Neural Networks Embedded in Wearable Devices: a Preliminary Digital vs. Analog Comparison
Daniele Maria Crafa, Stefano di Giacomo, Carlo Ettore Fiorini, Marco Carminati · 2023
A fully connected artificial neural network (NN) for color identification (with 125 neurons and 11 color classes with inputs from an RGB sensor) was developed to study the challenges of real-time, low-power and low-latency machine learning to be embedded in wearable devices. Digital and analog approaches were compared in terms of miniaturization, power consumption, accuracy and speed. A prototype was built using a Nordic nRF52840 microcontroller with BLE, where the NN runs with an energy consumption of 6$\mu$J/class and sub-ms time. The feasibility of an alternative analog implementation of the NN in a dedicated integrated circuit by means of switched capacitors was evaluated through simulations, focusing in particular on the impact of weights quantization and the potentially significant saving in terms of energy dissipation (2 nJ/class),