Fully Mixed Max-Average Pooling: A Comparative Study for a Convolutional Neural Network
Brahim Ait Skourt, Nikola S. Nikolov, Aicha Majda · 2024
As an integral part of a convolutional neural network, the pooling layers are responsible for the down-sampling operation which aims at preventing overfitting. Besides the conventional pooling methods (max and average), various methods that involve mixing max pooling and average pooling have been proposed. In this work, we propose a new mixed pooling method, called fully mixed max-average pooling (FMMAP), and evaluate its performance within a comparative study of various conventional and state-of-the-art pooling methods. FMMAP consists of fully mixing max pooling and average pooling features instead of stochastically selecting them as done in other popular mixedpooling methods. The experimental results suggest that FMMAP outperforms the conventional pooling methods in accuracy, and while being very close (within 0.05%) to the other mixed-pooling methods accuracy-wise, it significantly outperforms them in terms of running time, being at least 1.7 times and up to 2.7 times faster.