| DATE | Lecture | Topic | Scribe |
| 4/3/2002 | Lecture 1/pdf | Introduction, notation, pattern recognition,inference in pattern recognition, intro to graphical models, taxonomy | Salvador Ruiz Correa | 4/5/2002 | Lecture 2/pdf | DGMs, Independence, Cond. Indep, Canonical Bayes nets, Bayes-ball | Chris Bartels | 4/10/2002 | Lecture 3/pdf | Bayes ball algorithm and examples, graphical models describe families, the general problem of inference, the basics of "elimination" seen algebraically and graphically. | Arindam Mandel | 4/11/2002 | Lecture 4/pdf | GMs & HMMs, Undirected models (UGMs) and their representations, Start Linear Gaussian Models | Jeff Weinschenk | 4/17/2002 | Lecture 5/pdf | recap linear Gaussian models, steepest descent, simple convergence analysis, probabilstic interpretation of LMS, why linear Gaussian models are useful, simple multi-variate and econometric generalizations, PCA, | Manyuan Shen | 4/19/2002 | Lecture 6/pdf | factor analysis, static Kalman models and ML estimation, learning with hidden variables, begin EM algorithm. | Huaning Niu | 4/25/2002 | Lecture 7/pdf | Overview of alternating minimization view of EM, EM and DGMs, Finite Mixture models, Start Gaussians as DGMs and UGMs | Xiao Li | 4/26/2002 | Lecture 8/pdf | More on multi-variate Gaussian as graphical models, sub-partitions of Gaussian vectors, Gaussians as undirected models, Gaussians as directed models, from mixture models to hidden Markov models. | Gang Ji | 5/1/2002 | Lecture 9/pdf | Hidden Markov Models and filtering, prediction, smoothing. HMMs and EM. The Kalman model (i.e., Linear Gaussian HMMs). | Richard Oxford | 5/3/2002 | Lecture 10/pdf | LG-HMMs, the filtering problem, the smoothing problem, Rauch-Tung-Striebel (RTS) Recursion, conditional expectation, extensions of factor analysis (FA), mixtures of FA, independent FA, independent component analysis (ICA), LDA/HLDA/MDA/HMDA. Begin of review of Lauritzen terminology and notation for graphical models for their formal treatment. | Costas Boulis | 5/8/2002 | Lecture 11/pdf | Formal notation and terminology. Decomposable, triangulated or chordal, start equivalent conditions for graphs proof (decomposable == chordal == every minimal alpha-beta sep is complete). | Richard Oxford | 5/9/2002 | Lecture 12/pdf | More on decomposable graphs, proof that if every min alpha-beta seperator is complete, then we have a chordal graph. Junction tree definition, junction tree exists if and only if graph is chordal, running intersection property, junction tree construction. | Handsang Cho | 5/10/2002 | Lecture 13/pdf | Graph theory and running intersection property review, constructing a junction tree from cliques in r.i.p. order. perfect DAGs and numberings and chordality, simplicial nodes, fully explained version of maximum cardinality search, finding the cliques of a chordal graph in r.i.p. order using maximum cardinality search. | Arindam Mandel | 5/22/2002 | Lecture 14/pdf | Look-ahead triangulation, properties of triangulation and elimination, triangulation heuristics, Junction trees and maximum spanning trees over the graph of cliques | Karim Filali | 5/23/2002 | Lecture 15/pdf | Conditional independence, properties of conditional independence the 5 axioms, conditional independence analogy, graph semantics, Markov properties on graphs, local Markov property (L), pairwise Markov property (P), global Markov property (G), proof that (G) implies (L) implies (P), and other implications if positivity holds. | Chris Bartels | 5/24/2002 | Lecture 16/pdf | Conditional independence and factorization, theorem about factorization and the global Markov property on undirected graphs, Möbius Inversion Lemma, The Hammersley and Clifford Theorem, factorization and decomposability. | Jeff Weinshenk | 5/29/2002 | Lecture 17/pdf | Formal properties on directed graphs, e.g., directed global Markov property, etc. Begin of general inference on graphs. More on moralization. | Salvador Ruiz Correa | 3/30/2002 | Lecture 18/pdf | Moving towards inference, indroduction of evidence, clique potentials and the goal of potentials as marginals, local consistency of clique potentials. | Huaning Niu | 5/31/2002 | Lecture 19/pdf | - | Karim Filali | 6/5/2002 | Lecture 20/pdf | Junction trees and message passing, putting it all together using HMMs as an example. Begin structure learning, scientific modeling, graphs that are Markov equivalence (definitions), | Costas Boulis | 6/6/2002 | Lecture 21/pdf | - | Salvador Ruiz Correa | 6/7/2002 | Lecture 22/pdf | Naive Bayes Classifiers, discriminative score functions, TAM models (tree augmented Naive Bayes Classifiers), the idea of structural discriminability, bayesian multinetworks and TAM, learning BNs with unknown structure and missing data, SEM algorithm. | Jay Kim |