CGP-NAS: real-based solutions encoding for multi-objective evolutionary neural architecture search.

Convolutional Neural Networks (CNNs) have had a remarkable performance in difficult computer vision tasks. In previous years, human experts have developed a number of specialized CNN architectures to deal with complex image datasets. However, the automatic design of CNN through Neural Architecture Search (NAS) has gained importance to reduce and possibly avoid human expert intervention. One of the main challenges in NAS is to design less complex and yet highly precise CNNs when both objectives conflict. This study extends Cartesian Genetic Programming (CGP) for CNNs representation in NAS through multi-objective evolutionary optimization for image classification tasks. The proposed CGP-NAS algorithm is built on CGP by combining real-based solutions representation and the well-established Non-dominated Sorting Genetic Algorithm II (NSGA-II). A detailed empirical assessment shows CGP-NAS achieved competitive performance when compared to other state-of-the-art proposals while significantly reduced the evolved CNNs architecture's complexity as well as GPU-days.