Soulat, Hugo; (2025) Probabilistic Modeling and Sensory Representations. Doctoral thesis (Ph.D), UCL (University College London).

Text Main.pdf - Accepted Version Access restricted to UCL open access staff until 1 May 2026. Download (59MB)

Abstract

Brains and neuroscientists ironically face a similar challenge: to extract and explain structure from high dimensional, unlabeled and noisy observations. As such, unsupervised probabilistic approaches are instrumental in analyzing increasingly large datasets of neural recordings but also as models of perception and learning. In particular, they are robust to noise and offer principled ways to reason about the structure of the world and include prior knowledge about the observations. I hereby present two such methods developed with collaborators in the course of my doctoral studies. First, I describe a new tensor decomposition that finds a probabilistic low- rank factorization of neural spike-counts arranged in multidimensional arrays. The model is applied to neural recordings taken under conditions of visual-vestibular sensory integration and finds a description of the data which segregates the influence of time dynamics and sensory information available to the animal. Albeit tailored for spike train analysis, it nonetheless relies on strong assumptions regarding the statistical structure of observed data. I then present extensions of new approaches to probabilistic unsupervised learning, based on recognition-parametrized models, that tackle such pitfalls. In this setting, a semi-parametric description of the observation distribution bypasses the need to specify and learn a generative model. Nonlinear feature extraction is therefore solely based on the assumption that patterns of covariation across conditionally independent observations arise from complex dependencies on unobserved latent variables. The model infers low-dimensional descriptions shared across multiple modalities with no restrictive hypotheses on how observations are generated, which proved well-suited for studying neural representation of high level visual cortex. Overall, both methods disentangle the influence of multiple factors on high dimensional observations, account for the stochastic nature of recordings and computations, and show promising applications for neural data analysis.

Download activity - last month

Export as XMLCSVJSON

Download activity - last 12 months

Export as XMLCSVJSON

Downloads by country - last 12 months

Export as JSONCSVXML

Archive Staff Only

View Item