Events
FDS Postdoctoral Applicants
Recent Advances in Robust Learning of Multi-Index Models
 |
Speaker: Nikolaos Zarifis (Wisconsin-Madison) PhD Student University of Wisconsin-Madison, Computer Sciences Monday, January 6, 2025 10:00AM - 11:00AM and via Webcast: https://yale.zoom.us/j/96298650015 |
Abstract: Multi-index models (MIMs) incorporate the idea that the target function depends only on a few relevant directions. Multi-index models encompass a wide range of commonly studied function classes, including constant-depth neural networks, and Boolean concept classes such as intersections of halfspaces.
In the first part of this talk, I will talk about the power of query access for the fundamental task of agnostically learning Multi-Index models under the Gaussian distribution. Our main result shows that query access gives significant runtime improvements over random examples for agnostically learning MIMs.
In the second part of the talk, I will talk about the problem of learning halfspaces with Massart noise under the Gaussian distribution. In this noise model, an adversary is allowed to flip the label of each point with an unknown probability at most 1/2. We give the first efficient algorithm for this problem whose complexity is nearly matched by Statistical Query lower bounds.
This talk is based on the following works: Agnostically Learning Multi-index Models with Queries (FOCS’24) and Learning General Halfspaces with General Massart Noise under the Gaussian Distribution (STOC’22) and are joint works with I. Diakonikolas, D. Kane, V. Kontonis and C. Tzamos.
Speaker bio: Nikos is a PhD student at the Computer Sciences Department of University of Wisconsin-Madison, advised by Ilias Diakonikolas. His research focuses on developing efficient noise-tolerant learning algorithms for classical classification and regression problems, in the presence of random examples with corrupted labels, and characterizing the associated statistical-computational tradeoffs.
Add To: Google Calendar | Outlook | iCal File
Submit an Event
Interested in creating your own event, or have an event to share? Please fill the form if you’d like to send us an event you’d like to have added to the calendar.