Analysis of cortical function

National Institute Of Diabetes And Digestive And Kidney Diseases

Investigators

Linked publications, trials & patents

Abstract

For the past three decades, I have been developing biophysically-based cortical circuit models that can account for a comprehensive set of psychophysical and electrophysiological data pertaining to many neural phenomena including perceptual rivalry, stimulus disambiguation, neural activity normalization, synchronization, and decision making. Initially, the architecture of these models were specified a priori to produce the desired behavior. Now we train networks using modern machine learning approaches to match data. For example, we developed a scheme to train the synaptic input and spiking rate dynamics of individual neurons in a network averaged over a short time window to follow complex target patterns. We showed that the learning capacity grows linearly with network size and the learning rate is modulated by the synaptic decay and target correlation time. Our findings demonstrate that recurrent spiking networks possess a vast computational capability that can support the diverse activity patterns in the brain. However, the neurons in trained networks do not automatically maintain Poisson-like spiking observed in electrophysiological data because they do not have strong enough synapses to maintain a balanced-state between excitation and inhibition. We thus have learning schemes that maintain balance with strong synapses and also showed that starting with a balanced -state with dense connectivity, we could learn an arbitrary pattern for the mean firing rates by just training a small number of synapses, which we called subset-learning. Moreover, when only a subset of neurons are trained, the learned activity spreads across the entire network due to the strong synaptic connections. These results provide a mechanistic explanation as to why multi- or within-regional brain recordings show that task-related activities are distributed across brain areas.