Claire Sturgill: Investigating the Latent Dynamics of Neuronal Populations in Rat Orbitofrontal Cortex During Perceptual Decision Making

BCCN Berlin / Technische Universität Berlin

Abstract

Technological advances have made it feasible to record from large numbers of neurons at once, allowing researchers to investigate both complex interactions between neurons and population-level encoding. However, larger neural datasets have a higher dimensionality, which makes statistical analysis and interpretability more difficult. To reduce population activity to a manageable size, various dimensionality reduction methods are commonly used, each with limitations and assumptions about what information should be preserved in a low-dimensional subspace. This thesis uses electrophysiological data from the orbitofrontal cortex of rats performing a decision-making task to compare two dimensionality reduction techniques: Principal Component Analysis (PCA) and CEBRA, a contrastive learning-based technique.

As a linear method, PCA was found to produce much more interpretable embeddings, particularly when given the average neural activity across many trials as input. PCA trajectories were shown to reflect both trial events and behavioral differences between trial conditions. CEBRA, on the other hand, created idealized embeddings that maximized the separation between trajectories throughout the trial, but in doing so failed to reflect the time-specific nature of the underlying activity differences. However, the increased separation between trial conditions in CEBRA embeddings was shown to have advantages for effectively representing individual trials and for decoding behavioral labels. Based on these peculiarities, the advantages of each method are discussed and ideal use cases for each are recommended.

 

 

Additional information:

Master thesis defense

 

Organized by:

Dr. Torben Ott & Prof. Dr. Klaus Obermayer