Dr Charles Sutton – University of Edinburgh
Machine Learning for Computer System Performance
Probabilistic modelling is central to modern statistical machine learning, because it allows us to connect prior knowledge to data and uncertainty. New applications of probabilistic modelling have the potential to not only be useful in their own right, but also to motivate new methodological and theoretical advances. In this talk I will discuss the problem of predicting the performance of large, distributed computer systems, motivated by distributed applications such as the data centre applications of Google, Yahoo!, and Amazon. I will describe how these applications raise interesting new problems for machine learning, and how a probabilistic modelling perspective is opening up new techniques for debugging these systems when their performance is poor. I will aim for the talk to be relevant not only to researchers in artificial intelligence and machine learning, but also to researchers in performance modelling and autonomic computing.
School of Computing, Robert Gordon University, St Andrew Street, Aberdeen, Lecture Room A12, 14:00 – 15:00.