Learning Discrete Graphical Models: Exact & Neural Network Assisted Methods

Friday, September 18, 2020 - 10:40am - 11:25am
Marc Vuffray (Los Alamos National Laboratory)
Undirected graphical models are widely used in science to represent structured high-dimensional
joint probabilities. We focus our attention on the inverse problem of learning a discrete graphical
model given i.i.d. samples from its distribution. In the first part of the presentation, we show that the
inverse problem can be solved exactly and efficiently for common model parameterizations using a
method known as Generalized Regularized Interaction Screening Estimator (GRISE). In the second
part of the presentation, we address the situation when a good guess of a sparse or more generally
structured parameterization is lacking. We show how to construct a variation of the method GRISE
which integrates organically neural networks into the learning process. This resulting method, called
NN-GRISE, is especially efficient for models composed of multiple high-order interactions or with a
high degree of symmetry.