PhD Seminar: Chao Chen: "Robust Learning with Complex and Noisy Data – a Topological View"

Dates: 
Friday, October 9, 2020 - 2:40pm to 3:40pm
Location: 
Zoom
Event Description: 

Abstract: Modern machine learning faces new challenges. We are analyzing highly complex data with unknown noise and potentially poisonous information from attackers. These noise or attacks are especially dangerous to deep neural nets with strong memorization power. In this talk, we discuss several recent results on how to train deep learning models that are robust against noise in labels, or potentially robust against poisoning backdoor attacks. We also discuss how to develop graph neural networks that are robust against graph structural attacks, i.e., perturbation of graph edges with malicious intention. Under-the-hood is a unified geometric and topological view of the data in the latent representational space. We show that advanced topological and geometric tools, if tightly coupled with deep neural networks, can provide novel and powerful prior for robust learning. Results have been published in ICLR’20, ICML’20 and NeurIPS’20.

Contact events [at] cs.stonybrook.edu for Zoom info.
 

Computed Event Type: 
Mis
Event Title: 
PhD Seminar: Chao Chen: "Robust Learning with Complex and Noisy Data – a Topological View"