Bohórquez Giraldo, M. Felipe (2024) Modified Restricted Boltzmann Machines and their function in cortical recordings PFE - Project Graduation, ENSTA.

This is the latest version of this item.

[img]
Preview
PDF
6Mb

Abstract

In neuroscience, the continuous generation of new, larger, and more complex datasets presents significant challenges in extracting meaningful information. While several complex machine learning algorithms are used and developed to analyze neural data, they often present the ‘black box’ phenomena and the entanglement of data representations within them. Therefore, we implemented Restricted Boltzmann Machines (RBMs), one of the simplest representation-based generative models, and provided a comprehensive progression from classical RBMs to Conditional RBMs and Modified RBMs. Used on calcium imaging data from the primary visual cortex (V1) of mice and electrophysiological recordings from the primary and secondary visual cortices (V1 and V2) of non-human primates, RBMs correctly learn data structures and high-order dependencies, and Conditional and Modified RBMs improved interpretability and the later disentangled representations. We modeled data from orientation selectivity experiments and enhanced explainability by conditioning hidden units on specific stimuli to allow the identification of a particular label preference or by applying constraints that concentrated label-related information into designated ones. Hidden units’ receptive fields obtained from Conditional RBMs models trained on all datasets presented preference towards the selected stimuli, same for the Modified RBMs, plus the isolation of label-related information at a reasonable compromise of model performance. RBMs can successfully learn and capture the distribution and high-order correlations inherent in neural datasets, despite their simplicity, and Conditional and Modified RBMs offer explainability in contrast to the black box problem of more complex machine learning models and data disentanglement.

Item Type:Thesis (PFE - Project Graduation)
Uncontrolled Keywords:Restricted Boltzmann Machines
Subjects:Information and Communication Sciences and Technologies
Life Sciences and Engineering
ID Code:10485
Deposited By:Felipe BOHÓRQUEZ GIRALDO
Deposited On:28 nov. 2024 14:22
Dernière modification:28 nov. 2024 14:22

Available Versions of this Item

Repository Staff Only: item control page