March 11, 2022 09:33
EPFL CIS-RIKEN AIP Joint Seminar #10 20220309 thumbnails

Description

The 10th Seminar

Date and Time: March 9th 6:00pm – 7:00pm(JST) 10:00am-11:00pm(CET)
Venue:Zoom webinar
Language: English
Speaker: Ichiro Takeuchi, RIKEN AIP
Title: Selective Inference for Deep Learning Model-driven Hypotheses

Abstract:
Conditional selective inference (SI) framework was recently introduced as a new statistical inference method for Lasso. This framework allows us to derive the exact conditional sampling distribution of the selected test statistic when the selection event is characterized by a polyhedron. In fact, this framework is not only useful for Lasso but also generally applicable to a certain class of data-driven hypotheses. A common limitation of existing SI methods is that the hypothesis selection event must be characterized in a simple tractable form such as a set of linear or quadratic inequalities. To overcome this limitation, we propose a new computational method for conditional SI based on parametric programming, which we call PP-based SI. In this talk, after briefly reviewing the conditional SI framework, we show that the proposed PP-based SI is more powerful than (vanilla) SI and applicable to a wider class of problems. Furthermore, as a demonstration of the PP-based SI, we present our recent work on conditional SI for data-driven hypotheses derived from deep learning models.

Bio:
Ichiro Takeuchi is a professor at Nagoya Institute of Technology and the team leader of Data-Driven Biomedical Science team in RIKEN AIP. He received B. Eng., M. Eng., and D. Eng. degrees from Nagoya University, Japan, in 1996, 1998, and 2000, respectively. After he worked as a post-doctoral researcher in Montreal, Canada, under Prof. Yoshua Bengio’s supervision, he got tenured assistant professor position at Mie University, Japan in 2001, and associate and full professor positions at Nagoya Institute of Technology in 2008 and 2015, respectively. His research interests include theory and algorithm of machine learning and its application to bio-medical and material sciences.