Evolving Long Short-Term Memory Networks

Machine learning techniques have been massively employed in the last years over a wide variety of applications, especially those based on deep learning, which obtained state-of-the-art results in several research fields. Despite the success, such techniques still suffer from some shortcomings, such as the sensitivity to their hyperparameters, whose proper selection is context-dependent, i.e., the model may perform better over each dataset when using a specific set of hyperparameters. Therefore, we propose an approach based on evolutionary optimization techniques for fine-tuning Long Short-Term Memory networks. Experiments were conducted over three public word-processing datasets for part-of-speech tagging. The results showed the robustness of the proposed approach for the aforementioned task.