Architecture Optimization of Convolutional Neural Networks by Micro Genetic Algorithms

Convolutional Neural Networks (CNN) are novel techniques with significant performance in object detection and classification. An open research problem on CNN is the automatic finding of adequate architectures, which is usually done by hand. Metaheuristic algorithms are techniques that find optimal solutions in heuristic manner to problems where the knowledge is limited or almost nonexistent, such as finding optimal CNN architectures. In this chapter, we propose a framework that utilizes the micro genetic algorithm to find CNN architectures in the shortest possible time. The proposal is tested over three simple study cases known by the research community (MNIST, MNIST-Fashion, and MNIST-RB), and compared against two different frameworks from the literature: psoCNN, and simple genetic algorithm. The results show a better performance of the architectures found by our framework in terms of accuracy and processing time.