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

How biological brains can learn so quickly to become operational and achieve complex behavior remains an unresolved issue. Here we introduce a ‘neuromorphic’ learning strategy that resembles how immature biological brains learn by consisting of continual random activations of a complex mechanically coupled system with rich, dynamic intrinsic sensor dependencies, in this regard reminiscent of a biological body. Using a dynamic model of biological skin tissue with embedded sensors, we trained small, recurrent networks that emulated a primordial cortex and featured excitatory and inhibitory neurons with simultaneous independent learning in both types of synapses. Training with non-repetitive, random activations of the skin, where the recurrent network activity state was inherited between activations, autonomously led to rapid acquisition of remarkably generalizable representations of a predictive nature. The network could separate inputs and solve a kinematics task that had never been encountered, also after substantial parts of the sensor population were deleted. This strategy of focussing learning on the dominant regularities in dynamic sensory information can potentially explain complex brain operation and efficient learning.