BEGIN:VCALENDAR VERSION:2.0 PRODID:IEEE vTools.Events//EN CALSCALE:GREGORIAN BEGIN:VTIMEZONE TZID:America/Los_Angeles BEGIN:DAYLIGHT DTSTART:20240310T030000 TZOFFSETFROM:-0800 TZOFFSETTO:-0700 RRULE:FREQ=YEARLY;BYDAY=2SU;BYMONTH=3 TZNAME:PDT END:DAYLIGHT BEGIN:STANDARD DTSTART:20231105T010000 TZOFFSETFROM:-0700 TZOFFSETTO:-0800 RRULE:FREQ=YEARLY;BYDAY=1SU;BYMONTH=11 TZNAME:PST END:STANDARD END:VTIMEZONE BEGIN:VEVENT DTSTAMP:20240315T160126Z UID:3070351B-6F78-4BB6-A0BB-FB25DC95C0E0 DTSTART;TZID=America/Los_Angeles:20240308T110000 DTEND;TZID=America/Los_Angeles:20240308T120000 DESCRIPTION:Despite the success of data driven approaches in transfer learn ing\, the theoretical understanding of transfer learning from statistics o r information theory is somewhat behind. A key challenge when conducting t heoretical analyses for transfer learning problems is that the statistical correlation between source and target tasks are often unknown\, and hence traditional mathematical tools in statistics and estimation theory are di fficult to be applied. In this talk\, we address this issue by formulating the transfer learning as an optimization problem for the testing loss ove r certain similarity constraints of the source and target tasks. Specifica lly\, we first show that when the similarity between both tasks measured b y a certain distance metric is given\, the optimal linear transfer model c an be computed. The optimal coefficient in such a model reveals how the sa mple complexity\, model complexity\, and the task similarity affect the kn owledge transferring in transfer learning. Moreover\, when the task simila rity cannot be well estimated due to insufficient samples\, we propose a m inimax formulation\, which only requires the similarity being bounded\, an d the resulting distribution estimator is robust against sample insufficie ncy. We show an approximately optimal distribution estimator for the minim ax problem from the bounded normal mean problem\, and develop similar know ledge transferring insights as in the linear transferring model. Finally\, some experimental results validate the algorithms led by our theoretical approach.\n\nSpeaker(s): Shao-Lun Huang\, \n\nRoom: 4127\, Bldg: Earth Sci ence Building\, 2207 Main Mall\, Vancouver\, British Columbia\, Canada\, V 6T1Z4 LOCATION:Room: 4127\, Bldg: Earth Science Building\, 2207 Main Mall\, Vanco uver\, British Columbia\, Canada\, V6T1Z4 ORGANIZER:lelewang@ece.ubc.ca SEQUENCE:3 SUMMARY:Some Guidance of Transfer Learning Algorithm Designs From Statistic al Inference Perspectives - IEEE DLT\, 8 Mar 2024 11:00 AM URL;VALUE=URI:https://events.vtools.ieee.org/m/409942 X-ALT-DESC:Description: <br /><p>Despite the success of data driven approac hes in transfer learning\, the theoretical understanding of transfer learn ing from statistics or information theory is somewhat behind. A key challe nge when conducting theoretical analyses for transfer learning problems is that the statistical correlation between source and target tasks are ofte n unknown\, and hence traditional mathematical tools in statistics and est imation theory are difficult to be applied. In this talk\, we address this issue by formulating the transfer learning as an optimization problem for the testing loss over certain similarity constraints of the source and ta rget tasks. Specifically\, we first show that when the similarity between both tasks measured by a certain distance metric is given\, the optimal li near transfer model can be computed. The optimal coefficient in such a mod el reveals how the sample complexity\, model complexity\, and the task sim ilarity affect the knowledge transferring in transfer learning. Moreover\, when the task similarity cannot be well estimated due to insufficient sam ples\, we propose a minimax formulation\, which only requires the similari ty being bounded\, and the resulting distribution estimator is robust agai nst sample insufficiency. We show an approximately optimal distribution es timator for the minimax problem from the bounded normal mean problem\, and develop similar knowledge transferring insights as in the linear transfer ring model. Finally\, some experimental results validate the algorithms le d by our theoretical approach.</p> END:VEVENT END:VCALENDAR