Online Transformers with Spiking Neurons for Fast Prosthetic Hand Control

Fast and accurate online processing is essential for smooth prosthetic hand control with Surface Electromyography signals (sEMG). Although transformers are state-of-the-art deep learning models in signal processing, the self-attention mechanism at the core of their operations requires accumulating data for large time-windows. They are therefore not suited for online signal processing. In this paper, we use an attention mechanism with sliding windows that allows a transformer to process sequences element-by-element. Moreover, we increase the sparsity of the network using spiking neurons. We test the model on the NinaproDB8 finger position regression dataset. Our model sets its new state-of-the-art in terms of accuracy on NinaproDB8, while requiring only very short time windows of 3.5 ms at each inference step, and reducing the number of synaptic operations up to a factor of ×5.3 thanks to the spiking neurons. Our results hold great promises for wearable online sEMG processing systems for prosthetic hand control.