BEGIN:VCALENDAR VERSION:2.0 PRODID:-//wp-events-plugin.com//7.2.3.1//EN TZID:America/New_York X-WR-TIMEZONE:America/New_York BEGIN:VEVENT UID:827@fds.yale.edu DTSTART;TZID=America/New_York:20250106T100000 DTEND;TZID=America/New_York:20250106T110000 DTSTAMP:20250916T142146Z URL:https://fds.yale.edu/events/fds-seminar-nikos-zarifis-wisconsin-madiso n/ SUMMARY:FDS Seminar: Nikos Zarifis (Wisconsin-Madison)\, "Recent Advances i n Robust Learning of Multi-Index Models" DESCRIPTION:\nAbstract: Multi-index models (MIMs) incorporate the idea that the target function depends only \;on a few relevant directions. Mult i-index models encompass a wide range of commonly studied \;function c lasses\, including constant-depth neural networks\, and Boolean concept cl asses \;such as intersections of halfspaces. \n\n\n\nIn the first par t of this talk\, I will talk about the power of query access for the funda mental task of agnostically learning Multi-Index models under the Gaussian distribution. Our main result shows that query access gives significant r untime improvements over random examples for agnostically learning MIMs.&n bsp\;\n\n\n\nIn the second part of the talk\, I will talk about the proble m of learning halfspaces with Massart noise under the Gaussian distributio n. 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&nb sp\;efficient algorithm for this problem whose complexity is nearly matche d by Statistical Query lower bounds.\n\n\n\nThis talk is based on the foll owing works: Agnostically Learning Multi-index Models with Queries (FOCS'2 4) and Learning General Halfspaces with General Massart Noise under the Ga ussian Distribution (STOC'22) and are joint works with I. Diakonikolas\, D . Kane\, V. Kontonis and C. Tzamos.\n\n\n\nSpeaker bio: Nikos is a PhD stu dent at the Computer Sciences Department of University of Wisconsin-Madiso n\, advised by Ilias Diakonikolas.  His research focuses on developing ef ficient noise-tolerant learning algorithms for classical classification an d regression problems\, in the presence of random examples with corrupted labels\, and characterizing the associated statistical-computational trade offs.\n CATEGORIES:FDS Events,Postdoctoral Applicants END:VEVENT BEGIN:VTIMEZONE TZID:America/New_York X-LIC-LOCATION:America/New_York BEGIN:STANDARD DTSTART:20241103T010000 TZOFFSETFROM:-0400 TZOFFSETTO:-0500 TZNAME:EST END:STANDARD END:VTIMEZONE END:VCALENDAR