Convolutional Neural Network Structure Transformations for Complexity Reduction and Speed Improvement

Abstract

Two methods of convolution-complexity reduction, and therefore acceleration of convolutional neural network processing, are introduced. Convolutional neural networks (CNNs) are widely used in computer vision problems. In the first method, we propose to change the structure of the convolutional layer of the neural network into a separable one, which is more computationally simple. It is shown experimentally that the proposed structure makes it possible to achieve up to a 5.6-fold increase in the operating speed of the convolutional layer for 11 × 11-sized convolutional filters without loss in recognition accuracy. The second method uses 1 × 1 fusing convolutions to increase the number of convolution outputs along with decreasing the number of filters. It decreases the computational complexity of convolution and provides an experimental processing speed increase of 11% in the case of large convolutional filters. It is shown that both proposed methods preserve accuracy when tested with the recognition of Russian letters, CIFAR-10, and MNIST images.