Personalized Large Language Model and Deployment on Mobile Devices

The School of EECS is hosting the following PhD Progress Review 1 Confirmation Seminar: 

Personalized Large Language Model and Deployment on Mobile Devices

Speaker: Zhaofeng Zhong
Host: Dr. Xin Yu

Abstract:

With the advancement of large language models (LLMs), significant progress has been made in various Natural Language Processing (NLP) tasks. Recent research has primarily focused on either developing cloud-based personalized LLMs or exploring the on-device deployment of general LLMs. As the scale of LLMs grows, there is an urgent need to develop personalized LLMs on user-memory-restrained devices that address user-specific requirements while operating locally on edge devices. We identify three challenges obstructing the pathway to efficiently deploy personalized LLMs on user devices: 1) Alleviating user data sparsity to enhance the performance of personalized on-device models. 2) Decreasing the risk of user data privacy leakage while enhancing on-device model performance with server LLM assistance. 3) Integrating multitask capability with on-device LLM to generate personalized text on different tasks. 

To solve the first challenge, we introduces a cloud-device collaborative data augmentation for personalized language model (CDCDA-PLM), a framework for deploying a personalized LLM on user devices with the assistance of a powerful cloud-based LLM while satisfying personalized user privacy requirements. Specifically, to overcome the data sparsity of on-device personal data, users have the flexibility to selectively share personal data with the server-side LLM to generate more synthetic personal data. By combining this synthetic data with locally stored user data, we fine-tune the personalized parameter-efficient fine-tuning (PEFT) modules of the small on-device model to capture user personas effectively. Our experiments demonstrate the effectiveness of CDCDA-PLM across six tasks in a widely used personalization benchmark.

Bio:

Zhaofeng Zhong is a second-year Ph.D. student at the University of Queensland, Australia. Prior to his candidacy, he received his bachelor's degree in Electrical and Computer Engineering from The University of Queensland. He is currently working on the research topic of personalized and on-device LLM, advised by Prof. Hongzhi Yin and Dr. Rocky Tong Chen.