Comparing Industry Frameworks with Deeply Quantized Neural Networks on Microcontrollers

Differently to the common belief, the industry quest for ultra-low-power neural networks is just at the beginning. Examples are the efforts carried by open communities such as TinyML and TinyMLPerf currently focusing on deep learning frameworks for Machine Learning (ML) and associated applications targeting micro-controllers (MCUs). However little attention has been put on deep learning frameworks and applications to enable ultra-low precision ML. These are enabling technologies to target uW hardware implementations. This work aims to compare two Deep Learning frameworks with support to deep quantization, QKeras and Larq, that abstract Tensorflow and Keras frameworks. Currently, Tensorflow is one of the most used deep learning tools by the research and industry communities aimed at deploying ML on the field. Two applications are presented with associated deeply quantized neural networks: Human Activity Recognition (HAR) exploiting a Hybrid Binary Neural Network (HBN) and Anomaly Detection for Industry 4.0 based on a Hybrid Binary AutoEncoder (HBAE). The pros and cons of the frameworks will be discussed during their usage on those applications. Results show an accuracy of up to 98.6% for the HBN and a PSNR up to 111.2 for the HBAE. The inference time has also been measured on different ARM-based architecture.