Computation reduction in convolution layer by pre-emptive Max-pooling for energy-efficient CNN

Abstract

This thesis proposes a method of computation reduction in convolution layer for energy-efficient Convolutional Neural Network (CNN) accelerator. CNN is the most promising algorithm for image recognition, then academia and industries focus on developing CNN. It still has energy-efficiency problem to be used in mobile platform and IoT, while it is utilized on various applications. It seems that CNN will be utilized mostly in the two platforms in the future, hence it is important to solve the energy-efficiency problem. To solve the problem, I suggest reducing energy consumption with accuracy degradation through approximate computing. Since the most workload of CNN is concentrated on convolution layers, reducing computation in convolution layers by pre-emptive max-pooling results in energy reduction of CNN accelerator. Post synthesis simulation of proposed processing element is performed for MNIST and CIFAR-10. It shows 27% and 20% energy reduction with 0.47%p and 3.35%p accuracy degradation respectively for MNIST and CIFAR-10.