The BRADLEY DEPARTMENT of ELECTRICAL and COMPUTER ENGINEERING

ECE 6424 Probabilistic Graphical Models and Structured Prediction | ECE | Virginia Tech

Graduate PROGRAMS

Course Information

Description

Advanced concepts in machine learning; focus on probabilistic graphical models and structured output prediction. topics include directed models (Bayes Nets), undirected models (Markov/Conditional Random Fields), exact inference (junction tree), approximate inference (belief propagation, dual decomposition), parameter learning (MLE, MAP< EM, max-margin), structure learning.

Why take this course?

We are witnessing an explosion in data - from billions of images shared online to Petabytes of tweets, medical records and GPS tracks, generated by companies, users and scientific communities. Unfortunately, real-world data are often noisy, ambiguous, and exhibit nontrivial or complex structure. Probabilistic and structured models provide a principled framework for dealing with uncertainty and for converting evidence from multiple nosy sources into a posteriori belief about the world.

Prerequisites

A grade of C- or better in 5424G or CS 5824.

5424G and CS 5824: basic concepts of statistical learning I(likelihood, maximum likelihood, bias and variance, underfitting, overfitting, regularization, cross-validation); supervised and unsupervised learning algorithms.

Major Measurable Learning Objectives

  • Analyze and contrast board classes of probabilistic graphical models (Bayes Nets vs. Markov Nets; directed vs. undirected graphical models)
  • State and describe independence assumptions encoded by different models
  • Explain and implement algorithms for learning and inference (e.g. belief propagation)
  • Discuss and critique research papers on these topics
  • Identify open research questions in these areas

Course Topics

Topic

Percentage of Course

1. Bayesian Networks: a) Representation/Semantics; b) Parameter & Structure Learning; c) Inference 30%
2. Markov Random Fields: a) Representation/Semantics; b) Exact Inference; c) Approximate Inference; d) Parameter Learning 40%
3. Structured SVMs: a) Structured Hinge Loss; b) Cutting-Plane optimization (n-slack, 1-slack); c) Online optimization (Frank-Wolfe method) 30%