Analysis of Different Pooling Functions on a Convolution Neural Network Based Model

The common denominator of deep learning models used in many different fields today is the pooling functions used in their internal architecture. These functions not only directly affect the performance of the study, but also directly affect the training time. For this reason, it is extremely important to measure the performance of different pooling functions and share their success values. In this study, the performances of commonly used soft pooling, max pooling, spatial pyramid pooling and average pooling functions were measured on a dataset used as benchmarking in the literature. For this purpose, a new CNN based architecture was developed. Accuracy, F1 score, precision, recall and categorical cross entropy metrics used in many studies in the literature were used to measure the performance of the developed architecture. As a result of the performance metrics obtained, 97.79, 92.50, 91.60 and 89.09 values from best to worst for accuracy were obtained from soft pooling, max pooling, spatial pyramid pooling and average pooling functions, respectively. In the light of these results, the pooling functions used in this study have provided a better conceptual and comparative understanding of the impact of a CNN-based model.