On computational complexity, using randomized pooling for convolutional neural network

Abstract Randomization is a technique used in algorithms as a strategy that uses a random source as part of its logic. It is used in traditional algorithms to reduce time or space complexity. Many efforts have been made to increase the precision of convolutional neural networks (CNN) in various application domains, but less has been done to minimize the computational complexity of this model. In this work, we introduce randomized pooling (RPool) for CNN. RPool has reduced the number of operations of the CNN. Consequently, the computation time of the algorithm is reduced and the accuracy is improved. The (MNIST) dataset is used to demonstrate the pooling layer (PL) of CNN and compare the results of standard CNN with our proposed RPool for CNN. The simulation results show that as the number of epochs increases, the training and testing time of our proposed RPool decreases while the accuracy increases. We achieved 96.95% accuracy at epoch 10 and 8.85% decrease in training time, which demonstrates the superiority of our proposed RPool for CNN over its competitor average pooling (AvgPool).