While originally designed for natural language processing tasks, the self-attention mechanism has recently taken various computer vision areas by storm. However, the 2D nature of images brings three challenges for applying self-attention in computer vision. (1) Treating images as 1D sequences neglects their 2D structures. (2) The quadratic complexity is too expensive for high-resolution images. (3) It only captures spatial adaptability but ignores channel adaptability. In this paper, we propose a novel linear attention named large kernel attention (LKA) to enable self-adaptive and long-range correlations in self-attention while avoiding its shortcomings. Furthermore, we present a neural network based on LKA, namely Visual Attention Network (VAN). While extremely simple, VAN surpasses similar size vision transformers(ViTs) and convolutional neural networks(CNNs) in various tasks, including image classification, object detection, semantic segmentation, panoptic segmentation, pose estimation, etc. For example, VAN-B6 achieves 87.8% accuracy on ImageNet benchmark and set new state-of-the-art performance (58.2 PQ) for panoptic segmentation. Besides, VAN-B2 surpasses Swin-T 4% mIoU (50.1 vs. 46.1) for semantic segmentation on ADE20K benchmark, 2.6% AP (48.8 vs. 46.2) for object detection on COCO dataset. It provides a novel method and a simple yet strong baseline for the community. Code is available at https://github.com/Visual-Attention-Network.
Vision transformers have shown great potential in various computer vision tasks owing to their strong capability to model long-range dependency using the self-attention mechanism. Nevertheless, they treat an image as a 1D sequence of visual tokens, lacking an intrinsic inductive bias (IB) in modeling local visual structures and dealing with scale variance, which is instead learned implicitly from large-scale training data with longer training schedules. In this paper, we propose a Vision Transformer Advanced by Exploring intrinsic IB from convolutions, i.e., ViTAE. Technically, ViTAE has several spatial pyramid reduction modules to downsample and embed the input image into tokens with rich multi-scale context using multiple convolutions with different dilation rates. In this way, it acquires an intrinsic scale invariance IB and can learn robust feature representation for objects at various scales. Moreover, in each transformer layer, ViTAE has a convolution block parallel to the multi-head self-attention module, whose features are fused and fed into the feed-forward network. Consequently, it has the intrinsic locality IB and is able to learn local features and global dependencies collaboratively. The proposed two kinds of cells are stacked in both isotropic and multi-stage manners to formulate two families of ViTAE models, i.e., the vanilla ViTAE and ViTAEv2. Experiments on the ImageNet dataset as well as downstream tasks on the MS COCO, ADE20K, and AP10K datasets validate the superiority of our models over the baseline transformer models and concurrent works. Besides, we scale up our ViTAE model to 644M parameters and obtain the state-of-the-art classification performance, i.e., 88.5% Top-1 classification accuracy on ImageNet validation set and the best 91.2% Top-1 accuracy on ImageNet real validation set, without using extra private data.
In this paper, we introduce data multiplexing (DataMUX), a technique that enables deep neural networks to process multiple inputs simultaneously using a single compact representation. DataMUX demonstrates that neural networks are capable of generating accurate predictions over mixtures of inputs, resulting in increased throughput with minimal extra memory requirements. Our approach uses two key components -- 1) a multiplexing layer that performs a fixed linear transformation to each input before combining them to create a mixed representation of the same size as a single input, which is then processed by the base network, and 2) a demultiplexing layer that converts the base network's output back into independent representations before producing predictions for each input. We show the viability of DataMUX for different architectures (Transformers, and to a lesser extent MLPs and CNNs) across six different tasks spanning sentence classification, named entity recognition and image classification. For instance, DataMUX for Transformers can multiplex up to $20$x/$40$x inputs, achieving $11$x/$18$x increase in throughput with minimal absolute performance drops of $<2\%$ and $<4\%$ respectively on MNLI, a natural language inference task. We also provide a theoretical construction for multiplexing in self-attention networks and analyze the effect of various design elements in DataMUX.
Discriminative self-supervised learning allows training models on any random group of internet images, and possibly recover salient information that helps differentiate between the images. Applied to ImageNet, this leads to object centric features that perform on par with supervised features on most object-centric downstream tasks. In this work, we question if using this ability, we can learn any salient and more representative information present in diverse unbounded set of images from across the globe. To do so, we train models on billions of random images without any data pre-processing or prior assumptions about what we want the model to learn. We scale our model size to dense 10 billion parameters to avoid underfitting on a large data size. We extensively study and validate our model performance on over 50 benchmarks including fairness, robustness to distribution shift, geographical diversity, fine grained recognition, image copy detection and many image classification datasets. The resulting model, not only captures well semantic information, it also captures information about artistic style and learns salient information such as geolocations and multilingual word embeddings based on visual content only. More importantly, we discover that such model is more robust, more fair, less harmful and less biased than supervised models or models trained on object centric datasets such as ImageNet.
Human visual attention has recently shown its distinct capability in boosting machine learning models. However, studies that aim to facilitate medical tasks with human visual attention are still scarce. To support the use of visual attention, this paper describes a novel deep learning model for visual saliency prediction on chest X-ray (CXR) images. To cope with data deficiency, we exploit the multi-task learning method and tackles disease classification on CXR simultaneously. For a more robust training process, we propose a further optimized multi-task learning scheme to better handle model overfitting. Experiments show our proposed deep learning model with our new learning scheme can outperform existing methods dedicated either for saliency prediction or image classification. The code used in this paper is available at https://github.com/hz-zhu/MT-UNet.
Progress in digital pathology is hindered by high-resolution images and the prohibitive cost of exhaustive localized annotations. The commonly used paradigm to categorize pathology images is patch-based processing, which often incorporates multiple instance learning (MIL) to aggregate local patch-level representations yielding image-level prediction. Nonetheless, diagnostically relevant regions may only take a small fraction of the whole tissue, and current MIL-based approaches often process images uniformly, discarding the inter-patches interactions. To alleviate these issues, we propose ScoreNet, a new efficient transformer that exploits a differentiable recommendation stage to extract discriminative image regions and dedicate computational resources accordingly. The proposed transformer leverages the local and global attention of a few dynamically recommended high-resolution regions at an efficient computational cost. We further introduce a novel mixing data-augmentation, namely ScoreMix, by leveraging the image's semantic distribution to guide the data mixing and produce coherent sample-label pairs. ScoreMix is embarrassingly simple and mitigates the pitfalls of previous augmentations, which assume a uniform semantic distribution and risk mislabeling the samples. Thorough experiments and ablation studies on three breast cancer histology datasets of Haematoxylin & Eosin (H&E) have validated the superiority of our approach over prior arts, including transformer-based models on tumour regions-of-interest (TRoIs) classification. ScoreNet equipped with proposed ScoreMix augmentation demonstrates better generalization capabilities and achieves new state-of-the-art (SOTA) results with only 50% of the data compared to other mixing augmentation variants. Finally, ScoreNet yields high efficacy and outperforms SOTA efficient transformers, namely TransPath and SwinTransformer.
The integration of artificial intelligence into clinical workflows requires reliable and robust models. Repeatability is a key attribute of model robustness. Repeatable models output predictions with low variation during independent tests carried out under similar conditions. During model development and evaluation, much attention is given to classification performance while model repeatability is rarely assessed, leading to the development of models that are unusable in clinical practice. In this work, we evaluate the repeatability of four model types (binary classification, multi-class classification, ordinal classification, and regression) on images that were acquired from the same patient during the same visit. We study the performance of binary, multi-class, ordinal, and regression models on four medical image classification tasks from public and private datasets: knee osteoarthritis, cervical cancer screening, breast density estimation, and retinopathy of prematurity. Repeatability is measured and compared on ResNet and DenseNet architectures. Moreover, we assess the impact of sampling Monte Carlo dropout predictions at test time on classification performance and repeatability. Leveraging Monte Carlo predictions significantly increased repeatability for all tasks on the binary, multi-class, and ordinal models leading to an average reduction of the 95\% limits of agreement by 16% points and of the disagreement rate by 7% points. The classification accuracy improved in most settings along with the repeatability. Our results suggest that beyond about 20 Monte Carlo iterations, there is no further gain in repeatability. In addition to the higher test-retest agreement, Monte Carlo predictions were better calibrated which leads to output probabilities reflecting more accurately the true likelihood of being correctly classified.
CNN architectures are generally heavy on memory and computational requirements which makes them infeasible for embedded systems with limited hardware resources. We propose dual convolutional kernels (DualConv) for constructing lightweight deep neural networks. DualConv combines 3$\times$3 and 1$\times$1 convolutional kernels to process the same input feature map channels simultaneously and exploits the group convolution technique to efficiently arrange convolutional filters. DualConv can be employed in any CNN model such as VGG-16 and ResNet-50 for image classification, YOLO and R-CNN for object detection, or FCN for semantic segmentation. In this paper, we extensively test DualConv for classification since these network architectures form the backbones for many other tasks. We also test DualConv for image detection on YOLO-V3. Experimental results show that, combined with our structural innovations, DualConv significantly reduces the computational cost and number of parameters of deep neural networks while surprisingly achieving slightly higher accuracy than the original models in some cases. We use DualConv to further reduce the number of parameters of the lightweight MobileNetV2 by 54% with only 0.68% drop in accuracy on CIFAR-100 dataset. When the number of parameters is not an issue, DualConv increases the accuracy of MobileNetV1 by 4.11% on the same dataset. Furthermore, DualConv significantly improves the YOLO-V3 object detection speed and improves its accuracy by 4.4% on PASCAL VOC dataset.
Despite impressive capabilities and outstanding performance, deep neural networks (DNNs) have captured increasing public concern about their security problems, due to their frequently occurred erroneous behaviors. Therefore, it is necessary to conduct a systematical testing for DNNs before they are deployed to real-world applications. Existing testing methods have provided fine-grained metrics based on neuron coverage and proposed various approaches to improve such metrics. However, it has been gradually realized that a higher neuron coverage does \textit{not} necessarily represent better capabilities in identifying defects that lead to errors. Besides, coverage-guided methods cannot hunt errors due to faulty training procedure. So the robustness improvement of DNNs via retraining by these testing examples are unsatisfactory. To address this challenge, we introduce the concept of excitable neurons based on Shapley value and design a novel white-box testing framework for DNNs, namely DeepSensor. It is motivated by our observation that neurons with larger responsibility towards model loss changes due to small perturbations are more likely related to incorrect corner cases due to potential defects. By maximizing the number of excitable neurons concerning various wrong behaviors of models, DeepSensor can generate testing examples that effectively trigger more errors due to adversarial inputs, polluted data and incomplete training. Extensive experiments implemented on both image classification models and speaker recognition models have demonstrated the superiority of DeepSensor.
This paper focuses on filter-level network pruning. A novel pruning method, termed CLR-RNF, is proposed. We first reveal a "long-tail" long-tail pruning problem in magnitude-based weight pruning methods, and then propose a computation-aware measurement for individual weight importance, followed by a Cross-Layer Ranking (CLR) of weights to identify and remove the bottom-ranked weights. Consequently, the per-layer sparsity makes up of the pruned network structure in our filter pruning. Then, we introduce a recommendation-based filter selection scheme where each filter recommends a group of its closest filters. To pick the preserved filters from these recommended groups, we further devise a k-Reciprocal Nearest Filter (RNF) selection scheme where the selected filters fall into the intersection of these recommended groups. Both our pruned network structure and the filter selection are non-learning processes, which thus significantly reduce the pruning complexity, and differentiate our method from existing works. We conduct image classification on CIFAR-10 and ImageNet to demonstrate the superiority of our CLR-RNF over the state-of-the-arts. For example, on CIFAR-10, CLR-RNF removes 74.1% FLOPs and 95.0% parameters from VGGNet-16 with even 0.3\% accuracy improvements. On ImageNet, it removes 70.2% FLOPs and 64.8% parameters from ResNet-50 with only 1.7% top-5 accuracy drops. Our project is at https://github.com/lmbxmu/CLR-RNF.
Open-set recognition and adversarial defense study two key aspects of deep learning that are vital for real-world deployment. The objective of open-set recognition is to identify samples from open-set classes during testing, while adversarial defense aims to robustify the network against images perturbed by imperceptible adversarial noise. This paper demonstrates that open-set recognition systems are vulnerable to adversarial samples. Furthermore, this paper shows that adversarial defense mechanisms trained on known classes are unable to generalize well to open-set samples. Motivated by these observations, we emphasize the necessity of an Open-Set Adversarial Defense (OSAD) mechanism. This paper proposes an Open-Set Defense Network with Clean-Adversarial Mutual Learning (OSDN-CAML) as a solution to the OSAD problem. The proposed network designs an encoder with dual-attentive feature-denoising layers coupled with a classifier to learn a noise-free latent feature representation, which adaptively removes adversarial noise guided by channel and spatial-wise attentive filters. Several techniques are exploited to learn a noise-free and informative latent feature space with the aim of improving the performance of adversarial defense and open-set recognition. First, we incorporate a decoder to ensure that clean images can be well reconstructed from the obtained latent features. Then, self-supervision is used to ensure that the latent features are informative enough to carry out an auxiliary task. Finally, to exploit more complementary knowledge from clean image classification to facilitate feature denoising and search for a more generalized local minimum for open-set recognition, we further propose clean-adversarial mutual learning, where a peer network (classifying clean images) is further introduced to mutually learn with the classifier (classifying adversarial images).
Hyperspectral imagery is rich in spatial and spectral information. Using 3D-CNN can simultaneously acquire features of spatial and spectral dimensions to facilitate classification of features, but hyperspectral image information spectral dimensional information redundancy. The use of continuous 3D-CNN will result in a high amount of parameters, and the computational power requirements of the device are high, and the training takes too long. This letter designed the Faster selective kernel mechanism network (FSKNet), FSKNet can balance this problem. It designs 3D-CNN and 2D-CNN conversion modules, using 3D-CNN to complete feature extraction while reducing the dimensionality of spatial and spectrum. However, such a model is not lightweight enough. In the converted 2D-CNN, a selective kernel mechanism is proposed, which allows each neuron to adjust the receptive field size based on the two-way input information scale. Under the Selective kernel mechanism, it mainly includes two components, se module and variable convolution. Se acquires channel dimensional attention and variable convolution to obtain spatial dimension deformation information of ground objects. The model is more accurate, faster, and less computationally intensive. FSKNet achieves high accuracy on the IN, UP, Salinas, and Botswana data sets with very small parameters.
Recent works have demonstrated that transformer can achieve promising performance in computer vision, by exploiting the relationship among image patches with self-attention. While they only consider the attention in a single feature layer, but ignore the complementarity of attention in different levels. In this paper, we propose the broad attention to improve the performance by incorporating the attention relationship of different layers for vision transformer, which is called BViT. The broad attention is implemented by broad connection and parameter-free attention. Broad connection of each transformer layer promotes the transmission and integration of information for BViT. Without introducing additional trainable parameters, parameter-free attention jointly focuses on the already available attention information in different layers for extracting useful information and building their relationship. Experiments on image classification tasks demonstrate that BViT delivers state-of-the-art accuracy of 74.8\%/81.6\% top-1 accuracy on ImageNet with 5M/22M parameters. Moreover, we transfer BViT to downstream object recognition benchmarks to achieve 98.9\% and 89.9\% on CIFAR10 and CIFAR100 respectively that exceed ViT with fewer parameters. For the generalization test, the broad attention in Swin Transformer and T2T-ViT also bring an improvement of more than 1\%. To sum up, broad attention is promising to promote the performance of attention based models. Code and pre-trained models are available at https://github.com/DRL-CASIA/Broad_ViT.
Open-set domain adaptation (OSDA) has gained considerable attention in many visual recognition tasks. However, most existing OSDA approaches are limited due to three main reasons, including: (1) the lack of essential theoretical analysis of generalization bound, (2) the reliance on the coexistence of source and target data during adaptation, and (3) failing to accurately estimate the uncertainty of model predictions. We propose a Progressive Graph Learning (PGL) framework that decomposes the target hypothesis space into the shared and unknown subspaces, and then progressively pseudo-labels the most confident known samples from the target domain for hypothesis adaptation. Moreover, we tackle a more realistic source-free open-set domain adaptation (SF-OSDA) setting that makes no assumption about the coexistence of source and target domains, and introduce a balanced pseudo-labeling (BP-L) strategy in a two-stage framework, namely SF-PGL. Different from PGL that applies a class-agnostic constant threshold for all target samples for pseudo-labeling, the SF-PGL model uniformly selects the most confident target instances from each category at a fixed ratio. The confidence thresholds in each class are regarded as the 'uncertainty' of learning the semantic information, which are then used to weigh the classification loss in the adaptation step. We conducted unsupervised and semi-supervised OSDA and SF-OSDA experiments on the benchmark image classification and action recognition datasets. Additionally, we find that balanced pseudo-labeling plays a significant role in improving calibration, which makes the trained model less prone to over-confident or under-confident predictions on the target data. Source code is available at https://github.com/Luoyadan/SF-PGL.
When a clinician refers a patient for an imaging exam, they include the reason (e.g. relevant patient history, suspected disease) in the scan request; this appears as the indication field in the radiology report. The interpretation and reporting of the image are substantially influenced by this request text, steering the radiologist to focus on particular aspects of the image. We use the indication field to drive better image classification, by taking a transformer network which is unimodally pre-trained on text (BERT) and fine-tuning it for multimodal classification of a dual image-text input. We evaluate the method on the MIMIC-CXR dataset, and present ablation studies to investigate the effect of the indication field on the classification performance. The experimental results show our approach achieves 87.8 average micro AUROC, outperforming the state-of-the-art methods for unimodal (84.4) and multimodal (86.0) classification. Our code is available at https://github.com/jacenkow/mmbt.
Linear layers in neural networks (NNs) trained by gradient descent can be expressed as a key-value memory system which stores all training datapoints and the initial weights, and produces outputs using unnormalised dot attention over the entire training experience. While this has been technically known since the 1960s, no prior work has effectively studied the operations of NNs in such a form, presumably due to prohibitive time and space complexities and impractical model sizes, all of them growing linearly with the number of training patterns which may get very large. However, this dual formulation offers a possibility of directly visualising how an NN makes use of training patterns at test time, by examining the corresponding attention weights. We conduct experiments on small scale supervised image classification tasks in single-task, multi-task, and continual learning settings, as well as language modelling, and discuss potentials and limits of this view for better understanding and interpreting how NNs exploit training patterns. Our code is public.
Modern deep neural networks can easily overfit to biased training data containing corrupted labels or class imbalance. Sample re-weighting methods are popularly used to alleviate this data bias issue. Most current methods, however, require to manually pre-specify the weighting schemes as well as their additional hyper-parameters relying on the characteristics of the investigated problem and training data. This makes them fairly hard to be generally applied in practical scenarios, due to their significant complexities and inter-class variations of data bias situations. To address this issue, we propose a meta-model capable of adaptively learning an explicit weighting scheme directly from data. Specifically, by seeing each training class as a separate learning task, our method aims to extract an explicit weighting function with sample loss and task/class feature as input, and sample weight as output, expecting to impose adaptively varying weighting schemes to different sample classes based on their own intrinsic bias characteristics. Synthetic and real data experiments substantiate the capability of our method on achieving proper weighting schemes in various data bias cases, like the class imbalance, feature-independent and dependent label noise scenarios, and more complicated bias scenarios beyond conventional cases. Besides, the task-transferability of the learned weighting scheme is also substantiated, by readily deploying the weighting function learned on relatively smaller-scale CIFAR-10 dataset on much larger-scale full WebVision dataset. A performance gain can be readily achieved compared with previous SOAT ones without additional hyper-parameter tuning and meta gradient descent step. The general availability of our method for multiple robust deep learning issues, including partial-label learning, semi-supervised learning and selective classification, has also been validated.
One critical component in lossy deep image compression is the entropy model, which predicts the probability distribution of the quantized latent representation in the encoding and decoding modules. Previous works build entropy models upon convolutional neural networks which are inefficient in capturing global dependencies. In this work, we propose a novel transformer-based entropy model, termed Entroformer, to capture long-range dependencies in probability distribution estimation effectively and efficiently. Different from vision transformers in image classification, the Entroformer is highly optimized for image compression, including a top-k self-attention and a diamond relative position encoding. Meanwhile, we further expand this architecture with a parallel bidirectional context model to speed up the decoding process. The experiments show that the Entroformer achieves state-of-the-art performance on image compression while being time-efficient.
A well-known challenge associated with the multi-label classification problem is modelling dependencies between labels. Most attempts at modelling label dependencies focus on co-occurrences, ignoring the valuable information that can be extracted by detecting label subsets that rarely occur together. For example, consider customer product reviews; a product probably would not simultaneously be tagged by both "recommended" (i.e., reviewer is happy and recommends the product) and "urgent" (i.e., the review suggests immediate action to remedy an unsatisfactory experience). Aside from the consideration of positive and negative dependencies, the direction of a relationship should also be considered. For a multi-label image classification problem, the "ship" and "sea" labels have an obvious dependency, but the presence of the former implies the latter much more strongly than the other way around. These examples motivate the modelling of multiple types of bi-directional relationships between labels. In this paper, we propose a novel method, entitled Multi-relation Message Passing (MrMP), for the multi-label classification problem. Experiments on benchmark multi-label text classification datasets show that the MrMP module yields similar or superior performance compared to state-of-the-art methods. The approach imposes only minor additional computational and memory overheads.
In scoring systems used to measure the endoscopic activity of ulcerative colitis, such as Mayo endoscopic score or Ulcerative Colitis Endoscopic Index Severity, levels increase with severity of the disease activity. Such relative ranking among the scores makes it an ordinal regression problem. On the other hand, most studies use categorical cross-entropy loss function to train deep learning models, which is not optimal for the ordinal regression problem. In this study, we propose a novel loss function, class distance weighted cross-entropy (CDW-CE), that respects the order of the classes and takes the distance of the classes into account in calculation of the cost. Experimental evaluations show that models trained with CDW-CE outperform the models trained with conventional categorical cross-entropy and other commonly used loss functions which are designed for the ordinal regression problems. In addition, the class activation maps of models trained with CDW-CE loss are more class-discriminative and they are found to be more reasonable by the domain experts.