Exploiting Pixel Redundancy and Approximate Computing for Efficient Hardware–Software Co-design of CNN on IoT Edge Devices

Edge computing is an emerging paradigm in which speed is enhanced by deploying software programs on embedded systems in the vicinity of data collection. Edge computing is considered to be a vital concept for realizing the far-fetched dream of pervasively interconnecting millions of devices through Internet of things (IoT). In this paper, an approximate convolutional layer is proposed that is based on analysis of dataset before application to training and inference. The technique uses the similarity of the image pixels at the edges of the images in the dataset. At the same time, a novel approximate 8-bit fixed point multiplier is proposed that increases the energy efficiency without compromising much accuracy. The comparison results of exact and approximate CNN prove that the approximate CNN has 5.5% (about 8.8 million) less MAC operations as compared to the exact CNN with a minimal accuracy loss of 3.3%.