Summary
International Symposium on Nonlinear Theory and Its Applications
2022
Session Number:C3L-C
Session:
Number:C3L-C-03
Proposal of a New Zero-Shot Evaluation Index for Simple CNN
Chisato Takahashi , Kenya Jin'no,
pp.478-481
Publication Date:12/12/2022
Online ISSN:2188-5079
Summary:
Network Architecture Search (NAS), which aims to optimize the structure of neural networks themselves, has attracted much attention in recent years. The evaluation of the structure of a neural network in NAS is basically performed by actually training the neural network and measuring its performance. However, this method requires an enormous amount of computation. For this reason, the other zero-shot method that evaluates the structure without actually performing the training has begun to be proposed. The ultimate goal of this research is to create an evaluation index that can evaluate the structure in a zero-shot manner for NAS. In this article, we experimentally investigate the relationship between basic CNN structures and their performance, we create an index that can measure performance in a zero-shot environment.