Representational geometry: evaluating brain-computational models with neural activity data

University Of Texas At Austin, Austin TX

Investigators

Abstract

Project Summary Computational neuroscience has entered a new era of greatly increased capability to record neuronal activity and to construct large-scale computational models. The challenge before us is to connect theory and experiment by testing and comparing complex brain-computational models with massive measurements of neural activity. Con- necting experimental results to computational models is critical for deepening our understanding of the neural circuit computations that collectively give rise to perception, cognition, and behavior, basic brain functions with fundamental clinical implications. This project would launch a close collaboration between theorists and experi- mentalists to develop a rigorous and comprehensive methodology for evaluation and comparison of our new big models with our new big data. A central concept that has gained momentum over the past two decades is the concept of representational geometry. The representational geometry is the geometry of the points or trajectories in the multivariate neural population response space that are thought to represent the contents of brain compu- tations. Although a variety of estimators of representational distance have been proposed previously, none of them are well-suited for neural activity data. Adequate estimation of the representational geometry from neural response data requires carefully accounting for the marginal Poisson distribution of the noise, noise correlations, the context of the neural manifold, and biases caused by electrode sampling as well as by the sensitivity of es- timators to noise displacements. We will address key conceptual and technical challenges and conduct the first large-scale, objective evaluation of a broad range of options for evaluating representational models. Aim 1 is to develop a general-purpose methodology for testing theories implemented in neural network models with modern neurophysiological data. The methodology evaluates and inferentially compares models on the basis of their pre- dictions of neural representational geometries, and encompasses encoding models and representational similarity analysis, methods that allow different levels of flexibility in fitting model representations to neural data. Aim 2 is to validate the methodology developed in Aim 1 in the context of neuroscientific applications, where computational theories are evaluated with neurophysiological data. On the one hand, we will use model-based simulations and sub-sampling of neural recording data sets to validate the inferential methodology in scenarios where ground truth is known. On the other hand, we will seek empirical answers to exciting theoretical questions in the domains of vision and spatial navigation. Overall, the project will build a comprehensive methodology for linking theory to experiment in computational neuroscience in the new era of neural network models and modern neuroscience methods. The methodology involves formal inferential comparisons of brain-computational models that implement alternative theories. The project also launches the ongoing open-source development of this critical methodology through a series of workshops for theorists and experimentalists.