Research, Zehra Cataltepe
RESEARCH INTERESTS
- Test Inputs:
In some applications such as credit card approval, medical image
diagnosis, in addition to the labelled training examples,
we also have unlabelled test inputs. We show that using the
test inputs the generalization ability of the learning
system can be improved. Our method also extends to cases
where additional inputs that are not necessarily test inputs,
or not the identity but the density of the inputs are known.
( brief description,
postscript NIPS paper )
- Hints:
In learning-from-examples systems (such as neural networks)
we are given a set of training input-outputs and our goal
is to produce a system that performs well on out-sample data.
However sometimes we know some properties of the underlying
function. Learning-from-hints (@ Prof. Yaser S. Abu-Mostafa)
expresses the properties/rules as examples and teaches
them together with the training input-outputs to the neural
network. Teaching the hints in the best way is a nontrivial
problem. I have implemented simple schedules and also
worked on a method to incorporate invariance hints.
( brief description,
postscript CMSS paper)
- Applications:
- speech, video: I am especially interested in approaches
that merge speech and video and other information to achieve
better tools e.g. for teleconferencing.
- data mining: Combination of neural network and
knowledge based systems approaches can lead to better data mining
approaches.
- Regularization methods:
especially smoothing methods such as weight decay and early stopping of training.