6.438 Algorithms for Inference
Fall 2009
Carl Friedrich Gauss and Andrey Andreyevich Markov
Instructors: William T Freeman, Gregory W Wornell
TAs: Jin Choi, Matthew James Johnson
Lecture:
TR9.30-11
(32-124)
Recitation: F10 or F2
(36-372)
Information:
Introduction to statistical inference with probabilistic
graphical models.
Directed and undirected graphical models; factor graphs;
Gaussian
models. Hidden Markov models, linear dynamical systems.
Sum‐product and junction tree algorithm. Forward‐backward
algorithm;
Kalman filtering and smoothing. Variational methods, mean‐field
theory,
and loopy belief propagation. Particle methods and filtering.
Min‐sum
algorithm; Viterbi algorithm. Building graphical models from data;
parameter
estimation, learning structure. Selected special topics.
Announcements
Prob 9.2 Solution
Hi class,
I corrected the solution of Problem 9.2(d) in PS9. The denominator should be (26 + N(x1="E")), not (1+N(x1="E")) as in the previous version.
Jin
Announced on 07 December 2009 7:18 p.m. by Jin Choi
Extra Office Hours on Wednesday
Good luck studying for Thursday's quiz!
Matt
Announced on 07 December 2009 4:42 p.m. by Matthew James Johnson
EM examples posted
Oops, sorry - it doesn't show up if you select "Homework" from the menu. Go to "Materials" and find "Notes on EM" under the "Problem Sets" category.
Jin
Announced on 04 December 2009 9:37 p.m. by Jin Choi
EM examples posted
Hi all,
As I promised in recitations, I've posted notes with detailed derivations of the EM algorithm for the Naive Bayes and the homogeneous HMM (Baum-Welch). You can find it under the "Homework" category.
Jin
Announced on 04 December 2009 9:27 p.m. by Jin Choi
Baum-Welch Algorithm Reading
Hi 6.438'ers--In lecture today, I was more rushed at the end than I'd hoped, and thus skipped some details of our derivation of the Baum-Welch parameter estimation algorithm for HMMs. Since you will be using this algorithm in the homework, you might find a little additional reading helpful. For this I would recommend Section 12.8 of Jordan's notes. In anticipation of this, I used similar notation to Jordan's in class, so you should find it fairly easy to follow.
The pace of the past three lectures has been rather aggressive due to the problem set schedule. However, the pace should return to something more typical in the remaining lectures.
cheers,
Greg
Announced on 01 December 2009 5:28 p.m. by Gregory W Wornell
