- This event has passed.
Doctoral Dissertation Proposal Oral Presentation, Hongwu Peng
March 22 @ 1:30 pm - 2:30 pm EDT
Title: Improving Machine Learning Efficiency and Privacy — Algorithm and System Co-design Approaches
Ph.D. Candidate: Hongwu Peng
Major Advisors: Caiwen Ding
Associate Advisors: David Kaeli, Sanguthevar Rajasekaran, Yuan Hong, Wei Zhang
Committee Members: Chenghong Wang
Date/Time: Friday, March 22nd, 2024, 1:30pm
Location: HBL Class of 1947 Conference Room
Meeting link: https://uconn-cmr.webex.com/uconn-cmr/j.php?MTID=mdf31fd4f0826eb5e6d60a9fbf9a2d2a5
Meeting number: 2634 112 0380
Password: NHwMMmU329x
Abstract: In the pursuit of advancing machine learning, this doctoral dissertation proposal presents a unified examination of algorithm and system co-design approaches to enhance efficiency and privacy across various computational models. At the core of this thesis is a series of integrated methodologies that transform the landscape of machine learning by addressing the dual challenges of computational performance and data security in an increasingly connected world.
The research encapsulated within this proposal spans transformative solutions such as hardware-friendly sparse attention mechanisms and dynamic resource allocation on FPGA platforms, which streamline processing times for complex Transformer models. It further delves into the realm of private network inference, introducing automated ReLU replacement strategies with polynomial approximations that accelerate inference without sacrificing privacy or accuracy. The thesis also ventures into the optimization of privacy-preserving machine learning with homomorphic encryption, reducing computational overhead while maintaining robust security measures. Lastly, it explores the acceleration of graph neural network training through novel GPU kernel designs that improve computation speeds by orders of magnitude. Collectively, these contributions represent a cohesive strategy toward efficient and secure machine learning, showcasing the potential of co-design to revolutionize both the theory and practice of deep neural networks across various computing paradigms.
