Computational Tools for assessing mechanisms and functional relevance of divisive normalization

Project Summary Divisive normalization (DN) is a well-established theory of how interactions between neurons in a circuit modulate the activity of individual neurons. DN has been termed a canonical operation because it describes a wide range of empirical data across species and brain areas, and theory predicts that DN underlies behavioral gains of sensory integration and visual attention. Despite this progress, it has been difficult to tie DN to circuit and cellular mechanisms, and to quantify its impact on neural coding and behavior. Two main obstacles have limited progress. First, most models of DN have focused exclusively on trial-averaged, single-cell responses, neglecting other statistical properties of neural activity that determine the information encoded by neural populations, and thus perception. Second, single-trial normalization signals cannot be measured directly in experiments, limiting our ability to perturb and dissect mechanisms of DN and to test competing hypotheses about its functional role. This project aims to develop analytic and computational tools to estimate normalization signals in single neurons and populations from measured spiking activity with single-trial resolution, and to quantify DN’s influence on neural coding. To facilitate broad adoption by the research community, we will release software toolboxes in Matlab and Python including tutorials and example applications to data. In Aim 1 we will develop a statistical modeling framework that relates single-neuron response variability to across-trial fluctuations of normalization strength. We will assess the data requirements of the computational tools, and validate their usability and generality, by applying them to responses in primary visual cortex (V1) of mice and macaque monkeys, recorded with electrophysiology and calcium imaging. We will then apply our tools to simultaneous recordings from multiple excitatory and inhibitory cell types, to demonstrate how the tools allow the quantification of the role of those circuit elements in DN. In Aim 2, because the information encoded by neural populations is determined by shared variability between pairs of neurons, we will extend our framework to capture the effects of DN on pairwise responses, and validate it on V1 data. We will then demonstrate how these tools can be used to elucidate the relation between DN and information in visual adaptation and crowding in V1. In Aim 3 we will extend the framework to large populations and integrate it with latent dynamical system models, to estimate within-trial fluctuations of normalization signals shared between neurons, and quantify their influence on the dimensionality of population activity. We will validate these tools on simultaneous recordings from up to thousands of neurons. We will then use multiphoton holographic optogenetics in closed loop with our tools to inject population activity patterns that drive normalization signals precisely. Successful completion of this project will vastly expand the scope of the theory of DN, from single-neuron average activity to dynamic fluctuations in neurons and populations, enabling new studies of the mechanisms of normalization and its impact on neural coding.