전문가 초청 온라인 세미나를 개최하오니 관심있는 학생들의 많은 참여 부탁드립니다!

Title: Ridgelet Transform: Harmonic Analysis for Deep Learning Theory

Speaker: Dr. Sho Sonoda @ RIKEN

Time : 2023 Nov 08th 15:00 ~ 16:00

Location: Online https://xinics.zoom.us/j/81643257240 (passcode: 95315885)

Language: English speech & English slides

Abstract: 

Ridgelet transform is a pseudo-inverse operator of neural networks. Namely, given a function $f \in L^2(\mathbb{R}^m)$, the ridgelet transform $R[f]$ describes how the network parameters should be organized for the network to represent $f$. In this talk, I will explain two systematic schemes to derive the ridgelet transform. As applications, we investigate modern neural networks involving the ones on manifolds $G/K$ and Hilbert spaces $H$ as well as deep networks, and derive their associated ridgelet transforms.

Bio:

Sho Sonoda is a permanent research scientist at the deep learning theory team, RIKEN Center for Advance Intelligence Project (RIKEN AIP). His research interests lie in both theory and practice of machine learning, and in particular, he has continuously produced original research results in the ridgelet transform theory. The long term research goal is to turn deep learning from alchemy to principled science and/or technology. Before joining RIKEN, he earned his degree from Waseda university, which is in Tokyo and famous for drinking a lot.