Reproduce “Improving Routability Prediction via NAS Using a Smooth One-shot Augmented Predictor” on OpenResearch

About this paper

Arjun Sridhar, Chen-Chia Chang, Junyao Zhang, Yiran Chen

Duke University researchers developed SOAP-NAS (Smooth One-shot Augmented Predictor Neural Architecture Search), a framework for Design Rule Check (DRC) hotspot detection in integrated circuits. The method addresses the challenges of metric mismatch and high variance in ROC-AUC by integrating k-shot smoothing and data augmentation, achieving an ROC-AUC of 0.9802 while maintaining a fast query time of 0.461 ms.

Key insights

• ROC-AUC, the primary evaluation metric for DRC hotspot detection, exhibits significantly higher variance compared to accuracy, making direct optimization unstable for neural architecture search.
• Empirical evaluation across different search spaces revealed that NASBench101 is most suitable for DRC hotspot detection, yielding higher ROC-AUC scores and better performance correlation.
• Reducing ROC-AUC variance through k-shot smoothing (training multiple one-shot networks) and augmenting data to account for this variance significantly improves the correlation between predicted and actual architecture performance.

Summary by alphaXiv.