Bringing bio-inspired computing to bidirectional prosthesis control, by Dr. Corentin Piozin

Abstract Commercial hand prostheses face a double challenge. They generally lack real-time tactile feedback, forcing users to rely entirely on vision, and they require power-hungry, clock-driven processors to decode motor intent. This talk covers how we can use neuromorphic computing and Spiking Neural Networks (SNNs) to close this loop efficiently.

I will start by tracing the evolution of artificial neurons, moving from standard deep learning models to biophysical equations, and finally to the Leaky Integrate-and-Fire (LIF) model, which balances biological realism with engineering efficiency. From there, we will look at spike encoding methods to see how we translate continuous analog signals into sparse, event-driven spike trains.

The core of the presentation focuses on our team’s work in two main areas: first, Motor Decoding, where we use SNN regression to turn surface EMG recordings into smooth, continuous movements. Then, Sensory Encoding, for which we use SNNs to encode and classify raw data from tactile pressure sensors, allowing the system to identify different objects.

Finally, we will move from theory to physical implementation. I will show how these computational models translate onto actual neuromorphic hardware, highlight a few relevant chip design strategies, and share what our team has achieved in deploying these low-power, low-latency networks on silicon.