Report Title: Structured Matrices and Neural Networks

Speaker: Siyu Liao

Affiliation: Sun Yat-sen University

Time: 10:00 a.m., Thursday, June 12, 2025

Venue: Second Meeting Room, 1st Floor, Block A, Feicui Science and Education Building

Report Abstract:

The combination of circulant matrices and neural networks provides new ideas for efficient deep learning models. As a special type of structured matrix with cyclic shift characteristics, circulant matrices enable fast computation (O(n log n) complexity) through Fourier transform and significantly reduce the number of parameters (from n² to n). In neural networks, cyclic structures can be embedded in fully connected layers or convolutional layers, endowing the model with prior knowledge such as translation invariance, which is suitable for tasks such as image, time series and graph data. Current research has proven that lightweight networks based on circulant matrices greatly improve computational efficiency while maintaining performance, providing a feasible solution for edge computing and large-scale deployment.

Speaker Profile:

Siyu Liao has served as a Tenure-Track Assistant Professor (Associate Professor) at the School of Integrated Circuits, Sun Yat-sen University since June 2024. He received his Ph.D. from Rutgers, The State University of New Jersey, USA in 2020, his Master's degree from the City University of New York in 2018, and his Bachelor's degree from Hefei University of Technology in 2014. During his studies, he was a visiting scholar at New York University in the fall of 2017 and at Princeton University in the fall of 2019. He has interned at IBM, eBay and Facebook (now Meta), and worked as an Applied Scientist at Amazon for more than 3.5 years after obtaining his Ph.D., accumulating rich industrial experience. His main research interests lie in the design methods of high-performance neural network hardware and software. He has published a total of 27 papers in international academic journals and conferences, and holds 1 US invention patent.