Comparison of Pooling Methods for Handwritten Digit Recognition Problem

Convolutional Neural Networks (CNNs) are used as a current approach to the recognition of handwritten digits for the design of pattern recognition systems. The fact that Convolutional neural networks have a multilayered structure and a large number of items in each layer increases the level of complexity. Increasing the level of complexity makes extremely difficult to discover the optimum configuration for these networks. Modelling the layers of convolutional neural networks independently may be an effective solution to overcome this difficulty and successfully classify images. In this paper, we have investigated the design of the pool layer, which is one of the three basic layers of convoluted neural networks, with optimum techniques. For this purpose primarily a basic convolutional network structure is formed. Then effect of different pooling methods on the classification performance of these networks was investigated using the MNIST dataset. In this study, in addition to the average and maximum pooling methods commonly used in the literature, Mixed pooling, Stochastic pooling, random pooling, Gaussian pooling, median pooling, minimum pooling methods are also used. Experimental results show that the Mixed pooling method has highest accuracy and minimum pooling method has the lowest accuracy .