BEGIN:VCALENDAR VERSION:2.0 PRODID:IEEE vTools.Events//EN CALSCALE:GREGORIAN BEGIN:VTIMEZONE TZID:America/New_York BEGIN:DAYLIGHT DTSTART:20250309T030000 TZOFFSETFROM:-0500 TZOFFSETTO:-0400 RRULE:FREQ=YEARLY;BYDAY=2SU;BYMONTH=3 TZNAME:EDT END:DAYLIGHT BEGIN:STANDARD DTSTART:20241103T010000 TZOFFSETFROM:-0400 TZOFFSETTO:-0500 RRULE:FREQ=YEARLY;BYDAY=1SU;BYMONTH=11 TZNAME:EST END:STANDARD END:VTIMEZONE BEGIN:VEVENT DTSTAMP:20241222T013014Z UID:3517F639-6080-44A5-A815-C85D8DC44C9B DTSTART;TZID=America/New_York:20241218T100000 DTEND;TZID=America/New_York:20241218T120000 DESCRIPTION:Speaker: Arash Behboodi\, Director of Engineering at Qualcomm A I Research\n\nDate and Time: Wednesday\, December 18\, 2024\, 10-11am\n\nL ocation: BA-4164\n\nAbstract: Conformal Prediction (CP) is a distribution- free uncertainty estimation framework that constructs prediction sets guar anteed to contain the true answer with a user-specified probability. Intui tively\, the size of the prediction set encodes a general notion of uncert ainty\, with larger sets associated with higher degrees of uncertainty. In this work\, we leverage information theory to connect conformal predictio n to other notions of uncertainty. More precisely\, we prove three differe nt ways to upper bound the intrinsic uncertainty\, as described by the con ditional entropy of the target variable given the inputs\, by combining CP with information theoretical inequalities. Moreover\, we demonstrate two direct and useful applications of such connection between conformal predic tion and information theory: (i) more principled and effective conformal t raining objectives that generalize previous approaches and enable end-to-e nd training of machine learning models from scratch\, and (ii) a natural m echanism to incorporate side information into conformal prediction. We emp irically validate both applications in centralized and federated learning settings\, showing our theoretical results translate to lower inefficiency (average prediction set size) for popular CP methods.\n\nBio: Arash Behbo odi is a machine learning research scientist and Director of Engineering a t Qualcomm AI Research. He received the Ph.D. degree in information theory from Ecole Superieure d'Electricite (now CentraleSuplec)\, France\, in 20 12\, and a master’s degree in philosophy from Pantheon-Sorbonne universi ty\, 2011. Prior to Qualcomm\, Arash was a senior researcher at Institute for Theoretical Information Technology in RWTH Aachen University and TU Be rlin. He has been doing research on information\, machine learning and sig nal processing theory\, and recently focusing in particular on wireless AI \, inverse problems\, differentiable simulations\, and geometric deep lear ning. He has been a recipient of multiple best paper awards\, and organize d multiple workshops on machine learning and other related topics.\n\nCo-s ponsored by: Prof. Ashish Khisti \n\nSpeaker(s): Arash Behboodi\n\nRoom: BA-4164\, Bldg: BA-4164\, University of Toronto\, Toronto\, Ontario\, Cana da LOCATION:Room: BA-4164\, Bldg: BA-4164\, University of Toronto\, Toronto\, Ontario\, Canada ORGANIZER:peng.dai.ca@ieee.org SEQUENCE:11 SUMMARY:An Information Theoretic approach to Conformal Prediction URL;VALUE=URI:https://events.vtools.ieee.org/m/452043 X-ALT-DESC:Description: <br /><p><strong>Speaker</strong>: Arash Behboodi\, Director of Engineering at Qualcomm AI Research</p>\n<p><strong>Date and Time</strong>: Wednesday\, December 18\, 2024\, 10-11am</p>\n<p><strong>Lo cation</strong>: BA-4164</p>\n<p><strong>Abstract</strong>: Conformal Pred iction (CP) is a distribution-free uncertainty estimation framework that c onstructs prediction sets guaranteed to contain the true answer with a use r-specified probability. Intuitively\, the size of the prediction set enco des a general notion of uncertainty\, with larger sets associated with hig her degrees of uncertainty. In this work\, we leverage information theory to connect conformal prediction to other notions of uncertainty. More prec isely\, we prove three different ways to upper bound the intrinsic uncerta inty\, as described by the conditional entropy of the target variable give n the inputs\, by combining CP with information theoretical inequalities. Moreover\, we demonstrate two direct and useful applications of such conne ction between conformal prediction and information theory: (i) more princi pled and effective conformal training objectives that generalize previous approaches and enable end-to-end training of machine learning models from scratch\, and (ii) a natural mechanism to incorporate side information int o conformal prediction. We empirically validate both applications in centr alized and federated learning settings\, showing our theoretical results t ranslate to lower inefficiency (average prediction set size) for popular C P methods.</p>\n<p><strong>Bio</strong>: Arash Behboodi is a machine learn ing research scientist and Director of Engineering at Qualcomm AI Research . He received the Ph.D. degree in information theory from Ecole Superieure d'Electricite (now CentraleSuplec)\, France\, in 2012\, and a master&rsqu o\;s degree in philosophy from Pantheon-Sorbonne university\, 2011. Prior to Qualcomm\, Arash was a senior researcher at Institute for Theoretical I nformation Technology in RWTH Aachen University and TU Berlin. He has been doing research on information\, machine learning and signal processing th eory\, and recently focusing in particular on wireless AI\, inverse proble ms\, differentiable simulations\, and geometric deep learning. &nbsp\;He h as been a recipient of multiple best paper awards\, and &nbsp\;organized m ultiple workshops on machine learning and other related topics.</p> END:VEVENT END:VCALENDAR