Re: [theory students] [theory faculty] Theory seminar Thursday (April 10) with cookies! 9:45-10:45am


Date: Thu, 10 Apr 2025 14:50:36 +0000
From: Sandeep Silwal <silwal@xxxxxxxxxxx>
Subject: Re: [theory students] [theory faculty] Theory seminar Thursday (April 10) with cookies! 9:45-10:45am
Happening now! 

On Apr 10, 2025, at 9:30âAM, Sandeep Silwal <silwal@xxxxxxxxxxx> wrote:

ï
Happening soon!

On Apr 10, 2025, at 8:37âAM, Sandeep Silwal <silwal@xxxxxxxxxxx> wrote:

ï Hi everyone,

See you soon today at 9:45am in CS 3310!

Best,
Sandeep

On Apr 7, 2025, at 9:47 AM, Sandeep Silwal via Theory-faculty <theory-faculty@xxxxxxxxxxx> wrote:

Hi everyone, 

This Thursday (April 10) we will have a seminar at the usual time and place (CS 3310, 9:45-10:45 am) with our very own Ilias Diakonikolas.

Title: Learning Multi-Index Models

Abstract:
Multi-index models (MIMs) are functions that depend on the projection onto a low-dimensional subspace. These models provide a useful framework for analyzing a wide range of machine learning problems, such as multiclass linear classification, learning intersections of halfspaces, and more complex neural networks. Despite extensive investigation, there remains a significant gap in our understanding of the efficient learnability of MIMs.

In this talk, we will survey recent algorithmic work on learning MIMs, focusing on methods with provable performance guarantees. In particular, we will present a robust noise-tolerant learning algorithm that works for a broad class of well-behaved MIMs, under standard distributional assumptions. As applications, we will demonstrate how this framework leads to much faster noise-tolerant learning algorithms for multiclass linear classifiers and intersections of halfspaces.


See you then!
Sandeep
-------------------------------------------------------------------------------
This message was sent to the mailing list for the theory faculty at the
Computer Sciences Department of the University of Wisconsin in Madison.

[← Prev in Thread] Current Thread [Next in Thread→]