BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Memento EPFL//
BEGIN:VEVENT
SUMMARY:High-dimensional model selection: transfer learning and projected 
 computation
DTSTART:20260430T161500
DTEND:20260430T171500
DTSTAMP:20260601T073051Z
UID:26cec256d3a13c458221423b71f38fbd70aa28720e3f7f99d9b5475e
CATEGORIES:Conferences - Seminars
DESCRIPTION:David Rossel\, Universitat Pompeu Fabra (UPF) and Barcelona S
 chool of Economics (BSE). \nWe define as model selection the structural l
 earning problem where the goal is to learn the subset of truly zero parame
 ters in a probability model of interest\, such as canonical generalized li
 near models\, generalized additive models or graphical models. This is of 
 course one of the most classical and fundamental problems in Statistics an
 d Machine Learning. L0 penalization methods and their related Bayesian mod
 el selection counterparts have optimal mathematical properties for this ta
 sk\, yet much mainstream literature considers such methods to be either un
 necessary or impractical. This talk discusses these two objections\, and h
 ow to ameliorate the issues: is it useful to do this\, and can we do this 
 computationally?\nWe first discuss how these methods can be useful\, not o
 nly in theory but also in practice\, and how to improve their performance 
 via data integration (also called transfer learning). For example\, sparse
  model recovery methods enjoy excellent asymptotic properties when certain
  sparsity and signal strength (betamin) conditions hold\, but these assump
 tions often don't hold in some application domains. We show that data inte
 gration pushes the mathematical conditions under which consistent model re
 covery is possible.\nRegarding the second objection of computational impra
 cticality\, we review recent optimization and MCMC literature showing that
 \, under somewhat strict sparsity assumptions\, the computational cost sca
 les linearly with the problem dimension (with high probability\, asymptoti
 cally). A key practical issue is that such results assume that one can qui
 ckly score each candidate model (at constant cost)\, but even in least-squ
 ares the cost is (at least) quadratic in the model dimension and grows als
 o with the sample size n. We propose a new class of projected model select
 ion criteria that score models at constant cost\, after an initial pre-pro
 cessing step\, and which enjoy the same asymptotic and practical performan
 ce as the costlier exact model scores.\n\n\n 
LOCATION:CM 1 517 https://plan.epfl.ch/?room==CM%201%20517
STATUS:CONFIRMED
END:VEVENT
END:VCALENDAR