AIoT Solution Survey and Comparison in Machine Learning on Low-cost Microcontroller

Abstract

Neural Networks, especially Convolutional Neural Network [1] are becoming increasingly popular in IoT edge devices today for executing data analytics right at the source without transmitting to Cloud Computing centers. So that it will be reduced latency as well as energy consumption for data communication. In this paper, we will compare CMSIS-NN and uTensor: low energy consumption microcontrollers. Most classification tasks have always-on, and real-time requirements, which limits the total number of operations per neural network inference. So that, with image classification model, microcontrollers will execute lower frame per second than GPU and embedded CPU. CMSIS-NN is a collection of efficient kernels which was developed to maximize the performance and minimize the memory footprint of Neural Network applications. Allow deploy machine learning models on ARM Cortex-M processors for intelligent IoT edge devices.