Project requirements for CMPT 882-3:
Anoop Sarkar <anoop at cs.sfu.ca>

1. Based on readings done in this class (past and future). For other
   material talk to me first.

2. No re-implementations of an existing algorithm without
   modifications applied to a simple task. This is a grad seminar
   where you are expected to provide some state-of-the-art research. 

3. Try to build on previous implementations of an algorithm of
   interest. This is typically easy to do since you can sometimes
   simply take the output of an existing system and compare with your
   new and improved implementation (speed and accuracy considerations)
   or alternatively you can take the output of an existing system and
   do discriminative re-ranking (reduce its error rate post-hoc)

4. No joint projects. However, you can have parallel projects which
   are compared with each other in the end with some code sharing
   involved. There have to be two ideas/hypotheses in a parallel
   project. 

5. Use the standard data sets for the task that you choose to attack
   with a particular machine learning algorithm. Most of these are
   available from /cs/fac1/anoop/data or from links on the course web
   page. If you are in doubt as to what the dataset should be, send me
   email and I will put up a link on the course web page.

6. Artificial datasets are permitted only if you come up with a
   radically new learning algorithm or you solve some existing
   theoretical problem with a machine learning algorithm, e.g. if you
   come up with a new solution to parameter re-estimation in HMMs or
   provide an answer to the "label bias" problem in discriminative
   models.

7. No restriction on the programming language used (don't use
   Intercal, unlambda or other obfuscating languages)

8. Final project report should be no longer than 7 pages. The style
   files for LaTeX and Microsoft Word are linked on the course web
   page.