A Genetic Tuner for Fixed-skeleton TinyML Models

摘要

TinyML is a new paradigm of machine learning and inference on tiny devices. In many new-age applications such as Internet-of-Things (IoT), robotics, automotive embedded systems, etc., it is very important to process data close to the source for low-latency response, reduced data transfer, and privacy preservation. Despite the growing popularity of TinyML, it is very difficult to carry out these projects at scale. The reason for this is a critical dependency on skilled resources, who can design accurate and efficient models for very tiny devices, e.g. micro-controllers. In this paper, we address this shortage of TinyML skillsets by introducing a framework that can automatically generate, a tiny machine inference pipeline very fast. The framework is evaluated on diverse computer vision datasets, for which the models having state-of-the-art accuracy, less than 60KB in size are generated within 20 minutes on average.