Automated Neural Network Construction with Similarity Sensitive Evolutionary Algorithms

Abstract

Deep learning has been successfully applied to a wide variety of tasks. It generates reusable knowledge that allows transfer learning to significantly impact more scientific research areas. However, there is no automatic way to build a new model that guarantees an adequate performance. In this paper, we propose an automated neural network construction framework to overcome the limitations found in current approaches using transfer learning. Currently, researchers spend much time and effort to understand the characteristics of the data when designing a new network model. Therefore, the proposed method leverages the strength in evolutionary algorithms to automate the search and optimization process. Similarities between the individuals are also considered during the cycled evolutionary process to avoid sticking to a local optimal. Overall, the experimental results effectively reach optimal solutions proving that a time-consuming task could also be done by an automated process that exceeds the human ability to select the best hyperparameters.