Lightweight convolutional neural networks framework for really small tinyml devices

CÉSAR A. ESTREBOU; MARTÍN FLEMING; MARCOS D. SAAVEDRA; FEDERICO ADRA; ARMANDO E. DE GIUSTI

Quito

Conferencia; Second International Conference, SmartTech-IC 2021; 2021

Universidad Politécnica Salesiana

This paper presents a lightweight and compact framework designed to perform convolutional neural network inference on severely hardware constrained microcontrollers. A review of similar open source libraries is included and experiments are developed to compare their capabilities on several different microcontrollers. The proposed framework implementation shows at least a three-time improvement over the Google Tensorflow Lite Micro implementation with respect to memory usage and inference time.