Neural network implementation based on 'Introduction to the Math of Neural Networks' book. Not complete yet.

neural.js

Network = function(input, hidden, output) {
    if (arguments.length !== 3) {
        console.log('Check your arguments, quitting...');
        process.exit();
    }
    this.B1 = 1.0;
    this.B2 = 1.0;

    for (var i = 1; i <= input; i++) {
        this['I' + i] = 0;
    }

    for (var i = 1; i <= hidden; i++) {
        this['H' + i] = {
            sum: 0.0,
            out: 0.0,
        };
    }

    for (var i = 1; i <= output; i++) {
        this['O' + i] = {
            sum: 0.0,
            out: 0.0
        };
    }

    for (var i = 1; i <= hidden + output; i++) {
        this['w' + i] = [];
    }

    /*for (var i = 1; i <= input + 1; i++) {
        this.w1.push(Math.random() * (1 - (-1)) + (-1));
        this.w2.push(Math.random() * (1 - (-1)) + (-1));
    }

    for (var i = 1; i <= hidden + 1; i++) {
        this.w3.push(Math.random() * (1 - (-1)) + (-1));
    }*/

    this.w1 = [-0.07, 0.22, -0.46];
    this.w2 = [0.94, 0.46, 0.10];
    this.w3 = [-0.22, 0.58, 0.78];


};

Network.prototype.train = function(input, expected) {
    var input1 = []
    for (i = 1; i <= input.length; i++) {
        this['I' + i] = input[i - 1];
        input1.push(this['I' + i]);
    }
    input1.push(this.B1);

    this.H1.sum = this.run(input1, this.w1);
    this.H1.out = this.sigmoid(this.H1.sum);

    this.H2.sum = this.run(input1, this.w2);
    this.H2.out = this.sigmoid(this.H2.sum);

    input2 = [this.H1.out, this.H2.out, this.B2];

    this.O1.sum = this.run(input2, this.w3);
    this.O1.out = this.sigmoid(this.O1.sum);

    this.O1.delta = this.delta(this.O1.out, expected);
    this.H1.delta = this.deltaHidden(this.H1.out, this.O1.delta, this.w3[0]);
    this.H2.delta = this.deltaHidden(this.H2.out, this.O1.delta, this.w3[1]);

    this.gradientTable = {};

    this.gradientTable.H1O1 = this.gradient(this.H1.out, this.O1.delta);
    this.gradientTable.H2O1 = this.gradient(this.H2.out, this.O1.delta);
    this.gradientTable.B2O1 = this.gradient(this.B2, this.O1.delta);

    this.gradientTable.I1H1 = this.gradient(this.I1, this.H1.delta);
    this.gradientTable.I1H2 = this.gradient(this.I1, this.H2.delta);

    this.gradientTable.I2H1 = this.gradient(this.I2, this.H1.delta);
    this.gradientTable.I2H2 = this.gradient(this.I2, this.H2.delta);

    this.gradientTable.B1H1 = this.gradient(this.B1, this.H1.delta);
    this.gradientTable.B1H2 = this.gradient(this.B1, this.H2.delta);

    console.log(this);
};

Network.prototype.run = function(input, w) {
    var sum = 0.0;
    w.forEach(function(w, index) {
        sum += w * input[index];
    })
    return sum;
};

Network.prototype.sigmoid = function(x) {
    return 1.0 / (1.0 + Math.exp(-x))
};

Network.prototype.delta = function(output, expected) {
    var error = (output - expected);
    var derivative = output * (1.0 - output);
    var delta = -(error) * derivative;
    return delta;
};

Network.prototype.deltaHidden = function(output, delta, w) {
    var derivative = output * (1.0 - output);
    var deltaHidden = derivative * (delta * w);
    return deltaHidden;
};

Network.prototype.gradient = function(output, delta) {
    return output * delta;
};


var network = new Network(2, 2, 1);

network.train([1.0, 0], 1);