Zero-Shot Evaluation Index Based on Robustness of CNN Output

Abstract

Neural Architecture Search (NAS), which aims to automatically optimize the structure of a neural network for achieving excellent classification performance, has attracted considerable attention in recent years. Recently, zero-shot evaluation methods have been proposed for estimating classification performance without training to reduce the search time. However, these indices are still insufficient for finding the best-performing neural networks. In this study, we demonstrate that it is possible to evaluate convolutional neural networks (CNNs) using the robustness of the rectified linear unit (ReLU) output distribution to weights. We propose a new zero-shot CNN evaluation index based on this robustness index.