- HSSB 1173 and Zoom
- Department Seminar
Zoom Link: https://ucsb.zoom.us/j/82189331655
Title: Deep Learning meets Nonparametric Regression: Are Weight-decayed DNNs locally adaptive?
Abstract: They say deep learning is just curve fitting. But how good is it in curve fitting exactly? Are DNNs as good as, or even better than, classical curve-fitting tools, e.g., splines and wavelets? In this talk, I will cover my group's recent paper on this topic and some interesting insight. Specifically, I will provide new answers to "Why is DNN stronger than kernels?" "Why are deep NNs stronger than shallow ones?", "Why ReLU?", "How do DNNs generalize under overparameterization?", "What is the role of sparsity in deep learning?", "Is lottery ticket hypothesis real?" All in one package. Intrigued? Come to my talk and find out!
Link to the paper: arxiv.org/abs/2204.09664