BEGIN:VCALENDAR VERSION:2.0 PRODID:IEEE vTools.Events//EN CALSCALE:GREGORIAN BEGIN:VTIMEZONE TZID:Canada/Pacific BEGIN:DAYLIGHT DTSTART:20210314T030000 TZOFFSETFROM:-0800 TZOFFSETTO:-0700 RRULE:FREQ=YEARLY;BYDAY=2SU;BYMONTH=3 TZNAME:PDT END:DAYLIGHT BEGIN:STANDARD DTSTART:20211107T010000 TZOFFSETFROM:-0700 TZOFFSETTO:-0800 RRULE:FREQ=YEARLY;BYDAY=1SU;BYMONTH=11 TZNAME:PST END:STANDARD END:VTIMEZONE BEGIN:VEVENT DTSTAMP:20210602T064852Z UID:AD1B6F98-CCE0-4726-8F8A-C7EA9A514732 DTSTART;TZID=Canada/Pacific:20210616T160000 DTEND;TZID=Canada/Pacific:20210616T180000 DESCRIPTION:Recent advances in machine learning (ML) systems have made it i ncredibly easier to train ML models given a training set. However\, our un derstanding of the behavior of the model training process has not been imp roving at the same pace. Consequently\, a number of key questions remain: How can we systematically assign importance or value to training data with respect to the utility of the trained models\, may it be accuracy\, fairn ess\, or robustness? How does noise in the training data\, either injected by noisy data acquisition processes or adversarial parties\, have an impa ct on the trained models? How can we find the right data that can be clean ed and labeled to improve the utility of the trained models? Just when we start to understand these important questions for ML models in isolation r ecently\, we now have to face the reality that most real-world ML applicat ions are way more complex than a single ML model.\n\nIn this talk\, Profes sor Ce Zhang will revisit these questions for an end-to-end ML pipeline\, which consists of a noise model for data and a feature extraction pipeline \, followed by the training of an ML model. In the first part of this talk \, I will introduce some recent theoretical results in an abstract way: Ho w to calculate the Shapley value of a training example for ML models train ed over feature extractors\, modeled as a polynomial in the provenance sem iring? How to compute the entropy and expectation of ML models trained ove r data uncertainty\, modeled as a Codd Table? As we will see\, even these problems are #P-hard for general ML models\, though\, surprisingly\, we ca n obtain PTIME algorithms for a simpler proxy model (namely a K-nearest ne ighbor classifier)\, for a large family of polynomials\, input noise distr ibutions\, and utilities.\n\nProfessor Ce Zhang will then put these theore tical results into practice. Given a set of heuristics and a proxy model t o approximate a real- world end-to-end ML pipeline into these abstract pro blems\, I will provide a principled framework for three applications: (1) certifiable defence of backdoor attacks\, (2) targeted data cleaning for M L\, and (3) data valuation and debugging for end-to-end ML pipelines. Prof essor Ce Zhang will describe both our positive empirical results but also those cases that our current approach failed at.\n\nThe video recordings a nd the slides of the previous seminars in this series are now available at the webinar series webpage <http://data.cs.sfu.ca/tdsa.html>.\n\nThe semi nar is open to public for free. Zoom Registration is required for access t o the meeting link. This way we can ensure quality discussions of particip ants from industry and education\n\nVirtual: https://events.vtools.ieee.or g/m/273758 LOCATION:Virtual: https://events.vtools.ieee.org/m/273758 ORGANIZER:Bob_Gill@bcit.ca SEQUENCE:2 SUMMARY:Seminar Series on Trustworthy Data Science and AI: Toward Understan ding End-to-End Learning in the Context of Data by Professor Ce Zhang URL;VALUE=URI:https://events.vtools.ieee.org/m/273758 X-ALT-DESC:Description: <br /><p>Recent advances in machine learning (ML) s ystems have made it incredibly easier to train ML models given a training set. However\, our understanding of the behavior of the model training pro cess has not been improving at the same pace. Consequently\, a number of k ey questions remain: How can we systematically assign importance or value to training data with respect to the utility of the trained models\, may i t be accuracy\, fairness\, or robustness? How does noise in the training d ata\, either injected by noisy data acquisition processes or adversarial p arties\, have an impact on the trained models? How can we find the right d ata that can be cleaned and labeled to improve the utility of the trained models? Just when we start to understand these important questions for ML models in isolation recently\, we now have to face the reality that most r eal-world ML applications are way more complex than a single ML model.</p> \n<p>In this talk\, Professor Ce Zhang will revisit these questions for an end-to-end ML pipeline\, which consists of a noise model for data and a f eature extraction pipeline\, followed by the training of an ML model. In t he first part of this talk\, I will introduce some recent theoretical resu lts in an abstract way: How to calculate the Shapley value of a training e xample for ML models trained over feature extractors\, modeled as a polyno mial in the provenance semiring? How to compute the entropy and expectatio n of ML models trained over data uncertainty\, modeled as a Codd Table? As we will see\, even these problems are #P-hard for general ML models\, tho ugh\, surprisingly\, we can obtain PTIME algorithms for a simpler proxy mo del (namely a K-nearest neighbor classifier)\, for a large family of polyn omials\, input noise distributions\, and utilities.</p>\n<p>Professor Ce Z hang will then put these theoretical results into practice. Given a set of heuristics and a proxy model to approximate a real- world end-to-end ML p ipeline into these abstract problems\, I will provide a principled framewo rk for three applications: (1) certifiable defence of backdoor attacks\, ( 2) targeted data cleaning for ML\, and (3) data valuation and debugging fo r end-to-end ML pipelines. Professor Ce Zhang will describe both our posit ive empirical results but also those cases that our current approach faile d at.</p>\n<p>The video recordings and the slides of the previous seminars in this series are now available at the webinar series webpage &lt\;<a hr ef="http://data.cs.sfu.ca/tdsa.html">http://data.cs.sfu.ca/tdsa.html</a>&g t\;.</p>\n<div class="dc-content">\n<div class="dc-modules">\n<div class=" dc-modules__item">\n<div class="eds-l-mar-bot-12 eds-l-lg-mar-bot-14">\n<d iv class="dc-modules__item--text">\n<p><strong>The seminar is open to publ ic for free.&nbsp\;Zoom Registration is required for access to the meeting link.</strong>&nbsp\;This way we can ensure quality discussions of partic ipants from industry and education</p>\n</div>\n</div>\n</div>\n</div>\n</ div> END:VEVENT END:VCALENDAR