Select a calendar:
Filter April Events by Event Type:
Events for April 14, 2025
-
PhD Thesis Proposal - Changzhi Xie
Mon, Apr 14, 2025 @ 02:30 PM - 03:30 PM
Thomas Lord Department of Computer Science
University Calendar
Title of Presentation: On the Dynamics of Learning Linear Functinos with Neural Networks
Date and Time: 4.14 2:30-3:30PM
Location: EEB 203
Committee Members: Mahdi Soltanolkotabi(committee chair), Haipeng Luo, Robin Jia, Vatsal Sharan, Adel Javanmard.
Abstract: We study the gradient descent training dynamics of fitting a one-hidden-layer network with multi-dimensional outputs to linear target functions. That is, we focus on a realizable model where the inputs are drawn i.i.d. from a Gaussian distribution and the labels are generated according to a planted linear model with multiple outputs. This framework serves as a good model for a variety of interesting problems including end-to-end training in inverse problems and various auto-encoder models in machine learning. Despite the seemingly simple formulation, understanding training dynamics is a challenging unresolved problem. This is in part due to the fact that the training landscape contains multiple local optima and it is completely unclear why gradient descent from random initialization is able to escape such bad optima. In this work, we develop the first comprehensive analysis of the gradient descent dynamics for learning linear target functions with ReLU networks. We show that gradient descent with moderately small random initialization converges to a global minimizer at a linear rate. To rigorously show that GD avoids local optima, we develop intricate techniques to decompose the loss and control the GD trajectory, which may have broader implications for the analysis of non-convex optimization problems involving local optima. We corroborate our theoretical results with extensive experiments with various configurations.Location: Hughes Aircraft Electrical Engineering Center (EEB) - 203
Audiences: Everyone Is Invited
Contact: Changzhi Xie
This event is open to all eligible individuals. USC Viterbi operates all of its activities consistent with the University's Notice of Non-Discrimination. Eligibility is not determined based on race, sex, ethnicity, sexual orientation, or any other prohibited factor. -
PhD Thesis Proposal - Tejas Srinivasan
Mon, Apr 14, 2025 @ 04:00 PM - 05:00 PM
Thomas Lord Department of Computer Science
University Calendar
Title of Thesis Proposal: Facilitating Reliable Human-AI Collaboration Under Uncertainty
Date and Time: April 14, 2025, 4--5pm
Location: GCS 402C
Committee Members: Jesse Thomason (Chair), Robin Jia, Heather Culbertson, Morteza Dehghani, Diyi Yang
Abstract: AI systems are increasingly assisting humans with decision-making tasks. Effective human-AI collaboration requires AI assistants to be reliable by not only being accurate but also knowing when they don’t know and acting appropriately when uncertain. Popular strategies for handling uncertainty include abstaining from answering, providing prediction sets using conformal prediction, communicating uncertaintyto users, and asking clarification questions to resolve uncertainty. However, these mechanisms do not always facilitate appropriate reliance on and utilization of AI systems by users. In this thesis, we explore methods for proactively mitigating under- and over-reliance in human-AI collaboration under uncertainty. In selective prediction, always abstaining when uncertain can lead to under-utilization by the user, so we develop an algorithm to reduce over-abstention in multimodal selective prediction systems without increasing the error rate of the system’s predictions. When communicating uncertainty, we find that user trust can bias how users rely on AI confidence estimates and lead to inappropriate reliance, which we mitigate by adapting AI assistants’ behavior to user trust levels. Finally, we propose reducing over-reliance on LLM agents by modeling and proactively resolving uncertainty about user goals through frictive dialogue. Our works highlight the importance of modeling uncertainty about AI predictions and the user-AI interaction itself, and the benefits of responding to uncertainty through AI introspection and adaptive AI behaviorsLocation: Ginsburg Hall (GCS) - 402C
Audiences: Everyone Is Invited
Contact: Tejas Srinivasan
This event is open to all eligible individuals. USC Viterbi operates all of its activities consistent with the University's Notice of Non-Discrimination. Eligibility is not determined based on race, sex, ethnicity, sexual orientation, or any other prohibited factor.
