Regularized Evolution for Macro Neural Architecture Search

Neural Architecture Search is becoming an increasingly popular research field and method to design deep learning architectures. Most research focuses on searching for small blocks of deep learning operations, or micro-search. This method yields satisfactory results but demands prior knowledge of the macro architecture’s structure. Generally, methods that do not utilize macro structure knowledge perform worse but are able to be applied to datasets of completely new domains. In this paper, we propose a macro NAS methodology which utilizes concepts of Regularized Evolution and Macro Neural Architecture Search (DeepNEAT), and apply it to the Fashion-MNIST dataset. By utilizing our method, we are able to produce networks that outperform other macro NAS methods on the dataset, when the same post-search inference methods are used. Furthermore, we are able to achieve 94.46% test accuracy, while requiring considerably less epochs to fully train our network.