Beta Distribution of Human MTL Neuron Sparsity: A Sparse and Skewed Code

TL;DR

This paper introduces a beta distribution model for analyzing neural sparsity in human MTL neurons, revealing a highly skewed, sparse coding pattern similar to rat hippocampus activity.

Contribution

It develops a maximum likelihood statistical model to infer the distribution of neural response probabilities, demonstrating a skewed, sparse coding pattern in human MTL neurons.

Findings

Most neurons have extremely low response probabilities.

Distribution closely follows a power law at low sparsity.

Similar skewed distributions are observed in rat hippocampus.

Abstract

Single unit recordings in the human medial temporal lobe (MTL) have revealed a population of cells with conceptually based, highly selective activity, indicating the presence of a sparse neural code. Building off previous work by the author and J.C. Collins, this paper develops a statistical model for analyzing this data, based on maximum likelihood analysis. The goal is to infer the underlying distribution of neural response probabilities across the population of MTL cells. The response probability, or neuronal sparsity, is defined as the total probability that the neuron produces an above-threshold firing rate during the presentation of a randomly selected stimulus. Applying the method, it is shown that a beta-distributed neuronal sparsity across the cells of the MTL is consistent with the data. The resulting fits reveal a sparse and highly skewed code, with a huge majority of neurons exhibiting extremely low response probabilities, and a smaller minority possessing considerably higher response probabilities. The distributions are closely approximated by a power law at low sparsity values. Strikingly similar skewed distributions have been found in the statistics of place cell activity in rats, suggesting similar underlying coding dynamics between the human MTL and the rat hippocampus.