Understanding and Explaining the Root Causes of Performance Faults with Causal AI: A Path towards Building Dependable Computer Systems
Speaker: Pooyan Jamshidi, University of South Carolina
In this talk, I will present our recent progress in employing Causal AI (Causal Structure Learning and Inference, Counterfactual Reasoning, and Transfer Learning) in addressing several significant challenges in computer systems. After motivating the work, I will show how mainstream machine learning, which relies on spurious correlations, may become unreliable in certain situations. Next, I will present empirical observations that explain the underlying root causes of performance faults in several highly-configurable systems, including autonomous systems, robotics, on-device machine learning systems, and data analytics pipelines. I will then present our framework, Unicorn, and discuss how Unicorn fills the gap by employing a causal reasoning approach. In particular, I will discuss how Unicorn captures intricate interactions between configuration options across the software-hardware stack and how such interactions can impact performance variations. Finally, I will talk about our 2-years journey in a NASA-funded project called RASPBERRY-SI, developing a causal reasoning approach to enable synthesizing adaptation plans for reconfiguring autonomous systems to adapt to environmental uncertainties during operation.
For more information regarding the technical work and the people behind the work that I present, please refer to the following websites:
- The Unicorn framework: [ Ссылка ]
- The NASA-funded RASPBERRY-SI project: [ Ссылка ]
