Project for Machine Learning (308-761A)
Fall 2001
Goals
The main goal of the project is to allow you to study more in-depth
machine learning algorithms of your choice. A secondary goal is to
help you develop your research skills (reviewing research literature,
formulating questions, deciding on a theoretical or empirical approach
to answer them, writing up you findings and presenting them to your
peers). Ideally, the project topic you choose should be related to
your research interests.
There are three main categories of projects:
-
Applying machine learning to interesting problem domains, thereby
allowing you to gain in-depth experience with specific algorithms.
-
Studying machine learning algorithms not covered in class.
-
Investigating theoretical issues.
Ideally, the projects would be individual. However, if the topic is
very big, teams of two people would be accepted too. If you choose to
team up, there should be a clear delimitation between the work of that
each person does. Both students should turn in separate write-ups and
each should present a part of the work.
I strongly encourage you to perform some experiments as part of your
project, in order to gain practical experience with machine learning
algorithms. Many algorithms are available free from universities and
research labs, so you would not have to code them. The resources web
page should help you find both papers and software as needed.
Project themes
If you are already have a research topic in mind, feel free to propose
it. Otherwise, you can choose a topic from the list below. If no
topic suits your interest, please make an appointment to discuss
alternatives.
Proposed topics:
-
Machine learning for bioinformatics tasks
A good starting point for problems is Baldi and Brunak's book,
"Bioinformatics: The machine learning approach". The exact topic
depends on the data you have available here at McGill.
-
Machine learning for compiler optimization
Possible applications here include function inlining, instruction
scheduling, register allocation.
-
Comparing machine learning methods for face/object recognition
This is a topic that has been explored a lot. The main idea would be
to run an empirical study, mainly involving new algorithms
(e.g. ensembles, support vector machines).
-
Learning in active vision
The problem of active vision investigates where the focus of attention
should be in order to gather as much information as possible about the
task at hand. Reinforcement learning and Bayesian methods are
adequate for this task.
-
Links between machine learning and data compression
Machine learning algorithms can be viewed as compressing data into a
more compact representation. A theoretical project on this topic
would look at the relationship between Kolmogorov complexity and the
difficulty of machine learning tasks. An empirical project could look
into similarities between compression and learning algorithms.
-
Recurrent neural networks and their applications
Recurrent neural networks are typically used for processing time
series data. They use the outputs of units at time t as inputs for
other units at time t+1. They are briefly mentioned in Tom Mitchell's
book. A project could involve comparing different learning algorithms
for recurrent nets, and/or applying one such algorithm to
"interesting" data.
-
Hidden Markov Models and applications
HMMs are stochastic models for time series data, widely used in speech
recognition, text processing, bioinformatics. I have a couple of
great tutorials by Rabiner on learning for HMMs. The project involves
studying these, then working on an application.
-
Effect of exploration strategies in reinforcement learning.
There is a great variety of exploration strategies in RL, varying from
very straightforward methods (such as Boltzmann exploration or
epsilon-greedy) to methods very well-motivated, such as Singh and
Kearns' E3 algorithm. The project would involve surveying and
comparing existing methods.
-
Effect of eligibility traces in reinforcement learning
Eligibility traces can significantly speed up reinforcement learning.
In class we discuss only one version (accumulating traces) but
several have been proposed. The project will especially involve
looking a variable eligibility trace parameters.
-
Actor-critic methods for reinforcement learning
Our RL lectures discuss value-based RL only. Actor-critic methods
use an explicit representation of a policy and a value function. They
have better theoretical guarantees when used with function
approximation. The purpose of the project will be to study and
experiment 3with these methods.
- Learning orderings and preferences
There are many applications in which it is desirable to order rather
than classify instances. Of course, the desired ranks can be treated
as class labels, but that can make learning much harder. Several
special-purpose algorithms have been proposed for learning an
ordering directly.
-
Models of on-line learning.
When we studied computational learning theory, we assumed all the data
was available at the beginning. On-line learning studies the scenario
in which data keeps accumulating over time. Avrim Blum has a nice
survey paper on on-line learning, that you can use as a starting point.
The project would also involve experimenting with an on-line learning
algorithm.
- Blind source separation
This is a good project for people interested in auditory perception.
The problem considers a signal obtained by the superposition of
several different sources (e.g. an audio signal with several people
talking at once). The issue is to identify each source. Bayesian
methods help in solving this problem.
Format and dates
The first step is to complete a project proposal, which
is due in class on Tuesday, November 5. This should be a short
document (max. 2 pages) stating the title of your intended project,
the reason why you are interested in this topic, a rough plan
(e.g. what machine learning algorithms you would use, where you would
get the data), and a list of 5 papers that you think will be relevant
to your topic. You do not have to read the papers before the
proposal, nor do you have to include them in your final bibliography,
if they end up not being relevant. The main purpose of this document
is to inform me of your intentions, so I can give you feedback on the
scope of the work.
The project itself will include two major components:
-
Project report.
The report should be approximately 10 pages long (this requirement is
just to get you oriented; do not take it as a hard restriction). The
format should be similar to the research papers we have been reading
during the semester. It should contain the following information:
-
A description of the topic and why you were interested in it
-
A review of related literature; this should include at least 5 papers
relevant to your topic. If you want help seeking such papers,
let me know.
-
One or more specific questions that you wanted to investigate
-
The methodology you decided to follow (e.g. theory or experiments,
what kind of experiments, etc.)
-
An analysis of your findings
-
Conclusion and possible future work directions
The due date for the report is Monday, December 17, by 5pm.
-
Project presentation.
The presentations will take place in class during the lectures of
November 27, November 29, and December 4. You should plan for
a 15 minute talk, with 5 minutes for questions. You should assume
that everyone knows the material covered during the class, so cover
only the aspects that are particular to your topic. The presentations
will be evaluated by everyone present (although I will make the final
decision on the grade).
Note that the presentations will take place much earlier than the
project is due. Hence, you are not expected to have completed your
research by that time. However, you should have a clear picture of
what you are doing (which you should explain during your talk), and
ideally you should have some preliminary results (where applicable).
I would like to make the projects and presentations available from the
class web page, with your permission of course.
Prof. Doina PRECUP
Last modified: Fri Oct 26 16:07:33 EDT 2001