INCREMENTAL SVM LEARNING
This MATLAB package implements the methods for exact incremental/decremental
SVM learning, regularization parameter perturbation and kernel parameter perturbation
presented in "SVM Incremental Learning, Adaptation, and Optimization" by
Christopher Diehl and Gert Cauwenberghs.
This code is designed for training SVMs to solve binary classification problems. The code
can typically handle training problems with dataset sizes ranging up to 10,000 examples.
This constraint is driven primarily by the fact that this method is computing and updating
the inverse of the kernel matrix for the support vectors.
The primary benefits of this code are:
- Exact Incremental Learning - One or more examples can be exactly incremented
into the current SVM solution, resulting in a classifier that is valid for the entire
training set seen up to that point
- Regularization Parameter Perturbation - The current SVM valid for a given set of
regularization parameters C can be exactly perturbed to the valid SVM for a set of
regularization parameters C'
- Kernel Parameter Perturbation - The current SVM valid for a given kernel parameter
sigma can be exactly perturbed to the valid SVM for the kernel parameter sigma'
- Exact and Approximate Leave-One-Out (LOO) Error Estimation - The exact LOO error
estimate can be efficiently computed by exactly unlearning one example at a time and
testing the classifier on the example. An efficient LOO approximation is also implemented
that predicts the change in the example's margin based on the margin sensitivity.
This approximation is referred to as the span bound in the literature, as introduced
by Vapnik and Chapelle.
Download:
IncrementalSVM.zip
Questions/Comments:
Chris.Diehl at jhuapl dot edu