This paper presents a wearable, microfabricated accelerometer that integrates sensing and machine learning capabilities using in-sensor computing. Leveraging reservoir computing, the device identifies human gait patterns in real-time by coupling the displacement of suspended microstructures. Compared to a conventional system, the in-sensor approach demonstrates significantly better power efficiency, paving the way for more ubiquitous deployment of machine learning in edge computing devices. The device also offers enhanced data security by only transmitting classification labels, not raw acceleration data.

Guillaume Dion, Albert Tessier-Poirier, Laurent Chiasson-Poirier, Jean-François Morissette, Guillaume Brassard, Anthony Haman, Katia Turcot, Julien Sylvestre