# Best of Both Worlds: AutoML Codesign of a CNN and its Hardware Accelerator

**Authors:** Mohamed S. Abdelfattah, Łukasz Dudziak, Thomas Chau, Royson Lee, Hyeji Kim, Nicholas D. Lane  
**arXiv:** [2002.05022](https://arxiv.org/abs/2002.05022)  
**Date:** 2020-02-11  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2002.05022v2/

## Key points

- Neural architecture search (NAS) is a technique for designing CNNs that outperform human-designed models in accuracy.
- NAS-designed CNNs often have complex topologies, making it difficult to develop custom hardware accelerators for them.
- The authors propose Codesign-NAS, which automates the process of hardware-CNN codesign using NAS.
- Codesign-NAS aims to find the optimal model-accelerator pair that improves both accuracy and efficiency.
- The authors define the multiobjective optimization problem of Codesign-NAS and explore different strategies for navigating the codesign search space.
- They conduct experiments on CIFAR-10 image classification and enumerate nearly 4 billion model-accelerator pairs to identify the Pareto frontier within this vast search space.
- Three reinforcement-learning-based search strategies are employed to evaluate their approach.
- Results show a 1.3% improvement in CIFAR-100 classification accuracy compared to ResNet with its most optimal hardware accelerator, along with a 41% increase in performance/area.
- These improvements are achieved with just around 1000 GPU-hours of running Codesign-NAS.
- This study presents a novel approach for automating hardware-CNN codesign using NAS, highlighting its potential in enhancing both accuracy and efficiency while reducing manual effort.

## Summary

Neural architecture search (NAS) has emerged as a powerful technique for designing convolutional neural networks (CNNs) that outperform human-designed models in terms of accuracy. However, NAS-designed CNNs often have complex topologies, making it challenging to develop custom hardware accelerators for these networks. In this paper, the authors propose an approach called Codesign-NAS, which automates the process of hardware-CNN codesign using NAS. By considering parameters from both the CNN model and the hardware accelerator, they aim to find the optimal model-accelerator pair that improves both accuracy and efficiency. The authors focus on defining the multiobjective optimization problem of Codesign-NAS and explore different strategies for navigating the codesign search space. They conduct experiments on CIFAR-10 image classification and enumerate nearly 4 billion model-accelerator pairs to identify the Pareto frontier within this vast search space. To evaluate their approach, they employ three reinforcement-learning-based search strategies. The results demonstrate the effectiveness of Codesign-NAS. Compared to ResNet with its most optimal hardware accelerator within their design space, they achieve a 1.3% improvement in CIFAR-100 classification accuracy while simultaneously increasing performance/area by 41%. Impressively, these improvements are obtained with just around 1000 GPU-hours of running Codesign-NAS. Overall, this study presents a novel approach for automating hardware-CNN codesign using NAS. The findings highlight its potential in enhancing both accuracy and efficiency in CNN models while reducing manual effort in designing custom hardware accelerators.