This paper proposes a novel Unified Feature Optimization (UFO) paradigm for training and deploying deep models under real-world and large-scale scenarios, which requires a collection of multiple AI functions. UFO aims to benefit each single task with a large-scale pretraining on all tasks. Compared with existing foundation models, UFO has two different points of emphasis, i.e., relatively smaller model size and NO adaptation cost: 1) UFO squeezes a wide range of tasks into a moderate-sized unified model in a multi-task learning manner and further trims the model size when transferred to down-stream tasks. 2) UFO does not emphasize transfer to novel tasks. Instead, it aims to make the trimmed model dedicated for one or more already-seen task. To this end, it directly selects partial modules in the unified model, and thus requires completely NO adaptation cost. With these two characteristics, UFO provides great convenience for flexible deployment, while maintaining the benefits of large-scale pretraining. A key merit of UFO is that the trimming process not only reduces the model size and inference consumption, but also even improves the accuracy on certain tasks. Specifically, UFO considers the multi-task training and brings two-fold impact on the unified model: some closely related tasks have mutual benefits, while some tasks have conflicts against each other. UFO manages to reduce the conflicts and to preserve the mutual benefits through a novel Network Architecture Search (NAS) method. Experiments on a wide range of deep representation learning tasks (i.e., face recognition, person re-identification, vehicle re-identification and product retrieval) show that the model trimmed from UFO achieves higher accuracy than its single-task-trained counterpart and yet has smaller model size, validating the concept of UFO. Besides, UFO also supported the release of 17 billion parameters computer vision (CV) foundation model which is the largest CV model in the industry.
论文检索
输入标题、作者或关键词，从 1,645 篇学术成果中精准定位
CutMix is a popular augmentation technique commonly used for training modern convolutional and transformer vision networks. It was originally designed to encourage Convolution Neural Networks (CNNs) to focus more on an image’s global context instead of local information, which greatly improves the performance of CNNs. However, we found it to have limited benefits for transformer-based architectures that naturally have a global receptive field. In this paper, we propose a novel data augmentation technique TokenMix to improve the performance of vision transformers. TokenMix mixes two images at token level via partitioning the mixing region into multiple separated parts. Besides, we show that the mixed learning target in CutMix, a linear combination of a pair of the ground truth labels, might be inaccurate and sometimes counter-intuitive. To obtain a more suitable target, we propose to assign the target score according to the content-based neural activation maps of the two images from a pre-trained teacher model, which does not need to have high performance. With plenty of experiments on various vision transformer architectures, we show that our proposed TokenMix helps vision transformers focus on the foreground area to infer the classes and enhances their robustness to occlusion, with consistent performance gains. Notably, we improve DeiT-T/S/B with +1% ImageNet top-1 accuracy. Besides, TokenMix enjoys longer training, which achieves 81.2% top-1 accuracy on ImageNet with DeiT-S trained for 400 epochs.
Proxy-based Deep Metric Learning (DML) learns deep metric spaces by embedding images and class representatives (proxies) close to one another during training, as commonly measured by the angle between them. However, this disregards the embedding norm, which can carry additional beneficial context such as class- or image-intrinsic uncertainty. In addition, proxy-based DML struggles to learn class-internal structures. To address both issues at once, we introduce non-isotropic probabilistic proxy-based DML. We model images as directional von Mises-Fisher (vMF) distributions on the hypersphere and motivate a non-isotropic von Mises-Fisher (nivMF) model for class proxies. This allows proxies to better represent complex class-specific variances in the embedding space. To cast these probabilistic models into proxy-to-image metrics, we further develop and investigate multiple distribution-to-point and distribution-to-distribution metrics. Each framework choice is motivated through a set of ablational studies, which showcase beneficial properties of our probabilistic approach to proxy-based DML. This comprises uncertainty-awareness, better behaved gradients during training and overall improved generalization performance. The latter is especially reflected in the competitive performance on the standard DML benchmarks. We find our proposed approach to compare favourably, suggesting that existing proxy-based DML can significantly benefit from a more probabilistic treatment.
Noisy label Facial Expression Recognition (FER) is more challenging than traditional noisy label classification tasks due to the inter-class similarity and the annotation ambiguity. Recent works mainly tackle this problem by filtering out large-loss samples. In this paper, we explore dealing with noisy labels from a new feature-learning perspective. We find that FER models remember noisy samples by focusing on a part of the features that can be considered related to the noisy labels instead of learning from the whole features that lead to the latent truth. Inspired by that, we propose a novel Erasing Attention Consistency (EAC) method to suppress the noisy samples during the training process automatically. Specifically, we first utilize the flip semantic consistency of facial images to design an imbalanced framework. We then randomly erase input images and use flip attention consistency to prevent the model from focusing on a part of the features. EAC significantly outperforms state-of-the-art noisy label FER methods and generalizes well to other tasks with a large number of classes like CIFAR100 and Tiny-ImageNet. The code is available at https://github.com/zyh-uaiaaaa/Erasing-Attention-Consistency.
Deep Metric Learning (DML) serves to learn an embedding function to project semantically similar data into nearby embedding space and plays a vital role in many applications, such as image retrieval and face recognition. However, the performance of DML methods often highly depends on sampling methods to choose effective data from the embedding space in the training. In practice, the embeddings in the embedding space are obtained by some deep models, where the embedding space is often with barren area due to the absence of training points, resulting in so called “missing embedding” issue. This issue may impair the sample quality, which leads to degenerated DML performance. In this work, we investigate how to alleviate the “missing embedding” issue to improve the sampling quality and achieve effective DML. To this end, we propose a Densely-Anchored Sampling (DAS) scheme that considers the embedding with corresponding data point as “anchor” and exploits the anchor’s nearby embedding space to densely produce embeddings without data points. Specifically, we propose to exploit the embedding space around single anchor with Discriminative Feature Scaling (DFS) and multiple anchors with Memorized Transformation Shifting (MTS). In this way, by combing the embeddings with and without data points, we are able to provide more embeddings to facilitate the sampling process thus boosting the performance of DML. Our method is effortlessly integrated into existing DML frameworks and improves them without bells and whistles. Extensive experiments on three benchmark datasets demonstrate the superiority of our method.
Graph-based models have achieved great success in person re-identification tasks recently, which compute the graph topology structure (affinities) among different people first and then pass the information across them to achieve stronger features. But we find existing graph-based methods in the visible-infrared person re-identification task (VI-ReID) suffer from bad generalization because of two issues: 1) train-test modality balance gap, which is a property of VI-ReID task. The number of two modalities data are balanced in the training stage, but extremely unbalanced in inference, causing the low generalization of graph-based VI-ReID methods. 2) sub-optimal topology structure caused by the end-to-end learning manner to the graph module. We analyze that the well-trained input features weaken the learning of graph topology, making it not generalized enough during the inference process. In this paper, we propose a Counterfactual Intervention Feature Transfer (CIFT) method to tackle these problems. Specifically, a Homogeneous and Heterogeneous Feature Transfer (H2FT) is designed to reduce the train-test modality balance gap by two independent types of well-designed graph modules and an unbalanced scenario simulation. Besides, a Counterfactual Relation Intervention (CRI) is proposed to utilize the counterfactual intervention and causal effect tools to highlight the role of topology structure in the whole training process, which makes the graph topology structure more reliable. Extensive experiments on standard VI-ReID benchmarks demonstrate that CIFT outperforms the state-of-the-art methods under various settings.
The subpopulation shifting challenge, known as some subpopulations of a category that are not seen during training, severely limits the classification performance of the state-of-the-art convolutional neural networks. Thus, to mitigate this practical issue, we explore incremental subpopulation learning (ISL) to adapt the original model via incrementally learning the unseen subpopulations without retaining the seen population data. However, striking a great balance between subpopulation learning and seen population forgetting is the main challenge in ISL but is not well studied by existing approaches. These incremental learners simply use a pre-defined and fixed hyperparameter to balance the learning objective and forgetting regularization, but their learning is usually biased towards either side in the long run. In this paper, we propose a novel two-stage learning scheme to explicitly disentangle the acquisition and forgetting for achieving a better balance between subpopulation learning and seen population forgetting: in the first “gain-acquisition” stage, we progressively learn a new classifier based on the margin-enforce loss, which enforces the hard samples and population to have a larger weight for classifier updating and avoid uniformly updating all the population; in the second “counter-forgetting” stage, we search for the proper combination of the new and old classifiers by optimizing a novel objective based on proxies of forgetting and acquisition. We benchmark the representative and state-of-the-art non-exemplar-based incremental learning methods on a large-scale subpopulation shifting dataset for the first time. Under almost all the challenging ISL protocols, we significantly outperform other methods by a large margin, demonstrating our superiority to alleviate the subpopulation shifting problem.
Knowledge distillation conditioned on intermediate feature representations always leads to significant performance improvements. Conventional feature distillation framework demands extra selecting/training budgets of teachers and complex transformations to align the features between teacher-student models. To address the problem, we analyze teacher roles in feature distillation and have an intriguing observation: additional teacher architectures are not always necessary. Then we propose Tf-FD, a simple yet effective Teacher-free Feature Distillation framework, reusing channel-wise and layer-wise meaningful features within the student to provide teacher-like knowledge without an additional model. In particular, our framework is subdivided into intra-layer and inter-layer distillation. The intra-layer Tf-FD performs feature salience ranking and transfers the knowledge from salient feature to redundant feature within the same layer. For inter-layer Tf-FD, we deal with distilling high-level semantic knowledge embedded in the deeper layer representations to guide the training of shallow layers. Benefiting from the small gap between these self-features, Tf-FD simply needs to optimize extra feature mimicking losses without complex transformations. Furthermore, we provide insightful discussions to shed light on Tf-FD from feature regularization perspectives. Our experiments conducted on classification and object detection tasks demonstrate that our technique achieves state-of-the-art results on different models with fast training speeds. Code is available at https://lilujunai.github.io/Teacher-free-Distillation/.
Comparing the functional behavior of neural network models, whether it is a single network over time or two (or more networks) during or post-training, is an essential step in understanding what they are learning (and what they are not), and for identifying strategies for regularization or efficiency improvements. Despite recent progress, e.g., comparing vision transformers to CNNs, systematic comparison of function, especially across different networks, remains difficult and is often carried out layer by layer. Approaches such as canonical correlation analysis (CCA) are applicable in principle, but have been sparingly used so far. In this paper, we revisit a (less widely known) from statistics, called distance correlation (and its partial variant), designed to evaluate correlation between feature spaces of different dimensions. We describe the steps necessary to carry out its deployment for large scale models -- this opens the door to a surprising array of applications ranging from conditioning one deep model w.r.t. another, learning disentangled representations as well as optimizing diverse models that would directly be more robust to adversarial attacks. Our experiments suggest a versatile regularizer (or constraint) with many advantages, which avoids some of the common difficulties one faces in such analyses.
Recent progress in 4D implicit representation focuses on globally controlling the shape and motion with low dimensional latent vectors, which is prone to missing surface details and accumulating tracking error. While many deep local representations have shown promising results for 3D shape modeling, their 4D counterpart does not exist yet. In this paper, we fill this blank by proposing a novel Local 4D implicit Representation for Dynamic clothed human, named LoRD, which has the merits of both 4D human modeling and local representation, and enables high-fidelity reconstruction with detailed surface deformations, such as clothing wrinkles. Particularly, our key insight is to encourage the network to learn the latent codes of local part-level representation, capable of explaining the local geometry and temporal deformations. To make the inference at test-time, we first estimate the inner body skeleton motion to track local parts at each time step, and then optimize the latent codes for each part via auto-decoding based on different types of observed data. Extensive experiments demonstrate that the proposed method has strong capability for representing 4D human, and outperforms state-of-the-art methods on practical applications, including 4D reconstruction from sparse points, non-rigid depth fusion, both qualitatively and quantitatively.
Contrastive-based self-supervised learning methods achieved great success in recent years. However, self-supervision requires extremely long training epochs (e.g., 800 epochs for MoCo v3) to achieve promising results, which is unacceptable for the general academic community and hinders the development of this topic. This work revisits the momentum-based contrastive learning frameworks and identifies the inefficiency in which two augmented views generate only one positive pair. We propose Fast-MoCo - a novel framework that utilizes combinatorial patches to construct multiple positive pairs from two augmented views, which provides abundant supervision signals that bring significant acceleration with neglectable extra computational cost. Fast-MoCo trained with 100 epochs achieves 73.5% linear evaluation accuracy, similar to MoCo v3 (ResNet-50 backbone) trained with 800 epochs. Extra training (200 epochs) further improves the result to 75.1%, which is on par with state-of-the-art methods. Experiments on several downstream tasks also confirm the effectiveness of Fast-MoCo.
Deep neural networks (DNNs) for point clouds have achieved superior performance on a range of 3D vision tasks. However, generalization to out-of-distribution 3D point clouds remains challenging for DNNs. Due to the expensive cost or even infeasibility of annotating large-scale point clouds, it is critical but yet has not been fully explored to design methods to generalize DNN models to unseen domains of point clouds without any access to them during the training process. In this paper, we propose to learn 3D point cloud representation on a seen source domain and generalize to an unseen target domain via adversarial learning. Specifically, we unify several geometric transformations in a manifold-based framework under which distance between transformations is well-defined. Measured by the distance, adversarial samples are mined to form intermediate domains and retained in an adaptive replay-based memory. We further provide theoretical justification for the intermediate domains to reduce the generalization error of the DNN models. Experimental results on synthetic-to-real datasets illustrate that our method outperforms existing 3D deep learning models for domain generalization.
Continual learning requires incremental compatibility with a sequence of tasks. However, the design of model architecture remains an open question: In general, learning all tasks with a shared set of parameters suffers from severe interference between tasks; while learning each task with a dedicated parameter subspace is limited by scalability. In this work, we theoretically analyze the generalization errors for learning plasticity and memory stability in continual learning, which can be uniformly upper-bounded by (1) discrepancy between task distributions, (2) flatness of loss landscape and (3) cover of parameter space. Then, inspired by the robust biological learning system that processes sequential experiences with multiple parallel compartments, we propose Cooperation of Small Continual Learners (CoSCL) as a general strategy for continual learning. Specifically, we present an architecture with a fixed number of narrower sub-networks to learn all incremental tasks in parallel, which can naturally reduce the two errors through improving the three components of the upper bound. To strengthen this advantage, we encourage to cooperate these sub-networks by penalizing the difference of predictions made by their feature representations. With a fixed parameter budget, CoSCL can improve a variety of representative continual learning approaches by a large margin (e.g., up to 10.64% on CIFAR-100-SC, 9.33% on CIFAR-100-RS, 11.45% on CUB-200-2011 and 6.72% on Tiny-ImageNet) and achieve the new state-of-the-art performance. Our code is available at https://github.com/lywang3081/CoSCL.
DisCo: Remedying Self-Supervised Learning on Lightweight Models with Distilled Contrastive Learning
PDF ↗While Self-Supervised Learning (SSL) has received widespread attention from the community, recent researches argue that its performance often suffers a cliff fall when the model size decreases. Since current SSL methods mainly rely on contrastive learning to train the network, we propose a simple yet effective method termed Distilled Contrastive Learning (DisCo) to ease this issue. Specifically, we find that the final inherent embedding of the mainstream SSL methods contains the most important information and propose to distill the final embedding to maximally transmit a teacher’s knowledge to a lightweight model by constraining the last embedding of the student to be consistent with that of the teacher. In addition, we find that there exists a phenomenon termed Distilling BottleNeck and propose to enlarge the embedding dimension to alleviate this problem. Since the MLP only exists during the SSL phase, our method does not introduce any extra parameters to lightweight models for the downstream task deployment. Experimental results demonstrate that our method surpasses the state-of-the-art on many lightweight models by a large margin. Particularly, when ResNet-101/ResNet-50 is used respectively as a teacher to teach EfficientNet-B0, the linear result of EfficientNet-B0 on ImageNet is improved by 22.1% and 19.7%, respectively, which is very close to ResNet-101/ResNet-50 with much fewer parameters. Code is available at https://github.com/Yuting-Gao/DisCo-pytorch.
Continual learning is a learning paradigm that learns tasks sequentially with resources constraints, in which the key challenge is stability-plasticity dilemma, i.e., it is uneasy to simultaneously have the stability to prevent catastrophic forgetting of old tasks and the plasticity to learn new tasks well. In this paper, we propose a new continual learning approach, Advanced Null Space (AdNS), to balance the stability and plasticity without storing any old data of previous tasks. Specifically, to obtain better stability, AdNS makes use of low-rank approximation to obtain a novel null space and projects the gradient onto the null space to prevent the interference on the past tasks. To control the generation of the null space, we introduce a non-uniform constraint strength to further reduce forgetting. Furthermore, we present a simple but effective method, intra-task distillation, to improve the performance of the current task. Finally, we theoretically find that null space plays a key role in plasticity and stability, respectively. Experimental results show that the proposed method can achieve better performance compared to state-of-the-art continual learning approaches.
Deep metric learning (DML) attempts to learn a representation model as well as a metric function with a limited generalization gap, so that the model trained on finite known data can achieve similitude performance on infinite unseen data. While considerable efforts have been made to bound the generalization gap by enhancing the model architecture and training protocol a priori in the training phase, none of them notice that a lightweight posterior refinement operation on the trained metric matrix can significantly improve the generalization ability. In this paper, we attempt to fill up this research gap and theoretically analyze the impact of the refined metric matrix property on the generalization gap. Based on our theory, two principles, which suggest a smaller trace or a smaller Frobenius norm of the refined metric matrix, are proposed as guidance for the posterior refinement operation. Experiments on three benchmark datasets verify the correctness of our principles and demonstrate that a pluggable posterior refinement operation is potential to significantly improve the performance of existing models with negligible extra computation burden.
Despite the success of fully-supervised human skeleton sequence modeling, utilizing self-supervised pre-training for skeleton sequence representation learning has been an active field because acquiring task-specific skeleton annotations at large scales is difficult. Recent studies focus on learning video-level temporal and discriminative information using contrastive learning, but overlook the hierarchical spatial-temporal nature of human skeletons. Different from such superficial supervision at the video level, we propose a self-supervised hierarchical pre-training scheme incorporated into a hierarchical Transformer-based skeleton sequence encoder (Hi-TRS), to explicitly capture spatial, short-term, and long-term temporal dependencies at frame, clip, and video levels, respectively. To evaluate the proposed self-supervised pre-training scheme with Hi-TRS, we conduct extensive experiments covering three skeleton-based downstream tasks including action recognition, action detection, and motion prediction. Under both supervised and semi-supervised evaluation protocols, our method achieves the state-of-the-art performance. Additionally, we demonstrate that the prior knowledge learned by our model in the pre-training stage has strong transfer capability for different downstream tasks.
Federated Learning aims at training a global model from multiple decentralized devices (i.e. clients) without exchanging their private local data. A key challenge is the handling of non-i.i.d. (independent identically distributed) data across multiple clients that may induce disparities of their local features. We introduce the Hyperspherical Federated Learning (SphereFed) framework to address the non-i.i.d. issue by constraining learned representations of data points to be on a unit hypersphere shared by clients. Specifically, all clients learn their local representations by minimizing the loss with respect to a fixed classifier whose weights span the unit hypersphere. After federated training in improving the global model, this classifier is further calibrated with a closed-form solution by minimizing a mean squared loss. We show that the calibration solution can be computed efficiently and distributedly without direct access of local data. Extensive experiments indicate that our SphereFed approach is able to improve the accuracy of multiple existing federated learning algorithms by a considerable margin (up to 6% on challenging datasets) with enhanced computation and communication efficiency across datasets and model architectures.
In this paper, we propose a novel learning scheme for self-supervised video representation learning. Motivated by how humans understand videos, we propose to first learn general visual concepts then attend to discriminative local areas for video understanding. Specifically, we utilize static frame and frame difference to help decouple static and dynamic concepts, and respectively align the concept distributions in latent space. We add diversity and fidelity regularizations to guarantee that we learn a compact set of meaningful concepts. Then we employ a cross-attention mechanism to aggregate detailed local features of different concepts, and filter out redundant concepts with low activations to perform local concept contrast. Extensive experiments demonstrate that our method distills meaningful static and dynamic concepts to guide video understanding, and obtains state-of-the-art results on UCF-101, HMDB-51, and Diving-48.
A self-driving perception model aims to extract 3D semantic representations from multiple cameras collectively into the bird’s-eye-view (BEV) coordinate frame of the ego car in order to ground downstream planner. Existing perception methods often rely on error-prone depth estimation of the whole scene or learning sparse virtual 3D representations without the target geometry structure, both of which remain limited in performance and/or capability. In this paper, we present a novel end-to-end architecture for ego 3D representation learning from an arbitrary number of unconstrained camera views. Inspired by the ray tracing principle, we design a polarized grid of “imaginary eyes"" as the learnable ego 3D representation and formulate the learning process with the adaptive attention mechanism in conjunction with the 3D-to-2D projection. Critically, this formulation allows extracting rich 3D representation from 2D images without any depth supervision, and with the built-in geometry structure consistent w.r.t. BEV. Despite its simplicity and versatility, extensive experiments on standard BEV visual tasks (e.g., camera-based 3D object detection and BEV segmentation) show that our model outperforms all state-of-the-art alternatives significantly, with an extra advantage in computational efficiency from multi-task learning.