Mitigating mismatch in neuromorphic circuits using neuromodulation loops

Carver Mead and collaborators envisioned fast, low-power, and efficient computing inspired by biological systems.
This vision has been pursued by research groups and companies, leading to the development of multiple neuromorphic chips.
In recent years, neuromorphic chips have demonstrated their envisioned high energy-efficiency and have even shown promising results on real-world tasks.

A fundamental challenge hindering the deployment of analog neuromorphic chips is device mismatch.
Mismatch refers to the inherent variability in circuit parameters introduced during production. Designs that do not mitigate mismatch are impractical to deploy at scale; either each chip must be precisely calibrated during production or design parameters need to be adjusted individually for each chip.

Biological systems face similar variability yet operate reliably despite it, owing to homeostatic and neuromodulatory mechanisms.
The role of neuromodulation in real neural circuits is crucial, as blocking it leads to the collapse of the circuit's function.
Inspired by biology, we propose using neuromodulation-based regulation loops to stabilize the behavior of neuromorphic circuits in the presence of mismatch.
We demonstrate this approach using simulations performed with our Neuromorphic Control Toolbox, which provides all the necessary blocks (neurons, synapses, neuromodulation) as well as tools for mismatch emulation.