Learning Translation Invariant Kernels for Classification

This page is designed to help the researchers to repeat the experiments of the following publication:

K. Ghiasi-Shirazi, R. Safabakhsh, and M. Shamsi. Learning Translation Invariant Kernels for Classification. Journal of Machine Learning Research, 11:1353-1390, 2010. (pdf)

Datasets

Datasets extracted from MNIST

Task	Ready to use Matlab file
Odd vs. Even	mnistoe60K.mat
3 vs 8	mnist38_60K.mat
4 vs 7	mnist47_60K.mat

Datasets extracted from USPS

Task	Ready to use Matlab file
0-4 vs. 5-9	usps_0-4,5-9.mat

Software

The implementations of this paper are collected in the SIKL toolbox which contains a modifies version of libLBFGS. To use it you should also download MOSEK.

Installation

Compiling mex files

After unziping the SIKL toolbox into an appropriate folder, run the compileMexFiles.m file.

Generating log file

To generate log file, set the global parameter logfilenamePrefix to a non-empty string.

Results

Experiments on small-size benchmark datasets

To repeat the experiments with Gaussian Mixture (GM) method, uncomment the line  method = 'GaussianMixture'; in TestUCI.m file and run the file.
To repeat the experiments with  Cosine Mixture (CM) method, uncomment the line  method = 'CosineMixture'; in TestUCI.m file and run the file.
To repeat the experiments with  Cosine and Gaussian Mixture (CGM) method, uncomment the line  method = 'CosineAndGaussianMixture'; in TestUCI.m file and run the file.

Experiments on the USPS dataset

Run TestUSPS.m file.

Experiments on the MNIST dataset

Run TestMNISTAll.m file.