Coding in high dimensional state space? A novel interpretation of cortical processing

A hall mark of cortical architectures is the dense and specific reciprocal coupling among distributed feature specific neurons. This network engages in high dimensional non-linear dynamics that is characterized by oscillatory activity in widely differing frequency ranges and transient changes in correlation structure. Analysis of massive parallel recordings of neuronal responses in cat and monkey visual cortex suggests that the cerebral cortex exploits the high dimensional dynamic space offered by recurrent networks for the encoding, classification and storage of information. Evidence is presented that the recurrent connections among cortical neurons are susceptible to activity dependent modifications of their synaptic gain, which allows the network to store priors about the statistical contingencies of the outer world. Matching of sensory evidence with stored priors is associated with fast transitions towards substates of reduced dimensionality that are well classifiable by linear classifiers. In addition the network dynamics allow for the superposition and fast read out of information about sequentially presented stimuli, facilitating the encoding and storage of information about sequences. It is proposed that computations in high dimensional state space can account for the ultra-fast integration of sensory evidence with stored priors and the subsequent classification of the results of this matching operation.