private void stochasticUpdateStep(Pair<Integer, Set<Integer>> wordPlusContexts, int s) {
double eta = learningRateDecay(s);
int wordIndex = wordPlusContexts.getFirst(); // actual center word
// Set h vector equal to the kth row of weight matrix W1. h = x' * W = W[k,:] = v(input)
RealVector h = W1.getRowVector(wordIndex); // 1xN row vector
for (int contextWordIndex : wordPlusContexts.getSecond()) {
Set<Integer> negativeContexts;
if (sampleUnigram) {
negativeContexts = negativeSampleContexts(wordIndex, noiseSampler);
} else {
negativeContexts = negativeSampleContexts(wordIndex);
}
// wordIndex is the input word
// negativeContexts is the k negative contexts
// contextWordIndex is 1 positive context
// First update the output vectors for 1 positive context
RealVector vPrime_j = W2.getColumnVector(contextWordIndex); // Nx1 column vector
double u = h.dotProduct(vPrime_j); // u_j = vPrime(output) * v(input)
double t_j = 1.0; // t_j := 1{j == contextWordIndex}
double scale = sigmoid(u) - t_j;
scale = eta * scale;
RealVector gradientOut2Hidden = h.mapMultiply(scale);
vPrime_j = vPrime_j.subtract(gradientOut2Hidden);
W2.setColumnVector(contextWordIndex, vPrime_j);
// Next backpropagate the error to the hidden layer and update the input vectors
RealVector v_I = h;
u = h.dotProduct(vPrime_j);
scale = sigmoid(u) - t_j;
scale = eta * scale;
RealVector gradientHidden2In = vPrime_j.mapMultiply(scale);
v_I = v_I.subtract(gradientHidden2In);
h = v_I;
W1.setRowVector(wordIndex, v_I);
// Repeat update process for k negative contexts
t_j = 0.0; // t_j := 1{j == contextWordIndex}
for (int negContext : negativeContexts) {
vPrime_j = W2.getColumnVector(negContext);
u = h.dotProduct(vPrime_j);
scale = sigmoid(u) - t_j;
scale = eta * scale;
gradientOut2Hidden = h.mapMultiply(scale);
vPrime_j = vPrime_j.subtract(gradientOut2Hidden);
W2.setColumnVector(negContext, vPrime_j);
// Backpropagate the error to the hidden layer and update the input vectors
v_I = h;
u = h.dotProduct(vPrime_j);
scale = sigmoid(u) - t_j;
scale = eta * scale;
gradientHidden2In = vPrime_j.mapMultiply(scale);
v_I = v_I.subtract(gradientHidden2In);
h = v_I;
W1.setRowVector(wordIndex, v_I);
}
}
}
private static double sigmoid(RealVector x, RealVector y) {
double z = x.dotProduct(y);
return sigmoid(z);
}