Semantic segmentation arises as the backbone of many vision systems, spanning from self-driving cars and robot navigation to augmented reality and teleconferencing. Frequently operating under stringent latency constraints within a limited resource envelope, optimising for efficient execution becomes important. At the same time, the heterogeneous capabilities of the target platforms and diverse constraints of different applications require the design and training of multiple target-specific segmentation models, leading to excessive maintenance costs. To this end, we propose a framework for converting state-of-the-art segmentation CNNs to Multi-Exit Semantic Segmentation (MESS) networks: specially trained models that employ parametrised early exits along their depth to i) dynamically save computation during inference on easier samples and ii) save training and maintenance cost by offering a post-training customisable speed-accuracy trade-off. Designing and training such networks naively can hurt performance. Thus, we propose novel two-staged training scheme for multi-exit networks. Furthermore, the parametrisation of MESS enables co-optimising the number, placement and architecture of the attached segmentation heads along with the exit policy, upon deployment via exhaustive search in <1GPUh. This allows MESS to rapidly adapt to the device capabilities and application requirements for each target use-case, offering a train-once-deploy-everywhere solution. MESS variants achieve latency gains of up to 2.83x with the same accuracy, or 5.33 pp higher accuracy for the same computational budget, compared to the original backbone network. Lastly, MESS delivers orders of magnitude faster architectural customisation, compared to state-of-the-art techniques.
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
Well-known activation functions like ReLU or Leaky ReLU are non-differentiable at the origin. Over the years, many smooth approximations of ReLU have been proposed using various smoothing techniques. We propose new smooth approximations of a non-differentiable activation function by convolving it with approximate identities. In particular, we present smooth approximations of Leaky ReLU and show that they outperform several well-known activation functions in various datasets and models. We call this function Smooth Activation Unit (SAU). Replacing ReLU by SAU, we get 5.63%, 2.95%, and 2.50% improvement with ShuffleNet V2 (2.0x), PreActResNet 50 and ResNet 50 models respectively on the CIFAR100 dataset and 2.31% improvement with ShuffleNet V2 (1.0x) model on ImageNet-1k dataset.
In attention mechanism research, most existing methods are hard to utilize well the information of the neural network with high computing efficiency due to heavy feature compression in the attention layer. This paper proposes a simple and general approach named Bridge Attention to address this issue. As a new idea, BA-Net straightforwardly integrates features from previous layers and effectively promotes information interchange. Only simple strategies are employed for the model implementation, similar to the SENet. Moreover, after extensively investigating the effectiveness of different previous features, we discovered a simple and exciting insight that bridging all the convolution outputs inside each block with BN can obtain better attention to enhance the performance of neural networks. BA-Net is effective, stable, and easy to use. A comprehensive evaluation of computer vision tasks demonstrates that the proposed approach achieves better performance than the existing channel attention methods regarding accuracy and computing efficiency.
Convolutional Neural Networks (CNNs) compression is crucial to deploying these models in edge devices with limited resources. Existing channel pruning algorithms for CNNs have achieved plenty of success on complex models. They approach the pruning problem from various perspectives and use different metrics to guide the pruning process. However, these metrics mainly focus on the model’s ‘outputs’ or ‘weights’ and neglect its ‘interpretations’ information. To fill in this gap, we propose to address the channel pruning problem from a novel perspective by leveraging the interpretations of a model to steer the pruning process, thereby utilizing information from both inputs and outputs of the model. However, existing interpretation methods cannot get deployed to achieve our goal as either they are inefficient for pruning or may predict non-coherent explanations. We tackle this challenge by introducing a selector model that predicts real-time smooth saliency masks for pruned models. We parameterize the distribution of explanatory masks by Radial Basis Function (RBF)-like functions to incorporate geometric prior of natural images in our selector model’s inductive bias. Thus, we can obtain compact representations of explanations to reduce the computational costs of our pruning method. We leverage our selector model to steer the network pruning by maximizing the similarity of explanatory representations for the pruned and original models. Extensive experiments on CIFAR-10 and ImageNet benchmark datasets demonstrate the efficacy of our proposed method. Our implementations are available at \url{https://github.com/Alii-Ganjj/InterpretationsSteeredPruning}
Automated augmentation is an emerging and effective technique to search for data augmentation policies to improve generalizability of deep neural network training. Most existing work focuses on constructing a unified policy applicable to all data samples in a given dataset, without considering sample or class variations. In this paper, we propose a novel two-stage data augmentation algorithm, named Label-Aware AutoAugment (LA3), which takes advantage of the label information, and learns augmentation policies separately for samples of different labels. LA3 consists of two learning stages, where in the first stage, individual augmentation methods are evaluated and ranked for each label via Bayesian Optimization aided by a neural predictor, which allows us to identify effective augmentation techniques for each label under a low search cost. And in the second stage, a composite augmentation policy is constructed out of a selection of effective as well as complementary augmentations, which produces significant performance boost and can be easily deployed in typical model training. Extensive experiments demonstrate that LA3 achieves excellent performance matching or surpassing existing methods on CIFAR-10 and CIFAR-100, and achieves a new state-of-the-art ImageNet accuracy of 79.97% on ResNet-50 among auto-augmentation methods, while maintaining a low computational cost.
Filter pruning is one of the most effective methods to compress deep convolutional networks (CNNs). In this paper, as a key component in filter pruning, We first propose a feature discrimination based filter importance criterion, namely Receptive Field Criterion (RFC). It turns the maximum activation responses that characterize the receptive field into probabilities, then measure the filter importance by the distribution of these probabilities from a new perspective of feature discrimination. However, directly applying RFC to global threshold pruning may lead to some problems, because global threshold pruning neglects the differences between different layers. Hence, we propose Distinguishing Layer Pruning based on RFC (DLRFC), i.e., discriminately prune the filters in different layers, which avoids measuring filters between different layers directly against filter criteria. Specifically, our method first selects relatively redundant layers by hard and soft changes of the network output, and then prunes only at these layers. The whole process dynamically adjusts redundant layers through iterations. Extensive experiments conducted on CIFAR-10/100 and ImageNet show that our method achieves state-of-the-art performance in several benchmarks.
Downsampling is widely adopted to achieve a good trade-off between accuracy and latency for visual recognition. However, the commonly used pooling layers are not learned, which causes possible loss of information. As another dimension reduction method, adaptive sampling weights and processes regions that are relevant to the task, which can better preserve useful information. However, the use of adaptive sampling has been limited to certain layers. In this paper, we show that using adaptive sampling as the main component in a deep neural network can improve network efficiency. In particular, we propose SSBNet which is built by inserting sampling layers into existing networks like ResNet. The proposed SSBNet achieved competitive results in the ImageNet and COCO datasets. For example, the SSB-ResNet-RS-200 achieved 82.6% accuracy in the ImageNet dataset, which is 0.6% higher than the baseline ResNet-RS-152 with similar complexity. Visualization shows the advantage of SSBNet in allowing different layers to focus on different positions, and ablation studies further validate the advantage of adaptive sampling over uniform methods.
Batch normalization (BN) is comprised of a normalization component followed by an affine transformation and has become essential for training deep neural networks. Standard initialization of each BN in a network sets the affine transformation scale and shift to 1 and 0, respectively. However, after training we have observed that these parameters do not alter much from their initialization. Furthermore, we have noticed that the normalization process can still yield overly large values, which is undesirable for training. We revisit the BN formulation and present a new initialization method and update approach for BN to address the aforementioned issues. Experiments are designed to emphasize and demonstrate the positive influence of proper BN scale initialization on performance, and use rigorous statistical significance tests for evaluation. The approach can be used with existing implementations at no additional computational cost. Source code is available at https://github.com/osu-cvl/revisiting-bn-init.
Knowledge distillation (KD) is an efficient approach to transfer the knowledge from a large “teacher” network to a smaller “student” network. Traditional KD methods require lots of labeled training samples and a white-box teacher (parameters are accessible) to train a good student. However, these resources are not always available in real-world applications. The distillation process often happens at an external party side where we do not have access to much data, and the teacher does not disclose its parameters due to security and privacy concerns. To overcome these challenges, we propose a black-box few-shot KD method to train the student with few unlabeled training samples and a black-box teacher. Our main idea is to expand the training set by generating a diverse set of out-of-distribution synthetic images using MixUp and a conditional variational auto-encoder. These synthetic images along with their labels obtained from the teacher are used to train the student. We conduct extensive experiments to show that our method significantly outperforms recent SOTA few/zero-shot KD methods on image classification tasks.
Dynamic Time Warping (DTW) is used for matching pairs of sequences and celebrated in applications such as forecasting the evolution of time series, clustering time series or even matching sequence pairs in few-shot action recognition. The transportation plan of DTW contains a set of paths; each path matches frames between two sequences under a varying degree of time warping, to account for varying temporal intra-class dynamics of actions. However, as DTW is the smallest distance among all paths, it may be affected by the feature uncertainty which varies across time steps/frames. Thus, in this paper, we propose to model the so-called aleatoric uncertainty of a differentiable (soft) version of DTW. To this end, we model the heteroscedastic aleatoric uncertainty of each path by the product of likelihoods from Normal distributions, each capturing variance of pair of frames. (The path distance is the sum of base distances between features of pairs of frames of the path.) The Maximum Likelihood Estimation (MLE) applied to a path yields two terms: (i) a sum of Euclidean distances weighted by the variance inverse, and (ii) a sum of log-variance regularization terms. Thus, our uncertainty-DTW is the smallest weighted path distance among all paths, and the regularization term (penalty for the high uncertainty) is the aggregate of log-variances along the path. The distance and the regularization term can be used in various objectives. We showcase forecasting the evolution of time series, estimating the Fréchet mean of time series, and supervised/unsupervised few-shot action recognition of the articulated human 3D body joints.
Developing neural models that accurately understand objects in 3D point clouds is essential for the success of robotics and autonomous driving. However, arguably due to the higher-dimensional nature of the data (as compared to images), existing neural architectures exhibit a large variety in their designs, including but not limited to the views considered, the format of the neural features, and the neural operations used. Lack of a unified framework and interpretation makes it hard to put these designs in perspective, as well as systematically explore new ones. In this paper, we begin by proposing a unified framework of such, with the key idea being factorizing the neural networks into a series of view transforms and neural layers. We demonstrate that this modular framework can reproduce a variety of existing works while allowing a fair comparison of backbone designs. Then, we show how this framework can easily materialize into a concrete neural architecture search (NAS) space, allowing a principled NAS-for-3D exploration. In performing evolutionary NAS on the 3D object detection task on the Waymo Open Dataset, not only do we outperform the state-of-the-art models, but also report the interesting finding that NAS tends to discover the same macro-level architecture concept for both the vehicle and pedestrian classes.
Vision transformers (ViTs) inherited the success of NLP but their structures have not been sufficiently investigated and optimized for visual tasks. One of the simplest solutions is to directly search the optimal one via the widely used neural architecture search (NAS) in CNNs. However, we empirically find this straightforward adaptation would encounter catastrophic failures and be frustratingly unstable for the training of superformer. In this paper, we argue that since ViTs mainly operate on token embeddings with little inductive bias, imbalance of channels for different architectures would worsen the weight-sharing assumption and cause the training instability as a result. Therefore, we develop a new cyclic weight-sharing mechanism for token embeddings of the ViTs, which enables each channel could more evenly contribute to all candidate architectures. Besides, we also propose identity shifting to alleviate the many-to-one issue in superformer and leverage weak augmentation and regularization techniques for more steady training empirically. Based on these, our proposed method, ViTAS, has achieved significant superiority in both DeiT- and Twins-based ViTs. For example, with only $1.4$G FLOPs budget, our searched architecture has $3.3\%$ ImageNet-$1$k accuracy than the baseline DeiT. With $3.0$G FLOPs, our results achieve $82.0\%$ accuracy on ImageNet-$1$k, and $45.9\%$ mAP on COCO$2017$ which is $2.4\%$ superior than other ViTs.
State-of-the-art methods for optical flow estimation rely on deep learning, which require complex sequential training schemes to reach optimal performances on real-world data. In this work, we introduce the COMBO deep network that explicitly exploits the brightness constancy (BC) model used in traditional methods. Since State-of-the-art methods for optical flow estimation rely on deep learning, which require complex sequential training schemes to reach optimal performances on real-world data. In this work, we introduce the COMBO deep network that explicitly exploits the brightness constancy (BC) model used in traditional methods. Since BC is an approximate physical model violated in several situations, we propose to train a physically-constrained network complemented with a data-driven network. We introduce a unique and meaningful flow decomposition between the physical prior and the data-driven complement, including an uncertainty quantification of the BC model. We derive a training scheme for learning the different components of the decomposition, in a supervised but also in a semi-supervised context. Experiments show that COMBO can improve performances over state-of-the-art supervised networks, eg RAFT, reaching state-of-the-art performances on several benchmarks. We highlight how COMBO can leverage the BC model and adapt to its limitations. Finally, we show that our semi-supervised method can significantly simplify the training procedure.
Recently, community has paid increasing attention on model scaling and contributed to developing a model family with a wide spectrum of scales. Current methods either simply resort to a one-shot NAS manner to construct a non-structural and non-scalable model family or rely on a manual yet fixed scaling strategy to scale an unnecessarily best base model. In this paper, we bridge both two components and propose ScaleNet to jointly search base model and scaling strategy so that the scaled large model can have more promising performance. Concretely, we design a super-supernet to embody models with different spectrum of sizes (e.g., FLOPs). Then, the scaling strategy can be learned interactively with the base model via a Markov chain-based evolution algorithm and generalized to develop even larger models. To obtain a decent super-supernet, we design a hierarchical sampling strategy to enhance its training sufficiency and alleviate the disturbance. Experimental results show our scaled networks enjoy significant performance superiority on various FLOPs, but with at least 2.53x reduction on search cost.
Many problems in computer vision and machine learning can be cast as learning on hypergraphs that represent higher-order relations. Recent approaches for hypergraph learning extend graph neural networks based on message passing, which is simple yet fundamentally limited in modeling long-range dependencies and expressive power. On the other hand, tensor-based equivariant neural networks enjoy maximal expressiveness, but their application has been limited in hypergraphs due to heavy computation and strict assumptions on fixed-order hyperedges. We resolve these problems and present Equivariant Hypergraph Neural Network (EHNN), the first attempt to realize maximally expressive equivariant layers for general hypergraph learning. We also present two practical realizations of our framework based on hypernetworks (EHNN-MLP) and self-attention (EHNN-Transformer), which are easy to implement and theoretically more expressive than most message passing approaches. We demonstrate their capability in a range of hypergraph learning problems, including synthetic k-edge identification, semi-supervised classification, and visual keypoint matching, and report improved performances over strong message passing baselines. Our implementation is available at https://github.com/jw9730/ehnn.
Vision transformer (ViT) recently has drawn great attention in computer vision due to its remarkable model capability. However, most prevailing ViT models suffer from huge number of parameters, restricting their applicability on devices with limited resources. To alleviate this issue, we propose TinyViT, a new family of tiny and efficient small vision transformers pretrained on large-scale datasets with our proposed fast distillation framework. The central idea is to transfer knowledge from large pretrained models to small ones, while enabling small models to get the dividends of massive pretraining data. More specifically, we apply distillation during pretraining for knowledge transfer. The logits of large teacher models are sparsified and stored in disk in advance to save the memory cost and computation overheads. The tiny student transformers are automatically scaled down from a large pretrained model with computation and parameter constraints. Comprehensive experiments demonstrate the efficacy of TinyViT. It achieves a top-1 accuracy of 84.8% on ImageNet-1k with only 21M parameters, being comparable to Swin-B pretrained on ImageNet-21k while using 4.2 times fewer parameters. Moreover, increasing image resolutions, TinyViT can reach 86.5% accuracy, being slightly better than Swin-L while using only 11% parameters. Last but not the least, we demonstrate a good transfer ability of TinyViT on various downstream tasks. Code and models are available at https://github.com/microsoft/Cream/tree/main/TinyViT.
Vision Transformers have shown state-of-the-art results for various visual recognition tasks. The dot-product self-attention mechanism that replaces convolution to mix spatial information is commonly recognized as the indispensable ingredient behind the success of vision Transformers. In this paper, we thoroughly investigate the key differences between vision Transformers and recent all-MLP models. Our empirical results show the superiority of vision Transformers mainly comes from the data-dependent token mixing strategy and the multi-head scheme instead of query-key interactions. Inspired by this observation, we propose a computationally and parametrically efficient operation named adaptive weight mixing to generate attention weights without token-token interactions. Based on this operation, we develop a new architecture named as AMixer to capture both long-term and short-term spatial dependencies without self-attention. Extensive experiments demonstrate that our adaptive weight mixing is more efficient and effective than previous weight generation methods and our AMixer can achieve a better trade-off between accuracy and complexity than vision Transformers and MLP models on both ImageNet and downstream tasks.
Recently, transformer and multi-layer perceptron (MLP) architectures have achieved impressive results on various vision tasks. However, how to effectively combine those operators to form high-performance hybrid visual architectures still remains a challenge. In this work, we study the learnable combination of convolution, transformer, and MLP by proposing a novel unified architecture search approach. Our approach contains two key designs to achieve the search for high-performance networks. First, we model the very different searchable operators in a unified form and thus enable the operators to be characterized with the same set of configuration parameters. In this way, the overall search space size is significantly reduced, and the total search cost becomes affordable. Second, we propose context-aware downsampling modules (DSMs) to mitigate the gap between the different types of operators. Our proposed DSMs are able to better adapt features from different types of operators, which is important for identifying high-performance hybrid architectures. Finally, we integrate configurable operators and DSMs into a unified search space and search with a Reinforcement Learning-based search algorithm to fully explore the optimal combination of the operators. To this end, we search a baseline network and scale it up to obtain a family of models, named UniNets, which achieve much better accuracy and efficiency than previous ConvNets and Transformers. In particular, our UniNet-B5 achieves 84.9% top-1 accuracy on ImageNet, outperforming EfficientNet-B7 and BoTNet-T7 with 44% and 55% fewer FLOPs respectively. By pretraining on the ImageNet-21K, our UniNet-B6 achieves 87.4%, outperforming Swin-L with 51% fewer FLOPs and 41% fewer parameters.
Image-to-image translation has played an important role in enabling synthetic data for computer vision. However, if the source and target domains have a large semantic mismatch, existing techniques often suffer from source content corruption aka semantic flipping. To address this problem, we propose a new paradigm for image-to-image translation using Vector Symbolic Architectures (VSA), a theoretical framework which defines algebraic operations in a high-dimensional vector (hypervector) space. We introduce VSA-based constraints on adversarial learning for source-to-target translations by learning a hypervector mapping that inverts the translation to ensure consistency with source content. We show both qualitatively and quantitatively that our method improves over other state-of-the-art techniques.
Label noise is prevalent in real-world visual learning applications and correcting all label mistakes can be prohibitively costly. Training neural network classifiers on such noisy datasets may lead to significant performance degeneration. Active label correction (ALC) attempts to minimize the re-labeling costs by identifying examples for which providing correct labels will yield maximal performance improvements. Existing ALC approaches typically select the examples that the classifier is least confident about (\eg with the largest entropies). However, such confidence estimates can be unreliable as the classifier itself is initially trained on noisy data. Also, naively selecting a batch of low confidence examples can result in redundant labeling of spatially adjacent examples. We present a new ALC algorithm that addresses these challenges. Our algorithm robustly estimates label confidence values by regulating the contributions of individual examples in the parameter update of the network. Further, our algorithm avoids redundant labeling by promoting diversity in batch selection through propagating the confidence of each newly labeled example to the entire dataset. Experiments involving four benchmark datasets and two types of label noise demonstrate that our algorithm offers a significant improvement in re-labeling efficiency over state-of-the-art ALC approaches.