15-854 Presentations and Projects
Here is a list of presentation and project topics that
come to mind. It is not an exhaustive list: if somthing else
strikes you, come talk with me. Roughly, a
presentation is expected to require reading a research paper and
then discussing it in class. A project might involve programming
up and testing some algorithms, or it could be something larger
and more open-ended. In any case, come talk with me before
committing your presentation or project. If you do a project,
then you would be expected to give a short (5-15 min) presentation
of it in class.
Ideas for presentations:
- Expansion on "sample complexity" and VC-dimension discussion. For
example, maximizing the "margin", probabilistic concepts, or
distribution-specific or bayesian refinements.
- MDPs, Q-learning.
- the EM algorithm and hidden Markov Models
- Other topics in on-line learning
- Further analysis of decision tree or neural net learning.
Ideas for projects:
- Set up a program to repeatedly play some 2-player zero sum game
against a human opponent. Try out several algorithms,
including weighted-majority with "sharing".
- Implement a multi-layer version of winnow. (make negated copies of
inputs so effectively have negative weights too, and then can use
threshold of 0). How to train...?
Ideas for presentations, more specifically:
Note: I've attempted to choose the most readable papers I could
find on each topic. But, it's probably worth
your while looking up other papers on the same topic (e.g., see
the references in these papers) in case there is another that's
- Relations between decision tree algorithms and
- Models/algorithms for learning probabilistic functions:
- Online portfolio selection:
- Extensions to VC-dimension: bounds that depend on the "amount" of separation acheived (aka the "margin"):
- Extensions to VC-dimension: bounds that depend on the input distribution and "error shells":
- Extensions to VC-dimension: bounds based on metric entropy:
- Exploring an unknown environment:
Last modified: Wed Mar 11 16:05:22 EST 1998