Evaluating geometric similarity between neural activity and behavior using multi-level representational similarity analysis

Abstract

Population-level neural activity across time can be represented as a trajectory on a neural manifold. As neural circuits have evolved to be efficient, we hypothesize that similarity in the neural population activity (neural trajectories) across conditions/trials implies similarity in the behavior executed. To study this, we present an unsupervised approach to determine the relation between the geometry of neural activity and behavior. Specifically, we introduce a multi-level Representation Similarity Analysis (RSA) metric and present results on neural & behavioral data obtained from non-human primates performing hand-reach tasks in 108 different maze settings [Churchland et al. 2012]. In the neural domain, we apply hierarchical clustering using pairwise distances between neural trajectories and organize the corresponding behaviors based on the resulting dendrogram. At every node in the dendrogram, we calculate the similarity between the neural activity and behavior distance matrices using RSA, considering only the conditions/trials (root nodes) that are clustered under that node. We used data comprising simultaneous recording of neural activity (137 neurons in primary motor cortex and premotor cortex) and the corresponding behavior (hand velocity and position) obtained across 2295 trials. We performed this analysis on both trial-averaged data and single-trial data. Single-trial analysis allowed the introduction of complexity in the data due to inherent variability in behavior and neural activity. The results obtained remained consistent across both single-trial and trial-averaged analysis. Our approach identified that similar neural activity corresponds to hand reaches that are similar to each other and also grouped reaches made to similar targets together. The RSA between the neural activity and behavior decreased as we traversed higher and higher nodes in the dendrogram. At the lower nodes the RSA values were closer to 1 in both single-trial and trial-averaged analysis, and at the root node, the RSA values were 0.52 and 0.68 respectively. We also noted that the RSA between the neural activity and hand velocity was consistently higher than that between neural activity and hand position. This confirms with the observation that motor cortical activity correlates better to velocity than position during arm movements. Our approach provides a framework for analyzing and quantifying the relationship between the geometry of neural activity and behavior using multi-level RSA, and a principled approach to evaluate neural latent representations and identify ones that best captures the similarity between the neural activity and behavior.