3rd Year Candidacy Requirement: Analyzing Neural Networks via Low-Rank Approximations for Real-Time Tsunami Prediction

Abstract: Tsunamis present formidable natural hazards, requiring efficient early warning to mitigate their devastating impact. Deep learning models offer real-time predictions, reducing the time needed for early warning to a mere fraction of a second when compared to the numerical models that simulate the inundation. However, instabilities in neural network models can result in inaccurate results that can have significant consequences. This talk will focus on a novel approach to analyzing neural networks using low-rank approximations, with a focus on perturbative instabilities called adversarial examples. Our analysis reveals that a local basis near the input, encompassing the most severe instabilities, can be identified, and filtering the components in this basis direction effectively stabilizes the neural network. We will discuss future research directions that can potentially provide stabilization techniques that are applicable for tsunami prediction.

Advisor: Donsub Rim