A Tale of EDA's Long Tail: Long-Tailed Distribution Learning for Electronic Design Automation

Abstract

Long-tailed distribution is a common and critical issue in machine learning. While prior work addressed data imbalance in several electronic design automation tasks, insufficient attention has been paid to long-tailed distributions in real-world EDA problems. This paper shows that conventional performance metrics can be misleading in EDA contexts and demonstrates, through public EDA tasks with convolutional and graph neural networks, that simple model-agnostic methods can effectively alleviate long-tail induced performance degradation.