A Delta Recurrent Neural Network over TinyML Framework for Arithmetic Coding on the Edge

In this paper, we use IoT devices at the edge to perform the data compression with arithmetic coding algorithm on the Tiny Machine Learning (TinyML) framework. Our Arithmetic Codec combines a deep learning model, i.e., delta recurrent neural network (DRNN), to improve the compression ratio. In addition, a customized neural co-processor (NCP) is integrated with IoT device to construct a very low power TinyML system, in which all weights and features are stored. The experimental results show that the entire system achieved the lowest power consumption, with similar compression ratio and compressing speed compared with other competitors.