How to Do Spectral Learning at Scale for Science and Engineering
Abstract: Spectral decompositions such as singular value decompositions (SVDs) and eigenvalue decompositions (EVDs) are central tools across a vast swath of scientific computing and machine learning, with abundant engineering applications. Yet many modern methods for learning such decompositions in high dimensions struggle with instability, bias, and poor scalability, even when approximation power is not the limiting factor. I argue that these difficulties are not intrinsic to spectral problems, but instead arise from a shared reliance on Rayleigh-quotient-based constrained optimization, which forces explicit orthogonality handling through penalties, normalization, or whitening.
To address these challenges, I present a reformulation based on unconstrained variational objectives that implicitly encode spectral structure, eliminating the need for orthogonalization and ad-hoc regularization. This perspective leads to a conceptually simpler and scalable parametric framework for learning ordered spectral representations via nested optimization. The resulting framework is well matched to diverse settings in science and engineering. As examples, I demonstrate its effectiveness on eigenvalue problems for linear PDEs such as the Schrödinger equation, spectral (Koopman) analysis of nonlinear dynamical systems such as molecular dynamics, and structured representation learning with deep neural nets. Collectively, these examples illustrate how abandoning Rayleigh-quotient-based formulations resolves long-standing optimization pathologies across domains.
Bio: Jongha (Jon) Ryu is a postdoctoral associate at MIT EECS. He received his Ph.D. in Electrical and Computer Engineering from UC San Diego. His research develops statistical and mathematical foundations for scientific machine learning, with a focus on scalable spectral methods, efficient generative modeling, and reliable uncertainty quantification for scientific and engineering systems.
Location: NCS 120
