For pursuing accurate skeleton-based action recognition, many existing graph-based approaches deploy the higher-order polynomials of the skeletal adjacency matrix to model the node correlations of distant neighbours. To further capture robust graphical patterns, a novel multi-scale graph convolution operator is proposed, which enables to aggregate multi-scale dependencies and capture long-range joint relationships on human skeleton graph. Additionally, a novel corrective contrastive learning strategy is proposed, which aims to distinguish the representative clues and calibrate the confused action clips in the feature space. Comprehensive experiments validate the effectiveness and superiority of our proposed method over state-of-the-art approaches on three large-scale datasets: NTU RGB+D 60, NTU RGB+D 120, and Kinetics Skeleton 400.

The existing approaches for skeleton-based action recognition based on graph convolutional networks (GCNs) primarily emphasize the construction of human skeletal structure by leveraging inherent connections. However, the static skeletal topology used across all action categories fails to capture discriminative relationships between joint pairs, while current graph structures struggle to model dynamic motion information, limiting their ability to represent both temporal and motion-specific dependencies. To address this limitation, we propose the decoupled static-dynamic co-occurrence graph convolution (DSDC-GConv), which specifically aims to learn and adapt the graph topology by refining the inter-frame and intra-frame joint dependencies through decomposed manner. Additionally, a multi-level context-aware module is proposed to comprehensively model the latent saliencies of multiple domains in skeletal sequences. This module refines the spatial nodes, temporal dynamics, channel-wise characteristics, and motional dependencies within the graph convolution block. Furthermore, a hierarchical densely connected temporal convolution is proposed to enhance the representation of local features through partial dense connections and enrich the temporal information during the convolution process. Findings from our evaluations on five large-scale benchmark datasets (i.e., NTU RGB+D 60, NTU RGB+D 120, Kinetics Skeleton 400, Northwestern-UCLA, PKU-MMD) demonstrate the effectiveness and superiority of our proposed method over competing approaches, with an recognition accuracy of 93.0% and 97.1% on NTU RGB+D 60, 89.9% and 90.6% on NTU RGB+D 120, 38.6% and 63.4% on Kinetics Skeleton 400, 97.4% on Northwestern-UCLA, 97.6% and 63.6% on PKU-MMD.

In the area of medical image segmentation, the spatial information can be further used to enhance the image segmentation performance. And the 3D convolution is mainly used to better utilize the spatial information. However, how to better utilize the spatial information in the 2D convolution is still a challenging task. In this paper, we propose an image segmentation network based on reinforcement learning (RLSegNet), which can translate the image segmentation process into a serial of decision-making problem. The proposed RLSegNet is a U-shaped network, which is composed of three components: the feature extraction network, the Mask Prediction Network (MPNet), and the up-sampling network with the cascade attention module. The deep semantic feature in the image is first extracted by adopting the feature extraction network. Then, the Mask Prediction Network (MPNet) is proposed to generate the prediction mask for the current frame based on the prior knowledge (segmentation result). And the proposed cascade attention module is mainly used to generate the weighted feature mask so that the up-sampling network pays more attention to the interesting region. Specifically, the state, action and reward used in the reinforcement learning are redesigned in the proposed RLSegNet to translate the segmentation process as the decision-making process, which performs as the reinforcement learning to realize the brain tumor segmentation. Extensive experiments are conducted on the BRATS 2015 dataset to evaluate the proposed RLSegNet. The experimental results demonstrate that the proposed method can achieve a better segmentation performance, in comparison with other state-of-the-art methods.

Mitigating adversarial deep learning attacks remains challenging, partly because of the ease and low cost in carrying out such attacks. Therefore, in this paper, we focus on the understanding of universal adversarial example attack on image classification models. Specifically, we seek to understand the difference(s) between adversarial examples in two adversarial datasets (DAmageNet and PGD dataset) and clean examples in ImageNet learned by the classification model, and whether we can use such findings to resist adversarial example attacks. We also seek to determine if we can retrain a discriminator to discriminate whether the input image is an adversarial example, using adversarial training. We then design a number of experiments (e.g., class activation map (CAM) analysis, feature map analysis, feature maps/filters changing, adversarial training, and binary classification model) to help us determine whether the universal adversarial dataset can be successfully used to attack the classification model. This, in turn, contributes to a better understanding of adversarial defenses over pretrained classification model from an interpretation perspective. To the best of our knowledge, this work is one of the earliest works to systematically investigate the interpretation of universal adversarial example attack on image classification models, both visually and quantitatively.

Domain adaptation methods train a model to find similar feature representations between a source and target domain. Recent methods leverage self-supervised learning to discover the analogous representations of the two domains. However, prior self-supervised methods have three significant drawbacks: (1) leveraging pretext tasks that are susceptible to learning low-level representations, (2) aligning the two domains using adversarial loss without considering if the extracted features are low-level representations, (3) the models are not flexible to accommodate various proportions of target labels, i.e., they assume target labels are always available. This paper presents a Generic Domain Adaptation Network (GDAN) to address these issues. First, we introduce a criterion based on instance discrimination to select appropriate pretext tasks to learn high-level domain invariant representations. Then, we propose a semantic neighbor cluster to align the two domain features. The semantic neighbor cluster implements a clustering technique in a feature embedding space to form clusters according to high-level semantic similarities. Finally, we present a weighted target loss function to balance the model weights according to the target labels. This loss function makes GDAN flexible for semi-supervised scenarios, i.e., partly labeled target data. We evaluate the proposed methods on four domain adaptation benchmark datasets. The experiment findings show that the proposed methods align the two domains well and achieve competitive results.

Medical practitioners generally rely on multimodal brain images, for example based on the information from the axial, coronal, and sagittal views, to inform brain tumor diagnosis. Hence, to further utilize the 3D information embedded in such datasets, this paper proposes a multi-view dynamic fusion framework (hereafter, referred to as MVFusFra) to improve the performance of brain tumor segmentation. The proposed framework consists of three key building blocks. First, a multi-view deep neural network architecture, which represents multi learning networks for segmenting the brain tumor from different views and each deep neural network corresponds to multi-modal brain images from one single view. Second, the dynamic decision fusion method, which is mainly used to fuse segmentation results from multi-views into an integrated method. Then, two different fusion methods (i.e., voting and weighted averaging) are used to evaluate the fusing process. Third, the multi-view fusion loss (comprising segmentation loss, transition loss, and decision loss) is proposed to facilitate the training process of multi-view learning networks, so as to ensure consistency in appearance and space, for both fusing segmentation results and the training of the learning network. We evaluate the performance of MVFusFra on the BRATS 2015 and BRATS 2018 datasets. Findings from the evaluations suggest that fusion results from multi-views achieve better performance than segmentation results from the single view, and also implying effectiveness of the proposed multi-view fusion loss. A comparative summary also shows that MVFusFra achieves better segmentation performance, in terms of efficiency, in comparison to other competing approaches.

Network alignment (NA) - i.e., linking entities from different networks (also known as identity linkage) - is a fundamental problem in many application domains. Recent advances in deep graph learning have inspired various auspicious approaches for tackling the NA problem. However, most of the existing works suffer from efficiency and generalization, due to complexities and redundant computations.We approach the NA from a different perspective, tackling it via meta-learning in a semi-supervised manner, and propose an effective and efficient approach called Meta-NA - a novel, conceptually simple, flexible, and general framework. Specifically, we reformulate NA as a one-shot classification problem and address it with a graph meta-learning framework. Meta-NA exploits the meta-metric learning from known anchor nodes to obtain latent priors for linking unknown anchor nodes. It contains multiple sub-networks corresponding to multiple graphs, learning a unified metric space, where one can easily link entities across different graphs. In addition to the performance lift, Meta-NA greatly improves the anchor linking generalization, significantly reduces the computational overheads, and is easily extendable to multi-network alignment scenarios. Extensive experiments conducted on three real-world datasets demonstrate the superiority of Meta-NA over several state-of-the-art baselines in terms of both alignment accuracy and learning efficiency.

Meta-learning has received a tremendous recent attention as a possible approach for mimicking human intelligence, i.e., acquiring new knowledge and skills with little or even no demonstration. Most of the existing meta-learning methods are proposed to tackle few-shot learning problems such as image and text, in rather Euclidean domain. However, there are very few works applying meta-learning to non-Euclidean domains, and the recently proposed graph neural networks (GNNs) models do not perform effectively on graph few-shot learning problems. Towards this, we propose a novel graph meta-learning framework -- Meta-GNN -- to tackle the few-shot node classification problem in graph meta-learning settings. It obtains the prior knowledge of classifiers by training on many similar few-shot learning tasks and then classifies the nodes from new classes with only few labeled samples. Additionally, Meta-GNN is a general model that can be straightforwardly incorporated into any existing state-of-the-art GNN. Our experiments conducted on three benchmark datasets demonstrate that our proposed approach not only improves the node classification performance by a large margin on few-shot learning problems in meta-learning paradigm, but also learns a more general and flexible model for task adaption.