Vision Transformers (ViTs) have recently dominated a range of computer vision tasks, yet it suffers from low training data efficiency and inferior local semantic representation capability without appropriate inductive bias. Convolutional neural networks (CNNs) inherently capture regional-aware semantics, inspiring researchers to introduce CNNs back into the architecture of the ViTs to provide desirable inductive bias for ViTs. However, is the locality achieved by the micro-level CNNs embedded in ViTs good enough? In this paper, we investigate the problem by profoundly exploring how the macro architecture of the hybrid CNNs/ViTs enhances the performances of hierarchical ViTs. Particularly, we study the role of token embedding layers, alias convolutional embedding (CE), and systemically reveal how CE injects desirable inductive bias in ViTs. Besides, we apply the optimal CE configuration to 4 recently released state-of-the-art ViTs, effectively boosting the corresponding performances. Finally, a family of efficient hybrid CNNs/ViTs, dubbed CETNets, are released, which may serve as generic vision backbones. Specifically, CETNets achieve 84.9% Top-1 accuracy on ImageNet-1K (training from scratch), 48.6% box mAP on the COCO benchmark, and 51.6% mIoU on the ADE20K, substantially improving the performances of the corresponding state-of-the-art baselines.
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
Activation functions play a central role in deep learning since they form an essential building stone of neural networks. In the last few years, the focus has been shifting towards investigating new types of activations that outperform the classical Rectified Linear Unit (ReLU) in modern neural architectures. Most recently, rational activation functions (RAFs) have awakened interest because they were shown to perform on par with state-of-the-art activations on image classification. Despite their apparent potential, prior formulations are either not safe, not smooth or not ""true"" rational functions, and they only work with careful initialisation. Aiming to mitigate these issues, we propose a novel, enhanced rational function, ERA, and investigate how to better accommodate the specific needs of these activations, to both network components and training regime. In addition to being more stable, the proposed function outperforms other standard ones across a range of lightweight network architectures on two different tasks: image classification and 3d human pose and shape reconstruction.
Dataset bias and spurious correlations can significantly impair generalization in deep neural networks. Many prior efforts have addressed this problem using either alternative loss functions or sampling strategies that focus on rare patterns. We propose a new direction: modifying the network architecture to impose inductive biases that make the network robust to dataset bias. Specifically, we propose OccamNets, which are biased to favor simpler solutions by design. OccamNets have two inductive biases. First, they are biased to use as little network depth as needed for an individual example. Second, they are biased toward using fewer image locations for prediction. While OccamNets are biased toward simpler hypotheses, they can learn more complex hypotheses if necessary. In experiments, OccamNets outperform or rival state-of-the-art methods using architectures that do not incorporate these inductive biases. Furthermore, we demonstrate that when these methods are combined with OccamNets results further improve.
Neural Architecture Search (NAS) aims to automatically produce network architectures suitable to specific tasks on given datasets. Unlike previous NAS strategies based on reinforcement learning, genetic algorithm, Bayesian optimization, and differential programming method, we formulate the NAS task as a Max-Flow problem on search space consisting of Directed Acyclic Graph (DAG) and thus propose a novel NAS approach, called MF-NAS, which defines the search space and designs the search strategy in a fully graphic manner. In MF-NAS, parallel edges with capacities are induced by combining different operations, including skip connection, convolutions, and pooling, and the weights and capacities of the parallel edges are updated iteratively during the search process. Moreover, we interpret MF-NAS from the perspective of nonparametric density estimation and show the relationship between the flow of a graph and the corresponding classification accuracy of neural network architecture. We evaluate the competitive efficacy of our proposed MF-NAS across different datasets with different search spaces that are used in DARTS/ENAS and NAS-Bench-201.
Training CNN for detection is time-consuming due to the large dataset and complex network modules, making it hard to search architectures on detection datasets directly, which usually requires vast search costs (usually tens and even hundreds of GPU-days). In contrast, this paper introduces an efficient framework, named EAutoDet, that can discover practical backbone and FPN architectures for object detection in 1.4 GPU-days. Specifically, we construct a supernet for both backbone and FPN modules and adopt the differentiable method. To reduce the GPU memory requirement and computational cost, we propose a kernel reusing technique by sharing the weights of candidate operations on one edge and consolidating them into one convolution. A dynamic channel refinement strategy is also introduced to search channel numbers. Extensive experiments show significant efficacy and efficiency of our method. In particular, the discovered architectures surpass state-of-the-art object detection NAS methods and achieve 40.1 mAP with 120 FPS and 49.2 mAP with 41.3 FPS on COCO test-dev set. We also transfer the discovered architectures to rotation detection task, which achieve 77.05 mAP on DOTA-v1.0 test set with 21.1M parameters. The code is publicly available at https://github.com/vicFigure/EAutoDet.
Mobile devices run deep learning models for various purposes, such as image classification and speech recognition. Due to the resource constraints of mobile devices, researchers have focused on either making a lightweight deep neural network (DNN) model using model pruning or generating an efficient code using compiler optimization. We found that the straightforward integration between model compression and compiler auto-tuning often does not produce the most efficient model for a target device. We propose CPrune, a compiler-informed model pruning for efficient target-aware DNN execution to support an application with a required target accuracy. CPrune makes a lightweight DNN model through informed pruning based on the structural information of programs built during the compiler tuning process. Our experimental results show that CPrune increases the DNN execution speed up to 2.73x compared to the state-of-the-art TVM auto-tune while satisfying the accuracy requirement.
Online continual learning (online CL) studies the problem of learning sequential tasks from an online data stream without task boundaries, aiming to adapt to new data while alleviating catastrophic forgetting on the past tasks. This paper proposes a framework Contrastive Vision Transformer (CVT), which designs a focal contrastive learning strategy based on a transformer architecture, to achieve a better stability-plasticity trade-off for online CL. Specifically, we design a new external attention mechanism for online CL that implicitly captures previous tasks’ information. Besides, CVT contains learnable focuses for each class, which could accumulate the knowledge of previous classes to alleviate forgetting. Based on the learnable focuses, we design a focal contrastive loss to rebalance contrastive learning between new and past classes and consolidate previously learned representations. Moreover, CVT contains a dual-classifier structure for decoupling learning current classes and balancing all observed classes. The extensive experimental results show that our approach achieves state-of-the-art performance with even fewer parameters on online CL benchmarks and effectively alleviates the catastrophic forgetting.
In monocular depth estimation, disturbances in the image context, like moving objects or reflecting materials, can easily lead to erroneous predictions. For that reason, uncertainty estimates for each pixel are necessary, in particular for safety-critical applications such as automated driving. We propose a post hoc uncertainty estimation approach for an already trained and thus fixed depth estimation model, represented by a deep neural network. The uncertainty is estimated with the gradients which are extracted with an auxiliary loss function. To avoid relying on ground-truth information for the loss definition, we present an auxiliary loss function based on the correspondence of the depth prediction for an image and its horizontally flipped counterpart. Our approach achieves state-of-the-art uncertainty estimation results on the KITTI and NYU Depth V2 benchmarks without the need to retrain the neural network. Models and code are publicly available at https://github.com/jhornauer/GrUMoDepth.
Neural fields have emerged as a new data representation paradigm and have shown remarkable success in various signal representations. Since they preserve signals in their network parameters, the data transfer by sending and receiving the entire model parameters prevents this emerging technology from being used in many practical scenarios. We propose streamable neural fields, a single model that consists of executable sub-networks of various widths. The proposed architectural and training techniques enable a single network to be streamable over time and reconstruct different qualities and parts of signals. For example, a smaller sub-network produces smooth and low-frequency signals, while a larger sub-network can represent fine details. Experimental results have shown the effectiveness of our method in various domains, such as 2D images, videos, and 3D signed distance functions. Finally, we demonstrate that our proposed method improves training stability, by exploiting parameter sharing.
One-shot Neural Architecture Search (NAS) has been widely used to discover architectures due to its efficiency. However, previous studies reveal that one-shot performance estimations of architectures might not be well correlated with their performances in stand-alone training because of the excessive sharing of operation parameters (i.e., large sharing extent) between architectures. Thus, recent methods construct even more over-parameterized supernets to reduce the sharing extent. But these improved methods introduce a large number of extra parameters and thus cause an undesirable trade-off between the training costs and the ranking quality. To alleviate the above issues, we propose to apply Curriculum Learning On Sharing Extent (CLOSE) to train the supernet both efficiently and effectively. Specifically, we train the supernet with a large sharing extent (an easier curriculum) at the beginning and gradually decrease the sharing extent of the supernet (a harder curriculum). To support this training strategy, we design a novel supernet (CLOSENet) that decouples the parameters from operations to realize a flexible sharing scheme and adjustable sharing extent. Extensive experiments demonstrate that CLOSE can obtain a better ranking quality across different computational budget constraints than other one-shot supernets, and is able to discover superior architectures when combined with various search strategies. Code is available at https://github.com/walkerning/aw_nas.
Large Vision & Language models pretrained on web-scale data provide representations that are invaluable for numerous V&L problems. However, it is unclear how they can be extended to reason about user-specific visual concepts in unstructured language. This problem arises in multiple domains, from personalized image retrieval to personalized interaction with smart devices. We introduce a new learning setup called Personalized Vision & Language (PerVL) with two new benchmark datasets for retrieving and segmenting user-specific (""""personalized"""") concepts “in the wild"""". In PerVL, one should learn personalized concepts (1) independently of the downstream task (2) allowing a pretrained model to reason about them with free language, and (3) without providing personalized negative examples. We propose an architecture for solving PerVL that operates by expanding the input vocabulary of a pretrained model with new word embeddings for the personalized concepts. The model can then simply employ them as part of a sentence. We demonstrate that our approach learns personalized visual concepts from a few examples and effectively applies them in image retrieval and semantic segmentation using rich textual queries. For example the model improves MRR by 51.1% (28.4% vs 18.8%) compared to the strongest baseline.
Few-shot learning performs classification tasks and regression tasks on scarce samples. As one of the most representative few-shot learning models, Prototypical Network represents each class as sample average, or a prototype, and measures the similarity of samples and prototypes by Euclidean distance. In this paper, we propose a framework of spectral filtering (shrinkage) for measuring the difference between query samples and prototypes, or namely the relative prototypes, in a reproducing kernel Hilbert space (RKHS). In this framework, we further propose a method utilizing Tikhonov regularization as the filter function for few-shot classification. We conduct several experiments to verify our method utilizing different kernels based on the miniImageNet dataset, tiered-ImageNet dataset and CIFAR-FS dataset. The experimental results show that the proposed model can perform the state-of-the-art. In addition, the experimental results show that the proposed shrinkage method can boost the performance. Source code is available at https://github.com/zhangtao2022/DSFN.
This paper presents a transformer framework for few-shot learning, termed TransVLAD, with one focus showing the power of locally aggregated descriptors for few-shot learning. Our TransVLAD model is simple: a standard transformer encoder following a NeXtVLAD aggregation module to output the locally aggregated descriptors. In contrast to the prevailing use of CNN as part of the feature extractor, we are the first to prove self-supervised learning like masked autoencoders (MAE) can deal with the overfitting of transformers in few-shot image classification. Besides, few-shot learning can benefit from this general-purpose pre-training. Then, we propose two methods to mitigate few-shot biases, supervision bias and simple-characteristic bias. The first method is introducing masking operation into fine-tuning, by which we accelerate fine-tuning (by more than 3x) and improve accuracy. The second one is adapting focal loss into soft focal loss to focus on hard characteristics learning. Our TransVLAD finally tops 10 benchmarks on five popular few-shot datasets by an average of more than 2%.
We study few-shot semantic segmentation that aims to segment a target object from a query image when provided with a few annotated support images of the target class. Several recent methods resort to a feature masking (FM) technique to discard irrelevant feature activations which eventually facilitates the reliable prediction of segmentation mask. A fundamental limitation of FM is the inability to preserve the fine-grained spatial details that affect the accuracy of segmentation mask, especially for small target objects. In this paper, we develop a simple, effective, and efficient approach to enhance feature masking (FM). We dub the enhanced FM as hybrid masking (HM). Specifically, we compensate for the loss of fine-grained spatial details in FM technique by investigating and leveraging a complementary basic input masking method. Experiments have been conducted on three publicly available benchmarks with strong few-shot segmentation (FSS) baselines. We empirically show improved performance against the current state-of-the-art methods by visible margins across different benchmarks. Our code and trained models are available at: https://github.com/moonsh/HM-Hybrid-Masking
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.
Inductive and Transductive Few-Shot Video Classification via Appearance and Temporal Alignments
PDF ↗We present a novel method for few-shot video classification, which performs appearance and temporal alignments. In particular, given a pair of query and support videos, we conduct appearance alignment via frame-level feature matching to achieve the appearance similarity score between the videos, while utilizing temporal order-preserving priors for obtaining the temporal similarity score between the videos. Moreover, we introduce a few-shot video classification framework that leverages the above appearance and temporal similarity scores across multiple steps, namely prototype-based training and testing as well as inductive and transductive prototype refinement. To the best of our knowledge, our work is the first to explore transductive few-shot video classification. Extensive experiments on both Kinetics and Something-Something V2 datasets show that both appearance and temporal alignments are crucial for datasets with temporal order sensitivity such as Something-Something V2. Our approach achieves similar or better results than previous methods on both datasets.
In this paper, we mainly focus on the problem of how to learn additional feature representations for few-shot image classification through pretext tasks (e.g., rotation or color permutation and so on). This additional knowledge generated by pretext tasks can further improve the performance of few-shot learning (FSL) as it differs from human-annotated supervision (i.e., class labels of FSL tasks). To solve this problem, we present a plug-in Hierarchical Tree Structure-aware (HTS) method, which not only learns the relationship of FSL and pretext tasks, but more importantly, can adaptively select and aggregate feature representations generated by pretext tasks to maximize the performance of FSL tasks. A hierarchical tree constructing component and a gated selection aggregating component is introduced to construct the tree structure and find richer transferable knowledge that can rapidly adapt to novel classes with a few labeled images. Extensive experiments show that our HTS can significantly enhance multiple few-shot methods to achieve new state-of-the-art performance on four benchmark datasets. The code is available at: https://github.com/remiMZ/HTS-ECCV22.
In this paper, we consider the framework of multi-task representation (MTR) learning where the goal is to use source tasks to learn a representation that reduces the sample complexity of solving a target task. We start by reviewing recent advances in MTR theory and show that they can provide novel insights for popular meta-learning algorithms when analyzed within this framework. In particular, we highlight a fundamental difference between gradient-based and metric-based algorithms in practice and put forward a theoretical analysis to explain it. Finally, we use the derived insights to improve the performance of meta-learning methods via a new spectral-based regularization term and confirm its efficiency through experimental studies on few-shot classification benchmarks. To the best of our knowledge, this is the first contribution that puts the most recent learning bounds of MTR theory into practice for the task of few-shot classification.
Dual Contrastive Learning with Anatomical Auxiliary Supervision for Few-Shot Medical Image Segmentation
PDF ↗Few-shot semantic segmentation is a promising solution for scarce data scenarios, especially for medical imaging challenges with limited training data. However, most of the existing few-shot segmentation methods tend to over rely on the images containing target classes, which may hinder its utilization of medical imaging data. In this paper, we present a few-shot segmentation model that employs anatomical auxiliary information from medical images without target classes for dual contrastive learning. The dual contrastive learning module performs comparison among vectors from the perspectives of prototypes and contexts, to enhance the discriminability of learned features and the data utilization. Besides, to distinguish foreground features from background features more friendly, a constrained iterative prediction module is designed to optimize the segmentation of the query image. Experiments on two medical image datasets show that the proposed method achieves performance comparable to state-of-the-art methods.
Mutually Reinforcing Structure with Proposal Contrastive Consistency for Few-Shot Object Detection
PDF ↗Few-shot object detection is based on the base set with abundant labeled samples to detect novel categories with scarce samples. The majority of former solutions are mainly based on meta-learning or transfer-learning, neglecting the fact that images from the base set might contain unlabeled novel-class objects, which easily leads to performance degradation and poor plasticity since those novel objects are served as the background. Based on the above phenomena, we propose a Mutually Reinforcing Structure Network (MRSN) to make rational use of unlabeled novel class instances in the base set. In particular, MRSN consists of a mining model which unearths unlabeled novel-class instances and an absorbed model which learns variable knowledge. Then, we design a Proposal Contrastive Consistency (PCC) module in the absorbed model to fully exploit class characteristics and avoid bias from unearthed labels. Furthermore,we propose a simple and effective data synthesis method undirectional-CutMix (UD-CutMix) to improve the robustness of model mining novel class instances, urge the model to pay attention to discriminative parts of objects and eliminate the interference of background information. Extensive experiments illustrate that our proposed approach achieves state-of-the-art results on PASCAL VOC and MS-COCO datasets. Our code will be released at https://github.com/MMatx/MRSN.