Fri, Dec 10, 2021 @ 03:00 PM - 05:00 PM
Thomas Lord Department of Computer Science
PhD Thesis Proposal - SÃ©b Arnold
Friday, Dec 10, 2021 @ 03:00 PM - 05:00 PM
Committee members: Chair: Prof. Maja Mataric, Prof. Fei Sha, Prof. Yan Liu, Prof. Stefanos Nikolaidis, Prof. Jesse Thomason, Prof. Salman Avestimehr (ECE)
Quickly solving new tasks, with meta-learning and without.
This thesis proposal seeks to answer how learning systems can reuse and adapt their knowledge to quickly solve new test tasks. We first show how to improve the test task performance of meta-learning algorithms (eg, MAML) by carefully choosing which tasks to train on -- even when these test tasks are unknown a priori. We then zero in on these algorithms and uncover modeling pitfalls that completely prevent fast adaptation; fortunately, there exist simple remedies. Leveraging those insights, we conclude with the challenge of quickly solving new tasks using off-the-shelf models, which were trained without meta-learning.
Zoom link: https://usc.zoom.us/j/94965325337
WebCast Link: https://usc.zoom.us/j/94965325337
Audiences: Everyone Is Invited
Contact: Lizsl De Leon