Power-Law SNE

Perform an unsupervised dimensionality reduction of data using power-law SNE, generalized SNE, spherical SNE, and t-SNE, where the generalized SNE allows us to adjust the loss function from the Kullback–Leibler divergence to a general alpha-divergence.

Getting Started

1. Download the repository.
2. Add path (both pSNE-master & subfolder techniques) to the MATLAB.
3. Download 3 image datasets, MNIST, COIL-20, Olivetti faces, for demonstration.
4. Open main script SNE_display.m & set path for the 3 datasets in the beginning section. For example, placing folder Dataset inside GeneralizedSNE:

   MNIST_file = './Dataset/MNIST/train-images-idx3-ubyte' ;        % MNIST images
   MNIST_label_file = './Dataset/MNIST/train-labels-idx1-ubyte' ;  % MNIST labels
   COIL20_folder = './Dataset/coil-20-proc' ;                      % COIL-20
   Olivetti_folder = './Dataset' ;                                 % Olivetti faces

4. Run several SNEs in SNE_display.m.

Some Reference

• Power-law stochastic neighbor embedding
• Visualizing Data using t-SNE

Note

The source code of the pSNE & generalized SNE are mostly based on that of Laurens van der Maaten