Neural response variability in the navigational system of the brain

Electrophysiological population recordings contain high dimensional spike trains that are generally not interpretable to the naked eye. To progress in making sense of neural data, a fundamental statistic is the instantaneous firing rate of a neuron, which has helped reach groundbreaking insights into how the brain processes and represents information. Many statistical methods have been developed to estimate this quantity, leading to powerful concepts such as tuning curves and population dynamics. However, neural recordings show highly variable activity that is structured beyond the rate-based perspective, even in repeated trials using identical experimental conditions. This phenomenon has been quantified rigorously using trial-structured studies, in particular for sensory brain areas. However, many brain areas are involved in encoding naturalistic behaviour that contains no inherent trial structure, in particular the areas that are part of the navigational system of the brain. Traditional statistical methods designed for the trial-structured setting are not applicable to naturalistic settings, and as a consequence there have been way fewer attempts to characterize neural variability in this context. To remedy the situation, we develop two novel statistical methods for analyzing neural responses that allow analysis of variability under general experimental settings. We name our approach for spike counts the Universal Count Model (UCM), and for spike trains the Nonparametric Non-renewal (NPNR) process. By leveraging Bayesian machine learning, we obtain a data-efficient yet flexible model for neural activity that can flexibly model neural response statistics beyond the conventional rate-based Poisson framework. We apply the UCM and NPNR process to various navigation datasets, revealing rich variability patterns that are modulated by various covariates. Firing rates and variability measures appear to be decoupled in many cases, suggesting they may act as separate information channels. We also analyze hippocampal awake replay data, and find a small but significant reduction in count variability during replay episodes compared to locomotion periods. However, analyzing information-theoretic encoding and population decoding reveals that modeling the rich non-Poisson aspect generally does not yield significant improvement in information extracted. To understand the origins of variability patterns observed \emph{in vivo}, we train spiking neural networks to perform a simple vestibular integration task. Such networks develop neural representations reminiscent of the head direction circuit, even when deterministic, and exhibit variability patterns that are similar when quantified with the UCM. These results suggest that rich variability patterns may simply arise as a byproduct of underlying circuit dynamics, rather than being a critical component of neural coding.

Name dc:type.qualificationname

Doctor of Philosophy (PhD)

Level dc:type.qualificationlevel

Doctoral

Grantor dc:publisher.institution

University of Cambridge

Year dc:date.issued

2024

Author dc:creator

• Liu, David

Advisor dc:contributor.advisor

• Lengyel, Máté

• Gaussian processes
• machine learning
• neural data analysis
• neural variability
• neuroscience

Licence

• https://apollo8-f-pro.lib.cam.ac.uk/bitstreams/c24e9eee-95b6-413f-b430-34321489d24a/download
• https://creativecommons.org/licenses/by/4.0/

Language dc:language

eng

DOI dc:identifier.doi

https://doi.org/10.17863/CAM.116246

OAI identifier oai:identifier

oai:www.repository.cam.ac.uk:1810/380756