A novel evolutionary algorithm for hierarchical neural architecture search

Abstract

In this thesis, we propose a new neural architecture search algorithm that performs network discovery in global search spaces. We introduce a novel network representation that organizes the topology on multiple hierarchical levels of varying abstraction and develop an evolution based search process that exploits this structure to explore the search space. Our approach involved a curation system that selects well performing network components and uses them in subsequent generations to build better networks. Next, we investigate how the proposed method performs on different types of data. First, we apply our method on an activity recognition time series dataset and manage to discover a topology with impressive performance. We also test the method on two image classification datasets, Fashion-MNIST and NAS-Bench-101 and achieve accuracies of 93.2% and 94.8% respectively in a small amount of time.