|
Abstract:
Existing kernel learning methods using class information based on Kernel
Alignment with semidefinite programming are often memory intensive and
impractical for problems with fair-size dataset. I'll describe two
efficient approaches to learning kernels using class information.
The first approach makes use of class information to scale the training
part of a given kernel matrix to form the training part of a new kernel
matrix. The test part of the new kernel matrix is estimated based on a
linear transformation in a reduced feature space and can be calculated
computationally efficiently.
The second approach makes use of neural networks (encoder and autoencoder)
to learn a desired kernel consistent with the class information of the
training data directly. Preliminary experimental results on handwritten
digit classification and protein remote homology detection will be shown.