Acceleration of Neural Architecture Search by Policy Gradient with Credit Assignment

Abstract

Neural architecture search is gathering attention as an approach to automatically designing the architecture of deep neural networks. Aiming at accelerating neural architecture search without compromising its performance, we propose a novel one-shot architecture search algorithm that optimizes the weights and the architecture of a network simultaneously. Our algorithm is inspired by the notion of credit assignment in policy gradient. An advantage of each architecture component is defined and the gradient of the architecture parameter is computed using the advantage values. In our experiments, we observe that the proposed method achieves, with lower computational time, the final performance comparable to recently proposed gradient-based one-shot architecture search algorithms.