Analysis of visual processing capabilities and neural coding strategies of a detailed model for laminar cortical microcircuits in mouse V1
AbstractThe neocortex is a network of rather stereotypical cortical microcircuits that share an exquisite genetically encoded architecture: Neurons of a fairly large number of different types are distributed over several layers (laminae), with specific probabilities of synaptic connections that depend on the neuron types involved and their spatial locations. Most available knowledge about this structure has been compiled into a detailed model [Billeh et al., 2020] for a generic cortical microcircuit in the primary visual cortex, consisting of 51,978 neurons of 111 different types. We add a noise model to the network that is based on experimental data, and analyze the results of network computations that can be extracted by projection neurons on layer 5. We show that the resulting model acquires through alignment of its synaptic weights via gradient descent training the capability to carry out a number of demanding visual processing tasks. Furthermore, this weight-alignment induces specific neural coding features in the microcircuit model that match those found in the living brain: High dimensional neural codes with an arguably close to optimal power-law decay of explained variance of PCA components, specific relations between signal- and noise-coding dimensions, and network dynamics in a critical regime. Hence these important features of neural coding and dynamics of cortical microcircuits in the brain are likely to emerge from aspects of their genetically encoded architecture that are captured by this data-based model in combination with learning processes. In addition, the model throws new light on the relation between visual processing capabilities and details of neural coding.