SpikyControl aims at developing a control theory of neuromorphic machines. Neuromorphic machines take inspiration from the regulation mechanisms of neural circuits. Most importantly, they use “spikes” rather than “bits” as units of information, resulting in an event-based interaction between the machine and its physical environment. Event-based sensing and event-based actuation opens a new era for control theory. The question becomes how to close the loop between event trajectories instead of closing the loop between continuous or discrete trajectories.
The angle of attack of the proposed novel control theory is to leverage the existing theory by moving from continuous trajectories to events. Starting from the classical paradigm of the feedback amplifier, regulation theory is regarded as a theory of negative feedback, automation theory is regarded as a theory of positive feedback, and neuromorphic theory is regarded as a theory of mixed feedback. Mixing positive feedback and negative feedback results in excitability, regarded as "transient" decision making: the decision corresponds to the event, but the event nature of the decision makes it transient, enabling the return to equilibrium.
The methodology of SpikyControl is to generalize the convex optimization framework associated to negative feedback of monotone operators to a "convex-concave" optimization framework associated to mixed feedback of monotone operators. The circuit representation of biophysical neural networks is exploited to split a complex neuromorphic system into a circuit of elementary monotone operators corresponding to the memristive elements of the circuit, that is, the machine analog of ion channels and synapses in biophysical circuits. The physical model of the memristive elements is constrained to make the overal theory scalable, with the objective of designing complex machines mimicking the organization of neural circuits to solve novel engineering tasks.
A key illustration of SpikyControl is to develop novel active sensing filters for neuromorphic sensors such as event cameras, mimicking the exquisite detection and selection capabilities of animal vision.
