CIS & MINDS Seminar - Surbhi Goel

ZOOM meeting Link:<br><a href="">... Surbhi Goel, Ph.D.</p><p>       Assistant Professor</p><p>       University of Pennsylvania</p><p>Title: <b>“Beyond Worst-case Guarantees for SequentialPrediction:  Robustness via Abstention”</b></p><p><b>Abstract: </b><span>In this talk, we will focus on the problem ofsequential prediction over a stochastic sequence with an adversary that isallowed to inject clean-label adversarial (or out-of-distribution) examples asand when they desire. Traditional algorithms designed to handle purelystochastic data tend to fail in the presence of such adversarial examples,leading to erroneous predictions. Whereas, assuming fully adversarial dataleads to very pessimistic bounds that are often vacuous in practice. To movebeyond these pessimistic guarantees while allowing for arbitrarily manyadversarial examples, we will propose a new model that allows the learner toabstain from making a prediction at no cost on adversarial examples, therebyasking the learner to make predictions only when it is certain. In this newmodel, we will design learners that can handle any number of adversarialexamples, while ensuring their regret scales as in the purely stochasticsetting. We will conclude with several exciting open questions that our newmodel posits. </span></p><p><i>This talk is based onjoint work with Steve Hanneke, Shay Moran, and Abhishek Shetty.</i></p><p><b>Biography:</b><span> Surbhi Goel is a Magerman Term AssistantProfessor of Computer and Information Science at University ofPennsylvania. Surbhi is associated with the theory group, the ASSETCenter on safe, explainable, and trustworthy AI systems, andthe Warren Center for network and data sciences.</span></p><p><span>Her research interests lie at the intersection oftheoretical computer science and machine learning, with a focus on developingtheoretical foundations for modern machine learning paradigms especially deeplearning.</span><br></p><p><span>Prior to this, Surbhi was a postdoctoral researcher atMicrosoft Research NYC in the Machine Learning group. She obtained her Ph.D. inthe Computer Science department at the University of Texas at Austin advisedby Adam Klivans. Surbhi’s dissertation was awarded UTCS’s Bert KayDissertation award. Her Ph.D. research was generously supported by the JPMorgan AI Fellowship and several fellowships from UT Austin.</span></p><p></p><p> </p>

Tuesday, March 12, 2024 - 16:00 to 17:00

Clark, 110