Spectral Inference Networks: Unifying Deep and Spectral Learning

  • 2020-01-16 17:02:46
  • David Pfau, Stig Petersen, Ashish Agarwal, David G. T. Barrett, Kimberly L. Stachenfeld
  • 0

Abstract

We present Spectral Inference Networks, a framework for learningeigenfunctions of linear operators by stochastic optimization. SpectralInference Networks generalize Slow Feature Analysis to generic symmetricoperators, and are closely related to Variational Monte Carlo methods fromcomputational physics. As such, they can be a powerful tool for unsupervisedrepresentation learning from video or graph-structured data. We cast trainingSpectral Inference Networks as a bilevel optimization problem, which allows foronline learning of multiple eigenfunctions. We show results of trainingSpectral Inference Networks on problems in quantum mechanics and featurelearning for videos on synthetic datasets. Our results demonstrate thatSpectral Inference Networks accurately recover eigenfunctions of linearoperators and can discover interpretable representations from video in a fullyunsupervised manner.

 

Quick Read (beta)

loading the full paper ...