Biography

I am a tenure-track assistant professor of Computer Science and UMIACS at University of Maryland, College Park. My research interests are in machine learning, optimization, and natural language processing. I have published ~100 papers at NeurIPS, ICML, ICLR, AISTATS, ECML/PKDD, ACL, EMNLP, NAACL, COLING, CVPR, ICCV, KDD, ICDM, AAAI, IJCAI, ISIT, Machine Learning (Springer), IEEE TPAMI/TIP/TNNLS/TKDE, etc.

Our recent works study (1) How, why, and when to transfer human learning (e.g., curriculum, retention, sub-tasking, curiosity, exemplar selection, collaboration, etc.) to improve machine learning in the wild (e.g., with unlabeled, biased, noisy, redundant or distributed data, unseen tasks/environments, distribution shift); (2) Controllable AI in both training and inference/adaptation; and (3) Human-AI alignment and AI personalization. And Yes we are developing these methods for LLMs, multi-modality foundation models, and RL. Our goal is to develop efficient, versatile, trustworthy, and environmentally-friendly hybrid-intelligence based on coevolution between human and machine. A list of our research topics can be found at the bottom of this webpage.

I was a visiting research scientist at Google between 2021-2022. Before that, I received my Ph.D. (thesis) from Computer Science of University of Washington, where I was a member of MELODI lab led by Prof. Jeff A. Bilmes. I have been working with Prof. Dacheng Tao as a research assistant at University of Technology, Sydney (UTS) and Nanyang Technological University. I was a research intern at Yahoo! Labs, mentored by Dr. Hua Ouyang (Apple) and Prof. Yi Chang (Jilin University), and a research intern at Microsoft Research, mentored by Dr. Lin Xiao (Meta AI).

News

• 2024/07: We initialize TurningPoint AI, a research team from multiple universities and industry (UMD+UCLA+PSU+Google) investigating Muiltimodal Agents, with the goals of building Trustworthy Embodied AI, Self-Evolving Machines, Compositional Agents, and Controllable AIGC. We already released 8 projects with several ICML and ECCV publications and new datasets.
• 2024/07: 2 papers of diffusion models (analysis of negative prompts, extracting discriminative features from generative models) have been accepted by ECCV 2024.
• 2024/05: 4 ICLR + 4 ICML + 6 ACL + 2 NAACL + 2 CVPR have been accepted, featuring our works on controllable AIGC, personalized AI, data-efficient training of LLMs, RLHF, prompt optimization, multi-modal hallucinations, multi-modal and embodied agent, and curriculum reinforcement learning.
• 2023/11: I will give a talk “Towards Controllable and Personalized AI Models” at UMD CS department seminar on 11/03.
• 2023/10: We release HallusionBench focusing on the Language Hallucination and Visual Illusion of GPT-4V(ision), Llava-1.5, and other multi-modality models. Analyses and insights can be found in the preprint.
• 2023/10: Data recycling and filtering improves instruction-tuning of LLMs, leading to recycled LLMs outperforming other larger LLMs trained on new data and RLHF. We release Reflection-Tuning preprint, codebase, and the model.
• 2023/10: Two papers (How Many Demonstrations Do You Need for In-context Learning, Merging Mixture-of-Experts into One) have been accepted by EMNLP 2023.
• 2023/09: Two papers (multi-modality model distillation for task adaptation, clustered additive modeling for structured federated learning) have been accepted by NeurIPS 2023.
• 2023/07: Two papers (model-adaptive data augmentation curriculum, subclass balancing for long-tail recognition) have been accepted by ICCV 2023.
• 2023/06: How to efficiently optimize the textual instructions applied to API black-box LLMs (e.g., ChatGPT) for solving downstream tasks? Please check our recent work InstructZero, paper and code have been released.
• 2023/06: Invited talk at Purdue Statistics on “Structured Decentralized Learning”.
• 2023/06: Two papers (Meta-Vote Pruning and Eigensubspace of Temporal-Difference Dynamics) have been accepted by ECML/PKDD 2023.
• 2023/05: I will teach CMSC-421 on “Introduction to Artificial Intelligence” in Fall 2023.
• 2023/04: Three papers (training dynamics of continual learning, continual RL via sparse prompting, structured cooperative learning) have been accepted by ICML 2023. See you at Hawaii in July!
• 2023/04: One paper about personalization in federated recommendation system has been accepted by IJCAI 2023.
• 2022/12: I will teach CMSC-828A on “Fantastic Machine Learning Paradigms and Where to use Them” in Spring 2023.
• 2022/12: I will serve as an SPC (meta-reviewer) for IJCAI 2023.
• 2022/11: One XAI paper on extracting local reasoning chains for subtasks from neural networks such as ResNet and ViT has been accepted by TMLR.
• 2022/10: One paper (adversarial attacks to Question-Answer models) has been accepted by EMNLP 2022.
• 2022/09: Three papers (adversarial augmentation for continual learning, adversarial augmentation for representation learning, federated learning from pre-trained models) have been accepted by NeurIPS 2022.

Research Topics

• Machine Learning (2008-present)
  1. Learning over time: Curriculum Learning, Continual Learning
  2. Learning via interactions: Reinforcement Learning, Online Learning
  3. Learning across tasks/domains: Multi-task Learning, Meta-Learning, Domain Adaptation/Generalization
  4. Learning multiple models: Mixture-of-Experts (MoE), Collaborative/Cooperative Learning, Federated/Decentralized Learning
  5. Learning under noises: Noisy-Label Learning, Adversarial Learning
  6. Learning representations: Self-Supervised Learning, Dimension Reduction
  7. Sparse Learning: Compressed Sensing, Matrix Factorization, Spectral Method
  8. Optimization: Continuous, Combinatorial, Multi-Objective, Black-Box
• Natural Language Processing (2016-present)
  1. Attention mechanisms
  2. Toxicity and Bias in NLP models
  3. Adversarial textual attack and defense
  4. Large language models (LLMs) (Reflection-Tuning, Alpagasus, and Cherry LLM created by our group)
  5. Personalization
• Multi-modality Models (2021-present)
  1. Vision-Language Models
  2. Human-AI alignment
  3. VLM/LLM + RL and Multi-modality Embodied-AI