Cross-layer CNN approximations for hardware implementation

Convolution Neural Networks (CNNs) are widely used for image classification and object detection applications. The deployment of these architectures in embedded applications is a great challenge. This challenge arises from CNNs’ high computation complexity that is required to be implemented on platforms with limited hardware resources like FPGA. Since these applications are inherently error-resilient, approximate computing (AC) offers an interesting trade-off between resource utilization and accuracy. In this paper, we study the impact on CNN performances when several approximation techniques are applied simultaneously. We focus on two of the widely used approximation techniques, namely quantization and pruning. Our experimental results showed that for CNN networks of different parameter sizes and 3% loss in accuracy, we can obtain up to 27.9%–47.2% reduction in computation complexity in terms of FLOPs for CIFAR-10 and MNIST datasets.