Development and Applications of a Memory-Traffic-Efficient Convolutional Neural Network

Built upon circuit insight, HarDNet is introduced as an innovative and efficient convolutional neural network architecture aimed at minimizing DRAM access. This design approach leverages the faster speed and lower energy consumption of arithmetic operations compared to DRAM, resulting in significant improvements in accuracy, speed, and energy efficiency. HarDNet’s optimized architecture makes it well-suited for diverse applications including object detection, semantic segmentation, and medical image analysis. As an open-source initiative, HarD-Net is accessible for global use and adaptation. Its impact spans various domains and regions, from autonomous driving and industrial automation to vehicle safety, environmental monitoring, colonoscopy polyp segmentation, and MRI imaging. Drawing from deployment insights, this presentation outlines challenges and opportunities in circuit and system design to enable robust and cost-effective deep learning solutions. Additionally, it includes case studies of hardware implementations of HarDNet.