Course positioning
CS 228 is the classic PGM course, covering Bayesian and Markov networks, HMMs, exact and approximate inference, and parameter/structure learning. Its way of encoding uncertainty with graph structure has long-term value for NLP, computational biology, causal inference, and robotics.
What it covers
- Representation: Bayesian networks, Markov networks, factor graphs
- Exact inference: variable elimination, belief propagation, junction trees
- Approximate inference: variational inference, MCMC, importance sampling
- Parameter learning: MLE, EM, latent-variable models
- Structure learning: score-based search, constraint-based methods
- Temporal models: HMMs, dynamic Bayesian networks
- Applications: speech, biology, medical diagnosis, vision, robot planning
Prerequisites
Basic probability; algorithm design and analysis.
Assessment & workload
Assignments plus a midterm and a project; a mix of math and implementation. Check the current syllabus for details.
Who it's for
People working on probabilistic modeling, causality, or Bayesian deep learning; researchers in NLP, computational biology, or robotics who need to handle uncertainty.
Study advice
It complements CS 229: 229 is prediction-oriented, 228 is about structured uncertainty. Derive variable elimination and VI by hand once.
Related courses
Official course description
Probabilistic graphical modeling languages for representing complex domains, algorithms for reasoning using these representations, and learning these representations from data. Topics include: Bayesian and Markov networks, extensions to temporal modeling such as hidden Markov models and dynamic Bayesian networks, exact and approximate probabilistic inference algorithms, and methods for learning models from data. Also included are sample applications to various domains including speech recognition, biological modeling and discovery, medical diagnosis, message encoding, vision, and robot motion planning. Prerequisites: basic probability theory and algorithm design and analysis.