Learning-based Approaches for Controlling Neural Spiking.

We consider the problem of controlling populations of interconnected neurons using extrinsic stimulation. Such a problem, which is relevant to applications in both basic neuroscience as well as brain medicine, is challenging due to the nonlinearity of neuronal dynamics and the highly unpredictable structure of underlying neuronal networks. Compounding this difficulty is the fact that most neurostimulation technologies offer a single degree of freedom to actuate tens to hundreds of interconnected neurons. To meet these challenges, here we consider an adaptive, learning-based approach to controlling neural spike trains. Rather than explicitly modeling neural dynamics and designing optimal controls, we instead synthesize a so-called control network (CONET) that interacts with the spiking network by maximizing the Shannon mutual information between it and the realized spiking outputs. Thus, the CONET learns a representation of the spiking network that subsequently allows it to learn suitable control signals through a reinforcement-type mechanism. We demonstrate feasibility of the approach by controlling networks of stochastic spiking neurons, wherein desired patterns are induced for neuron-to-actuator ratios in excess of 10 to 1.