New learning principles emerge from biomimetic computational primitives

Abstract

To examine the biological building blocks of thought and action, we created biologically realistic local circuits based on detailed well-cited physiological and anatomical characteristics. These biomimetic circuits were then integrated into a large-scale model of cortical-striatal interactions for category learning. The model was not trained on, but nonetheless displayed properties similar to, neurophysiological recordings from non-human primates (NHPs) performing the same task. The model had learning curves similar to the NHPs. It showed how synaptic modifications could induce changes in spiking and synchrony in the brain. The model even made novel predictions that were subsequently found in the brain, including “bad-idea neurons” that signaled impending incorrect choices after learning. This demonstrates how key computational principles can be discovered by modeling local circuitry that mimics the brain.