CS839 Probabilistic Graphical Models

COMP SCI 1325 on TuTh 2:30-3:45pm

Description

The class will cover topics such as Directed/Undirected graphical models, template models, Inference (variable elimination and sum-product message passing), Learning (Maximum Likelihood Estimation, Generalized Linear Models, learning over fully/partially observed data etc.), approximate inference (MCMC methods, Gibbs sampling). We will also dive to more research oriented topics such as scalable implementations of graphical models, connections of graphical models and relational representation learning, and applications of graphical models to problems in data management (such as data integration and data cleaning).

Class Logistics

Text books

Prerequisites

  • You should have taken an introductory machine learning course. You should understand basic probability and statistics, and college-level algebra and calculus. For example it is expected that you know about standard probability distributions (Gaussians, Poisson), and also how to calculate derivatives.

Assignments

  • There will be three homework assigments. All homework assignments should be turned in electronically via Canvas.
  • There will be one midterm exam
  • A research class project: You will need to form groups of up to three people. The project will be broken down to three assignments: (1) written initial proposal, (2) proposal presentation, (3) final report.
  • There will be no final exam.

Misc

  • Class time may be adjusted to accomodate external talks related to the class.

Tentative Lecture Plan (Subject to Change)


# Date Topic Lecture Materials Reading Material Assignments
Introduction and Class Overview
1 9/6 Introduction to Graphical Models Lecture 1
Representation
2 9/11 Directed Graphical Models: Bayesian Networks Lecture 2
3 9/13 Undirected Graphical Models Lecture 3
Exact Inference
4 9/18 Variable Elimination Lecture 4
  • PGM textbook, Ch. 9
5 9/20 Clique Trees and Message Passing Lecture 5
Learning
6 9/25 Learning over Generalized Linear Models Lecture 6

Homework 1: Due Oct 2nd by 2:30 p.m. (beginning of class)

7 9/27 Learning BNs Lecture 7
  • PGM textbook, Ch. 17
8 10/2 Learning Undirected Graphical Models Lecture 8
  • PGM textbook, Ch. 20
9 10/4 Structure Learning Lecture 9
  • PGM textbook, Ch. 18, Ch.20.7
10 10/9 Learning with Partially Observed Data-The Expectation Maximization Algorithm Lecture 10
  • PGM textbook, Ch. 19, Ch.20.3.3
Approximate Inference
11 10/11 Loopy Belief Propagation Lecture 11

Homework 2: Due Oct 25th by 2:30 p.m. (beginning of class)

12 10/16 Mean Field Approximation Lecture 12
13 10/18 Monte Carlo Methods Lecture 13
14 10/23 Markov Chain Monte Carlo Methods Lecture 14
15 10/25 Review before Midterm Lecture 15
16 10/30 Midterm Midterm

Grading
Homework Assignments20%
Midterm Exam30%
Project Proposal10%
Project Proposal Presentation10%
Project Final Report30%
Office Hours

Theo: by appointment @ Room CS4361

Late Policy and Deliverables

You are not allowed any late dates for the project deliverables.

You will be allowed 2 total homework late days without penalty for the entire semester. You may be late by 1 day on two different homeworks or late by 2 days on one homework. Once those days are used, you will be allowed no late days.

Honor Code and Collaboration Policy

You are encouraged to discuss the homework assignments with other students; it's fine to discuss overall strategy and collaborate with a partner or in a small group, as both giving and receiving advice will help you to learn.

However, you must write your own solutions to all of the assignemtns, and you must cite all people you worked with. If you consult any resources outside of the materials provided in class, you must cite these sources.

If you do not do so, we will consider this a violation of the University of Wisconsin Honor Code.