Hong Kong Baptist University

Selected work

Representative Papers

ConjNorm: Tractable Density Estimation for Out-of-Distribution Detection

Feb 27, 2024International Conference on Learning Representations

Existing post-hoc out-of-distribution (OOD) detection methods rely on logits, distance metrics, or strong distributional assumptions, limiting their ability to accurately model true data density. To address this, we propose a unified density modeling framework based on Bregman divergences, reformulating density estimation as a differentiable optimization problem for the optimal norm coefficient $ p $. We first uncover a novel paradigm for exponential-family density modeling under conjugate constraints, leading to ConjNorm—a method that achieves unbiased, analytically differentiable density estimation without restrictive distributional assumptions. ConjNorm integrates Bregman divergence theory, exponential-family modeling, and Monte Carlo importance sampling. On CIFAR-100 and ImageNet-1K, it reduces false positive rate at 95% true positive rate (FPR95) by 13.25% and 28.19%, respectively, over prior state-of-the-art methods, significantly improving both OOD detection accuracy and robustness.

18 citations1 influentialRead paper

Robust Categorical Data Clustering Guided by Multi-Granular Competitive Learning

Jul 23, 2024IEEE International Conference on Distributed Computing Systems

Categorical data pose significant clustering challenges due to the absence of a well-defined distance metric, particularly when they exhibit multi-granular nested cluster structures. To address this, this work proposes a Multi-Granular Competitive Penalty Learning (MGCPL) mechanism that adaptively refines cluster structures in stages, integrated with a Cluster Aggregation and Metric Embedding (CAME) strategy based on learned distributions to enable robust clustering in the embedding space. This approach is the first to incorporate multi-granular competitive learning into categorical data modeling, offering both automatic granularity discovery and linear time complexity, thereby supporting scalability to large-scale datasets and compatibility with distributed pre-partitioning. Extensive experiments on multiple real-world datasets demonstrate its significant superiority over existing methods.

16 citationsRead paper

Accurate Forgetting for Heterogeneous Federated Continual Learning

Feb 20, 2025International Conference on Learning Representations

To address statistical bias and noise interference arising from client data/task heterogeneity—or even adversarial behavior—in federated continual learning (FCL), this paper introduces the “Accurate Forgetting” (AF) paradigm: proactively identifying and discarding unreliable feature representations induced by skewed distributions and noise prior to knowledge reuse. Methodologically, we propose the first probability-based credibility assessment framework built upon normalized flows, enabling quantifiable, knowledge-granular filtering. Further, we integrate generative replay with selective knowledge inheritance to dynamically enhance global model robustness within the federated architecture. Evaluated on multiple heterogeneous FCL benchmarks, AF achieves an average accuracy improvement of 12.3%, significantly boosting generalization and noise resilience. Our approach provides a novel, interpretable, and computationally tractable pathway for bias mitigation in FCL.

5 citationsRead paper

Towards Robust Learning to Optimize with Theoretical Guarantees

Jun 16, 2024Computer Vision and Pattern Recognition

Existing learned optimization (L2O) methods demonstrate strong empirical performance in wireless communications, networking, and electronic design automation (EDA), yet lack theoretical guarantees on out-of-distribution (OOD) robustness and convergence. This paper establishes the first unified theoretical framework for analyzing both convergence and OOD robustness of L2O. We derive sufficient conditions for uniform in-distribution (InD) convergence of L2O models; introduce a problem-transformation paradigm—“OOD alignment to InD”—to bridge distributional shifts; and quantitatively characterize the intrinsic relationship between convergence-rate degradation and input feature deviation. Methodologically, we propose a gradient-driven lightweight feature engineering module and a historical state modeling mechanism. Experiments demonstrate consistent and significant improvements over state-of-the-art methods under both InD and OOD settings, achieving up to 10× convergence acceleration. Our code is publicly available.

3 citationsRead paper

Learning to adapt unknown noise for hyperspectral image denoising

Dec 09, 2022

Existing variational models for hyperspectral image denoising lack noise adaptivity due to fixed weights in the data-fidelity term, rendering them inadequate for complex, unknown mixed noise (e.g., impulse, stripe, and coupled noise). To address this, we propose a learnable pixel-wise weighted data-fidelity term and design a Hyper-Weight Network (HWnet) that dynamically predicts spatially varying noise intensity maps. Within a bi-level optimization framework, weight prediction and denoising are decoupled. This work is the first to formulate noise intensity estimation as a hypernetwork learning problem, introduces a model-level noise knowledge transfer mechanism, and provides preliminary theoretical analysis of generalizability. Experiments demonstrate consistent PSNR improvements of 2.1–4.7 dB over mainstream model-driven frameworks (e.g., LRMR, LRTV). Moreover, HWnet exhibits strong cross-model and cross-noise-type generalization capability.

3 citationsRead paper

Recent publications