Dissertation Talk: Towards Trustworthy Machine Learning

Seminar | November 23 | 11 a.m.-12 p.m. | WeWork Berkeley, Meeting Room A, 4th Floor | Note change in location

Location:  2120 University Avenue, Berkeley, CA 94709

Speaker:  Adam Gleave, PhD Candidate, UC Berkeley

Sponsor:  Electrical Engineering and Computer Sciences (EECS)

Machine learning has made remarkable progress towards building automated systems that achieve high average-case performance on procedurally specified objectives. However, real-world tasks often have complex objectives that are difficult to specify procedurally, and safety-critical tasks often demand worst-case guarantees. In this talk, I will first discuss how agent objectives can be inferred from human feedback, with a focus on how to test and validate the learned objective. I will then introduce methods for adversarially testing agents, concluding with methods to make agents more robust.

This is a hybrid event and may be also be attended via:

Zoom Location: https://berkeley.zoom.us/j/97650504829?pwd=QkhSUXg3UURiRytId3VFT1VtNHFuQT09
Zoom Meeting ID: 976 5050 4829
Zoom Passcode: 879671

Event Contact:  gleave@berkeley.edu

Access Coordinator:  Roxana Infante,  roxana@eecs.berkeley.edu,  510-643-3257