Professor Harry Wechsler
Department of Computer Science
e-mail : wechsler@cs.gmu.edu
web : http://cs.gmu.edu/~wechsler/
(703) 993-1533 (office)
(703) 993-1530 (sec)
(703)993-1710 (fax)
SPRING '2005
CS
844 --- PATTERN
RECOGNITION
001 13474 T
[except
on March 29 and April 5 when the class is held in ST2 260]
IT 844 --- PATTERN RECOGNITION
001 11307
T
[except on March 29
and April 5 when the class is held in ST2 260]
Office Hours
T
Textbook
1. A. Webb, Statistical Pattern Recognition
(2nd ed), Wiley, 2004.
References
1. C. Bishop, Neural
Networks for Pattern Recognition,
2. V. Cherkassky and
F. Mulier, Learning from Data
: Concepts, Theory, and Methods,
Wiley, 1999.
3. N. Cristianini and J. Shawe-Taylor, An
Introduction to Support Vector Machines,
4. R. Duda, P. Hart
and D. Stork, Pattern Classification, Wiley, 2002.
5. S. Haykin, Neural Networks, Prentice-Hall, 1999.
6. B. Scholkopf and A. J. Smola, Learning with Kernels :
Support Vector Machines, Regularization, Optimization, and Beyond, MIT
Press, 2002.
7. J. Shawe-Taylor and N. Cristianini, , Kernel Methods for Pattern Analysis,
8.
V. Vapnik, Statistical Learning Theory, Wiley, 1998.
Course
Description
The course covers the Statistical Pattern Recognition (SPR),
the Neural Network (NN), and the Statistical
Learning Theory (SLT) approaches
for Pattern Recognition (PR) Topics include decision theory and Bayes theorem,
density (parametric and non-parametric)
estimation, linear (MSE and LMS) and non-linear
discriminant analysis, SVM (support vector machines) and kernel
methods, SRM (structural risk
minimization) and model selection, performance
evaluation, mixture of experts (AdaBoost),
feature selection and extraction, clustering. Experimental design,
applications,
and performance evaluation are emphasized
throughout the course.
Schedule
1st day of classes:
Spring Break:
Last Day of Classes:
Grading
1. Homework Assignments: 50 % (#1: 12.50; #2: 12.50; #
assignment #1 -- due February 15 -- elementary decision
theory (see Sect. 1.5.1 in
textbook pp. 6 16) includes {Bayes
minimum error and risk, reject option, and
Neyman Pearson decision rule} task: roll [enough time] two pairs of dice, one
pair fair (1 6) and
one pair not fair (3 8); roll one pair at a time and
report the sum x = s for the faces of two dices. for
the fair pair s = 2 ..12 while for the non-fair pair s
= 6 .. 16. The decision required is to guess
in an optimal fashion the pair rolled. Simulate (on
computer) the task,
duplicate the derivations found in
Sect. 1.5.1 to minimize error and/or risk, and graph the results
accordingly.
assignment #2 -- due February 22 EM Consider
three Gaussians pdfs N (1.0, 0.1), N (1.5, 0.1) and
N (2.0, 0.2). Generate 500 samples according to the following rule. The first
two samples are generated from the 2nd Gaussian, the 3rd
sample from the 1st Gaussian, and the 4th sample from the
last Gaussian. This rule repeats until all 500 samples have been generated. The
pdf underlying the random samples is modeled as a
mixture SUM (i = 1..3) N (m(i), sigma**2(i)) P(i). Use the EM algorithm on the generated samples to
estimate the unknown parameters {m(i),
sigma**2(i) and P(i)}.
Display/graph the three (3) original Gaussians and the mixture approximation.
Discuss your results in terms of accuracy and convergence (number of steps).
assignment #3 -- due March 29 (Data) Structure < Representation > and
Algorithm
<Classifier> - Use 2 (two) Data Sets to Assess Different
{representations, classifiers} combinations in Terms of Overall Performance for
Multi-Class (c > 2) Classification. First data set should be 2D to allow for
display and visualization, while the second data set should be
multidimensional. Representations
include {PCA <principal component analysis>, LDA <linear discriminant analysis>, NPLDA <non-parametric LDA>,
EPP <exploratory projection pursuit> / EP <evolutionary pursuit [Liu
and Wechsler, IEEE on PAMI, 2000]}. Classifiers include {RBF <radial basis
functions>, BP <back propagation>}. You can include additional
representations, e.g., MDS <multidimensional scaling> and/or classifiers,
e.g., SVM <support vector machines>. Define briefly each of the methods
used, experimental design regarding data acquisition and software used, comparative
(using raw data vs. transformed data) performance observed, limitations and sensitivity {to parameter choice and/or
noise}, and conclusions. Power point presentation expected. The presentation should
be made available to the other students on the web.
3.
PROJECT: 50%
- (3.1) Literature Survey, (3.2) Method, (3.3)
Experimental Set Up {data acquisition + software}, Results, and Performance
Evaluation, (3.4) Analysis and Future, and (3.5) Class Presentation [ April 26 and May 3].
Topic and Scope for the project to be agreed with the
instructor.