Comparison of Deep Convolutional Neural Network Structures The effect of layer counts and kernel sizes

Abstract

Deep learning algorithms have become popular methods for pattern recognition due to their advantages over traditional methods such as providing deep representations of data, high-level semantic features. Deep convolutional neural network is one of the deep learning technique used in computer vision. Deep convolutional neural network consists of alternating convolution and pooling layers, and feedforward layers after them. It has not a fixed structure hence determining the optimal structure such as number of convolution and pooling layers, kernel size of these layers is crucial for faster and high performance implementations. Hence, in this work different convolutional neural network structures were established and tested on recognition of 28x28 MINST handwritten digits. According to the test results, kernels should cover at least 2 neighbor pixels of the current pixel from each side. Moreover, increasing the number of layers provide better results at the same time leads to decreases in the kernel size which may lead to worse performance. Hence, while the number of layers are increased, kernel size must be considered.