论文检索

Deep Learning

Zixiang Chen, Junkai Zhang, Yiwen Kou, Xiangning Chen, Cho-Jui Hsieh, Quanquan Gu

The challenge of overfitting, in which the model memorizes the training data and fails to generalize to test data, has become increasingly significant in the training of large neural networks. To tackle this challenge, Sharpness-Aware Minimization (SAM) has emerged as a promising training method, which can improve the generalization of neural networks even in the presence of label noise. However, a deep understanding of how SAM works, especially in the setting of nonlinear neural networks and classification tasks, remains largely missing. This paper fills this gap by demonstrating why SAM generalizes better than Stochastic Gradient Descent (SGD) for a certain data model and two-layer convolutional ReLU networks. The loss landscape of our studied problem is nonsmooth, thus current explanations for the success of SAM based on the Hessian information are insufficient. Our result explains the benefits of SAM, particularly its ability to prevent noise learning in the early stages, thereby facilitating more effective learning of features. Experiments on both synthetic and real data corroborate our theory.

Miscellaneous Aspects of Machine Learning/Sequential, Network, and Time Series Modeling

Giovanni De Felice, John Goulermas, Vladimir Gusev

Kernel design is a pivotal but challenging aspect of time series analysis, especially in the context of small datasets. In recent years, Reservoir Computing (RC) has emerged as a powerful tool to compare time series based on the underlying dynamics of the generating process rather than the observed data. However, the performance of RC highly depends on the hyperparameter setting, which is hard to interpret and costly to optimize because of the recurrent nature of RC. Here, we present a new kernel for time series based on the recently established equivalence between reservoir dynamics and Nonlinear Vector AutoRegressive (NVAR) processes. The kernel is non-recurrent and depends on a small set of meaningful hyperparameters, for which we suggest an effective heuristic. We demonstrate excellent performance on a wide range of real-world classification tasks, both in terms of accuracy and speed. This further advances the understanding of RC representation learning models and extends the typical use of the NVAR framework to kernel design and representation of real-world time series data.

Chenfeng Xu, Shijia Yang, Tomer Galanti, Bichen Wu, Xiangyu Yue, Bohan Zhai, Wei Zhan, Peter Vajda, Kurt Keutzer, Masayoshi Tomizuka

3D point-clouds and 2D images are different visual representations of the physical world. While human vision can understand both representations, computer vision models designed for 2D image and 3D point-cloud understanding are quite different. Our paper explores the potential of transferring 2D model architectures and weights to understand 3D point-clouds, by empirically investigating the feasibility of the transfer, the benefits of the transfer, and shedding light on why the transfer works. We discover that we can indeed use the same architecture and pretrained weights of a neural net model to understand both images and point-clouds. Specifically, we transfer the image-pretrained model to a point-cloud model by copying or inflating the weights. We find that finetuning the transformed image-pretrained models (FIP) with minimal efforts --- only on input, output, and normalization layers --- can achieve competitive performance on 3D point-cloud classification, beating a wide range of point-cloud models that adopt task-specific architectures and use a variety of tricks. When finetuning the whole model, the performance gets further improved. Meanwhile, FIP improves data efficiency, reaching up to 10.0 top-1 accuracy percent on few-shot classification. It also speeds up training of point-cloud models by up to 11.1x for a target accuracy (e.g., 90 % accuracy). Lastly, we provide an explanation of the image to point-cloud transfer from the aspect of neural collapse.

Dingkang Yang, Shuai Huang, Shunli Wang, Yang Liu, Peng Zhai, Liuzhen Su, Mingcheng Li, Lihua Zhang

Understanding emotion in context is a rising hotspot in the computer vision community. Existing methods lack reliable context semantics to mitigate uncertainty in expressing emotions and fail to model multiple context representations complementarily. To alleviate these issues, we present a context-aware emotion recognition framework that combines four complementary contexts. The first context is multimodal emotion recognition based on facial expression, facial landmarks, gesture and gait. Secondly, we adopt the channel and spatial attention modules to obtain the emotion semantics of the scene context. Inspired by sociology theory, we explore the emotion transmission between agents by constructing relationship graphs in the third context. Meanwhile, we propose a novel agent-object context, which aggregates emotion cues from the interactions between surrounding agents and objects in the scene to mitigate the ambiguity of prediction. Finally, we introduce an adaptive relevance fusion module for learning the shared representations among multiple contexts. Extensive experiments show that our approach outperforms the state-of-the-art methods on both EMOTIC and GroupWalk datasets. We also release a dataset annotated with diverse emotion labels, Human Emotion in Context (HECO). In practice, we compare with the existing methods on the HECO, and our approach obtains a higher classification average precision of 50.65% and a lower regression mean error rate of 0.7. The project is available at https://heco2022.github.io/.

Matan Levy, Rami Ben-Ari, Dani Lischinski

Chart question answering (CQA) is a task used for assessing chart comprehension, which is fundamentally different from understanding natural images. CQA requires analyzing the relationships between the textual and the visual components of a chart, in order to answer general questions or infer numerical values. Most existing CQA datasets and models are based on simplifying assumptions that often enable surpassing human performance. In this work, we address this outcome and propose a new model that jointly learns classification and regression. Our language-vision setup uses co-attention transformers to capture the complex real-world interactions between the question and the textual elements. We validate our design with extensive experiments on the realistic PlotQA dataset, outperforming previous approaches by a large margin, while showing competitive performance on FigureQA. Our model is particularly well suited for realistic questions with out-of-vocabulary answers that require regression.

Jingcheng Ni, Nan Zhou, Jie Qin, Qian Wu, Junqi Liu, Boxun Li, Di Huang

Contrastive learning has shown great potential in video representation learning. However, existing approaches fail to sufficiently exploit short-term motion dynamics, which are crucial to various down-stream video understanding tasks. In this paper, we propose Motion Sensitive Contrastive Learning (MSCL) that injects the motion information captured by optical flows into RGB frames to strengthen feature learning. To achieve this, in addition to clip-level global contrastive learning, we develop Local Motion Contrastive Learning (LMCL) with frame-level contrastive objectives across the two modalities. Moreover, we introduce Flow Rotation Augmentation (FRA) to generate extra motion-shuffled negative samples and Motion Differential Sampling (MDS) to accurately screen training samples. Extensive experiments on standard benchmarks validate the effectiveness of the proposed method. With the commonly-used 3D ResNet-18 as the backbone, we achieve the top-1 accuracies of 91.5\% on UCF101 and 50.3\% on Something-Something v2 for video classification, and a 65.6\% Top-1 Recall on UCF101 for video retrieval, notably improving the state-of-the-art.

Md Mohaiminul Islam, Gedas Bertasius

Most modern video recognition models are designed to operate on short video clips (e.g., 5-10s in length). Thus, it is challenging to apply such models to long movie understanding tasks, which typically require sophisticated long-range temporal reasoning. The recently introduced video transformers partially address this issue by using long-range temporal self-attention. However, due to the quadratic cost of self-attention, such models are often costly and impractical to use. Instead, we propose ViS4mer, an efficient long-range video model that combines the strengths of self-attention and the recently introduced structured state-space sequence (S4) layer. Our model uses a standard Transformer encoder for short-range spatiotemporal feature extraction, and a multi-scale temporal S4 decoder for subsequent long-range temporal reasoning. By progressively reducing the spatiotemporal feature resolution and channel dimension at each decoder layer, ViS4mer learns complex long-range spatiotemporal dependencies in a video. Furthermore, ViS4mer is 2.63x faster and requires 8x less GPU memory than the corresponding pure self-attention-based model. Additionally, ViS4mer achieves state-of-the-art results in 6 out of 9 long-form movie video classification tasks on the Long Video Understanding (LVU) benchmark. Furthermore, we show that our approach successfully generalizes to other domains, achieving competitive results on the Breakfast and the COIN procedural activity datasets. The code is publicly available.

Junqiang Huang, Xiangwen Kong, Xiangyu Zhang

We focus on better understanding the critical factors of augmentation-invariant representation learning. We revisit MoCo v2 and BYOL and try to prove the authenticity of the following assumption: different frameworks bring about representations of different characteristics even with the same pretext task. We establish the first benchmark for fair comparisons between MoCo v2 and BYOL, and observe: (i) sophisticated model configurations enable better adaptation to pre-training dataset; (ii) mismatched optimization strategies of pre-training and fine-tuning hinder model from achieving competitive transfer performances. Given the fair benchmark, we make further investigation and find asymmetry of network structure endows contrastive frameworks to work well under the linear evaluation protocol, while may hurt the transfer performances on long-tailed classification tasks. Moreover, negative samples do not make models more sensible to the choice of data augmentations, nor does the asymmetric network structure. We believe our findings provide useful information for future work.

Huseyin Coskun, Alireza Zareian, Joshua L. Moore, Federico Tombari, Chen Wang

Clustering is a ubiquitous tool in unsupervised learning. Most of the existing self-supervised representation learning methods typically cluster samples based on visually dominant features. While this works well for image-based selfsupervision, it often fails for videos, which require understanding motion rather than focusing on background. Using optical flow as complementary information to RGB can alleviate this problem. However, we observe that a na¨ıve combination of the two modalities does not provide meaningful gains. In this paper, we propose a principled way to combine two modalities. Specifically, we propose a novel clustering strategy where we use the initial cluster assignment of each modality as prior to guide the final cluster assignment of the other modality. This idea will enforce similar cluster structures for both modalities, and the formed clusters will be semantically abstract and robust to noisy inputs coming from each individual modality. Additionally, we propose a novel regularization strategy to address the feature collapse problem, which is common in cluster-based self-supervised learning methods. Our extensive evaluation shows the effectiveness of our learned representations on downstream tasks, e.g., video retrieval and action recognition. Specifically, we outperform the state of the art by 7% on UCF and 4% on HMDB for video retrieval as well as 5% on UCF and 6% on HMDB for linear video classification.

Xiao-Juan Li, Jie Yang, Fang-Lue Zhang

Deep learning-based approaches for shape understanding and processing tasks have attracted considerable attention. Despite the great progress that has been made, the existing approaches fail to efficiently capture sophisticated structure information and critical part features simultaneously, limiting their capability of providing discriminative deep shape features. To address the above issue, we proposed a novel deep learning framework, Laplacian Mesh Transformer, to extract the critical structure and geometry features. We introduce a dual attention mechanism, where the $1^{\rm st}$ level self-attention mechanism is used to capture the structure and critical partial geometric information on the entire mesh, and the $2^{\rm nd}$ level is to learn the importance of structure and geometric information. More particularly, Laplacian spectral decomposition is adopted as our basic structure representation given its ability to describe shape topology. Our approach builds a hierarchical structure to process shape features from fine to coarse using the dual attention mechanism, which is stable under the isometric transformations. It enables an effective feature extraction that can tackle 3D meshes with complex structure and geometry efficiently in various shape analysis tasks, such as shape segmentation and classification. Extensive experiments on the standard benchmarks show that our method outperforms state-of-the-art methods.

Lingzhi Zhang, Shenghao Zhou, Simon Stent, Jianbo Shi

Egocentric videos offer fine-grain information for high-fidelity modeling of human behaviors. Hands and interacting objects are one crucial aspect of understanding viewer’s behaviors and intentions. We provide a labeled dataset consisting of 11,235 egocentric images with per-pixel segmentation labels of hands and the interacting objects in diverse daily activities. Our dataset is the first to label detailed interacting hand-object contact boundaries. We introduce a context-aware compositional data augmentation technique to adapt to out-of-the-distribution YouTube egocentric video. We show that our robust hand-object segmentation model and dataset can serve as a foundation tool to boost or enable several downstream vision applications, such as: Hand state classification, video activity recognition, 3D mesh reconstruction of hand-object interaction, and Seeing through the hand with video inpainting in egocentric videos. All of our data and code will be released to the public.

Chuhui Xue, Jiaxing Huang, Wenqing Zhang, Shijian Lu, Changhu Wang, Song Bai

Most existing scene text detectors focus on detecting characters or words that only capture partial text messages due to missing contextual information. For a better understanding of text in scenes, it is more desired to detect contextual text blocks (CTBs) which consist of one or multiple integral text units (e.g., characters, words, or phrases) in natural reading order and transmit certain complete text messages. This paper presents contextual text detection, a new setup that detects CTBs for better understanding of texts in scenes. We formulate the new setup by a dual detection task which first detects integral text units and then groups them into a CTB. To this end, we design a novel scene text clustering technique that treats integral text units as tokens and groups them (belonging to the same CTB) into an ordered token sequence. In addition, we create two datasets SCUT-CTW-Context and ReCTS-Context to facilitate future research, where each CTB is well annotated by an ordered sequence of integral text units. Further, we introduce three metrics that measure contextual text detection in local accuracy, continuity, and global accuracy. Extensive experiments show that our method accurately detects CTBs which effectively facilitates downstream tasks such as text classification and translation. The project is available at \url{https://sg-vilab.github.io/publication/xue2022contextual/}.

Xudong Xie, Ling Fu, Zhifei Zhang, Zhaowen Wang, Xiang Bai

Artistic text recognition is an extremely challenging task with a wide range of applications. However, current scene text recognition methods mainly focus on irregular text while have not explored artistic text specifically. The challenges of artistic text recognition include the various appearance with special-designed fonts and effects, the complex connections and overlaps between characters, and the severe interference from background patterns. To alleviate these problems, we propose to recognize the artistic text at three levels. Firstly, corner points are applied to guide the extraction of local features inside characters, considering the robustness of corner structures to appearance and shape. In this way, the discreteness of the corner points cuts off the connection between characters, and the sparsity of them improves the robustness for background interference. Secondly, we design a character contrastive loss to model the character-level feature, improving the feature representation for character classification. Thirdly, we utilize Transformer to learn the global feature on image-level and model the global relationship of the corner points, with the assistance of a corner-query cross-attention mechanism. Besides, we provide an artistic text dataset to benchmark the performance. Experimental results verify the significant superiority of our proposed method on artistic text recognition and also achieve state-of-the-art performance on several blurred and perspective datasets.

Jaesung Choe, Chunghyun Park, Francois Rameau, Jaesik Park, In So Kweon

MLP-Mixer has newly appeared as a new challenger against the realm of CNNs and Transformer. Despite its simplicity compared to Transformer, the concept of channel-mixing MLPs and token-mixing MLPs achieves noticeable performance in image recognition tasks. Unlike images, point clouds are inherently sparse, unordered and irregular, which limits the direct use of MLP-Mixer for point cloud understanding. To overcome these limitations, we propose PointMixer, a universal point set operator that facilitates information sharing among unstructured 3D point cloud. By simply replacing token-mixing MLPs with Softmax function, PointMixer can “mix"" features within/between point sets. By doing so, PointMixer can be broadly used for intra-set, inter-set, and hierarchical-set mixing. We demonstrate that various channel-wise feature aggregation in numerous point sets is better than self-attention layers or dense token-wise interaction in a view of parameter efficiency and accuracy. Extensive experiments show the competitive or superior performance of PointMixer in semantic segmentation, classification, and reconstruction against Transformer-based methods.

Apostolos Modas, Rahul Rade, Guillermo Ortiz-Jiménez, Seyed-Mohsen Moosavi-Dezfooli, Pascal Frossard

Despite their impressive performance on image classification tasks, deep networks have a hard time generalizing to unforeseen corruptions of their data. To fix this vulnerability, prior works have built complex data augmentation strategies, combining multiple methods to enrich the training data. However, introducing intricate design choices or heuristics makes it hard to understand which elements of these methods are indeed crucial for improving robustness. In this work, we take a step back and follow a principled approach to achieve robustness to common corruptions. We propose PRIME, a general data augmentation scheme that relies on simple yet rich families of max-entropy image transformations. PRIME outperforms the prior art in terms of corruption robustness, while its simplicity and plug-and-play nature enable combination with other methods to further boost their robustness. We analyze PRIME to shed light on the importance of the mixing strategy on synthesizing corrupted images, and to reveal the robustness-accuracy trade-offs arising in the context of common corruptions. Finally, we show that the computational efficiency of our method allows it to be easily used in both on-line and off-line data augmentation schemes.

Otniel-Bogdan Mercea, Thomas Hummel, A. Sophia Koepke, Zeynep Akata

Audio-visual generalised zero-shot learning for video classification requires understanding the relations between the audio and visual information in order to be able to recognise samples from novel, previously unseen classes at test time. The natural semantic and temporal alignment between audio and visual data in video data can be exploited to learn powerful representations that generalise to unseen classes at test time. We propose a multi-modal and Temporal Cross-attention Framework for audio-visual generalised zero-shot learning. Its inputs are temporally aligned audio and visual features that are obtained from pre-trained networks. Encouraging the framework to focus on cross-modal correspondence across time instead of self-attention within the modalities boosts the performance significantly. We show that our proposed framework that ingests temporal features yields state-of-the-art performance on the UCF-GZSL, VGGSound-GZSL, and ActivityNet-GZSL benchmarks for (generalised) zero-shot learning.

Jindong Gu, Volker Tresp, Yao Qin

Recent advances in Vision Transformer (ViT) have demonstrated its impressive performance in image classification, which makes it a promising alternative to Convolutional Neural Network (CNN). Unlike CNNs, ViT represents an input image as a sequence of image patches. The patch-based input image representation makes the following question interesting: How does ViT perform when individual input image patches are perturbed with natural corruptions or adversarial perturbations, compared to CNNs? In this work, we study the robustness of ViT to patch-wise perturbations. Surprisingly, we {find} that ViTs are more robust to naturally corrupted patches than CNNs, whereas they are more vulnerable to adversarial patches. Furthermore, we discover that the attention mechanism greatly affects the robustness of vision transformers. Specifically, the attention module can help improve the robustness of ViT by effectively ignoring natural corrupted patches. However, when ViTs are attacked by an adversary, the attention mechanism can be easily fooled to focus more on the adversarially perturbed patches and cause a mistake. Based on our analysis, we propose a simple temperature-scaling based method to {improve} the robustness of ViT against adversarial patches. Extensive qualitative and quantitative experiments are performed to support our findings, understanding, and improvement of ViT robustness to patch-wise perturbations across a set of transformer-based architectures.

Wuyang Chen, Xianzhi Du, Fan Yang, Lucas Beyer, Xiaohua Zhai, Tsung-Yi Lin, Huizhong Chen, Jing Li, Xiaodan Song, Zhangyang Wang 等

This work presents a simple vision transformer design as a strong baseline for object localization and instance segmentation tasks. Transformers recently demonstrate competitive performance in image classification tasks. To adopt ViT to object detection and dense prediction tasks, many works inherit the multistage design from convolutional networks and highly customized ViT architectures. Behind this design, the goal is to pursue a better trade-off between computational cost and effective aggregation of multiscale global contexts. However, existing works adopt multistage architectural design as a black-box solution without a clear understanding of its true benefits. In this paper, we comprehensively study three architecture design choices on ViT -- spatial reduction, doubled channels, and multiscale features -- and demonstrate that a vanilla ViT architecture can fulfill this goal without handcrafting multiscale features, maintaining the original ViT design philosophy. We further complete a scaling rule to optimize our model’s trade-off on accuracy and computation cost / model size. By leveraging a constant feature resolution and hidden size throughout the encoder blocks, we propose a simple and compact ViT architecture called Universal Vision Transformer (UViT) that achieves strong performance on COCO object detection and instance segmentation benchmark. Our code will be released upon acceptance.

Jang Hyun Cho, Philipp Krähenbühl

Large-scale object detection and instance segmentation faces a severe data imbalance. The finer-grained object classes become, the less frequent they appear in our datasets. However at test-time, we expect a detector that performs well for all classes and not just the most frequent ones. In this paper, we provide a theoretical understanding of the long-trail detection problem. We show how the commonly used mean average precision evaluation metric on an unknown test-set is bound by a margin-based binary classification error on a long-tailed object-detection training set. We optimize margin-based binary classification error with a novel surrogate objective called Effective Class-Margin Loss (ECM). The ECM loss is simple, theoretically well-motivated, and outperforms other heuristic counterparts on LVIS v1 benchmark over a wide range of architecture and detectors. Code is available at https://github.com/janghyuncho/ECM-Loss.

Johannes Theodoridis, Jessica Hofmann, Johannes Maucher, Andreas Schilling

Do deep learning models for instance segmentation generalize to novel objects in a systematic way? For classification, such behavior has been questioned. In this study, we aim to understand if certain design decisions such as framework, architecture or pre-training contribute to the semantic understanding of instance segmentation. To answer this question, we consider a special case of robustness and compare pre-trained models on a challenging benchmark for object-centric out-of-distribution texture. We do not introduce another method in this work. Instead, we take a step back and evaluate a broad range of existing literature. This includes Cascade and Mask R-CNN, Swin Transformer, BMask, YOLACT(++), DETR, BCNet, SOTR and SOLOv2. We find that YOLACT++, SOTR and SOLOv2 are significantly more robust to out-of-distribution texture than other frameworks. In addition, we show that deeper and dynamic architectures improve robustness whereas training schedules, data augmentation and pre-training have only a minor impact. In summary we evaluate 68 models on 61 versions of MS COCO for a total of 4148 evaluations.