Energy-aware learning cuts DBS power 80% on SynSense neuromorphic chip

A team led by Binh Nguyen (UC Santa Cruz) has developed a neuromorphic energy-aware learning framework that dramatically reduces power consumption in adaptive deep brain stimulation (DBS) for Parkinson’s disease. Unlike prior work focusing only on inference efficiency, this approach incorporates actuator energy directly into the reinforcement learning reward. The deep spiking Q-network learned to suppress pathological alpha-beta oscillations by 45.2% in a biophysical cortico-basal ganglia-thalamic circuit model, while reducing stimulation charge by 80% relative to continuous DBS. This was achieved by co-optimizing both the neural controller and the stimulation output.

The compressed policy was deployed on the SynSense XyloAudio 3 neuromorphic processor, consuming just 0.52 mW during inference. Compared to an equivalent artificial neural network running on conventional edge hardware, this represents a 28.1x reduction in energy per inference. The paper, published on arXiv (2606.28600), demonstrates that in closed-loop neuromodulation, the actuator often dominates total power—making co-optimization essential. The framework was tested in simulation but points toward longer-lasting implantable devices that can adapt in real time without frequent battery changes.