A Compact Neural Architecture Search for Accelerating Image Classification Models

Nowadays, Automated Machine Learning (AutoML) has gradually become an inevitable trend providing automatic and suitable solutions to address AI tasks without needing more efforts from experts. Neural Architecture Search (NAS), a subfield of AutoML, has generated automated models solving fundamental problems in computer vision such as image recognition, objects detection. NAS with differentiable search strategies has reduced significantly the GPU time that occupancy on calculation. In this paper, we present an effective algorithm that allows expanding search spaces by selecting operation candidates from the initial set with different ways in concurrent execution. The extended search space makes NAS having more opportunities to find good architectures simultaneously by running the group of search spaces in overlapping time periods instead of sequentially. Our approach, is called Accelerated NAS, shortens 1.8x searching time when comparing to previous works. In addition, the Accelerated NAS generates potential neural architectures having comparable performances with the low inference time.

I. Introduction

Image classification is one of various fundamental tasks with a variety of image datasets in computer science, especially computer vision factors. During the artificial intelligence era, manual deep neural network (DNN) reached state-of-the-art about accuracy performance by stacking deeper layers like ResNet [1] or applying squeeze-excitation such as SE-Net [2]. Existing DNN models have required a lot of technical knowledge from experts with a long duration for the trial phase. Nowadays, NAS has experimented the various improvements in searching the adaptive models by changing the approach with automated building instead of manual design. The automated architecture was optimized for accuracy under efficient constraints (e.g., memory, searching time, or latency). NAS has been applied to numerous tasks in computer vision and brought the optimal performances on image classification [3], generative adversarial network (GAN) [4], object detection section [5].