The effect of pooling parameters on the performance of convolution neural network

Abstract Pooling is a crucial aspect of Convolutional Neural Networks (CNNs), a prominent machine learning technique. It plays an essential role in the learning process by reducing the spatial dimensions of feature maps and minimizing computational costs. Common pooling methods, such as Max Pooling (MaxPool) and Average Pooling (AvgPool), have notable limitations. While MaxPool efficiently extracts important features, it often discards useful information, which can negatively impact CNN performance. In contrast, AvgPool preserves background information (maintaining global detail) but treats all inputs equally (assigning the same importance to every pixel in a region of an image), potentially leading to inefficiencies. We propose enhancing pooling adaptability by introducing learnable parameters to address these issues. These parameters (pooling kernels) are incorporated into the pooling methods to make them learnable. The study also explores how this modification affects the image classification performance of CNNs across various widely used datasets. The findings suggest that integrating parameters into the pooling process produces better results compared to the traditional pooling techniques previously discussed.