|
I am a PhD student in Department of Electrical and Computer Engineering, University of Texas at Austin. I am co-advised by Prof. Sandeep Chinchali and Prof. Ufuk Topcu. I received my BS degree in Department of Electrical Engineering at National Taiwan University, where I worked with Prof. Wanjiun Liao.
Email: pohan[last name] [at] utexas [dot] edu |
|
News
- May 2025: Check out my new preprint about evaluating video captions (video-to-text evaluation) without human supervision here .
- Jan 2025: My first-authored paper and second-authored paper were accepted to ICLR 2025!
- Nov 2024: Check out my new preprint about Any2Any, a retrieval framework for incomplete multimodal data here .
- Sep 2024: I finished my Ph.D. progress review!
- Aug 2024: I finished my summer internship at Meta!
- Feb 2024: Check out my new preprint about foundation model selection here .
- Jan 2024: I became a Ph.D. candidate of UT ECE!
- Sep 2023: Check out my new preprint about NDPCA, a dynamic compression framework for efficient multi-sensor data transmission under varying bandwidth conditions here .
Research
My research centers on multimodal machine learning, with a focus on foundation models, modality adaptation, inference-time optimization, and robust evaluation pipelines. I work on handling missing or noisy modalities and integrating multiple modalities to extend the frontier of foundation models. My projects emphasize integration of multimodal information and the development of scalable multimodal ML systems for both research and industry applications. I recently focus on video language models (VLMs) and the evaluation of their generated text.
Information Extraction across Modalities
My research focuses on advancing multimodal machine learning by integrating information from diverse modalities, each offering unique perspectives or signals. I work on adapting foundation models to handle missing or noisy modalities, optimizing inference, and building robust evaluation pipelines. These efforts improve the efficiency, reliability, and scalability of multimodal systems, with recent work emphasizing video-language models (VLMs) and the evaluation of their generated text.The projects concentrate on three key areas: (1) enhanced multi-modal data fusion with novel post-training/post-processing techniques, (2) implementing differential privacy mechanisms for information exchanging, and (3) mitigating the effects of adversarial attacks on false information sharing.
Publications
Po-han Li, Sandeep P. Chinchali, Ufuk Topcu
ICLR, 2025
Blog / Code
Mohammad Omama, Po-han Li, Sandeep P. Chinchali
ICLR, 2025
Shenghui Chen*, Po-han Li*, Sandeep P. Chinchali, Ufuk Topcu
Under Review, 2025
Po-han Li, Sandeep P. Chinchali, Ufuk Topcu
Under Review, 2024
Po-han Li, Sravan Kumar Ankireddy, Ruihan Zhao, Hossein Nourkhiz Mahjoub, Ehsan Moradi Pari, Ufuk Topcu, Sandeep P. Chinchali, Hyeji Kim
Advances in Neural Information Processing Systems (Neurips), 2023
Blog / Code
Po-han Li, Sandeep P Chinchali, Ufuk Topcu
American Control Conference (ACC), 2023
Blog / Code
Po-han Li, Ufuk Topcu, Sandeep P Chinchali
58th Allerton Conference on Communication, Control, and Computing, 2022
Decentralized Data Sharing
These projects aim to improve federated learning (FL), which typically requires all devices to have identical neural network structures. My projects explore sharing raw data, not gradients, among devices to overcome this limitation. Additionally, I focus on challenges related to limited network bandwidth and privacy preservation when sharing valuable data. My work covers topics such as out-of-distribution detection, data valuation, active learning, differential privacy, and distributed optimization.Publications
Oguzhan Akcin, Po-han Li, Shubhankar Agarwal, Sandeep P. Chinchali
Conference on Robot Learning (CoRL), 2022
Yuchong Geng, Dongyue Zhang, Po-han Li, Oguzhan Akcin, Ao Tang, Sandeep P Chinchali
Conference on Robot Learning (CoRL), 2021
Large Language Model Selection
Language models like ChatGPT and Bard have become integral to our daily lives, yet their performance hinges on user context, training data, and model architecture, which users typically have limited knowledge of. Interactions with these models occur through internet APIs, treating them as black boxes. My project tackles a key question: How can we select the right large language model on the Internet for a specific task? How do we even define tasks of our chatbot conversation? To answer this, I seek to uncover the correlation between model performance and user context of conversation. Finally, I want create a decision-making algorithm for choosing the best model.Publications
Po-han Li, Oyku Selin Toprak, Aditya Narayanan, Ufuk Topcu, Sandeep Chinchali
Under review, 2024
Blog / Code
Miscellaneous
I enjoy working on projects with brilliant minds across labs, even if they are not directly related to my research. The followings are some of my side, but cool projects.Learning for Dynamics & Control Conference, 2025
IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2024
AAAI Conference on Artificial Intelligence, 2023
Conference on Decision and Control (CDC), 2023