Data-Driven Discovery of Functional Cell Types that Improve Models of Neural Activity

Daniel Zdeblick, Eric Shea-Brown, Daniela Witten, Michael Buice

TL;DR: We developed an algorithm that takes as input recordings of neural activity and returns clusters of neurons by cell type and models of neural activity constrained by these clusters.

Keywords: cell types, GLM, computational neuroscience, neural models

Abstract: Computational neuroscience aims to fit reliable models of in vivo neural activity and interpret them as abstract computations. Recent work has shown that functional diversity of neurons may be limited to that of relatively few cell types; other work has shown that incorporating constraints into artificial neural networks (ANNs) can improve their ability to mimic neural data. Here we develop an algorithm that takes as input recordings of neural activity and returns clusters of neurons by cell type and models of neural activity constrained by these clusters. The resulting models are both more predictive and more interpretable, revealing the contributions of functional cell types to neural computation and ultimately informing the design of future ANNs.