras and many other imaging devices. The current state-of-the-art methods are supervised deep neural architectures that have large numbers of parameters and a signal number of hyperparameters. Replication of results and benchmarking of new models is critical for scientific progress. However, as with many other applications of deep learning, reliable codebases are not easy to find. We present a comprehensive toolbox, rPPG-Toolbox, containing code for training and evaluating unsupervised and supervised rPPG models: https://github.com/ubicomplab/rPPG-Toolbox

Deep Physiological Sensing Toolbox

Deep generative models parametrized up to a normalizing constant (e.g. energy-based models) are difficult to train by maximizing the likelihood of the data because the likelihood and/or gradients thereof cannot be explicitly or efficiently written down. Score matching is a training method, whereby instead of fitting the likelihood $\log p(x)$ for the training data, we instead fit the score function $\nabla_x \log p(x)$ -- obviating the need to evaluate the partition function. Though this estimator is known to be consistent, its unclear whether (and when) its statistical efficiency is comparable to that of maximum likelihood -- which is known to be (asymptotically) optimal. We initiate this line of inquiry in this paper, and show a tight connection between statistical efficiency of score matching and the isoperimetric properties of the distribution being estimated -- i.e. the Poincar\'e, log-Sobolev and isoperimetric constant -- quantities which govern the mixing time of Markov processes like Langevin dynamics. Roughly, we show that the score matching estimator is statistically comparable to the maximum likelihood when the distribution has a small isoperimetric constant. Conversely, if the distribution has a large isoperimetric constant -- even for simple families of distributions like exponential families with rich enough sufficient statistics -- score matching will be substantially less efficient than maximum likelihood. We suitably formalize these results both in the finite sample regime, and in the asymptotic regime. Finally, we identify a direct parallel in the discrete setting, where we connect the statistical properties of pseudolikelihood estimation with approximate tensorization of entropy and the Glauber dynamics.

Statistical Efficiency of Score Matching: The View from Isoperimetry

Automated machine vision pipelines do not need the exact visual content to perform their tasks. Therefore, there is a potential to remove private information from the data without significantly affecting the machine vision accuracy. We present a novel method to create a privacy-preserving latent representation of an image that could be used by a downstream machine vision model. This latent representation is constructed using adversarial training to prevent accurate reconstruction of the input while preserving the task accuracy. Specifically, we split a Deep Neural Network (DNN) model and insert an autoencoder whose purpose is to both reduce the dimensionality as well as remove information relevant to input reconstruction while minimizing the impact on task accuracy. Our results show that input reconstruction ability can be reduced by about 0.8 dB at the equivalent task accuracy, with degradation concentrated near the edges, which is important for privacy. At the same time, 30% bit savings are achieved compared to coding the features directly.

Privacy-Preserving Feature Coding for Machines

Graph Convolutional Neural Networks (GCNs) has been generally accepted to be an effective tool for node representations learning. An interesting way to understand GCNs is to think of them as a message passing mechanism where each node updates its representation by accepting information from its neighbours (also known as positive samples). However, beyond these neighbouring nodes, graphs have a large, dark, all-but forgotten world in which we find the non-neighbouring nodes (negative samples). In this paper, we show that this great dark world holds a substantial amount of information that might be useful for representation learning. Most specifically, it can provide negative information about the node representations. Our overall idea is to select appropriate negative samples for each node and incorporate the negative information contained in these samples into the representation updates. Moreover, we show that the process of selecting the negative samples is not trivial. Our theme therefore begins by describing the criteria for a good negative sample, followed by a determinantal point process algorithm for efficiently obtaining such samples. A GCN, boosted by diverse negative samples, then jointly considers the positive and negative information when passing messages. Experimental evaluations show that this idea not only improves the overall performance of standard representation learning but also significantly alleviates over-smoothing problems.

Learning from the Dark: Boosting Graph Convolutional Neural Networks with Diverse Negative Samples

In this paper, we look at the impact of Environment, Social and Governance related news articles and social media data on the stock market performance. We pick four stocks of companies which are widely known in their domain to understand the complete effect of ESG as the newly opted investment style remains restricted to only the stocks with widespread information. We summarise live data of both twitter tweets and newspaper articles and create a sentiment index using a dictionary technique based on online information for the month of July, 2022. We look at the stock price data for all the four companies and calculate the percentage change in each of them. We also compare the overall sentiment of the company to its percentage change over a specific historical period.

Sentiment Analysis of ESG disclosures on Stock Market

Infinitesimal gradient boosting is defined as the vanishing-learning-rate limit of the popular tree-based gradient boosting algorithm from machine learning (Dombry and Duchamps, 2021). It is characterized as the solution of a nonlinear ordinary differential equation in a infinite-dimensional function space where the infinitesimal boosting operator driving the dynamics depends on the training sample. We consider the asymptotic behavior of the model in the large sample limit and prove its convergence to a deterministic process. This infinite population limit is again characterized by a differential equation that depends on the population distribution. We explore some properties of this population limit: we prove that the dynamics makes the test error decrease and we consider its long time behavior.

A large sample theory for infinitesimal gradient boosting

Federated Learning (FL) effectively protects client data privacy. However, client absence or leaving during training can seriously degrade model performances, particularly for unbalanced and non-IID client data. We address this issue by generating data digests from the raw data and using them to guide training at the FL moderator. The proposed FL framework, called FedDig, can tolerate unexpected client absence in cross-silo scenarios while preserving client data privacy because the digests de-identify the raw data by mixing encoded features in the features space. We evaluate FedDig using EMNIST, CIFAR-10, and CIFAR-100; the results consistently outperform against three baseline algorithms (FedAvg, FedProx, and FedNova) by large margins in various client absence scenarios.

FedDig: Robust Federated Learning Using Data Digest to Represent Absent Clients

For tackling the task of 2D human pose estimation, the great majority of the recent methods regard this task as a heatmap estimation problem, and optimize the heatmap prediction using the Gaussian-smoothed heatmap as the optimization objective and using the pixel-wise loss (e.g. MSE) as the loss function. In this paper, we show that optimizing the heatmap prediction in such a way, the model performance of body joint localization, which is the intrinsic objective of this task, may not be consistently improved during the optimization process of the heatmap prediction. To address this problem, from a novel perspective, we propose to formulate the optimization of the heatmap prediction as a distribution matching problem between the predicted heatmap and the dot annotation of the body joint directly. By doing so, our proposed method does not need to construct the Gaussian-smoothed heatmap and can achieve a more consistent model performance improvement during the optimization of the heatmap prediction. We show the effectiveness of our proposed method through extensive experiments on the COCO dataset and the MPII dataset.

Heatmap Distribution Matching for Human Pose Estimation

Designing proper training pairs is critical for super-resolving the real-world low-quality (LQ) images, yet suffers from the difficulties in either acquiring paired ground-truth HQ images or synthesizing photo-realistic degraded observations. Recent works mainly circumvent this by simulating the degradation with handcrafted or estimated degradation parameters. However, existing synthetic degradation models are incapable to model complicated real degradation types, resulting in limited improvement on these scenarios, \eg, old photos. Notably, face images, which have the same degradation process with the natural images, can be robustly restored with photo-realistic textures by exploiting their specific structure priors. In this work, we use these real-world LQ face images and their restored HQ counterparts to model the complex real degradation (namely ReDegNet), and then transfer it to HQ natural images to synthesize their realistic LQ ones. Specifically, we take these paired HQ and LQ face images as inputs to explicitly predict the degradation-aware and content-independent representations, which control the degraded image generation. Subsequently, we transfer these real degradation representations from face to natural images to synthesize the degraded LQ natural images. Experiments show that our ReDegNet can well learn the real degradation process from face images, and the restoration network trained with our synthetic pairs performs favorably against SOTAs. More importantly, our method provides a new manner to handle the unsynthesizable real-world scenarios by learning their degradation representations through face images within them, which can be used for specifically fine-tuning. The source code is available at https://github.com/csxmli2016/ReDegNet.

From Face to Natural Image: Learning Real Degradation for Blind Image Super-Resolution

Current breakthroughs in natural language processing have benefited dramatically from neural language models, through which distributional semantics can leverage neural data representations to facilitate downstream applications. Since neural embeddings use context prediction on word co-occurrences to yield dense vectors, they are inevitably prone to capture more semantic association than semantic similarity. To improve vector space models in deriving semantic similarity, we post-process neural word embeddings through deep metric learning, through which we can inject lexical-semantic relations, including syn/antonymy and hypo/hypernymy, into a distributional space. We introduce hierarchy-fitting, a novel semantic specialization approach to modelling semantic similarity nuances inherently stored in the IS-A hierarchies. Hierarchy-fitting attains state-of-the-art results on the common- and rare-word benchmark datasets for deriving semantic similarity from neural word embeddings. It also incorporates an asymmetric distance function to specialize hypernymy's directionality explicitly, through which it significantly improves vanilla embeddings in multiple evaluation tasks of detecting hypernymy and directionality without negative impacts on semantic similarity judgement. The results demonstrate the efficacy of hierarchy-fitting in specializing neural embeddings with semantic relations in late fusion, potentially expanding its applicability to aggregating heterogeneous data and various knowledge resources for learning multimodal semantic spaces.

Lexical semantics enhanced neural word embeddings

Small target detection is an essential yet challenging task in defense applications, since differentiating low-contrast targets from natural textured and noisy environment remains difficult. To better take into account the contextual information, we propose to explore deep learning approaches based on attention mechanisms. Specifically, we propose a customized version of TransUnet including channel attention, which has shown a significant improvement in performance. Moreover, the lack of annotated data induces weak detection precision, leading to many false alarms. We thus explore a contrario methods in order to select meaningful potential targets detected by a weak deep learning training. -- La d\'etection de petites cibles est une probl\'ematique d\'elicate mais essentielle dans le domaine de la d\'efense, notamment lorsqu'il s'agit de diff\'erencier ces cibles d'un fond bruit\'e ou textur\'e, ou lorsqu'elles sont de faible contraste. Pour mieux prendre en compte les informations contextuelles, nous proposons d'explorer diff\'erentes approches de segmentation par apprentissage profond, dont certaines bas\'ees sur les m\'ecanismes d'attention. Nous proposons \'egalement d'inclure un module d'attention par canal au TransUnet, r\'eseau \`a l'\'etat de l'art, ce qui permet d'am\'eliorer significativement les performances. Par ailleurs, le manque de donn\'ees annot\'ees induit une perte en pr\'ecision lors des d\'etections, conduisant \`a de nombreuses fausses alarmes non pertinentes. Nous explorons donc des m\'ethodes a contrario afin de s\'electionner les cibles les plus significatives d\'etect\'ees par un r\'eseau entra\^in\'e avec peu de donn\'ees.

D\'etection de petites cibles par apprentissage profond et crit\`ere a contrario

Perceiving the surrounding environment is essential for enabling autonomous or assisted driving functionalities. Common tasks in this domain include detecting road users, as well as determining lane boundaries and classifying driving conditions. Over the last few years, a large variety of powerful Deep Learning models have been proposed to address individual tasks of camera-based automotive perception with astonishing performances. However, the limited capabilities of in-vehicle embedded computing platforms cannot cope with the computational effort required to run a heavy model for each individual task. In this work, we present CERBERUS (CEnteR Based End-to-end peRception Using a Single model), a lightweight model that leverages a multitask-learning approach to enable the execution of multiple perception tasks at the cost of a single inference. The code will be made publicly available at https://github.com/cscribano/CERBERUS

CERBERUS: Simple and Effective All-In-One Automotive Perception Model with Multi Task Learning

Prototype learning and decoder construction are the keys for few-shot segmentation. However, existing methods use only a single prototype generation mode, which can not cope with the intractable problem of objects with various scales. Moreover, the one-way forward propagation adopted by previous methods may cause information dilution from registered features during the decoding process. In this research, we propose a rich prototype generation module (RPGM) and a recurrent prediction enhancement module (RPEM) to reinforce the prototype learning paradigm and build a unified memory-augmented decoder for few-shot segmentation, respectively. Specifically, the RPGM combines superpixel and K-means clustering to generate rich prototype features with complementary scale relationships and adapt the scale gap between support and query images. The RPEM utilizes the recurrent mechanism to design a round-way propagation decoder. In this way, registered features can provide object-aware information continuously. Experiments show that our method consistently outperforms other competitors on two popular benchmarks PASCAL-${{5}^{i}}$ and COCO-${{20}^{i}}$.

Few-Shot Segmentation via Rich Prototype Generation and Recurrent Prediction Enhancement

Since the release of various large-scale natural language processing (NLP) pre-trained models, parameter efficient transfer learning (PETL) has become a popular paradigm capable of achieving impressive performance on various downstream tasks. PETL aims at making good use of the representation knowledge in the pre-trained large models by fine-tuning a small number of parameters. Recently, it has also attracted increasing attention to developing various PETL techniques for vision tasks. Popular PETL techniques such as Prompt-tuning and Adapter have been proposed for high-level visual downstream tasks such as image classification and video recognition. However, Prefix-tuning remains under-explored for vision tasks. In this work, we intend to adapt large video-based models to downstream tasks with a good parameter-accuracy trade-off. Towards this goal, we propose a framework with a unified view called visual-PETL (V-PETL) to investigate the different aspects affecting the trade-off. Specifically, we analyze the positional importance of trainable parameters and differences between NLP and vision tasks in terms of data structures and pre-training mechanisms while implementing various PETL techniques, especially for the under-explored prefix-tuning technique. Based on a comprehensive understanding of differences between NLP and video data, we propose a new variation of prefix-tuning module called parallel attention (PATT) for video-based downstream tasks. An extensive empirical analysis on two video datasets via different frozen backbones has been carried and the findings show that the proposed PATT can effectively contribute to other PETL techniques. An effective scheme Swin-BAPAT derived from the proposed V-PETL framework achieves significantly better performance than the state-of-the-art AdaptFormer-Swin with slightly more parameters and outperforms full-tuning with far less parameters.

Towards a Unified View on Visual Parameter-Efficient Transfer Learning

Background: Drug synergy occurs when the combined effect of two drugs is greater than the sum of the individual drugs' effect. While cell line data measuring the effect of single drugs are readily available, there is relatively less comparable data on drug synergy given the vast amount of possible drug combinations. Thus, there is interest to use computational approaches to predict drug synergy for untested pairs of drugs. Methods: We introduce a Graph Neural Network (GNN) based model for drug synergy prediction, which utilizes drug chemical structures and cell line gene expression data. We use information from the largest drug combination database available (DrugComb), combining drug synergy scores in order to construct high confidence benchmark datasets. Results: Our proposed solution for drug synergy predictions offers a number of benefits: 1) It is trained on high confidence benchmark dataset. 2) It utilizes 34 distinct drug synergy datasets to learn on a wide variety of drugs and cell lines representations. 3) It learns task-specific drug representations, instead of relying on generalized and pre-computed chemical drug features. 4) It achieves similar or better prediction performance (AUPR scores ranging from 0.777 to 0.964) compared to state-of-the-art baseline models when tested on various benchmark datasets. Conclusions: We demonstrate that a GNN based model can provide state-of-the-art drug synergy predictions by learning task-specific representations of drugs.

DDoS: A Graph Neural Network based Drug Synergy Prediction Algorithm

We study the training of deep neural networks by gradient descent where floating-point arithmetic is used to compute the gradients. In this framework and under realistic assumptions, we demonstrate that it is highly unlikely to find ReLU neural networks that maintain, in the course of training with gradient descent, superlinearly many affine pieces with respect to their number of layers. In virtually all approximation theoretical arguments which yield high order polynomial rates of approximation, sequences of ReLU neural networks with exponentially many affine pieces compared to their numbers of layers are used. As a consequence, we conclude that approximating sequences of ReLU neural networks resulting from gradient descent in practice differ substantially from theoretically constructed sequences. The assumptions and the theoretical results are compared to a numerical study, which yields concurring results.

Limitations of neural network training due to numerical instability of backpropagation

Video frame interpolation (VFI) is a fundamental research topic in video processing, which is currently attracting increased attention across the research community. While the development of more advanced VFI algorithms has been extensively researched, there remains little understanding of how humans perceive the quality of interpolated content and how well existing objective quality assessment methods perform when measuring the perceived quality. In order to narrow this research gap, we have developed a new video quality database named BVI-VFI, which contains 540 distorted sequences generated by applying five commonly used VFI algorithms to 36 diverse source videos with various spatial resolutions and frame rates. We collected more than 10,800 quality ratings for these videos through a large scale subjective study involving 189 human subjects. Based on the collected subjective scores, we further analysed the influence of VFI algorithms and frame rates on the perceptual quality of interpolated videos. Moreover, we benchmarked the performance of 28 classic and state-of-the-art objective image/video quality metrics on the new database, and demonstrated the urgent requirement for more accurate bespoke quality assessment methods for VFI. To facilitate further research in this area, we have made BVI-VFI publicly available at https://github.com/danielism97/BVI-VFI-database.

BVI-VFI: A Video Quality Database for Video Frame Interpolation

We have gained access to vast amounts of multi-omics data thanks to Next Generation Sequencing. However, it is challenging to analyse this data due to its high dimensionality and much of it not being annotated. Lack of annotated data is a significant problem in machine learning, and Self-Supervised Learning (SSL) methods are typically used to deal with limited labelled data. However, there is a lack of studies that use SSL methods to exploit inter-omics relationships on unlabelled multi-omics data. In this work, we develop a novel and efficient pre-training paradigm that consists of various SSL components, including but not limited to contrastive alignment, data recovery from corrupted samples, and using one type of omics data to recover other omic types. Our pre-training paradigm improves performance on downstream tasks with limited labelled data. We show that our approach outperforms the state-of-the-art method in cancer type classification on the TCGA pan-cancer dataset in semi-supervised setting. Moreover, we show that the encoders that are pre-trained using our approach can be used as powerful feature extractors even without fine-tuning. Our ablation study shows that the method is not overly dependent on any pretext task component. The network architectures in our approach are designed to handle missing omic types and multiple datasets for pre-training and downstream training. Our pre-training paradigm can be extended to perform zero-shot classification of rare cancers.

Self-omics: A Self-supervised Learning Framework for Multi-omics Cancer Data

Accurately predicting road networks from satellite images requires a global understanding of the network topology. We propose to capture such high-level information by introducing a graph-based framework that simulates the addition of sequences of graph edges using a reinforcement learning (RL) approach. In particular, given a partially generated graph associated with a satellite image, an RL agent nominates modifications that maximize a cumulative reward. As opposed to standard supervised techniques that tend to be more restricted to commonly used surrogate losses, these rewards can be based on various complex, potentially non-continuous, metrics of interest. This yields more power and flexibility to encode problem-dependent knowledge. Empirical results on several benchmark datasets demonstrate enhanced performance and increased high-level reasoning about the graph topology when using a tree-based search. We further highlight the superiority of our approach under substantial occlusions by introducing a new synthetic benchmark dataset for this task.

Mastering Spatial Graph Prediction of Road Networks

We study reinforcement learning for continuous-time Markov decision processes (MDPs) in the finite-horizon episodic setting. We present a learning algorithm based on the methods of value iteration and upper confidence bound. We derive an upper bound on the worst-case expected regret for the proposed algorithm, and establish a worst-case lower bound, both bounds are of the order of square-root on the number of episodes. Finally, we conduct simulation experiments to illustrate the performance of our algorithm.

Square-root regret bounds for continuous-time episodic Markov decision processes

Low-overhead visual place recognition (VPR) is a highly active research topic. Mobile robotics applications often operate under low-end hardware, and even more hardware capable systems can still benefit from freeing up onboard system resources for other navigation tasks. This work addresses lightweight VPR by proposing a novel system based on the combination of binary-weighted classifier networks with a one-dimensional convolutional network, dubbed merger. Recent work in fusing multiple VPR techniques has mainly focused on increasing VPR performance, with computational efficiency not being highly prioritized. In contrast, we design our technique prioritizing low inference times, taking inspiration from the machine learning literature where the efficient combination of classifiers is a heavily researched topic. Our experiments show that the merger achieves inference times as low as 1 millisecond, being significantly faster than other well-established lightweight VPR techniques, while achieving comparable or superior VPR performance on several visual changes such as seasonal variations and viewpoint lateral shifts.

Merging Classification Predictions with Sequential Information for Lightweight Visual Place Recognition in Changing Environments

The recent observation of neural power-law scaling relations has made a significant impact in the field of deep learning. A substantial amount of attention has been dedicated as a consequence to the description of scaling laws, although mostly for supervised learning and only to a reduced extent for reinforcement learning frameworks. In this paper we present an extensive study of performance scaling for a cornerstone reinforcement learning algorithm, AlphaZero. On the basis of a relationship between Elo rating, playing strength and power-law scaling, we train AlphaZero agents on the games Connect Four and Pentago and analyze their performance. We find that player strength scales as a power law in neural network parameter count when not bottlenecked by available compute, and as a power of compute when training optimally sized agents. We observe nearly identical scaling exponents for both games. Combining the two observed scaling laws we obtain a power law relating optimal size to compute similar to the ones observed for language models. We find that the predicted scaling of optimal neural network size fits our data for both games. This scaling law implies that previously published state-of-the-art game-playing models are significantly smaller than their optimal size, given the respective compute budgets. We also show that large AlphaZero models are more sample efficient, performing better than smaller models with the same amount of training data.

Scaling Laws for a Multi-Agent Reinforcement Learning Model

This research investigates the effective incorporation of the human factor and user perception in text-based interactive media. In such contexts, the reliability of user texts is often compromised by behavioural and emotional dimensions. To this end, several attempts have been made in the state of the art, to introduce psychological approaches in such systems, including computational psycholinguistics, personality traits and cognitive psychology methods. In contrast, our method is fundamentally different since we employ a psychoanalysis-based approach; in particular, we use the notion of Lacanian discourse types, to capture and deeply understand real (possibly elusive) characteristics, qualities and contents of texts, and evaluate their reliability. As far as we know, this is the first time computational methods are systematically combined with psychoanalysis. We believe such psychoanalytic framework is fundamentally more effective than standard methods, since it addresses deeper, quite primitive elements of human personality, behaviour and expression which usually escape methods functioning at "higher", conscious layers. In fact, this research is a first attempt to form a new paradigm of psychoanalysis-driven interactive technologies, with broader impact and diverse applications. To exemplify this generic approach, we apply it to the case-study of fake news detection; we first demonstrate certain limitations of the well-known Myers-Briggs Type Indicator (MBTI) personality type method, and then propose and evaluate our new method of analysing user texts and detecting fake news based on the Lacanian discourses psychoanalytic approach.

Concepts and Experiments on Psychoanalysis Driven Computing

Many material response functions depend strongly on microstructure, such as inhomogeneities in phase or orientation. Homogenization presents the task of predicting the mean response of a sample of the microstructure to external loading for use in subgrid models and structure-property explorations. Although many microstructural fields have obvious segmentations, learning directly from the graph induced by the segmentation can be difficult because this representation does not encode all the information of the full field. We develop a means of deep learning of hidden features on the reduced graph given the native discretization and a segmentation of the initial input field. The features are associated with regions represented as nodes on the reduced graph. This reduced representation is then the basis for the subsequent multi-level/scale graph convolutional network model. There are a number of advantages of reducing the graph before fully processing with convolutional layers it, such as interpretable features and efficiency on large meshes. We demonstrate the performance of the proposed network relative to convolutional neural networks operating directly on the native discretization of the data using three physical exemplars.

Deep learning and multi-level featurization of graph representations of microstructural data

In this paper, we introduce an emerging quantum machine learning (QML) framework to assist classical deep learning methods for biosignal processing applications. Specifically, we propose a hybrid quantum-classical neural network model that integrates a variational quantum circuit (VQC) into a deep neural network (DNN) for electroencephalogram (EEG), electromyogram (EMG), and electrocorticogram (ECoG) analysis. We demonstrate that the proposed quantum neural network (QNN) achieves state-of-the-art performance while the number of trainable parameters is kept small for VQC.

quEEGNet: Quantum AI for Biosignal Processing

Python was used to download and format NewsAPI article data relating to 400 publicly traded, low cap. Biotech companies. Crowd-sourcing was used to label a subset of this data to then train and evaluate a variety of models to classify the public sentiment of each company. The best performing models were then used to show that trading entirely off public sentiment could provide market beating returns.

Multiclass Sentiment Prediction for Stock Trading

In addition to being familiar with policies, high investment returns also require extensive knowledge of relevant industry knowledge and news. In addition, it is necessary to leverage relevant theories for investment to make decisions, thereby amplifying investment returns. A effective investment return estimate can feedback the future rate of return of investment behavior. In recent years, deep learning are developing rapidly, and investment return prediction based on deep learning has become an emerging research topic. This paper proposes an embedding-based dual branch approach to predict an investment's return. This approach leverages embedding to encode the investment id into a low-dimensional dense vector, thereby mapping high-dimensional data to a low-dimensional manifold, so that highdimensional features can be represented competitively. In addition, the dual branch model realizes the decoupling of features by separately encoding different information in the two branches. In addition, the swish activation function further improves the model performance. Our approach are validated on the Ubiquant Market Prediction dataset. The results demonstrate the superiority of our approach compared to Xgboost, Lightgbm and Catboost.

Embedding-based neural network for investment return prediction

Cryptocurrencies have become a trendy topic recently, primarily due to their disruptive potential and reports of unprecedented returns. In addition, academics increasingly acknowledge the predictive power of Social Media in many fields and, more specifically, for financial markets and economics. In this paper, we leverage the predictive power of Twitter and Reddit sentiment together with Google Trends indexes and volume to forecast the log returns of ten cryptocurrencies. Specifically, we consider $Bitcoin$, $Ethereum$, $Tether$, $Binance Coin$, $Litecoin$, $Enjin Coin$, $Horizen$, $Namecoin$, $Peercoin$, and $Feathercoin$. We evaluate the performance of LASSO-VAR using daily data from January 2018 to January 2022. In a 30 days recursive forecast, we can retrieve the correct direction of the actual series more than 50% of the time. We compare this result with the main benchmarks, and we see a 10% improvement in Mean Directional Accuracy (MDA). The use of sentiment and attention variables as predictors increase significantly the forecast accuracy in terms of MDA but not in terms of Root Mean Squared Errors. We perform a Granger causality test using a post-double LASSO selection for high-dimensional VARs. Results show no "causality" from Social Media sentiment to cryptocurrencies returns

Forecasting Cryptocurrencies Log-Returns: a LASSO-VAR and Sentiment Approach

Human health is closely associated with their daily behavior and environment. However, keeping a healthy lifestyle is still challenging for most people as it is difficult to recognize their living behaviors and identify their surrounding situations to take appropriate action. Human activity recognition is a promising approach to building a behavior model of users, by which users can get feedback about their habits and be encouraged to develop a healthier lifestyle. In this paper, we present a smart light wearable badge with six kinds of sensors, including an infrared array sensor MLX90640 offering privacy-preserving, low-cost, and non-invasive features, to recognize daily activities in a realistic unmodified kitchen environment. A multi-channel convolutional neural network (MC-CNN) based on data and feature fusion methods is applied to classify 14 human activities associated with potentially unhealthy habits. Meanwhile, we evaluate the impact of the infrared array sensor on the recognition accuracy of these activities. We demonstrate the performance of the proposed work to detect the 14 activities performed by ten volunteers with an average accuracy of 92.44 % and an F1 score of 88.27 %.

Smart-Badge: A wearable badge with multi-modal sensors for kitchen activity recognition

Sign language recognition could significantly improve the user experience for d/Deaf people with the general consumer technology, such as IoT devices or videoconferencing. However, current sign language recognition architectures are usually computationally heavy and require robust GPU-equipped hardware to run in real-time. Some models aim for lower-end devices (such as smartphones) by minimizing their size and complexity, which leads to worse accuracy. This highly scrutinizes accurate in-the-wild applications. We build upon the SPOTER architecture, which belongs to the latter group of light methods, as it came close to the performance of large models employed for this task. By substituting its original third-party pose estimation module with the MediaPipe library, we achieve an overall state-of-the-art result on the WLASL100 dataset. Significantly, our method beats previous larger architectures while still being twice as computationally efficient and almost $11$ times faster on inference when compared to a relevant benchmark. To demonstrate our method's combined efficiency and precision, we built an online demo that enables users to translate sign lemmas of American sign language in their browsers. This is the first publicly available online application demonstrating this task to the best of our knowledge.

Combining Efficient and Precise Sign Language Recognition: Good pose estimation library is all you need

This paper studies the problem of injecting factual knowledge into large pre-trained language models. We train adapter modules on parts of the ConceptNet knowledge graph using the masked language modeling objective and evaluate the success of the method by a series of probing experiments on the LAMA probe. Mean P@K curves for different configurations indicate that the technique is effective, increasing the performance on subsets of the LAMA probe for large values of k by adding as little as 2.1% additional parameters to the original models.

The Effectiveness of Masked Language Modeling and Adapters for Factual Knowledge Injection

Standard approaches to hate speech detection rely on sufficient available hate speech annotations. Extending previous work that repurposes natural language inference (NLI) models for zero-shot text classification, we propose a simple approach that combines multiple hypotheses to improve English NLI-based zero-shot hate speech detection. We first conduct an error analysis for vanilla NLI-based zero-shot hate speech detection and then develop four strategies based on this analysis. The strategies use multiple hypotheses to predict various aspects of an input text and combine these predictions into a final verdict. We find that the zero-shot baseline used for the initial error analysis already outperforms commercial systems and fine-tuned BERT-based hate speech detection models on HateCheck. The combination of the proposed strategies further increases the zero-shot accuracy of 79.4% on HateCheck by 7.9 percentage points (pp), and the accuracy of 69.6% on ETHOS by 10.0pp.

Hypothesis Engineering for Zero-Shot Hate Speech Detection

Prevalent state-of-the-art instance segmentation methods fall into a query-based scheme, in which instance masks are derived by querying the image feature using a set of instance-aware embeddings. In this work, we devise a new training framework that boosts query-based models through discriminative query embedding learning. It explores two essential properties, namely dataset-level uniqueness and transformation equivariance, of the relation between queries and instances. First, our algorithm uses the queries to retrieve the corresponding instances from the whole training dataset, instead of only searching within individual scenes. As querying instances across scenes is more challenging, the segmenters are forced to learn more discriminative queries for effective instance separation. Second, our algorithm encourages both image (instance) representations and queries to be equivariant against geometric transformations, leading to more robust, instance-query matching. On top of four famous, query-based models ($i.e.,$ CondInst, SOLOv2, SOTR, and Mask2Former), our training algorithm provides significant performance gains ($e.g.,$ +1.6 - 3.2 AP) on COCO dataset. In addition, our algorithm promotes the performance of SOLOv2 by 2.7 AP, on LVISv1 dataset.

Learning Equivariant Segmentation with Instance-Unique Querying

Domain generalization (DG) has been a hot topic in image recognition, with a goal to train a general model that can perform well on unseen domains. Recently, federated learning (FL), an emerging machine learning paradigm to train a global model from multiple decentralized clients without compromising data privacy, brings new challenges, also new possibilities, to DG. In the FL scenario, many existing state-of-the-art (SOTA) DG methods become ineffective, because they require the centralization of data from different domains during training. In this paper, we propose a novel domain generalization method for image recognition under federated learning through cross-client style transfer (CCST) without exchanging data samples. Our CCST method can lead to more uniform distributions of source clients, and thus make each local model learn to fit the image styles of all the clients to avoid the different model biases. Two types of style (single image style and overall domain style) with corresponding mechanisms are proposed to be chosen according to different scenarios. Our style representation is exceptionally lightweight and can hardly be used for the reconstruction of the dataset. The level of diversity is also flexible to be controlled with a hyper-parameter. Our method outperforms recent SOTA DG methods on two DG benchmarks (PACS, OfficeHome) and a large-scale medical image dataset (Camelyon17) in the FL setting. Last but not least, our method is orthogonal to many classic DG methods, achieving additive performance by combined utilization.

Federated Domain Generalization for Image Recognition via Cross-Client Style Transfer

Previous virtual try-on methods usually focus on aligning a clothing item with a person, limiting their ability to exploit the complex pose, shape and skin color of the person, as well as the overall structure of the clothing, which is vital to photo-realistic virtual try-on. To address this potential weakness, we propose a fill in fabrics (FIFA) model, a self-supervised conditional generative adversarial network based framework comprised of a Fabricator and a unified virtual try-on pipeline with a Segmenter, Warper and Fuser. The Fabricator aims to reconstruct the clothing image when provided with a masked clothing as input, and learns the overall structure of the clothing by filling in fabrics. A virtual try-on pipeline is then trained by transferring the learned representations from the Fabricator to Warper in an effort to warp and refine the target clothing. We also propose to use a multi-scale structural constraint to enforce global context at multiple scales while warping the target clothing to better fit the pose and shape of the person. Extensive experiments demonstrate that our FIFA model achieves state-of-the-art results on the standard VITON dataset for virtual try-on of clothing items, and is shown to be effective at handling complex poses and retaining the texture and embroidery of the clothing.

Fill in Fabrics: Body-Aware Self-Supervised Inpainting for Image-Based Virtual Try-On

Scene Graphs are widely applied in computer vision as a graphical representation of relationships between objects shown in images. However, these applications have not yet reached a practical stage of development owing to biased training caused by long-tailed predicate distributions. In recent years, many studies have tackled this problem. In contrast, relatively few works have considered predicate similarities as a unique dataset feature which also leads to the biased prediction. Due to the feature, infrequent predicates (e.g., parked on, covered in) are easily misclassified as closely-related frequent predicates (e.g., on, in). Utilizing predicate similarities, we propose a new classification scheme that branches the process to several fine-grained classifiers for similar predicate groups. The classifiers aim to capture the differences among similar predicates in detail. We also introduce the idea of transfer learning to enhance the features for the predicates which lack sufficient training samples to learn the descriptive representations. The results of extensive experiments on the Visual Genome dataset show that the combination of our method and an existing debiasing approach greatly improves performance on tail predicates in challenging SGCls/SGDet tasks. Nonetheless, the overall performance of the proposed approach does not reach that of the current state of the art, so further analysis remains necessary as future work.

Unbiased Scene Graph Generation using Predicate Similarities

Self-attention is of vital importance in semantic segmentation as it enables modeling of long-range context, which translates into improved performance. We argue that it is equally important to model short-range context, especially to tackle cases where not only the regions of interest are small and ambiguous, but also when there exists an imbalance between the semantic classes. To this end, we propose Masked Supervised Learning (MaskSup), an effective single-stage learning paradigm that models both short- and long-range context, capturing the contextual relationships between pixels via random masking. Experimental results demonstrate the competitive performance of MaskSup against strong baselines in both binary and multi-class segmentation tasks on three standard benchmark datasets, particularly at handling ambiguous regions and retaining better segmentation of minority classes with no added inference cost. In addition to segmenting target regions even when large portions of the input are masked, MaskSup is also generic and can be easily integrated into a variety of semantic segmentation methods. We also show that the proposed method is computationally efficient, yielding an improved performance by 10\% on the mean intersection-over-union (mIoU) while requiring $3\times$ less learnable parameters.

Masked Supervised Learning for Semantic Segmentation

While PAC-Bayes is now an established learning framework for bounded losses, its extension to the case of unbounded losses (as simple as the squared loss on an unbounded space) remains largely uncharted and has attracted a growing interest in recent years. We contribute to this line of work by developing an extention of Markov's inequality for supermartingales, which we use to establish a novel PAC-Bayesian generalisation bound holding for unbounded losses. We show that this bound extends, unifies and even improves on existing PAC-Bayesian bounds.

PAC-Bayes with Unbounded Losses through Supermartingales

No-reference image quality assessment (NR-IQA) aims to quantify how humans perceive visual distortions of digital images without access to their undistorted references. NR-IQA models are extensively studied in computational vision, and are widely used for performance evaluation and perceptual optimization of man-made vision systems. Here we make one of the first attempts to examine the perceptual robustness of NR-IQA models. Under a Lagrangian formulation, we identify insightful connections of the proposed perceptual attack to previous beautiful ideas in computer vision and machine learning. We test one knowledge-driven and three data-driven NR-IQA methods under four full-reference IQA models (as approximations to human perception of just-noticeable differences). Through carefully designed psychophysical experiments, we find that all four NR-IQA models are vulnerable to the proposed perceptual attack. More interestingly, we observe that the generated counterexamples are not transferable, manifesting themselves as distinct design flows of respective NR-IQA methods.

Perceptual Attacks of No-Reference Image Quality Models with Human-in-the-Loop

Unsupervised multimodal change detection is a practical and challenging topic that can play an important role in time-sensitive emergency applications. To address the challenge that multimodal remote sensing images cannot be directly compared due to their modal heterogeneity, we take advantage of two types of modality-independent structural relationships in multimodal images. In particular, we present a structural relationship graph representation learning framework for measuring the similarity of the two structural relationships. Firstly, structural graphs are generated from preprocessed multimodal image pairs by means of an object-based image analysis approach. Then, a structural relationship graph convolutional autoencoder (SR-GCAE) is proposed to learn robust and representative features from graphs. Two loss functions aiming at reconstructing vertex information and edge information are presented to make the learned representations applicable for structural relationship similarity measurement. Subsequently, the similarity levels of two structural relationships are calculated from learned graph representations and two difference images are generated based on the similarity levels. After obtaining the difference images, an adaptive fusion strategy is presented to fuse the two difference images. Finally, a morphological filtering-based postprocessing approach is employed to refine the detection results. Experimental results on five datasets with different modal combinations demonstrate the effectiveness of the proposed method.

Unsupervised Multimodal Change Detection Based on Structural Relationship Graph Representation Learning