In video object tracking, there exist rich temporal contexts among successive frames, which have been largely overlooked in existing trackers. this work, we bridge the individual frames and explore across them via a transformer architecture for robust tracking. Different from classic usage of natural language processing tasks, separate its encoder decoder into two parallel branches carefully design within Siamese-like tracking pipelines. The promotes target templates attention-based feature...

Knowledge graph embedding, which aims to represent entities and relations as low dimensional vectors (or matrices, tensors, etc.), has been shown be a powerful technique for predicting missing links in knowledge graphs. Existing embedding models mainly focus on modeling relation patterns such symmetry/antisymmetry, inversion, composition. However, many existing approaches fail model semantic hierarchies, are common real-world applications. To address this challenge, we propose novel...

State-of-the-art neural style transfer methods have demonstrated amazing results by training feed-forward convolutional networks or using an iterative optimization strategy. The image representation used in these methods, which contains two components: and content representation, is typically based on high-level features extracted from pretrained classification networks. Because the are originally designed for object recognition, often focus central neglect other details. As a result,...

Graph convolutional networks (GCNs)—which are effective in modeling graph structures—have been increasingly popular knowledge completion (KGC). GCN-based KGC models first use GCNs to generate expressive entity representations and then embedding (KGE) capture the interactions among entities relations. However, many fail outperform state-of-the-art KGE though introducing additional computational complexity. This phenomenon motivates us explore real effect of KGC. Therefore, this paper, we...

Deep graph neural networks (GNNs) have been shown to be expressive for modeling graph-structured data. Nevertheless, the overstacked architecture of deep models makes it difficult deploy and rapidly test on mobile or embedded systems. To compress over-stacked GNNs, knowledge distillation via a teacher-student turns out an effective technique, where key step is measure discrepancy between teacher student with predefined distance functions. However, using same graphs various structures may...

Sample efficiency remains a key challenge for the deployment of deep reinforcement learning (RL) in real-world scenarios. A common approach is to learn efficient representations through future prediction tasks, facilitating agent make farsighted decisions that benefit its long-term performance. Existing methods extract predictive features by predicting multi-step state signals. However, they do not fully exploit structural information inherent sequential signals, which can potentially...

Tensor factorization based models have shown great power in knowledge graph completion (KGC). However, their performance usually suffers from the overfitting problem seriously. This motivates various regularizers -- such as squared Frobenius norm and tensor nuclear while limited applicability significantly limits practical usage. To address this challenge, we propose a novel regularizer namely, DUality-induced RegulArizer (DURA) which is not only effective improving of existing but widely...

Node representation learning on attributed graphs-whose nodes are associated with rich attributes (e.g., texts and protein sequences)-plays a crucial role in many important downstream tasks. To encode the graph structures simultaneously, recent studies integrate pre-trained models neural networks (GNNs), where serve as node encoders (NEs) to attributes. As jointly training large NEs GNNs large-scale graphs suffers from severe scalability issues, methods propose train separately....

This paper presents a joint head pose and facial landmark regression method with input from depth images for realtime application. Our main contributions are: firstly, optimization to estimate landmarks, i.e., the result provides supervised initialization cascaded regression, while landmarks can also help further refine at each stage. Secondly, we classify space into 9 sub-spaces, then use random forest global shape constraint training in specific space. classification-guided effectively...

Semantic matching models-which assume that entities with similar semantics have embeddings-have shown great power in knowledge graph embeddings (KGE). Many existing semantic models use inner products embedding spaces to measure the plausibility of triples and quadruples static temporal graphs. However, vectors same another vector can still be orthogonal each other, which implies may dissimilar embeddings. This property significantly limits performance models. To address this challenge, we...

Three-dimensional convolutional networks (3D CNNs) are used efficiently in various video recognition applications. Compared to traditional 2D CNNs, extra temporal dimension causes 3D CNNs more computationally intensive and have a larger memory footprint. Therefore, the optimization is extremely crucial this case. This paper presents design space exploration of access for FPGA-based CNN accelerator. We present non-overlapping data tiling method contiguous off-chip explore on-chip reuse...

With the success of generative adversarial networks (GANs) on various real-world applications, controllability and security GANs have raised more concerns from community. Specifically, understanding latent space GANs, i.e., obtaining completely decoupled space, is essential for applications in some secure scenarios. At present, there no quantitative method to measure decoupling which not conducive development In this article, we propose two methods sensitivity dimensions: one a sequential...

Knowledge Graphs (KGs) provide human knowledge with nodes and edges being entities relations among them, respectively.Multihop question answering over KGs-which aims to find answer of given questions through reasoning paths in KGs-has attracted great attention from both academia industry recently.However, this task remains challenging, as it requires accurately identify answers a large candidate entity set, which the size grows exponentially number hops.To tackle problem, we propose novel...

Medical Visual Question Answering (MedVQA) has gained increasing attention at the intersection of computer vision and natural language processing. Its capability to interpret radiological images deliver precise answers clinical inquiries positions MedVQA as a valuable tool for supporting diagnostic decision-making physicians alleviating workload on radiologists. While recent approaches focus using unified pre-trained large models multi-modal fusion like cross-modal Transformers, research...

We investigate the explainability of graph neural networks (GNNs) as a step toward elucidating their working mechanisms. While most current methods focus on explaining nodes, edges, or features, we argue that, inherent functional mechanism GNNs, message flows are more natural for performing explainability. To this end, propose novel method here, known FlowX, to explain GNNs by identifying important flows. quantify importance flows, follow philosophy Shapley values from cooperative game...

Generalization across different environments with the same tasks is critical for successful applications of visual reinforcement learning (RL) in real scenarios. However, distractions---which are common scenes---from high-dimensional observations can be hurtful to learned representations RL, thus degrading performance generalization. To tackle this problem, we propose a novel approach, namely Characteristic Reward Sequence Prediction (CRESP), extract task-relevant information by reward...

Knowledge graph embedding, which aims to represent entities and relations as low dimensional vectors (or matrices, tensors, etc.), has been shown be a powerful technique for predicting missing links in knowledge graphs. Existing embedding models mainly focus on modeling relation patterns such symmetry/antisymmetry, inversion, composition. However, many existing approaches fail model semantic hierarchies, are common real-world applications. To address this challenge, we propose novel --...

In industrial production and engineering operations, the health state of complex systems is critical, predicting it can ensure normal operation.Complex have many monitoring indicators, coupling structures, non-linear time-varying characteristics, so a challenge to establish reliable prediction model.The belief rule base (BRB) fuse observed data expert knowledge nonlinear relationship between input output has well modeling capabilities.Since each indicator system reflect some extent, BRB...

Cutting planes (cuts) play an important role in solving mixed-integer linear programs (MILPs), which formulate many real-world applications. Cut selection heavily depends on (P1) cuts to prefer and (P2) how select. Although modern MILP solvers tackle (P1)-(P2) by human-designed heuristics, machine learning carries the potential learn more effective heuristics. However, existing learning-based methods prefer, neglecting importance of Moreover, we observe that (P3) what order selected...