About Me

I am an undergraduate student majoring in Computer Engineering and Computer Science, and a Machine Learning researcher in Prof. Kangwook Lee's group in University of Wisconsin-Madison. My research interests include Large Language Model, Multimodal, Continual Federated Learning, Visual-Language Model, Image Generation, among others.

I am currently applying for a PhD program for Fall 2024.

Ongoing Project

Benchmarking Visual In-Context Learning for Multimodal Large Language Models
• Summary: We aim to explore the foundation model's in-context learning capabilities for text-to-image generation. This involves generating high-dimensional image data from low-dimensional textual input, a process that potentially exhibits distinct characteristics specific to multimodal in-context learning and has not yet been thoroughly explored.

Publications

FedGP: Buffer-based Gradient Projection for Continual Federated Learning
Shenghong Dai, Yicong Chen, Jy-yong Sohn, S M Iftekharul Alam, Ravikumar Balakrishnan, Suman Banerjee, Nageen Himayat, Kangwook Lee
Federated Learning Systems (FLSys) Workshop @ MLSys 2023 • Oral Presentation • Best Paper Award
• Summary: We introduce a buffer-based Gradient Projection method for Continual Federated Learning that combats catastrophic forgetting using local buffer samples and aggregated buffer gradients.

Zero-shot Improvement of Object Counting with CLIP
Ruisu Zhang*, Yicong Chen*, Kangwook Lee
Robustness of Few-shot and Zero-shot Learning in Foundation Models (R0-FoMo) Workshop @ NeurIPS 2023
• Summary: We mitigate CLIP's object counting limitations by introducing a zero-shot method that manipulates its text embedding space, enhancing counting accuracy and also improving the performance of text-to-image generative models.

Coded Prompts for Large Language Models
Ziqian Lin, Yicong Chen, Yuchen Zeng, Kangwook Lee
Robustness of Few-shot and Zero-shot Learning in Foundation Models (R0-FoMo) Workshop @ NeurIPS 2023
• Summary: We introduce coded prompts, inspired by coding theory, to process multiple inputs simultaneously in Large Language Models (LLMs), enhancing task performance. Viewing LLM inference as a noisy communication channel, a coded prompt has the potential to protect against information lost.