Uncovering Neural Encoding Variability with Infinite Gaussian Process Factor Analysis

Changmin Yu, Máté Lengyel

Keywords: Neural variability, GPFA, Bayesian nonparametrics.

TL;DR: We propose infinite GPFA, a fully Bayesian nonparametric extension of GPFA model, for investigating the nature of neural variability from a novel perspective.

Abstract:

Gaussian Process Factor Analysis (GPFA) is a powerful factor analysis model
for extracting low-dimensional latent processes underlying population neural
activities. However, one limitation of standard GPFA models is that the number
of latent factors needs to be pre-specified or selected through
heuristic-based approaches. We propose the infinite GPFA model, a Bayesian
non-parametric extension of the classical GPFA model by incorporating an
Indian Buffet Process (IBP) prior, such that we are able to infer the
potentially infinite set of likely latent factors active at each time points,
in a probabilistically principled manner. Learning and inference in the
infinite GPFA model is performed through variational expectation-maximisation,
and we additionally propose a scalable extension based on sparse variational
Gaussian Process methods. We empirically demonstrate that the infinite GPFA
model correctly infers dynamically changing activations of latent factors on
synthetic dataset. Through fitting the infinite GPFA model to simultaneously
recorded population neural activities, we identify non-trivial and behavioural
meaningful variability in neural encoding process, and addressing an important
gap in existing interpretations of the nature of neural variability.

Submission Number: 24