# XNAS: Neural Architecture Search with Expert Advice

**Authors:** Niv Nayman, Asaf Noy, Tal Ridnik, Itamar Friedman, Rong Jin, Lihi Zelnik-Manor  
**arXiv:** [1906.08031](https://arxiv.org/abs/1906.08031)  
**Date:** 2019-06-19  
**Summary page:** https://www.summarizepaper.com/en/paper/1906.08031v1/

## Key points

- Novel optimization method for differential neural architecture search
- Leveraging prediction with expert advice theory to minimize regret in architecture selection
- Dynamically eliminates inferior architectures while enhancing superior ones
- Avoids hard pruning of architectures
- Recommends multiple learning rates based on backward gradients
- Achieved impressive error rates of 1.6% for CIFAR-10 and 24% for ImageNet under mobile settings
- Outperformed existing techniques on three other datasets, establishing state-of-the-art results
- Niv Nayman, Asaf Noy, Tal Ridnik, Itamar Friedman, Rong Jin, and Lihi Zelnik-Manor
- Incorporates expert advice into the optimization process
- Enhances model performance and efficiency in complex tasks like image classification

## Summary

The paper "XNAS: Neural Architecture Search with Expert Advice" introduces a novel optimization method for differential neural architecture search. This approach leverages the theory of prediction with expert advice to minimize regret associated with sub-optimal choices of operations in architecture selection. Unlike previous methods that require hard pruning of architectures, XNAS dynamically eliminates inferior architectures while enhancing superior ones. It also achieves an optimal worst-case regret bound and recommends the use of multiple learning rates based on backward gradients. Experimental results demonstrate its effectiveness across various image classification datasets, achieving impressive error rates of 1.6% for CIFAR-10 and 24% for ImageNet under mobile settings. Additionally, XNAS outperforms existing techniques on three other datasets, establishing state-of-the-art results in those domains. Authored by Niv Nayman, Asaf Noy, Tal Ridnik, Itamar Friedman, Rong Jin, and Lihi Zelnik-Manor, this research presents a significant advancement in neural architecture search methodologies by incorporating expert advice into the optimization process. The findings showcase the potential of this approach to enhance model performance and efficiency in complex tasks such as image classification.