Maximizing Mutual Information Across Feature and Topology Views for Representing Graphs

Article Computer Science, Artificial Intelligence

Tasks Integrated Networks: Joint Detection and Retrieval for Image Search

Lei Zhang et al.

Summary: The traditional object retrieval task requires precise cropping of the objects and learning discriminative feature representations. However, in real-world scenarios, accurate detection or annotation of objects is often not available, leading to the need for image-level search with joint detection and retrieval. This paper proposes an Integrated Net and an improved version, DC-I-Net, for tackling this challenge.

IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE (2022)

Add to Collection

Article Engineering, Electrical & Electronic

Partial Alignment for Object Detection in the Wild

Zhenwei He et al.

Summary: Conventional object detectors often suffer from performance drops when faced with domain shift caused by environmental changes. Unsupervised domain adaptive object detection (DAOD) has gained attention as a solution to this problem. However, existing cross-domain object detectors with parameter-shared network architecture accumulate errors and distortions from the source domain, leading to model collapse. To address this issue, we propose a novel Asymmetric Tri-way Faster-RCNN (ATF) model with an ancillary net, which helps preserve the source distribution and enhance the discrimination of the target domain object detector. Additionally, we introduce a Partial Alignment based ATF (PA-ATF) model to remove source-specific knowledge and exploit informative domain-invariant knowledge. Our models demonstrate remarkable performance on benchmark datasets, surpassing other state-of-the-art approaches.

IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY (2022)

Add to Collection

Article Computer Science, Artificial Intelligence

Deep Neural Message Passing With Hierarchical Layer Aggregation and Neighbor Normalization

Xiaolong Fan et al.

Summary: The paper introduces a deep hierarchical layer aggregation (DHLA) strategy and neighbor normalization (NeighborNorm) strategy for addressing the training difficulties in deep MPNNs. Experimental results demonstrate the necessity and effectiveness of these proposed strategies for graph message-passing neural networks.

IEEE TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS (2022)

Add to Collection

Article Automation & Control Systems

Graph Convolutional Network-Based Method for Fault Diagnosis Using a Hybrid of Measurement and Prior Knowledge

Zhiwen Chen et al.

Summary: A new fault diagnosis method combining structural analysis and graph convolutional network is proposed, utilizing a hybrid of available measurement and prior knowledge. Experimental results demonstrate that this method achieves better diagnosis results compared to existing methods based on common evaluation indicators.

IEEE TRANSACTIONS ON CYBERNETICS (2022)

Add to Collection

Proceedings Paper Computer Science, Artificial Intelligence

Node Similarity Preserving Graph Convolutional Networks

Wei Jin et al.

Summary: The SimP-GCN framework balances information from graph structure and node features by preserving feature similarity during aggregation process, effectively maintaining node similarity.

WSDM '21: PROCEEDINGS OF THE 14TH ACM INTERNATIONAL CONFERENCE ON WEB SEARCH AND DATA MINING (2021)

Add to Collection

Proceedings Paper Computer Science, Artificial Intelligence

HDMI: High-order Deep Multiplex Infomax

Baoyu Jing et al.

Summary: In this study, a new framework called HDMI is proposed for self-supervised learning of node embedding on multiplex networks. The framework aims to address the limitations of existing methods and achieve better performance on multiple layers of networks.

PROCEEDINGS OF THE WORLD WIDE WEB CONFERENCE 2021 (WWW 2021) (2021)

Add to Collection

Proceedings Paper Computer Science, Artificial Intelligence

Dynamic Weighted Learning for Unsupervised Domain Adaptation

Ni Xiao et al.

Summary: This paper proposes Dynamic Weighted Learning (DWL) to address the issue of negative transfer in unsupervised domain adaptation, caused by insufficient balance between domain alignment and class discrimination. By dynamically weighting the learning losses and considering the problem of sample imbalance across domains, DWL demonstrates excellent performance on several benchmark datasets.

2021 IEEE/CVF CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION, CVPR 2021 (2021)

Add to Collection

Article Automation & Control Systems