A predictive, cell type-level model of primary visual cortex layer 2/3
Systems neuroscientists tend to believe that the richness of computation in the brain emerges primarily from the structure of its synaptic connections. Neural codes, then, can be understood as the solutions to a dynamical system, in which synapses couple the electrical activity of one cell type to another. By individually imaging genetically defined cell types during sensory stimulation, we seek to infer cell type-level local recurrent connectivity from patterns of activity.
We develop a data-driven recurrent network model of the four main neuronal cell types of layer 2/3, that explains their diverse tuning for visual textures of varying size and contrast. Comparing the effects of simulated and experimental optogenetic perturbations validates key predictions of the model. Along the way, we uncover how competition and cooperation of neural activity across cortical space emerges from a well-known disinhibitory circuit motif, and how network superlinearity shapes optogenetic effects.
A superlinear code for higher order stimulus features in the primary somatosensory and visual cortices
In collaboration with Evan Lyall. To help animals navigate the world, sensory systems associate behavioral and cognitive significance to arbitrary conjunctions of complex sensory stimuli. To understand how representations of arbitrary subsets of complex stimuli might be made linearly discriminable, we measured activity across two cortical layers and primary sensory cortical areas, during sensory stimulation spanning all possible subsets of five neighboring patches of space. We find that across layers and areas, neurons superlinearly encode diverse and specific patterns of sensory stimulation across space. As a population, neurons smoothly tile the full space of possible stimulus patterns. From such a code, a single linear classifier could be trained to detect any possible subset of the patterns tested.