Low-Shot Learning and Pattern Separation using Cellular Automata Integrated CNNs
Tej Pandit, Dhireesha Kudithipudi · 2022
Traditionally, deep convolutional neural networks are computationally expensive to train and require large amounts of data samples. In this article, we explore the use of pre-trained cellular automata as a substitute for convolutional layers. We propose a specialized cellular automata system with multilayered attributes and kernels to better harness its inherent spatio-temporal processing capabilities. An architecture search, combining deep Q-learning with a speciated genetic algorithm, is used to optimize the kernels and identify cell attributes. Experiments under low-shot conditions demonstrate that the cellular automata-integrated CNN outperforms compact state-of-the-art CNN models by 6-10% on static image datasets and 8-12% on temporal image sequence datasets.