Thesis Proposal - Runtian Zhai August 12, 2024 4:00pm — 5:30pm Location: In Person and Virtual - ET - Traffic21 Classroom, Gates Hillm 6501 and Zoom Speaker: RUNTIAN ZHAI, Ph.D. Student, Computer Science Department, Carnegie Mellon University https://www.runtianzhai.com/ Machine learning has shifted to a new paradigm driven by representation learning and foundation models, big encoders that extract useful features from data. This thesis studies how big models learn good representations, and especially focuses on two aspects: generalization and transferability. For generalization, the problem is why big foundation models wouldn’t overfit as classical theory suggests. My work proves a generalization bound that works for big models, by viewing them as algorithmic models instead of data models. For transferability, my work focuses on reweighting, the most popular class of methods. This talk will focus on one issue with reweighting that is its sensitivity to outliers, and propose a solution that significantly improves the performance and stability of reweighting. Finally, I will propose two future work, feature learning for tabular data, and combining multiple sources of prior knowledge. Thesis Committee: Pradeep Ravikumar (Co-chair) Zico Kolter (Co-chair) Andrej Risteski Yuandong Tian (Meta AI) Additional Information In Person and Zoom Participation. See announcement. Event Website: https://csd.cmu.edu/calendar/thesis-proposal-runtian-zhai Add event to Google Add event to iCal
