Strided Convolution Instead of Max Pooling for Memory Efficiency of Convolutional Neural Networks

Abstract

This paper describes a new optimization technique to perform an embedded implementation of convolutional neural networks (CNN). In this case, only the inference of convolutional neural networks is discussed. As known that both pooling layer and strided convolution can be used to summarize the data. So, the proposed technique aims to replace only max pooling layers by a strided convolution layers using the same filter size and stride of the old pooling layers in order to reduce the model size and improve the accuracy of a CNN. Also, pooling layer is parameter less. However, convolution layer has weights and biases to optimize. Then, the CNN can learn how to summarize the data. By replacing max pooling layers with strided convolution layers enhance the CNN accuracy and reduce the model size. This technique is proposed in order to build a CNN accelerator for real time application and embedded implementation.