|
Title
|
|
|
|
Predicting image classifier performance using the synthetic petri dish method
| |
|
Author
|
|
|
|
| |
|
Abstract
|
|
|
|
| An important part of any Neural Architecture Search (NAS) system is a good performance estimation strategy. This strategy should ideally have a high accuracy, a low initialization time and a low query time. This paper evaluates the Synthetic Petri Dish method of neural network performance estimation in an image classification context. The method is tested on the CIFAR-10 dataset using the data from the NASBench-101 dataset as ground truth. We examine different ways of constructing motifs and analyze the influence of these motif choices on the final performance. While the motif networks are capable of converging on the synthetic dataset, the outer loop losses show little improvement throughout the training process. |
| |
|
Language
|
|
|
|
English
| |
|
Source (book)
|
|
|
|
34th Benelux Conference on Artificial Intelligence and the 31 Belgium Dutch Conference on Machine Learning (BNAIC/BENELEARN 2022), 7-9 November, 2022, Mechelen, Belgium
| |
|
Publication
|
|
|
|
2022
| |
|
Volume/pages
|
|
|
|
p. 1-16
| |
|
Full text (open access)
|
|
|
|
| |
|