Low Precision Floating Point Spectral Feature Extraction Engine for Closed-Loop Neuromodulation

Closed loop neuromodulation system on chip enables treatment of neurological disorders through implantable biomed-ical devices. On-chip machine learning algorithms for such applications rely on low-power and form-factor implementations while still providing sufficient sensitivity and specificity. Spectral features are one of the most important feature sets in neurological data. This work presents a low-precision floating point spectral feature extraction engine. ASIC implementation of the same in a 65 nm CMOS prcoess along with the counterpart fixed point hardware shows that the floating point FEE performs poorly in terms of stop band attenuation, noise floor, and area performance. Furthermore, we demonstrate that fixed point hardware with truncation provides optimal performance in terms of area, power, noise-floor, and stop-band attenuation.