Friday Lunch Seminar (English)
September 22, 2023
12:15 〜 13:00

Sing-up for participation  by noon, September 21
from here
You will be notified of participation details by e-mail on September 21.

Talk Title: Identifiable nonlinear representation learning and some applications

Hiroshi Morioka
Researcher
Center for Advanced Intelligence Project
RIKEN

Host PI: Okito Yamashita

Abstract:
Revealing fundamental representation (latent components) generating observational data in a data-driven manner is called representation learning, and has a long history including such as principal component analysis (PCA) and independent component analysis (ICA). Many frameworks were proposed based on deep learning in recent years to somehow extend them to nonlinear cases, including variational autoencoders (VAEs) and generative adversarial networks (GANs). However, such nonlinear representation learning is in general ill-posed, and it is known that there is no theoretical guarantee that they can estimate the “true” components. In this talk, I will introduce our work on nonlinear independent component analysis (NICA), and explain how such problems can be solved and made identifiable by adding some assumptions on the latent components. I will also introduce some recent extensions of NICA to dynamical models and causal discovery, with some applications to neuroimaging data.

“