Yaser
Abu-Mostafa
Professor
of Electrical Engineering and Computer Science
B.Sc., Cairo University, 1979
M.S.E.E., Georgia Institute of Technology, 1981
Ph.D., California Institute
of Technology, 1983
1200
East California Boulevard, MS 136-93
Pasadena, CA 91125 USA
(626) 395-4842 (office)
(626) 395-2137 (fax)
yaser@caltech.edu
http://www.work.caltech.edu/yaser/index.html
Research
The
Learning Systems group at Caltech works on the theory, implementation,
and application of automated learning, pattern recognition, and
neural networks. We are an interdisciplinary group with students
coming from Electrical Engineering, Computer Science, Mathematics,
and Physics. We work on a variety of projects analyzing and synthesizing
systems that can be trained to perform their task.
Learning
is a topic that can be approached from different angles. For instance,
one project deals with the rigorous mathematical analysis of over-learning;
when a system memorizes too much detail about a limited experience.
Another project develops techniques for selecting high-level features
for pattern recognition. A third project applies learning to forecasting
in the financial markets.
We
have an on-going research effort in learning from hints which
generalizes the basic learning-from-examples paradigm. Learning
from examples assumes that the function we are trying to learn
is represented to us by a training set of input/output examples,
but is otherwise unknown. This poses a dilemma: We would like
to use a learning model that is sophisticated enough to have a
chance of implementing the unknown function, yet we want the model
to be simple enough that a limited set of examples will suffice
to tune it properly.
Hints
come in as a learning aid to improve the situation. We usually
have, in addition to the set of examples, some prior knowledge
of certain facts about the system. We wish to use this side information
to our advantage. For instance, we may know that is scale-invariant
and monotonic, and we want to put this information together with
the training set of examples to come up with a good implementation.
We want this process to be automated, so that the system is genuinely
learning from hints.
Our
group has pioneered the use of hints in learning, and we continue
to work on all aspects of the technique. The goal is to automate
the use of hints to a degree where we can effectively utilize
a large number of different hints that may be available in a practical
situation. We have also developed a theoretical analysis of learning
from hints. The analysis is based on the Vapnik-Chervonenkis dimension
provides an estimate for the number of examples the system will
need to learn.
The
Learning Systems group is committed to the understanding of the
fundamental components of automated learning, and to the development
of real-life systems that utilize learning to achieve state-of-the
art performance.
