LASVM Applied to Invariant Problems
In: Large Scale Kernel Machines, 9 Dec 2005, Whistler, Canada.
Because many patterns are insensitive to certain transformations such as rotations or translations,
it is widely admitted that the quality of a pattern recognition system can be improved by taking into account invariances.
In spite of numerous efforts, however,
the question on how to do it when learning from examples, still remains an open issue.
Indeed different ways have been investigated
including the proposal of a unifying framework for invariant pattern-recognition and the design of invariant features.
Another way of taking into account invariance %when learning from examples
is by creating new labeled examples deforming available input data.
In doing so large datasets can be generated.
As a matter of fact their size is not even limited.
For instance, in an experiment reported here,
the sample size of the training set of the MNIST OCR database has been increased
from 60,000 to 8 millions of examples by generating more than 130 random deformations per digit.
Because up to now almost no efficient learning algorithm was available to deal with the size of the resulting dataset,
no experiment have been reported so far using this approach.
Thanks to the use of the recently published LASVM algorithm
(a SVM training algorithm that can be used on such a large datasets)
it has been possible to realize experiments
on datasets incorporating invariant data.
Results reported are good regarding the recognition rate
and prove the ability of SVM to handle millions of examples.