NME130/Graphical models

Present: Andreas, Andy, Ufuk, Nader, Richard, John, Pablo, Javad

• Copy of Andreas's slides

CS/CNS/EE 155

• Course will be taught for the first time this fall; looking for feedback
• Key theme in graphical models: global insight from local observations
• Bayesian perspective; modeling, inference, learning
• Syllabus
  • Modeling and representation - 5 lectures
  • Inference - 6 lectures
  • Learning GMs from data - 5 lectures
  • Applications and case studies - 2 lectures
  • Current research directions 3 lectures
• Pre-requisites: machine learning (statistical models, generalization, distributions)

Connections to NME 130

• Information theory and coding (turbo codes, belief propogation)
• Stochastic optimal control (MDPs, POMDPs -> controlled Markov chains, HMMS)
• Dynamical systems (inference in hybrid systems, filtering)
• Robustness (Bayesian model averaging, reasoning about very uncertain data)
• Synthesis/hard limits (eg, how hard are certain decision problems)

Module in 6 lectures (???)

• GMs = probability/statistics meets algorithms/optimization
• Modeling - 2 lectures: factorization, structured distributions, factor graphs
• Inference - 2 lectures: inference as optimization, samping (Gibbs, MCMC)
• Learning - 2 lectures: parameter, structure

Discussion

• Pablo: much of the structure is purely algebraic; not really stochastic. Graphical structure allows you to be efficient. So lines up well with optimization.
• John: what might make sense for next year - take a set of classes that are going to be taught anyway. Some subset of people take all of those courses together. Then an additional course that everyone takes together (CDS 212/213).
  • Stochastic modeling: ACM 116 (f), ACM 216 (w), EE 156 (f), CS 155 (w)
  • Optimization: ACM 113 (w)
  • Controls: CDS 210 (f)
  • Networking: CS 2xx (s)