Title: Building Dependable Computer Systems with Modular Learning and Causal Reasoning

Abstract: We are witnessing a paradigm shift in AI system design, moving from monolithic models to modular-composed systems with multiple interdependent components. This shift raises new challenges in exploring a vast design space and making real-time multi-objective optimization decisions in uncertain environments. In this talk, I will discuss our recent work on Modular Learning and Optimization and Causal Reasoning to address these challenges, focusing on scalability, accuracy, and reliability in building dependable AI systems. I will share insights from our empirical studies on performance faults in various systems and highlight our innovative solutions, including Unicorn, CAMEO, CaRE, CURE, FlexiBO, InfAdapter, IPA, and Sponge. These solutions leverage modular learning, multi-objective Bayesian optimization, transfer learning, and causal reasoning for building efficient and reliable systems. Finally, I will share my research vision toward building AI accelerators for the next generation of modular-composed AI systems through strategic software-hardware co-design, specifically targeting Large Language Model (LLM) inference pipelines.