Sensory dependencies rapidly and autonomously yield generalizable representations in recurrent cortical-like networks

How do biological brains become operational so quickly? Here we introduce a ‘neuromorphic’ learning strategy that mimics the strategies of animals in early development. It consists of continual random activations of the body, which is a mechanically coupled system with rich, dynamic intrinsic sensor dependencies. Using a dynamic model of biological skin tissue with distributed sensors, we trained small, recurrent networks designed to emulate a primordial cortex with excitatory and inhibitory neurons and simultaneous independent learning in both types of synapses. Continual random activations of the skin, without resetting the network state, led to rapid acquisition of remarkably generalizable representations with emergent predictive capability. The network could separate inputs and solve a kinematics task it had never encountered, even when substantial parts of the sensor population were removed. This strategy of learning the dominant regularities in dynamic sensory information can explain efficient learning of complex operation in the brain.