Kang, Qiyukang0080@e.ntu.edu.sg
Research Interest and Biography
I am exploring a new paradigm in machine learning that integrates principles of physics dynamics and mathematics, aiming to address some of the pressing challenges in artificial intelligence, such as effectiveness, interpretability, and robustness. My current work focuses on three significant areas: the complexity of neural network design, the vulnerability of neural networks to perturbations, and the practical validation of this emerging paradigm. Through my research on dynamics-driven continuous graph neural networks, dynamics-guided robust neural networks, and perception-decision practices, I have had the opportunity to contribute to both theoretical insights and practical applications. Looking ahead, I am eager to further enhance our understanding of this innovative machine learning paradigm and explore its potential to transform the landscape of artificial intelligence.
I am currently a research fellow at the School of Electrical and Electronic Engineering in Nanyang Technological Universty (NTU). I received the B.S. degree in Electronic Information Science and Technology from University of Science and Technology of China (USTC) in 2015, and the Ph.D. degree from NTU in 2020, supervised by Prof. Tay Wee Peng.
Academic services: reviewer for NeurIPS, ICRL, IEEE TSP, AAAI, CVPR, IEEE TITS, TNNLS, IEEE L-CSS, IJCAI, PR, and others.
News
One paper is accepted to NeurIPS 2024 as Spotlight.
Our paper, titled "Distributed-Order Fractional Graph Operating Network", introduces a highly general continuous GNN framework inspired by differential equations. This framework features a learnable probability distribution over a range of real numbers for the derivative orders.
One paper is accepted to TITS.
Our paper, titled "PRFusion: Towards Effective and Robust Multi-Modal Place Recognition with Image and Point Cloud Fusion", proposes multi-modal place recognition models that utilize global fusion with manifold metric attention, achieving state-of-the-art performance on three large-scale benchmarks.
One paper is accepted to ICLR 2024 as Spotlight.
Our paper, titled "Unleashing the Potential of Fractional Calculus in Graph Neural Networks with FROND", introduces a novel continuous GNN framework that incorporates fractional differential equations.
One paper is accepted to TGRS.
Our paper, titled "PointDifformer: Robust Point Cloud Registration with Neural Diffusion and Transformer", proposes a robust point cloud registration approach using graph neural diffusion.
Three papers are accepted to AAAI 2024.
- Coupling Graph Neural Networks with Fractional Order Continuous Dynamics: A Robustness Study
- DistilVPR: Cross-Modal Knowledge Distillation for Visual Place Recognition
- PosDiffNet: Positional Neural Diffusion for Point Cloud Registration in a Large Field of View with Perturbations
One paper is accepted to NeurIPS 2023 as Spotlight.
Our paper, titled "Adversarial robustness in graph neural networks: A Hamiltonian approach", introduces a novel physics-driven GNN that leverages conservative Hamiltonian flows and Lyapunov stability to significantly enhance robustness against adversarial perturbations.
One paper is accepted to ICML 2023.
Our paper, titled "Node embedding from neural Hamiltonian orbits in graph neural networks", presents a novel approach for graph node embedding that addresses the challenge of varying embedding spaces for different data types. We introduce Hamiltonian orbits to model the embedding update of a node feature, allowing us to learn the underlying manifold of the graph during training.
One paper is accepted to IJCAI 2023.
Our paper, titled "Graph neural convection-diffusion with heterophily", introduces a novel graph neural network (GNN) approach to address the challenge of heterophilic graphs, where connected nodes may have different classes or dissimilar features. By incorporating the convection-diffusion equation (CDE), the proposed GNN captures both homophily and heterophily in information flow.
One paper is accepted to CVPR 2023.
Our paper, titled "HypLiLoc: Towards effective LiDAR pose regression with hyperbolic fusion", introduces a new model for LiDAR pose regression. The model incorporates two branched backbones to extract 3D features and 2D projection features, respectively. We propose multi-modal feature fusion in both Euclidean and hyperbolic spaces to improve feature representations.
One paper is accepted to AAAI 2023.
In this paper, named "RobustLoc: Robust Camera Pose Regression in Challenging Driving Environments", to deal with challenging driving environments that may have changing seasons, weather, illumination, and the presence of unstable objects, we propose RobustLoc, which derives its robustness against perturbations from neural differential equations.