Contents

Description of demo_unsupervised_tSNE.m

Learn swiss roll manifold using symmetric t-distribution Stochastic Neighbour Embedding algorithm (van der Maaten & Hinton [2008])

close all;
clear all;
generateData_swissRoll;

usage of t-SNE visualization

grid on;
grid minor;
scatter3(X(:,1), X(:,2), X(:,3), 100, col,'filled');
title('Manifold to Learn using t-SNE');
view(-10,170);
options = [];
% NOTE: kappa and gamma tune adaptive weights learning:
%       - if direction of gradient component switches from one iter to next
%         the algorithm scales that gradient weighting by (1-gamma)
%       - if direction of gradient component stays constant from one iter
%         to next, algorithm adds kappa to that gradient weighting
%       - See R.A. Jacobs: "Increased rates of convergence...", 1988
options.tSNEOptions = struct('initial_momentum', 0.5, ...
                             'final_momentum', 0.8, ...
                             'momentum_switch_iter', 250, ...
                             'eta', 100, ... % intitial learning rate
                             'exaggerate_factor', 4, ...
                             'last_exaggeration_iter', 75, ...
                             'descent_iters', 1000, ...
                             'gamma', 0.2, ...
                             'kappa', 0.6, ...
                             'min_rate', 0.1, ...
                             'verbose', 0, ... % print out progress
                             'final_dims', 2); % dim of learned points

Y = ml_visualize_tSNE(X,options);
figure;
scatter(Y(:,1),Y(:,2),30,col,'filled')
title('Learned Lower-Dimensional Representation');
grid on;
grid minor;

Published with MATLAB® R2015a