Bayesian Spike Train Inference via Non-Local Priors

TL;DR

This paper introduces a Bayesian method using non-local priors and stochastic search for more accurate and computationally efficient spike train inference from neural fluorescence data, providing uncertainty quantification.

Contribution

It proposes a novel Bayesian approach with non-local priors and a stochastic search algorithm that improves inference accuracy and efficiency over existing methods.

Findings

Outperforms L1 regularization-based methods in simulations

Achieves comparable accuracy to L0-based methods

Provides automatic uncertainty quantification

Abstract

Advances in neuroscience have enabled researchers to measure the activities of large numbers of neurons simultaneously in behaving animals. We have access to the fluorescence of each of the neurons which provides a first-order approximation of the neural activity over time. Determining the exact spike of a neuron from this fluorescence trace constitutes an active area of research within the field of computational neuroscience. We propose a novel Bayesian approach based on a mixture of half-non-local prior densities and point masses for this task. Instead of a computationally expensive MCMC algorithm, we adopt a stochastic search-based approach that is capable of taking advantage of modern computing environments often equipped with multiple processors, to explore all possible arrangements of spikes and lack thereof in an observed spike train. It then reports the highest posterior probability arrangement of spikes and posterior probability for a spike at each location of the spike train. Our proposals lead to substantial improvements over existing proposals based on L1 regularization, and enjoy comparable estimation accuracy to the state-of-the-art L0 proposal, in simulations, and on recent calcium imaging data sets. Notably, contrary to optimization-based frequentist approaches, our methodology yields automatic uncertainty quantification associated with the spike-train inference.