Spectral Inference Networks: Unifying Deep and Spectral Learning

• DOI: 10.48550/arxiv.1806.02215
• Publication Date: 2018-01-01
• Authors: David Pfau; Stig Petersen; Ashish Agarwal; David G. T. Barrett; Kimberly Stachenfeld
• Keywords: Spectral Analysis

We present Spectral Inference Networks, a framework for learning eigenfunctions of linear operators by stochastic optimization. Networks generalize Slow Feature Analysis to generic symmetric operators, and are closely related Variational Monte Carlo methods from computational physics. As such, they can be powerful tool unsupervised representation video or graph-structured data. cast training as bilevel optimization problem, which allows online multiple eigenfunctions. show results on problems in quantum mechanics feature videos synthetic datasets. Our demonstrate that accurately recover discover interpretable representations fully manner.