École Nationale Supérieure de
Techniques Avancées

Abstract

The development of deep learning has attracted more and more attention, because they can all be applied to a variety of tasks, such as image recognition, image segmentation etc. However, finding novel neural networks for some specific tasks developed manually with expert knowledge is a difficult task. Also, it is a time-consuming process. Thereby, it is crucial to study automated neural architecture search methods (NAS). In this paper, we study the application of morphological dilation and erosion operations on the Differentiable Architecture Search (DARTS) and Architecture Optimization algorithms (NAO) and we achieved better accuracy than the original author on CIFAR 10 dataset. In addition, in this work, we investigate morphological operations and NAO on a semantic segmentation task by constructing an architecture inspired by U-Net.