A Tiny CNN for Embedded Electronic Skin Systems

The quest for efficient Tiny Machine Learning on Microcontroller Units is increasing rapidly due to the vast application spectrum made possible with the advancement of Tiny ML. One application area that could benefit from such advancement is Electronic Skin systems, that are employed in several domains such as: wearable devices, robotics, prosthesis, etc. An e-skin system demands hard constraints including real-time processing, low energy consumption, and low memory footprint. This paper presents a tiny Convolution Neural Network (CNN) architecture suitable for the deployment on an off-the-shelf commercial microcontroller in compliance with the e-skin requirements. The training, optimization, and implementation of the proposed CNN are presented. The CNN implementation is optimized through layer fusion and buffer re-use strategies for efficient inference on edge devices. As a case study, experimental analysis of a touch modality classification task demonstrates that the proposed CNN-based system is capable of processing tactile data in real-time directly near the source while reducing the model size by up to 65% with respect to comparable existing solutions.