Master thesis : Neuromodulation Controller Enabling Robustness to Transistor Variability in Subthreshold Silicon Neurons

Abstract

In the brain, billions of neurons communicate through neuronal signaling based on both electrical and chemical processes, particularly action potentials. This communication process is implemented in analog neuromorphic neurons using electronic circuits that emulate the behavior of biological neurons. In such implementations, transistor-intrinsic mismatch arises, causing intra- and inter-neuron variability. One consequence of mismatch is the difficulty of implementing neuromodulation, the process by which neuronal activity is regulated, enabling neurons to adapt their behavior to context and to support different functions.

Therefore, this thesis investigates the feasibility of feedback-based regulation of neuronal activity, robust to mismatch, in an analog neuromorphic neuron.

A controller is designed using dynamic input conductances for neuronal activity tracking, allowing control of excitability instead of raw voltage trajectories. This approach is first validated with the computational model of the neuron chip and against artificial mismatch. Then, variability is characterized both globally and individually by incorporating mismatch models, providing insight into the impact of device variability. After controller adaptation and neuron calibration, the control loop is tested on ultra-low-power neurons implemented on the developed chip, while accounting for practical mismatch effects and electronic constraints.

The results show that neuronal activity can be robustly regulated despite variability. However, large-scale implementation appears impractical, as the parameters must be specifically calibrated for each individual neuron. The framework is therefore evaluated at higher power levels to reduce mismatch-induced variability. Under these conditions, neuronal dynamics become more consistent across neurons, enabling a common calibration strategy that is more suitable for large-scale integration.

Overall, this work demonstrates that neuromodulation is achievable in analog subthreshold neurons through activity-based feedback control, despite the pronounced effects of device mismatch, contributing to robust, adaptive behavior in scalable and biologically inspired neurons.