Chao Zhang

I develop data science and machine learning models to assist with task solving and decision-making. My research focuses on rapidly developing AI systems that are more data-efficient, trustworthy, and customized for specific domains.

Currently, I am working on the following themes:

1. Data-Efficient AI – Adapting Large Language Models for target domains by addressing data scarcity challenges through data-efficient methods such as feedback-based adaptation, weak supervision, and active learning.
2. Trustworthy AI – Making AI models more reliable and aligned with human values through uncertainty quantification, robust fine-tuning, scalable alignment, and decision-focused learning.
3. AI for Science – Leveraging data science and AI to accelerate scientific discovery in fields like material science, biomedical and life sciences, and urban science through close collaboration with domain experts.

Acknowledgment: My work has been generously supported by research funding/gift from NSF (IIS CAREER-2144338, IIS-2106961, IIS-2008334), ONR MURI , Kolon, HomeDepot, ADP, and Adobe. My work has also been recognized by an NSF CAREER Award, a Facebook Faculty Award, an Amazon AWS Machine Learning Research Award, a Google Faculty Research Award, a Kolon Faculty Fellowship, an ACM SIGKDD Dissertation Runner-up Award, and several paper awards from IMWUT (UbiComp), ECML/PKDD, and ML4H.

We aim to adapt LLMs to various domains and complex tasks. A key bottleneck when adapting LLMs is data scarcity – the lack of high-quality, representative data for the target domain. We are tackling this bottleneck by developing methods for data-efficient LLM adaptation, including:

• Enhancing LLM Reasoning Through Feedback: We investigate improving LLM reasoning and planning abilities by enabling them to evolve through interaction with external environments for feedback. The goal is to better adapt LLMs and enhance their reasoning capabilities without expensive manual curation of fine-tuning data.
  • BBox-Adapter: Lightweight Adapting for Black-Box Large Language Models, ICML'24
  • ARL2: Aligning Retrievers for Black-box Large Language Models via Self-guided Adaptive Relevance Labeling, ACL'24
  • Adapting LLM Agents with Universal Feedback Through Communication, preprint
  • ToolChain*: Efficient Action Space Navigation in Large Language Models with A* Search, ICLR'24
  • AdaPlanner: Adaptive Planning from Feedback with Language Models, NeurIPS'23
  • ToolQA: A Dataset for LLM Question Answering with External Tools, NeurIPS'23

• Fine-Tuning LLMs with Weak Supervision: To address the lack of data for LLM fine-tuning, we explore using weaker forms of supervision signals that are easier to collect (e.g., labeling rules, distant supervision, pseudo labels). We investigate 1) fine-tuning LLMs with weak supervision; and 2) discovering rules from language models for interactive weak supervision.
  • DyGen: Fine-Tuning Language Models with Noisy Labels by Dynamics-Enhanced Generative Modeling, KDD'23
  • PRBoost: Prompt-Based Rule Discovery and Boosting for Interactive Weakly-Supervised Learning, ACL'22
  • Fine-Tuning Pre-trained Language Model with Weak Supervision, NAACL'21
  • Sparse Conditional Hidden Markov Model for Weakly Supervised Named Entity Recognition, KDD'22
  • BERTifying Hidden Markov Models for Multi-Source Weakly Supervised Named Entity Recognition, ACL'21
  • Text Classification Using Label Names Only: A Language Model Self-Training Approach, EMNLP'20
  • Denoising Multi-Source Weak Supervision for Neural Text Classification, EMNLP'20
  • BOND: Bert-Assisted Open-Domain Named Entity Recognition with Distant Supervision, KDD'20
  • Weakly-Supervised Neural Text Classification, CIKM'18

• Active Learning for LLM Fine-Tuning: We develop techniques for fine-tuning LLMs with active learning, by strategically selecting data instances for annotation to improve fine-tuning performance.
  • Cold-start Data Selection for Better Few-shot Fine-tuning of Pretrained Language Models, ACL'23
  • AcTune: Uncertainty-Aware Active Self-Training for Active Fine-Tuning of Language Models, NAACL'22

We aim to develop AI systems that are not only capable but also trustworthy for deployment in various domains. We focus on several aspects of trustworthy AI: uncertainty quantification for deep learning, robust LLM fine-tuning and alignment, and decision-focused learning.

• Uncertainty Quantification for Deep Learning: Uncertainty-aware ML models are crucial for building trustworthy AI systems. Many deep learning models produce uncertainty-agnostic point estimates or miscalibrated distributions. We are addressing this issue by designing techniques for quantifying uncertainty in deep learning models:
  • SMURF-THP: Score Matching-based UnceRtainty quantiFication for Transformer Hawkes Process, ICML'23
  • When Rigidity Hurts: Soft Consistency Regularization for Probabilistic Hierarchical Time Series Forecasting, KDD'23
  • UnfoldML: Cost-Aware and Uncertainty-Based Dynamic 2D Prediction for Multi-Stage Classification, NeurIPS'22
  • CAMUL: Calibrated and Accurate Multi-view Time-Series Forecasting, WWW'22
  • When in Doubt: Neural Non-Parametric Uncertainty Quantification for Epidemic Forecasting, NeurIPS'21
  • SDE-Net: Equipping Deep Neural Networks with Uncertainty Estimates, ICML'20
  • RDeepSense: Reliable Deep Mobile Computing Models with Uncertainty Estimations, UbiComp'18

• Robust LLM Fine-Tuning and Alignment: We investigate robust fine-tuning methods for LLMs to improve generalization on out-of-distribution data, as well as identifying vulnerabilities in LLMs and scalably aligning them:
  • Aligning Large Language Models with Representation Editing: A Control Perspective, NeurIPS'24
  • Large Language Model as Attributed Training Data Generator: A Tale of Diversity and Bias, NeurIPS'23
  • COCO-DR: Combating the Distribution Shift in Zero-Shot Dense Retrieval with Contrastive and Distributional Robust Learning, EMNLP'22
  • Calibrated Language Model Fine-Tuning for In- and Out-of-Distribution Data, EMNLP'20

• Decision-Focused Learning: We study how to make learning informed by downstream decision-making tasks. We design end-to-end decision-focused learning methods that are scalable and can handle hard-to-specify constraints:
  • Diffusion Models as Constrained Samplers for Optimization with Unknown Constraints, preprint
  • DF2: Distribution-Free Decision-Focused Learning, preprint
  • End-to-end Stochastic Programming with Energy-based Model, NeurIPS'22

We aim to leverage data science and AI for scientific applications ranging from macro-level systems like transportation and epidemiology to micro-level systems such as molecules. We have been developing data-driven tools to advance scientific discovery in material design, biomedical and life science, and urban science:

• AI for Material Science:
  • Efficient Evolutionary Search Over Chemical Space with Large Language Models, preprint
  • LLMatDesign: Autonomous Materials Discovery with Large Language Models, preprint
  • May the Force be with You: Unified Force-Centric Pre-Training for 3D Molecular Conformations, NeurIPS'23
  • Autoregressive Diffusion Model for Graph Generation, ICML'23
  • MUBen: Benchmarking the Uncertainty of Pre-Trained Models for Molecular Property Prediction, TMLR'24
  • POLYIE: A Dataset of Information Extraction from Polymer Material Scientific Literature, NAACL'24

• AI for Biomedical and Life Science:
  • Contrastive Fitness Learning: Reprogramming Protein Language Models for Low-N Learning of Protein Fitness Landscape, RECOMB'24
  • TransEHR: Self-Supervised Transformer for Clinical Time Series Data, ML4H'23
  • When in Doubt: Neural Non-Parametric Uncertainty Quantification for Epidemic Forecasting, NeurIPS'21
  • ReSel: N-ary Relation Extraction from Scientific Text and Tables by Learning to Retrieve and Select, EMNLP'22
• AI for Urban Science:
  • T-GCN: A Temporal Graph Convolutional Network for Traffic Prediction, IEEE ITS'20
  • A Linear Time Approach to Computing Time Series Similarity based on Deep Metric Learning, TKDE'20
  • DeepMove: Predicting Human Mobility with Attentional Recurrent Networks, WWW'18

• 2024 GaTech CoC Outstanding Junior Faculty Award
• 2022 NSF Career Award
• 2022 ML4H Outstanding Paper Award
• 2021 Facebook Faculty Research Award
• 2021 Kolon Faculty Fellowship
• 2020 Amazon AWS Machine Learning Research Award
• 2020 Google Faculty Research Award
• 2019 ACM SIGKDD Dissertation Award Runner-up
• 2018 ACM IMWUT Distinguished Paper Award
• 2015 ECML/PKDD Best Student Paper Runner-up Award
• 2013 Chiang Chen Overseas Graduate Fellowship

(* denotes equal contribution)

• CX4240: Introduction to Computational Data Analysis
• CSE8803-DLT: Deep Learning for Text Data

Prospective students: I am always looking for strong and motivated students to join our group. If you are interested in working with me, you can either email me or fill out this form.

Current:

• Yue Yu: Ph.D. Student in CSE
• Rui Feng: Ph.D. Student in CS
• Yuchen Zhuang: Ph.D. Student in ML
• Yinghao Li: Ph.D. Student in ML
• Rongzhi Zhang: Ph.D. Student in ML
• Haotian Sun: Ph.D. Student in ML (co-advised with Bo Dai)
• Kuan Wang: Ph.D. Student in CSE
• Haorui Wang: Ph.D. Student in CSE
• Agam A. Shah: Ph.D. Student in ML (co-advised with Sudheer Chava)
• Rushi Qiang: Ph.D. Student in CSE (co-advised with Bo Dai)
• Changhao Li: Ph.D. Student in CSE (co-advised with Bo Dai)
• Jacob Wessell: M.S. student in CS
• Wenhao Mu: M.S. student in CS
• Shangqing Xu: M.S. student in CS

Alumni:

• Lingkai Kong: Ph.D., 2024 (–> Postdoc Fellow @ Harvard)
• Yanbo Xu: Ph.D., 2023 (–> Research Scienctist @ Microsoft Research)
• Binghong Chen: Ph.D., 2023 (–> Quant @ Citadel Capital, co-advised with Prof. Le Song)
• Pranav Shetty: Ph.D., 2023 (–> Research Scienctist @ JP Morgan Chase, JP Morgan AI Ph.D. Fellowship, co-advised with Prof. Rampi Ramprasad)
• Vidit Jain: M.S. Student in CS
• Mukund Rungta: M.S. Student in CS
• Junyang Zhang: M.S. Student in CS
• Piyush Patil: M.S. Student in CS
• Mengyang Liu: M.S. Student in CSE
• Isaac Rehg: M.S. in CS
• Wendi Ren: M.S. in CSE
• Ruijia Wang: M.S. in CSE
• Yi Rong: Visiting Ph.D. Student