BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//https://caida.ubc.ca//NONSGML iCalcreator 2.41.92//
CALSCALE:GREGORIAN
METHOD:PUBLISH
UID:65643734-6230-4636-b838-323838633564
X-WR-RELCALID:efc09d74-9c93-479e-a94f-485231ddccde
X-WR-TIMEZONE:America/Vancouver
X-WR-CALNAME:Bridge theory and practice: One-step full gradient could suffi
 ce for low-rank fine-tuning in LLMs\, provably and efficiently - Fanghui L
 iu\, Assistant Professor\, Warwick University
BEGIN:VTIMEZONE
TZID:America/Vancouver
BEGIN:STANDARD
TZNAME:PST
DTSTART:20241103T020000
TZOFFSETFROM:-0700
TZOFFSETTO:-0800
RDATE:20251102T020000
END:STANDARD
BEGIN:STANDARD
TZNAME:MST
DTSTART:20261101T020000
TZOFFSETFROM:-0700
TZOFFSETTO:-0700
END:STANDARD
BEGIN:DAYLIGHT
TZNAME:PDT
DTSTART:20250309T020000
TZOFFSETFROM:-0800
TZOFFSETTO:-0700
RDATE:20260308T020000
END:DAYLIGHT
END:VTIMEZONE
BEGIN:VEVENT
UID:7ca12fa6-d065-4510-bda6-c79d6b0533df
DTSTAMP:20260825T215531Z
CLASS:PUBLIC
CREATED:20250704T230846Z
DESCRIPTION:Abstract: In this talk\, I will talk about how theory can guide
  practice exemplified by low-rank fine-tuning (LoRA) in large language mod
 els. Our theory demonstrates that LoRA naturally aligns with a specific si
 ngular subspace of the one-step gradient from full fine-tuning. This cruci
 al insight motivated our development of a spectral initialization strategy
  that precisely targets and achieves this subspace alignment from the outs
 et. This strategy is not only theoretically optimal for subspace alignment
  but also allows for effective fine-tuning with just a single full gradien
 t step. We'll…
DTSTART;TZID=America/Vancouver:20250721T110000
DTEND;TZID=America/Vancouver:20250721T120000
LAST-MODIFIED:20250704T231259Z
LOCATION:UBC Vancouver Campus\, ICCS X836
SUMMARY:Bridge theory and practice: One-step full gradient could suffice fo
 r low-rank fine-tuning in LLMs\, provably and efficiently - Fanghui Liu\, 
 Assistant Professor\, Warwick University
TRANSP:OPAQUE
URL:https://caida.ubc.ca/index.php/event/bridge-theory-and-practice-one-ste
 p-full-gradient-could-suffice-low-rank-fine-tuning-llms
END:VEVENT
END:VCALENDAR