Abstract

Applying machine learning in Electronic Design Automation (EDA) and Test has received growing interests in recent years. In many applications, machine learning is employed in a flow which is largely domain-knowledge driven. In such a flow, the analyst (or the expert) plays a key part for its effectiveness and consequently, automation of the flow needs to automate the analyst layer as well. This automation essentially is asking machine to acquire and model the domain knowledge possessed by the analyst and with the knowledge, to perform the analyst part of the task. If such automation can be accomplished, the result will be an autonomous system for performing the particular engineering task. In this talk, I will discuss such an “AI” view for applying machine learning in EDA and Test. In this context, I will explain the inadequacy of adopting a traditional machine learning problem formulation view due to its various difficulties to achieve robustness. Then, I will present an alternative machine learning view for robust learning. The concept of “learnable” will be discussed between the two machine learning views. The theoretical and practical considerations for automating a machine learning based flow will be discussed in two selected applications, functional verification and production yield optimization, and experiment results will be presented based on industrial settings in practice.