Cellular automata integrated with deep learning methods for feature extraction

Neural networks and deep learning methods use feature extraction strategies to create embeddings and feature maps representing an input, which can be a word, signal, or image according to the problem domain. The most common feature extraction methodology in the image domain is convolutional filters, which are employed by Convolutional Neural Networks (CNN). However, a CNN learns those filters with a computationally expensive training process that requires several passes over the entire training dataset and the data utilized needs to be labeled. This thesis proposes an unsupervised Cellular Automata (CA) based framework to create the convolutional filters that extract features. The proposed framework accesses each data instance only twice, regardless of the number of layers in the model, and it requires no epoch-based training operation, unlike CNNs. Thus, the computational burden is significantly reduced compared to CNNs. Also, the model can enhance CNNs in terms of time and accuracy by initializing the parameters of CNN or by preprocessing the raw data. The proposed methodology is tested on different datasets in the image classification domain, and it obtains competitive results compared to CNNs regarding prediction success and computational complexity. Also, the results show that the performance of the CNN model can be improved by using the filters created by the proposed methodology.