We explore how learning progresses over time and varies across individual animals through brief moments of movement. Using temporal proximity as supervision, our self-supervised framework captures individuality and learning progress without manually defining any behavioral information. Through the learned representations, we reveal how changes in animal behavior unfold with learning.
@article{zhang2026representing,title={Representing short-term movement to reveal animal identity and learning progress},author={Zhang, Ping and Nguyen, Quang-Huy and Lei, Yingtie and Shridhar, Sameera and Leveroni, Zachary and Quattrocchi, Ellie and Yu, Jai Y and Chao, Wei-Lun},journal={arXiv preprint arXiv:2610.27140},year={2026},month=oct,}
We address OOD Object Detection by leveraging the inconsistency between generative and discriminative model outputs. We employ an off-the-shelf generative model as an auxiliary to the object detector and introduce a triplet similarity metric that captures both semantic and visual differences, enabling effective OOD object detection in a zero-shot manner.
@inproceedings{nguyen2026ronin,title={Detecting Out-of-Distribution Objects through Class-Conditioned Inpainting},author={Nguyen, Quang-Huy and Zhou, Jin and Liu, Zhenzhen and Bui, Huyen and Weinberger, Kilian Q. and Chao, Wei-Lun and Le, Dung D.},booktitle={IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},year={2026},}
We present a comprehensive study on Semi-Supervised Learning (SSL) using Vision Foundation Models (VFMs) and propose a simple yet effective baseline that leverages diverse predictions from multiple Parameter-Efficient Fine-Tuning (PEFT) strategies to enhance SSL performance.
@inproceedings{mai2025revisiting,title={Revisiting Semi-Supervised Learning in the Era of Foundation Models},author={Zhang, Ping and Mai, Zheda and Nguyen, Quang-Huy and Chao, Wei-Lun},booktitle={Advances in Neural Information Processing Systems (NeurIPS)},year={2025},}
We present a unified empirical study of Parameter-Efficient Fine-Tuning (PEFT) methods in visual recognition, offering complementary perspectives to deeply understand their behaviors under different regimes (low-shot, many-shot, domain shift), and highlight their complementary predictions and robustness trade-offs.
@inproceedings{mai2025lessons,title={Lessons and Insights from a Unifying Study of Parameter-Efficient Fine-Tuning (PEFT) in Visual Recognition},author={Mai, Zheda and Zhang, Ping and Tu, Cheng-Hao and Chen, Hong-You and Nguyen, Quang-Huy and Zhang, Li and Chao, Wei-Lun},booktitle={IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},year={2025},}
We investigate Expensive Multi-Objective Optimization by introducing the Stein Variational Hypernetwork for Pareto Set Learning, which alleviates fragmented and uncertain regions in surrogate models while preserving the diversity of learned solutions, demonstrating strong performance on expensive multi-objective optimization problems.
@inproceedings{nguyen2025improving,title={Improving Pareto Set Learning for Expensive Multi-objective Optimization via Stein Variational Hypernetworks},author={Nguyen, Minh-Duc and Dinh, Phuong Mai and Nguyen, Quang-Huy and Hoang, Long P. and Le, Dung D.},booktitle={AAAI Conference on Artificial Intelligence (AAAI)},year={2025},}
We explore Few-shot Image Classification by proposing a new cross-attention mechanism based on cosine similarity, without using softmax, to further emphasize the correlation between labeled support and unlabeled query representations, thus enhancing ViT-based few-shot algorithms across various settings and scenarios compared to conventional attention mechanisms.
@article{nguyen2023enhancing,title={Enhancing Few-shot Image Classification with Cosine Transformer},author={Nguyen, Quang-Huy and Nguyen, Cuong Q. and Le, Dung D. and Pham, Hieu H.},journal={IEEE Access},year={2023},}
