In this paper, we hypothesize that the effects of the degree of typicality in natural semantic categories can be generated based on the structure of artificial categories learned with deep learning models. Motivated by the human approach to representing natural semantic categories and based on the Prototype Theory foundations, we propose a novel Computational Prototype Model (CPM) to represent the internal structure of semantic categories. Unlike other prototype learning approaches, our mathematical framework proposes a first approach to provide deep neural networks with the ability to model abstract semantic concepts such as category central semantic meaning, typicality degree of an object's image, and family resemblance relationship. We proposed several methodologies based on the typicality's concept to evaluate our CPM-model in image semantic processing tasks such as image classification, a global semantic description, and transfer learning. Our experiments on different image datasets, such as ImageNet and Coco, showed that our approach might be an admissible proposition in the effort to endow machines with greater power of abstraction for the semantic representation of objects' categories.
Fetal ultrasound screening during pregnancy plays a vital role in the early detection of fetal malformations which have potential long-term health impacts. The level of skill required to diagnose such malformations from live ultrasound during examination is high and resources for screening are often limited. We present an interpretable, atlas-learning segmentation method for automatic diagnosis of Hypo-plastic Left Heart Syndrome (HLHS) from a single `4 Chamber Heart' view image. We propose to extend the recently introduced Image-and-Spatial Transformer Networks (Atlas-ISTN) into a framework that enables sensitising atlas generation to disease. In this framework we can jointly learn image segmentation, registration, atlas construction and disease prediction while providing a maximum level of clinical interpretability compared to direct image classification methods. As a result our segmentation allows diagnoses competitive with expert-derived manual diagnosis and yields an AUC-ROC of 0.978 (1043 cases for training, 260 for validation and 325 for testing).
Image classification models deployed in the real world may receive inputs outside the intended data distribution. For critical applications such as clinical decision making, it is important that a model can detect such out-of-distribution (OOD) inputs and express its uncertainty. In this work, we assess the capability of various state-of-the-art approaches for confidence-based OOD detection through a comparative study and in-depth analysis. First, we leverage a computer vision benchmark to reproduce and compare multiple OOD detection methods. We then evaluate their capabilities on the challenging task of disease classification using chest X-rays. Our study shows that high performance in a computer vision task does not directly translate to accuracy in a medical imaging task. We analyse factors that affect performance of the methods between the two tasks. Our results provide useful insights for developing the next generation of OOD detection methods.
Although transformers have become the neural architectures of choice for natural language processing, they require orders of magnitude more training data, GPU memory, and computations in order to compete with convolutional neural networks for computer vision. The attention mechanism of transformers scales quadratically with the length of the input sequence, and unrolled images have long sequence lengths. Plus, transformers lack an inductive bias that is appropriate for images. We tested three modifications to vision transformer (ViT) architectures that address these shortcomings. Firstly, we alleviate the quadratic bottleneck by using linear attention mechanisms, called X-formers (such that, X in {Performer, Linformer, Nystr\"omformer}), thereby creating Vision X-formers (ViXs). This resulted in up to a seven times reduction in the GPU memory requirement. We also compared their performance with FNet and multi-layer perceptron mixers, which further reduced the GPU memory requirement. Secondly, we introduced an inductive bias for images by replacing the initial linear embedding layer by convolutional layers in ViX, which significantly increased classification accuracy without increasing the model size. Thirdly, we replaced the learnable 1D position embeddings in ViT with Rotary Position Embedding (RoPE), which increases the classification accuracy for the same model size. We believe that incorporating such changes can democratize transformers by making them accessible to those with limited data and computing resources.
Though convolutional neural networks are widely used in different tasks, lack of generalization capability in the absence of sufficient and representative data is one of the challenges that hinder their practical application. In this paper, we propose a simple, effective, and plug-and-play training strategy named Knowledge Distillation for Domain Generalization (KDDG) which is built upon a knowledge distillation framework with the gradient filter as a novel regularization term. We find that both the ``richer dark knowledge" from the teacher network, as well as the gradient filter we proposed, can reduce the difficulty of learning the mapping which further improves the generalization ability of the model. We also conduct experiments extensively to show that our framework can significantly improve the generalization capability of deep neural networks in different tasks including image classification, segmentation, reinforcement learning by comparing our method with existing state-of-the-art domain generalization techniques. Last but not the least, we propose to adopt two metrics to analyze our proposed method in order to better understand how our proposed method benefits the generalization capability of deep neural networks.
Quantization based model compression serves as high performing and fast approach for inference that yields models which are highly compressed when compared to their full-precision floating point counterparts. The most extreme quantization is a 1-bit representation of parameters such that they have only two possible values, typically -1(0) or +1, enabling efficient implementation of the ubiquitous dot product using only additions. The main contribution of this work is the introduction of a method to smooth the combinatorial problem of determining a binary vector of weights to minimize the expected loss for a given objective by means of empirical risk minimization with backpropagation. This is achieved by approximating a multivariate binary state over the weights utilizing a deterministic and differentiable transformation of real-valued, continuous parameters. The proposed method adds little overhead in training, can be readily applied without any substantial modifications to the original architecture, does not introduce additional saturating nonlinearities or auxiliary losses, and does not prohibit applying other methods for binarizing the activations. Contrary to common assertions made in the literature, it is demonstrated that binary weighted networks can train well with the same standard optimization techniques and similar hyperparameter settings as their full-precision counterparts, specifically momentum SGD with large learning rates and $L_2$ regularization. To conclude experiments demonstrate the method performs remarkably well across a number of inductive image classification tasks with various architectures compared to their full-precision counterparts. The source code is publicly available at https://bitbucket.org/YanivShu/binary_weighted_networks_public.
Image classification models tend to make decisions based on peripheral attributes of data items that have strong correlation with a target variable (i.e., dataset bias). These biased models suffer from the poor generalization capability when evaluated on unbiased datasets. Existing approaches for debiasing often identify and emphasize those samples with no such correlation (i.e., bias-conflicting) without defining the bias type in advance. However, such bias-conflicting samples are significantly scarce in biased datasets, limiting the debiasing capability of these approaches. This paper first presents an empirical analysis revealing that training with "diverse" bias-conflicting samples beyond a given training set is crucial for debiasing as well as the generalization capability. Based on this observation, we propose a novel feature-level data augmentation technique in order to synthesize diverse bias-conflicting samples. To this end, our method learns the disentangled representation of (1) the intrinsic attributes (i.e., those inherently defining a certain class) and (2) bias attributes (i.e., peripheral attributes causing the bias), from a large number of bias-aligned samples, the bias attributes of which have strong correlation with the target variable. Using the disentangled representation, we synthesize bias-conflicting samples that contain the diverse intrinsic attributes of bias-aligned samples by swapping their latent features. By utilizing these diversified bias-conflicting features during the training, our approach achieves superior classification accuracy and debiasing results against the existing baselines on synthetic and real-world datasets.
Recently, self-supervised representation learning gives further development in multimedia technology. Most existing self-supervised learning methods are applicable to packaged data. However, when it comes to streamed data, they are suffering from a catastrophic forgetting problem, which is not studied extensively. In this paper, we make the first attempt to tackle the catastrophic forgetting problem in the mainstream self-supervised methods, i.e., contrastive learning methods. Specifically, we first develop a rehearsal-based framework combined with a novel sampling strategy and a self-supervised knowledge distillation to transfer information over time efficiently. Then, we propose an extra sample queue to help the network separate the feature representations of old and new data in the embedding space. Experimental results show that compared with the naive self-supervised baseline, which learns tasks one by one without taking any technique, we improve the image classification accuracy by 1.60% on CIFAR-100, 2.86% on ImageNet-Sub, and 1.29% on ImageNet-Full under 10 incremental steps setting. Our code will be available at https://github.com/VDIGPKU/ContinualContrastiveLearning.
We present a new framework to measure the intrinsic properties of (deep) neural networks. While we focus on convolutional networks, our framework can be extrapolated to any network architecture. In particular, we evaluate two network properties, namely, capacity, which is related to expressivity, and compression, which is related to learnability. Both these properties depend only on the network structure and are independent of the network parameters. To this end, we propose two metrics: the first one, called layer complexity, captures the architectural complexity of any network layer; and, the second one, called layer intrinsic power, encodes how data is compressed along the network. The metrics are based on the concept of layer algebra, which is also introduced in this paper. This concept is based on the idea that the global properties depend on the network topology, and the leaf nodes of any neural network can be approximated using local transfer functions, thereby, allowing a simple computation of the global metrics. We show that our global complexity metric can be calculated and represented more conveniently than the widely-used VC dimension. We also compare the properties of various state-of-the art architectures using our metrics and use the properties to analyze their accuracy on benchmark image classification datasets.
The classification of histopathological images is of great value in both cancer diagnosis and pathological studies. However, multiple reasons, such as variations caused by magnification factors and class imbalance, make it a challenging task where conventional methods that learn from image-label datasets perform unsatisfactorily in many cases. We observe that tumours of the same class often share common morphological patterns. To exploit this fact, we propose an approach that learns similarity-based multi-scale embeddings (SMSE) for magnification-independent histopathological image classification. In particular, a pair loss and a triplet loss are leveraged to learn similarity-based embeddings from image pairs or image triplets. The learned embeddings provide accurate measurements of similarities between images, which are regarded as a more effective form of representation for histopathological morphology than normal image features. Furthermore, in order to ensure the generated models are magnification-independent, images acquired at different magnification factors are simultaneously fed to networks during training for learning multi-scale embeddings. In addition to the SMSE, to eliminate the impact of class imbalance, instead of using the hard sample mining strategy that intuitively discards some easy samples, we introduce a new reinforced focal loss to simultaneously punish hard misclassified samples while suppressing easy well-classified samples. Experimental results show that the SMSE improves the performance for histopathological image classification tasks for both breast and liver cancers by a large margin compared to previous methods. In particular, the SMSE achieves the best performance on the BreakHis benchmark with an improvement ranging from 5% to 18% compared to previous methods using traditional features.
Weak supervision learning on classification labels has demonstrated high performance in various tasks, while a few pixel-level fine annotations are also affordable. Naturally a question comes to us that whether the combination of pixel-level (e.g., segmentation) and image level (e.g., classification) annotation can introduce further improvement. However in computational pathology this is a difficult task for this reason: High resolution of whole slide images makes it difficult to do end-to-end classification model training, which is challenging to research of weak or hybrid supervision learning in the past. To handle this problem, we propose a hybrid supervision learning framework for this kind of high resolution images with sufficient image-level coarse annotations and a few pixel-level fine labels. This framework, when applied in training patch model, can carefully make use of coarse image-level labels to refine generated pixel-level pseudo labels. Complete strategy is proposed to suppress pixel-level false positives and false negatives. A large hybrid annotated dataset is used to evaluate the effectiveness of hybrid supervision learning. By extracting pixel-level pseudo labels in initially image-level labeled samples, we achieve 5.2% higher specificity than purely training on existing labels while retaining 100% sensitivity, in the task of image-level classification to be positive or negative.
Hierarchical classification is significant for complex tasks by providing multi-granular predictions and encouraging better mistakes. As the label structure decides its performance, many existing approaches attempt to construct an excellent label structure for promoting the classification results. In this paper, we consider that different label structures provide a variety of prior knowledge for category recognition, thus fusing them is helpful to achieve better hierarchical classification results. Furthermore, we propose a multi-task multi-structure fusion model to integrate different label structures. It contains two kinds of branches: one is the traditional classification branch to classify the common subclasses, the other is responsible for identifying the heterogeneous superclasses defined by different label structures. Besides the effect of multiple label structures, we also explore the architecture of the deep model for better hierachical classification and adjust the hierarchical evaluation metrics for multiple label structures. Experimental results on CIFAR100 and Car196 show that our method obtains significantly better results than using a flat classifier or a hierarchical classifier with any single label structure.
Federated learning is promising for its capability to collaboratively train models with multiple clients without accessing their data, but vulnerable when clients' data distributions diverge from each other. This divergence further leads to a dilemma: "Should we prioritize the learned model's generic performance (for future use at the server) or its personalized performance (for each client)?" These two, seemingly competing goals have divided the community to focus on one or the other, yet in this paper we show that it is possible to approach both at the same time. Concretely, we propose a novel federated learning framework that explicitly decouples a model's dual duties with two prediction tasks. On the one hand, we introduce a family of losses that are robust to non-identical class distributions, enabling clients to train a generic predictor with a consistent objective across them. On the other hand, we formulate the personalized predictor as a lightweight adaptive module that is learned to minimize each client's empirical risk on top of the generic predictor. With this two-loss, two-predictor framework which we name Federated Robust Decoupling (Fed-RoD), the learned model can simultaneously achieve state-of-the-art generic and personalized performance, essentially bridging the two tasks.
Active learning has proven to be useful for minimizing labeling costs by selecting the most informative samples. However, existing active learning methods do not work well in realistic scenarios such as imbalance or rare classes, out-of-distribution data in the unlabeled set, and redundancy. In this work, we propose SIMILAR (Submodular Information Measures based actIve LeARning), a unified active learning framework using recently proposed submodular information measures (SIM) as acquisition functions. We argue that SIMILAR not only works in standard active learning, but also easily extends to the realistic settings considered above and acts as a one-stop solution for active learning that is scalable to large real-world datasets. Empirically, we show that SIMILAR significantly outperforms existing active learning algorithms by as much as ~5% - 18% in the case of rare classes and ~5% - 10% in the case of out-of-distribution data on several image classification tasks like CIFAR-10, MNIST, and ImageNet. SIMILAR is available as a part of the DISTIL toolkit: "https://github.com/decile-team/distil".
Deep neural network (DNN) is a popular model implemented in many systems to handle complex tasks such as image classification, object recognition, natural language processing etc. Consequently DNN structural vulnerabilities become part of the security vulnerabilities in those systems. In this paper we study the root cause of DNN adversarial examples. We examine the DNN response surface to understand its classification boundary. Our study reveals the structural problem of DNN classification boundary that leads to the adversarial examples. Existing attack algorithms can generate from a handful to a few hundred adversarial examples given one clean image. We show there are infinitely many adversarial images given one clean sample, all within a small neighborhood of the clean sample. We then define DNN uncertainty regions and show transferability of adversarial examples is not universal. We also argue that generalization error, the large sample theoretical guarantee established for DNN, cannot adequately capture the phenomenon of adversarial examples. We need new theory to measure DNN robustness.
We present CSWin Transformer, an efficient and effective Transformer-based backbone for general-purpose vision tasks. A challenging issue in Transformer design is that global self-attention is very expensive to compute whereas local self-attention often limits the field of interactions of each token. To address this issue, we develop the Cross-Shaped Window self-attention mechanism for computing self-attention in the horizontal and vertical stripes in parallel that form a cross-shaped window, with each stripe obtained by splitting the input feature into stripes of equal width. We provide a mathematical analysis of the effect of the stripe width and vary the stripe width for different layers of the Transformer network which achieves strong modeling capability while limiting the computation cost. We also introduce Locally-enhanced Positional Encoding (LePE), which handles the local positional information better than existing encoding schemes. LePE naturally supports arbitrary input resolutions, and is thus especially effective and friendly for downstream tasks. Incorporated with these designs and a hierarchical structure, CSWin Transformer demonstrates competitive performance on common vision tasks. Specifically, it achieves 85.4\% Top-1 accuracy on ImageNet-1K without any extra training data or label, 53.9 box AP and 46.4 mask AP on the COCO detection task, and 52.2 mIOU on the ADE20K semantic segmentation task, surpassing previous state-of-the-art Swin Transformer backbone by +1.2, +2.0, +1.4, and +2.0 respectively under the similar FLOPs setting. By further pretraining on the larger dataset ImageNet-21K, we achieve 87.5% Top-1 accuracy on ImageNet-1K and high segmentation performance on ADE20K with 55.7 mIoU. The code and models are available at https://github.com/microsoft/CSWin-Transformer.
Recently, pure transformer-based models have shown great potentials for vision tasks such as image classification and detection. However, the design of transformer networks is challenging. It has been observed that the depth, embedding dimension, and number of heads can largely affect the performance of vision transformers. Previous models configure these dimensions based upon manual crafting. In this work, we propose a new one-shot architecture search framework, namely AutoFormer, dedicated to vision transformer search. AutoFormer entangles the weights of different blocks in the same layers during supernet training. Benefiting from the strategy, the trained supernet allows thousands of subnets to be very well-trained. Specifically, the performance of these subnets with weights inherited from the supernet is comparable to those retrained from scratch. Besides, the searched models, which we refer to AutoFormers, surpass the recent state-of-the-arts such as ViT and DeiT. In particular, AutoFormer-tiny/small/base achieve 74.7%/81.7%/82.4% top-1 accuracy on ImageNet with 5.7M/22.9M/53.7M parameters, respectively. Lastly, we verify the transferability of AutoFormer by providing the performance on downstream benchmarks and distillation experiments. Code and models are available at https://github.com/microsoft/AutoML.
A set of novel approaches for estimating epistemic uncertainty in deep neural networks with a single forward pass has recently emerged as a valid alternative to Bayesian Neural Networks. On the premise of informative representations, these deterministic uncertainty methods (DUMs) achieve strong performance on detecting out-of-distribution (OOD) data while adding negligible computational costs at inference time. However, it remains unclear whether DUMs are well calibrated and can seamlessly scale to real-world applications - both prerequisites for their practical deployment. To this end, we first provide a taxonomy of DUMs, and evaluate their calibration under continuous distributional shifts. Then, we extend them to semantic segmentation. We find that, while DUMs scale to realistic vision tasks and perform well on OOD detection, the practicality of current methods is undermined by poor calibration under distributional shifts.
Recent advances in self-attention and pure multi-layer perceptrons (MLP) models for vision have shown great potential in achieving promising performance with fewer inductive biases. These models are generally based on learning interaction among spatial locations from raw data. The complexity of self-attention and MLP grows quadratically as the image size increases, which makes these models hard to scale up when high-resolution features are required. In this paper, we present the Global Filter Network (GFNet), a conceptually simple yet computationally efficient architecture, that learns long-term spatial dependencies in the frequency domain with log-linear complexity. Our architecture replaces the self-attention layer in vision transformers with three key operations: a 2D discrete Fourier transform, an element-wise multiplication between frequency-domain features and learnable global filters, and a 2D inverse Fourier transform. We exhibit favorable accuracy/complexity trade-offs of our models on both ImageNet and downstream tasks. Our results demonstrate that GFNet can be a very competitive alternative to transformer-style models and CNNs in efficiency, generalization ability and robustness. Code is available at https://github.com/raoyongming/GFNet
Recently, Vision Transformer and its variants have shown great promise on various computer vision tasks. The ability of capturing short- and long-range visual dependencies through self-attention is arguably the main source for the success. But it also brings challenges due to quadratic computational overhead, especially for the high-resolution vision tasks (e.g., object detection). In this paper, we present focal self-attention, a new mechanism that incorporates both fine-grained local and coarse-grained global interactions. Using this new mechanism, each token attends the closest surrounding tokens at fine granularity but the tokens far away at coarse granularity, and thus can capture both short- and long-range visual dependencies efficiently and effectively. With focal self-attention, we propose a new variant of Vision Transformer models, called Focal Transformer, which achieves superior performance over the state-of-the-art vision Transformers on a range of public image classification and object detection benchmarks. In particular, our Focal Transformer models with a moderate size of 51.1M and a larger size of 89.8M achieve 83.5 and 83.8 Top-1 accuracy, respectively, on ImageNet classification at 224x224 resolution. Using Focal Transformers as the backbones, we obtain consistent and substantial improvements over the current state-of-the-art Swin Transformers for 6 different object detection methods trained with standard 1x and 3x schedules. Our largest Focal Transformer yields 58.7/58.9 box mAPs and 50.9/51.3 mask mAPs on COCO mini-val/test-dev, and 55.4 mIoU on ADE20K for semantic segmentation, creating new SoTA on three of the most challenging computer vision tasks.