elmTrain.m

function [inW, bias, outW, scores] = elmTrain( X, Y, nHiddenNeurons, C )
% elmTrain trains the Extreme Learning Machine. The activation function is
% sigmoid which could be changed easily if needed.
%
%	[inW, bias, outW, scores] = elmTrain( X, Y, nHiddenNeurons, C );
%
% Inputs:
%	X				: data patterns (column vectors)
%	Y				: numeric labels for each pattern (1, ... )
%	nHiddenNeurons	: number of hidden neurons
%	C				: regularization parameter
%
% Outputs:
%	inW				: input weights matrix
%	bias			: bias vector
%	outW			: output weights matrix
%	scores			: scores on the own training data
%

nTrainData = size( X, 2 );
nInputNeurons = size(X, 1);
nClasses = 1;

targets = Y;

% generate random input weight matrix
inW = rand( nHiddenNeurons, nInputNeurons ) * 2 - 1;

% generate random hidden neuron vector
bias = rand( nHiddenNeurons, 1 );

% compute the pre-H matrix
preH = inW * X;

% build the bias matrix
biasM = repmat( bias, 1, nTrainData );

% update the H matrix
preH = preH + biasM;

% calculate hidden neuron output matrix H
H = 1 ./ (1 + exp(-preH));

% compute regularized output weights matrix
outW = ( eye(nHiddenNeurons)/C + H * H') \ H * targets';

% output for the training data
scores = (H' * outW)';