Research

Research Overview

Our research centers on Edge AI, aiming to fulfill various user demands on edge nodes (e.g., mobile and IoT devices) through scalable, trustworthy, and efficient AI/ML services. Grounded in both theoretical and empirical foundations, we develop new AI/ML algorithms (applicable to both vision and language tasks) tailored to the following research areas:

Collaborative/Federated Learning
Personalized and Safe AI
Efficient Learning and Inference
Understanding Neural Network Representations

Detailed descriptions of each research area are provided below.

Collaborative/Federated Learning

Personalized and Safe AI

Efficient Learning and Inference

1. Collaborative/Federated Learning

With the explosive growth in the number of smartphones, IoT devices, and data-holding institutions (e.g., companies, hospitals), a vast volume of data is being generated at the edge. However, these data owners may not want to share their privacy-sensitive data (e.g., medical data) with others. We focus on both the algorithmic and theoretical aspects of scalable and heterogeneous federated LLM fine-tuning.

2. Personalized and Safe AI

Providing personalized, robust, and trustworthy AI services to every edge user is one of the most rewarding challenges in the current AI/ML era. In particular, models should be continuously personalized to each user's preference while also effectively handling out-of-distribution scenarios (i.e., generalization/robustness to unseen domains and adversarial prompts). In this context, we focus on developing new strategies for continual learning, LLM safety, and unlearning.

3. Efficient Learning and Inference

In practice, enabling on-device AI on edge devices is crucial. In particular, each user must operate under constraints such as limited data, scarce labels, and restricted computation and memory resources during both training and inference. To address these challenges, our goal is to develop algorithms for efficient training, LLM compression, and collaborative reasoning/inference.

4. Understanding Neural Network Representations

Understanding how deep neural networks represent and process information is essential for building reliable and controllable AI systems. We study how representation structures influence model behavior, focusing on interpretable representations (e.g., SAEs) and mathematical foundations for the principled understanding and control of LLMs, VLMs, and MLLMs.

Finally, we have also been exploring the above aspects of AI/ML over edge/communication/wireless networks (e.g., edge computing assisted AI). This research direction has a significant demand for the delivery of intelligent AI/ML services in real-world network edge.

Google Sites

Report abuse