back_prop.java

//
// inside the main class
//
public void trainNetwork()
{
    for( int i = 0; i < 1000000; i++)
    {
        int r = (int) (Math.random() * 100000) % 4;
        int a = r % 2;
        int b = (r >> 1) % 2;
        int c = a ^ b;

        double y = output.output();
        double dErr = (y - c); //Err = 0.5*(y - c)*(y - c)

        input1.setInput( a); //ok
        input2.setInput( b); //ok

        output.backPropagate( dErr ); // ?
    }
}

//
// inside the neuron class
//
public void backPropagate( double dErr)
{
    if( isInputNeuron) 
    {
        return; 
    }

    double hj = inputSum(); //works well
    double yj = output();  //works well
    for( int i = 0; i < inputSynapses.size(); i++)
    {
        double xi = inputSynapses.get(i).output();
        double dErrj = activationDerivative( hj)*dErr;

        inputSynapses.get(i).backPropagate( dErrj*weights[i] ); //beware, recursion.
        weights[i] -= learningRate*dErrj*xi;
    }
}

private double activationDerivative( double x) //sigmoid
{
    return activation(x)*(1.0 - activation(x));
}

private double activation( double arg) //sigmoid
{
    return 1.0/(1 + Math.exp( -arg));
}