Video Not Available
Details
Abstract
The design of efficient hardware for Convolutional Neural Networks has always been a big challenge for resource-constrained devices such as IoT). This is due to the cost of the high computational complexity of CNNs. Accordingly, different methods are used to improve energy efficiency and throughput without affecting the accuracy or increasing hardware cost. This article introduced a method to reduce the storage and computation needed by DNNs by approximate the DNN weights linearly across different filters. That leads to a repetition of the weights across different kernels, which enables the reuse of the DNN computations of different output feature map.