The Case for Learning-and-System Co-design

Abstract

While decision-makings in systems are commonly solved with explicit rules and heuristics, machine learning (ML) and deep learning (DL) have been driving a paradigm shift in modern system design. Based on our decade of experience in operationalizing a large production cloud system, Web Search, learning fills the gap in comprehending and taming the system design and operation complexity. However, rather than just improving specific ML/DL algorithms or system features, we posit that the key to unlocking the full potential of learning-augmented systems is a principled methodology promoting learning-and-system co-design. On this basis, we present the AutoSys, a common framework for the development of learning-augmented systems.