Machine Learning Based Edge AI for IoT

Abstract

With the further development of deep neural network, more and more large-scale network structures are proposed, the large-scale network led to the increase in the number of parameters, and the amount of calculation. This will challenge devices’ hardware and greatly increase the power consumption [1]. So, it is unfavorable for using large-scale network on terminal devices. We will solve these problems through design special network structure and quantize the weight and activation function in our design. Make the large-scale network can be better used on the terminal equipment. The design is based on the ResNet structure [2] and used groups of convolutions [3] with both binarized weights and activations [4] to build lightweight deep neural networks. During the forward pass, our design drastically reduces memory size and accesses, and replace most arithmetic operations with bit-wise operations to simplify the computation and reduce memory size, which is expected to substantially improve power-efficiency and Computing speed. We designed a baseline and continued to adjust the network architecture and parameters for better tradeoff between latency and accuracy. Finally, we got satisfied results on CIFAR-10 dataset [5]. At last, we will compare the network we designed with state-of-the-art network in the Cifar-10 100k benchmark [6] to show the performance of our design.