Spiking control through local predictive optimization
Abstract: Neurons communicate with downstream systems with short-lived action potentials. Using such brief spikes, they must affect and control those downstream systems. One approach to understanding how neurons and networks do this is by using control theory to consider how they should spike. However, such approaches usually consider the effect of filtered spikes to approximate a more analogue control signal --- which ultimately means the neurons have to output a continuous control signal (either through synapses or on hardware by filtering the spikes). We instead consider how downstream linear dynamical systems could be controlled solely by brief spiking events, as if the neuron can only give brief 'kicks'. We formalize this by requiring that neurons only spike if the controlled system is brought closer to a target within some time-window, requiring each neuron to essentially predict the future effect of its spike. Surprisingly, starting from this principle, we arrive at a network of standard integrate and fire neuron that can successfully control linear systems through sparse spiking activity. The work gives insight both into how real neurons could control e.g. muscle fibers, and has applications in neuromorphic hardware design for control tasks where the control output has to be brief and sparse.
Format: I plan to give the talk somewhat in tutorial form and go through the mathematical approach quite in depth, and leave room for discussion.
