A novel cellular automata-based approach for generating convolutional filters

Image classification is a well-studied problem where the aim is to categorize given images into a predefined set of classes. Although there are different approaches for solving the problem, convolutional neural networks (CNNs) have achieved significant success in the domain. CNN uses convolutional layers to extract features from images, and these layers are usually created with a supervised training process. This training process requires a group of convolution operations and several passes over the dataset. Hence, the model possesses a heavy computational burden. In this work, a cellular automata-based unsupervised methodology is proposed to create convolutional filters. The proposed methodology accesses each data instance only twice regardless of the number of layers in the model, and it requires no backpropagation operation. Thus, the computational burden is significantly reduced compared to CNNs. The classification process can be carried out directly by using the model together with a multilayer perceptron. Also, the model can be used to enhance CNNs in terms of time and accuracy by initializing the parameters of CNN or by preprocessing the raw data. The proposed methodology creates competitive results compared to CNNs in terms of accuracy and computational complexity. Also, the results show that the performance of the CNN model can be increased by using the filters created by the proposed methodology.