Self-attention architectures have emerged as a recent advancement for improving the performance of vision tasks. Manual determination of the architecture for self-attention networks relies on the experience of experts and cannot automatically adapt to various scenarios. Meanwhile, neural architecture search (NAS) has significantly advanced the automatic design of neural architectures. Thus, it is appropriate to consider using NAS methods to discover a better self-attention architecture automatically. However, it is challenging to directly use existing NAS methods to search attention networks because of the uniform cell-based search space and the lack of long-term content dependencies. To address this issue, we propose a full-attention based NAS method. More specifically, a stage-wise search space is constructed that allows various attention operations to be adopted for different layers of a network. To extract global features, a self-supervised search algorithm is proposed that uses context auto-regression to discover the full-attention architecture. To verify the efficacy of the proposed methods, we conducted extensive experiments on various learning tasks, including image classification, fine-grained image recognition, and zero-shot image retrieval. The empirical results show strong evidence that our method is capable of discovering high-performance, full-attention architectures while guaranteeing the required search efficiency.
Teaching plays a fundamental role in human learning. Typically, a human teaching strategy would involve assessing a student's knowledge progress for tailoring the teaching materials in a way that enhances the learning progress. A human teacher would achieve this by tracing a student's knowledge over important learning concepts in a task. Albeit, such teaching strategy is not well exploited yet in machine learning as current machine teaching methods tend to directly assess the progress on individual training samples without paying attention to the underlying learning concepts in a learning task. In this paper, we propose a novel method, called Knowledge Augmented Data Teaching (KADT), which can optimize a data teaching strategy for a student model by tracing its knowledge progress over multiple learning concepts in a learning task. Specifically, the KADT method incorporates a knowledge tracing model to dynamically capture the knowledge progress of a student model in terms of latent learning concepts. Then we develop an attention pooling mechanism to distill knowledge representations of a student model with respect to class labels, which enables to develop a data teaching strategy on critical training samples. We have evaluated the performance of the KADT method on four different machine learning tasks including knowledge tracing, sentiment analysis, movie recommendation, and image classification. The results comparing to the state-of-the-art methods empirically validate that KADT consistently outperforms others on all tasks.
The state-of-the-art deep neural networks (DNNs) have been widely applied for various real-world applications, and achieved significant performance for cognitive problems. However, the increment of DNNs' width and depth in architecture results in a huge amount of parameters to challenge the storage and memory cost, limiting to the usage of DNNs on resource-constrained platforms, such as portable devices. By converting redundant models into compact ones, compression technique appears to be a practical solution to reducing the storage and memory consumption. In this paper, we develop a nonlinear tensor ring network (NTRN) in which both fullyconnected and convolutional layers are compressed via tensor ring decomposition. Furthermore, to mitigate the accuracy loss caused by compression, a nonlinear activation function is embedded into the tensor contraction and convolution operations inside the compressed layer. Experimental results demonstrate the effectiveness and superiority of the proposed NTRN for image classification using two basic neural networks, LeNet-5 and VGG-11 on three datasets, viz. MNIST, Fashion MNIST and Cifar-10.
Histopathological analysis is the present gold standard for precancerous lesion diagnosis. The goal of automated histopathological classification from digital images requires supervised training, which requires a large number of expert annotations that can be expensive and time-consuming to collect. Meanwhile, accurate classification of image patches cropped from whole-slide images is essential for standard sliding window based histopathology slide classification methods. To mitigate these issues, we propose a carefully designed conditional GAN model, namely HistoGAN, for synthesizing realistic histopathology image patches conditioned on class labels. We also investigate a novel synthetic augmentation framework that selectively adds new synthetic image patches generated by our proposed HistoGAN, rather than expanding directly the training set with synthetic images. By selecting synthetic images based on the confidence of their assigned labels and their feature similarity to real labeled images, our framework provides quality assurance to synthetic augmentation. Our models are evaluated on two datasets: a cervical histopathology image dataset with limited annotations, and another dataset of lymph node histopathology images with metastatic cancer. Here, we show that leveraging HistoGAN generated images with selective augmentation results in significant and consistent improvements of classification performance (6.7% and 2.8% higher accuracy, respectively) for cervical histopathology and metastatic cancer datasets.
Contrastive learning can largely enhance the feature discriminability in a self-supervised manner and has achieved remarkable success for various visual tasks. However, it is undesirably observed that the standard contrastive paradigm (features+$\ell_{2}$ normalization) only brings little help for domain adaptation. In this work, we delve into this phenomenon and find that the main reason is due to the class weights (weights of the final fully connected layer) which are vital for the recognition yet ignored in the optimization. To tackle this issue, we propose a simple yet powerful Probabilistic Contrastive Learning (PCL), which does not only assist in extracting discriminative features but also enforces them to be clustered around the class weights. Specifically, we break the standard contrastive paradigm by removing $\ell_{2}$ normalization and replacing the features with probabilities. In this way, PCL can enforce the probability to approximate the one-hot form, thereby reducing the deviation between the features and class weights. Benefiting from the conciseness, PCL can be well generalized to different settings. In this work, we conduct extensive experiments on five tasks and observe consistent performance gains, i.e., Unsupervised Domain Adaptation (UDA), Semi-Supervised Domain Adaptation (SSDA), Semi-Supervised Learning (SSL), UDA Detection, and UDA Semantic Segmentation. Notably, for UDA Semantic Segmentation on SYNTHIA, PCL surpasses the sophisticated CPSL-D by $>\!2\%$ in terms of mean IoU with a much smaller training cost (PCL: 1*3090, 5 days v.s. CPSL-D: 4*V100, 11 days). Code is available at https://github.com/ljjcoder/Probabilistic-Contrastive-Learning.
Spiking neural networks (SNNs) have become an interesting alternative to conventional artificial neural networks (ANN) thanks to their temporal processing capabilities and energy efficient implementations in neuromorphic hardware. However the challenges involved in training SNNs have limited their performance in terms of accuracy and thus their applications. Improving learning algorithms and neural architectures for a more accurate feature extraction is therefore one of the current priorities in SNN research. In this paper we present a study on the key components of modern spiking architectures. We design a spiking version of the successful residual network architecture and provide an in-depth study on the possible implementations of spiking residual connections. This study shows how, depending on the use case, the optimal residual connection implementation may vary. Additionally, we empirically compare different techniques in image classification datasets taken from the best performing networks. Our results provide a state of the art guide to SNN design, which allows to make informed choices when trying to build the optimal visual feature extractor. Finally, our network outperforms previous SNN architectures in CIFAR-10 (94.14%) and CIFAR-100 (74.65%) datasets and matches the state of the art in DVS-CIFAR10 (72.98%), with less parameters than the previous state of the art and without the need for ANN-SNN conversion. Code available at https://github.com/VicenteAlex/Spiking_ResNet
This paper shows that masked autoencoders (MAE) are scalable self-supervised learners for computer vision. Our MAE approach is simple: we mask random patches of the input image and reconstruct the missing pixels. It is based on two core designs. First, we develop an asymmetric encoder-decoder architecture, with an encoder that operates only on the visible subset of patches (without mask tokens), along with a lightweight decoder that reconstructs the original image from the latent representation and mask tokens. Second, we find that masking a high proportion of the input image, e.g., 75%, yields a nontrivial and meaningful self-supervisory task. Coupling these two designs enables us to train large models efficiently and effectively: we accelerate training (by 3x or more) and improve accuracy. Our scalable approach allows for learning high-capacity models that generalize well: e.g., a vanilla ViT-Huge model achieves the best accuracy (87.8%) among methods that use only ImageNet-1K data. Transfer performance in downstream tasks outperforms supervised pre-training and shows promising scaling behavior.
Supervised learning can learn large representational spaces, which are crucial for handling difficult learning tasks. However, due to the design of the model, classical image classification approaches struggle to generalize to new problems and new situations when dealing with small datasets. In fact, supervised learning can lose the location of image features which leads to supervision collapse in very deep architectures. In this paper, we investigate how self-supervision with strong and sufficient augmentation of unlabeled data can train effectively the first layers of a neural network even better than supervised learning, with no need for millions of labeled data. The main goal is to disconnect pixel data from annotation by getting generic task-agnostic low-level features. Furthermore, we look into Vision Transformers (ViT) and show that the low-level features derived from a self-supervised architecture can improve the robustness and the overall performance of this emergent architecture. We evaluated our method on one of the smallest open-source datasets STL-10 and we obtained a significant boost of performance from 41.66% to 83.25% when inputting low-level features from a self-supervised learning architecture to the ViT instead of the raw images.
Machine learning practitioners often have access to a spectrum of data: labeled data for the target task (which is often limited), unlabeled data, and auxiliary data, the many available labeled datasets for other tasks. We describe TAGLETS, a system built to study techniques for automatically exploiting all three types of data and creating high-quality, servable classifiers. The key components of TAGLETS are: (1) auxiliary data organized according to a knowledge graph, (2) modules encapsulating different methods for exploiting auxiliary and unlabeled data, and (3) a distillation stage in which the ensembled modules are combined into a servable model. We compare TAGLETS with state-of-the-art transfer learning and semi-supervised learning methods on four image classification tasks. Our study covers a range of settings, varying the amount of labeled data and the semantic relatedness of the auxiliary data to the target task. We find that the intelligent incorporation of auxiliary and unlabeled data into multiple learning techniques enables TAGLETS to match-and most often significantly surpass-these alternatives. TAGLETS is available as an open-source system at github.com/BatsResearch/taglets.
We present a neat yet effective recursive operation on vision transformers that can improve parameter utilization without involving additional parameters. This is achieved by sharing weights across the depth of transformer networks. The proposed method can obtain a substantial gain (~2%) simply using naive recursive operation, requires no special or sophisticated knowledge for designing principles of networks, and introduces minimal computational overhead to the training procedure. To reduce the additional computation caused by recursive operation while maintaining the superior accuracy, we propose an approximating method through multiple sliced group self-attentions across recursive layers which can reduce the cost consumption by 10~30% with minimal performance loss. We call our model Sliced Recursive Transformer (SReT), a novel and parameter-efficient vision transformer design that is compatible with a broad range of other designs for efficient ViT architectures. Our best model establishes significant improvement on ImageNet-1K over state-of-the-art methods while containing fewer parameters. The proposed weight sharing mechanism by sliced recursion structure allows us to build a transformer with more than 100 or even 1000 shared layers with ease while keeping a compact size (13~15M), to avoid optimization difficulties when the model is too large. The flexible scalability has shown great potential for scaling up models and constructing extremely deep vision transformers. Code is available at https://github.com/szq0214/SReT.
A widely used algorithm for transfer learning is fine-tuning, where a pre-trained model is fine-tuned on a target task with a small amount of labeled data. When the capacity of the pre-trained model is much larger than the size of the target data set, fine-tuning is prone to overfitting and "memorizing" the training labels. Hence, an important question is to regularize fine-tuning and ensure its robustness to noise. To address this question, we begin by analyzing the generalization properties of fine-tuning. We present a PAC-Bayes generalization bound that depends on the distance traveled in each layer during fine-tuning and the noise stability of the fine-tuned model. We empirically measure these quantities. Based on the analysis, we propose regularized self-labeling -- the interpolation between regularization and self-labeling methods, including (i) layer-wise regularization to constrain the distance traveled in each layer; (ii) self label-correction and label-reweighting to correct mislabeled data points (that the model is confident) and reweight less confident data points. We validate our approach on an extensive collection of image and text data sets using multiple pre-trained model architectures. Our approach improves baseline methods by 1.76% (on average) for seven image classification tasks and 0.75% for a few-shot classification task. When the target data set includes noisy labels, our approach outperforms baseline methods by 3.56% on average in two noisy settings.
This paper presents a novel multi-fake evolutionary generative adversarial network(MFEGAN) for handling imbalance hyperspectral image classification. It is an end-to-end approach in which different generative objective losses are considered in the generator network to improve the classification performance of the discriminator network. Thus, the same discriminator network has been used as a standard classifier by embedding the classifier network on top of the discriminating function. The effectiveness of the proposed method has been validated through two hyperspectral spatial-spectral data sets. The same generative and discriminator architectures have been utilized with two different GAN objectives for a fair performance comparison with the proposed method. It is observed from the experimental validations that the proposed method outperforms the state-of-the-art methods with better classification performance.
Annotating data, especially in the medical domain, requires expert knowledge and a lot of effort. This limits the amount and/or usefulness of available medical data sets for experimentation. Therefore, developing strategies to increase the number of annotations while lowering the needed domain knowledge is of interest. A possible strategy is the use of gamification, i.e. transforming the annotation task into a game. We propose an approach to gamify the task of annotating lung fluid cells from pathological whole slide images (WSIs). As the domain is unknown to non-expert annotators, we transform images of cells to the domain of flower images using a CycleGAN architecture. In this more assessable domain, non-expert annotators can be (t)asked to annotate different kinds of flowers in a playful setting. In order to provide a proof of concept, this work shows that the domain transfer is possible by evaluating an image classification network trained on real cell images and tested on the cell images generated by the CycleGAN network (reconstructed cell images) as well as real cell images. The classification network reaches an average accuracy of 94.73 % on the original lung fluid cells and 95.25 % on the transformed lung fluid cells, respectively. Our study lays the foundation for future research on gamification using CycleGANs.
We show that a deep learning model with built-in relational inductive bias can bring benefits to sample-efficient learning, without relying on extensive data augmentation. The proposed one-shot classification model performs relational matching of a pair of inputs in the form of local and pairwise attention. Our approach solves perfectly the one-shot image classification Omniglot challenge. Our model exceeds human level accuracy, as well as the previous state of the art, with no data augmentation.
Due to the unreliability of Internet workers, it's difficult to complete a crowdsourcing project satisfactorily, especially when the tasks are multiple and the budget is limited. Recently, meta learning has brought new vitality to few-shot learning, making it possible to obtain a classifier with a fair performance using only a few training samples. Here we introduce the concept of \emph{meta-worker}, a machine annotator trained by meta learning for types of tasks (i.e., image classification) that are well-fit for AI. Unlike regular crowd workers, meta-workers can be reliable, stable, and more importantly, tireless and free. We first cluster unlabeled data and ask crowd workers to repeatedly annotate the instances nearby the cluster centers; we then leverage the annotated data and meta-training datasets to build a cluster of meta-workers using different meta learning algorithms. Subsequently, meta-workers are asked to annotate the remaining crowdsourced tasks. The Jensen-Shannon divergence is used to measure the disagreement among the annotations provided by the meta-workers, which determines whether or not crowd workers should be invited for further annotation of the same task. Finally, we model meta-workers' preferences and compute the consensus annotation by weighted majority voting. Our empirical study confirms that, by combining machine and human intelligence, we can accomplish a crowdsourcing project with a lower budget than state-of-the-art task assignment methods, while achieving a superior or comparable quality.
Recent learning-based image classification and speech recognition approaches make extensive use of attention mechanisms to achieve state-of-the-art recognition power, which demonstrates the effectiveness of attention mechanisms. Motivated by the fact that the frequency and time information of modulated radio signals are crucial for modulation mode recognition, this paper proposes a time-frequency attention mechanism for a convolutional neural network (CNN)-based modulation recognition framework. The proposed time-frequency attention module is designed to learn which channel, frequency and time information is more meaningful in CNN for modulation recognition. We analyze the effectiveness of the proposed time-frequency attention mechanism and compare the proposed method with two existing learning-based methods. Experiments on an open-source modulation recognition dataset show that the recognition performance of the proposed framework is better than those of the framework without time-frequency attention and existing learning-based methods.
Current scene segmentation methods suffer from cumbersome model structures and high computational complexity, impeding their applications to real-world scenarios that require real-time processing. This paper proposes a novel Efficient Pyramid Representation Network (EPRNet), which strikes an innovative record on segmentation accuracy, model lightness and inference efficiency. Unlike existing methods delivering transfer learning based on pixel features of limited receptive fields encoded by shallow image classification backbones, EPRNet distributes multi-scale representations throughout the feature encoding flow to quickly enlarge and enrich receptive fields. Specifically, we introduce an extremely lightweight and efficient Multi-scale Processing Unit (MPU) that encodes multi-scale features through parallel convolutions of different kernels. By combining MPU and residual learning, we propose a core Pyramid Representation Module (PRM) to correctly acquire and aggregate region-based contexts in both shallow and deep layers. In this way, EPRNet can encode discriminative and comprehensive representations of multi-scale objects with a compact structure. We conduct extensive experiments on Cityscapes and CamVid datasets, demonstrating the superiority. Without any extra and coarse labeled data, EPRNet obtains mIoU 73.9% on the Cityscapes test set with only 0.9 million parameters at a speed of 42 FPS.
Recently, deep learning has drawn broad attention in the hyperspectral image (HSI) classification task. Many works have focused on elaborately designing various spectral-spatial networks, where convolutional neural network (CNN) is one of the most popular structures. To explore the spatial information for HSI classification, pixels with its adjacent pixels are usually directly cropped from hyperspectral data to form HSI cubes in CNN-based methods. However, the spatial land-cover distributions of cropped HSI cubes are usually complicated. The land-cover label of a cropped HSI cube cannot simply be determined by its center pixel. In addition, the spatial land-cover distribution of a cropped HSI cube is fixed and has less diversity. For CNN-based methods, training with cropped HSI cubes will result in poor generalization to the changes of spatial land-cover distributions. In this paper, an end-to-end fully convolutional segmentation network (FCSN) is proposed to simultaneously identify land-cover labels of all pixels in a HSI cube. First, several experiments are conducted to demonstrate that recent CNN-based methods show the weak generalization capabilities. Second, a fine label style is proposed to label all pixels of HSI cubes to provide detailed spatial land-cover distributions of HSI cubes. Third, a HSI cube generation method is proposed to generate plentiful HSI cubes with fine labels to improve the diversity of spatial land-cover distributions. Finally, a FCSN is proposed to explore spectral-spatial features from finely labeled HSI cubes for HSI classification. Experimental results show that FCSN has the superior generalization capability to the changes of spatial land-cover distributions.
In this extended abstract, we will present and discuss opportunities and challenges brought about by a new deep learning method by AUC maximization (aka \underline{\bf D}eep \underline{\bf A}UC \underline{\bf M}aximization or {\bf DAM}) for medical image classification. Since AUC (aka area under ROC curve) is a standard performance measure for medical image classification, hence directly optimizing AUC could achieve a better performance for learning a deep neural network than minimizing a traditional loss function (e.g., cross-entropy loss). Recently, there emerges a trend of using deep AUC maximization for large-scale medical image classification. In this paper, we will discuss these recent results by highlighting (i) the advancements brought by stochastic non-convex optimization algorithms for DAM; (ii) the promising results on various medical image classification problems. Then, we will discuss challenges and opportunities of DAM for medical image classification from three perspectives, feature learning, large-scale optimization, and learning trustworthy AI models.
We propose an accurate and efficient scene text detection framework, termed FAST (i.e., faster arbitrarily-shaped text detector). Different from recent advanced text detectors that used complicated post-processing and hand-crafted network architectures, resulting in low inference speed, FAST has two new designs. (1) We design a minimalist kernel representation (only has 1-channel output) to model text with arbitrary shape, as well as a GPU-parallel post-processing to efficiently assemble text lines with a negligible time overhead. (2) We search the network architecture tailored for text detection, leading to more powerful features than most networks that are searched for image classification. Benefiting from these two designs, FAST achieves an excellent trade-off between accuracy and efficiency on several challenging datasets, including Total Text, CTW1500, ICDAR 2015, and MSRA-TD500. For example, FAST-T yields 81.6% F-measure at 152 FPS on Total-Text, outperforming the previous fastest method by 1.7 points and 70 FPS in terms of accuracy and speed. With TensorRT optimization, the inference speed can be further accelerated to over 600 FPS. Code and models will be released at https://github.com/czczup/FAST.