Genetically-trained deep neural networks

Deep learning is a widely explored research area, as it established the state of the art in many fields. However, the effectiveness of deep neural networks (DNNs) is affected by several factors related with their training. The commonly used gradient-based back-propagation algorithm suffers from a number of shortcomings, such as slow convergence, difficulties with escaping local minima of the search space, and vanishing/exploding gradients. In this work, we propose a genetic algorithm assisted by gradient learning to improve the DNN training process. Our method is applicable to any DNN architecture or dataset, and the reported experiments confirm that the evolved DNN models consistently outperform those trained using a classical method within the same time budget.