March 3: students should submit an initial list of the five
papers they would most prefer to present (ranked in
order)
March 5: students should submit a one to three page description
of an implementation project they propose to undertake,
this description should indicate the mechanism the student
proposes to implement, a timetable for implementation,
and indicate (concretely) how they plan to test their
method
March 12: students should submit (if needed) a revised description
of their implementation project
April 14: students should submit a three page mid-project description
of their implementation projects
May 12: final report for term projects due. This report should have
an overall writeup of your solution and background information
(minimum 3, maximum 10 pages). It should also include a
discussion of experiments you did for validation (2-5 pages),
plus the actual code and print out of a representative set
of the output.
Talk Stuff:
A short list of suggestions on preparing your talk.
Implement a significant subunit to be added to the UMD ML code
library such as:
A decision tree module with multiple gain functions, the
ability to deal with unknown values, the ability to deal
with continuous functions and multiple pruning methods
A neural network method including an arbitrary network
format and the ability to constrain the network weights
in various ways
A support vector machine learner
Implement and test a genetic algorithm / neural network hybrid
Implement a known algorithm to extract understandable knowledge
from a trained neural network
Adapt an existing ILP system to be used in an ensemble learner (such
as bagging or boosting).
Build a visualization method for an SVM package such as SVM-Light that
would allow users to see (understand) the SVM decision surface.
Implement several bi-clustering algorithms as Java applets and compare
their results.
Suggested Papers to Present:
Functional Gradient Techniques for Combining Hypotheses, Mason,
Baxter, Bartlett and Frean, in Advances in Large Margin Classifiers,
Smola et al. (eds.), MIT Press