A neural network framework for small microcontrollers

CÉSAR A. ESTREBOU; MARTÍN FLEMING; MARCOS D. SAAVEDRA; FEDERICO ADRA

Salta

Congreso; XXVII Congreso Argentino de Ciencias de la Computación; 2021

Red de Universidades con Carreras en Informática

This paper presents a lightweight and compact library designed to perform convolutional neural network inference for microcontrollers with severe hardware limitations. A review of similar open source libraries is included and an experiment is developed to compare their performance on different microcontrollers. The proposed library shows at least a 9 times improvement over the implementation of Google Tensorflow Lite with respect to memory usage and inference time.