Accelerated CNN Training with Genetic Algorithm

This paper presents an accelerated genetic algorithm for training deep convolutional neural networks. By introducing parent-child connections that allow individuals in one generation to inherit knowledge from their ancestors, the algorithm reduces execution time while maintaining robust training. Additionally, a new dataset, DoubledMNIST, is introduced, poised to succeed the MNIST dataset, making it ideal for machine learning education and applications in handwriting recognition. The algorithm exhibits strong performance under various evolutionary scenarios, opening avenues for further improvements, such as enhancing ancestor-descendant relationships for more comprehensive knowledge transfer.