Improving kernel-energy trade-offs for machine learning in implantable and wearable biomedical applications