BEGIN:VCALENDAR VERSION:2.0 PRODID:-//UC Irvine//CML Seminars//EN CALSCALE:GREGORIAN METHOD:PUBLISH X-WR-CALNAME:Deep Learning Theory in the Age of Generative AI X-WR-TIMEZONE:America/Los_Angeles BEGIN:VTIMEZONE TZID:America/Los_Angeles BEGIN:DAYLIGHT TZOFFSETFROM:-0800 TZOFFSETTO:-0700 TZNAME:PDT DTSTART:20070311T020000 RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=2SU END:DAYLIGHT BEGIN:STANDARD TZOFFSETFROM:-0700 TZOFFSETTO:-0800 TZNAME:PST DTSTART:20071104T020000 RRULE:FREQ=YEARLY;BYMONTH=11;BYDAY=1SU END:STANDARD END:VTIMEZONE BEGIN:VEVENT UID:2025-03-11-sadhika-malladi@cml.ics.uci.edu DTSTAMP:20250311T000000Z SEQUENCE:57690 DTSTART;TZID=America/Los_Angeles:20250311T110000 DTEND;TZID=America/Los_Angeles:20250311T120000 SUMMARY:[CML Seminar] Sadhika Malladi: Deep Learning Theory in the Age of G enerative AI LOCATION:Donald Bren Hall 4011 DESCRIPTION:Sadhika Malladi\, PhD Student\, Department of Computer Science\ , Princeton University\n\nTitle: Deep Learning Theory in the Age of Genera tive AI\n\nAbstract: Modern deep learning has achieved remarkable results\ , but the design of training methodologies largely relies on guess-and-che ck approaches. Thorough empirical studies of recent massive language model s (LMs) is prohibitively expensive\, underscoring the need for theoretical insights\, but classical ML theory struggles to describe modern training paradigms. I present a novel approach to developing prescriptive theoretic al results that can directly translate to improved training methodologies for LMs. My research has yielded actionable improvements in model training across the LM development pipeline — for example\, my theory motivates the design of MeZO\, a fine-tuning algorithm that reduces memory usage by up to 12x and halves the number of GPU-hours required. Throughout the talk \, to underscore the prescriptiveness of my theoretical insights\, I will demonstrate the success of these theory-motivated algorithms on novel empi rical settings published after the theory.\n\nhttps://cml.ics.uci.edu/semi nars/2025-03-11-sadhika-malladi X-ALT-DESC;FMTTYPE=text/html:Sadhika Malladi\, PhD Stude nt\, Department of Computer Science\, Princeton University

Title : Deep Learning Theory in the Age of Generative AI

Abstract: Modern deep learning has achieved remarkable results\, but the design of training methodologies largely relies on guess-and-check approaches. T horough empirical studies of recent massive language models (LMs) is prohi bitively expensive\, underscoring the need for theoretical insights\, but classical ML theory struggles to describe modern training paradigms. I pre sent a novel approach to developing prescriptive theoretical results that can directly translate to improved training methodologies for LMs. My rese arch has yielded actionable improvements in model training across the LM d evelopment pipeline — for example\, my theory motivates the design of Me ZO\, a fine-tuning algorithm that reduces memory usage by up to 12x and ha lves the number of GPU-hours required. Throughout the talk\, to underscore the prescriptiveness of my theoretical insights\, I will demonstrate the success of these theory-motivated algorithms on novel empirical settings p ublished after the theory.

https://cml.ics.uci.edu/seminars/2025-03-11- sadhika-malladi URL:https://cml.ics.uci.edu/seminars/2025-03-11-sadhika-malladi END:VEVENT END:VCALENDAR