BEGIN:VCALENDAR
BEGIN:VEVENT
SUMMARY:PhD Thesis Proposal - Hikaru Ibayashi
DESCRIPTION:PhD Thesis Proposal - Hikaru Ibayashi\n
Wed, Dec 8, 2021 @ 02:00 PM - 03:30 PM\n
Committee members: Chair: Prof. Aiichiro Nakano, Prof. Satish Kumar Thittamaranahalli, Prof. Emilio Ferrara, Prof. Yan Liu, Prof. Paulo Branicio (Department of Chemical Engineering and Materials Science)\n
\n
Title:\n
Sharpness analysis of neural-networks for high-performance physics simulation\n
\n
Abstract:\n
In recent years, deep neural networks have witnessed tremendous success in a wide range of fields. Especially, in physics simulations, neural networks have achieved drastically efficient computation by approximating first-principles simulations. However, such practical successes have opened theoretical problems. One open question is why a simple optimization algorithm such as stochastic gradient descent (SGD) can find solutions that generalize well over non-convex loss surfaces. In this proposal, we leverage the second-order information of loss surface, i.e., sharpness, to lay a theoretical foundation of the generalizability of neural networks. First, we use a novel quasi-potential theory to prove that SGD avoids non-generalizing sharp minima. Secondly, we develop a scale-invariant sharpness measure named "minimum sharpness" to theoretically explain why sharp minima are not generalizing. Finally, as a practical application of the thus-developed theoretical framework, we propose a novel sharpness-regularization scheme for robust neural-network-based molecular dynamics simulations. This research will demonstrate the effectiveness of sharpness analysis to deepen the understanding of neural networks and their successful application in physics simulations.\n
\n
Zoom link: https://usc.zoom.us/j/7751892842
DTSTART:20211208T140000
LOCATION:
URL;VALUE=URI:
DTEND:20211208T153000
END:VEVENT
END:VCALENDAR