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

How biological brains become operational from development and onwards is an unresolved issue. Here we explored the emerging effects in a brain circuit model that received simulated touches through biological skin model, which featured the potentially critical aspect of sensory dependencies resulting from mechanical coupling across the tissue. Our skin model system was connected to a recurrent network a generic, primitive cortical-like subnetwork that was composed of fully connected excitatory and inhibitory neurons, where each individual synapse was subject to continuous, independent, Hebbian-like learning. We used continuous random mechanical activations of the skin model, in this regard mimicking behavioral patterns of early brain development observed in infants, and let the neuron-independent learning define the function that emerged in the network. Remarkably, we found that the network could rapidly learn to separate various naive dynamic skin inputs and solve a kinematics task it had never encountered, even when substantial parts of the sensor population or even network connections were removed post-training. We propose that autonomous learning from a sensor population with intrinsic dependencies could cause the extensively recursive cortical network to gradually adapt its intrinsic dynamics to better mirror the various dynamics of whatever body it is connected to, which result in many biologically useful features for the early acquisition of brain circuitry function.