Create neuronal networks

This framework creates some neuronal networks. It shows and provides you a collection of algorithms how to use the framework.

0. Preparations

This framework needs two external git repositories:

• js-analysis
• js-testing-framework

0.1 Checkout this project with submodules

user$ git clone git@github.com:friends-of-ai/create-neuronal-networks.git && cd create-neuronal-networks
user$ git submodule init
user$ git submodule update

1. The elements

1.1 Predictors

Coming soon..

1.2 Perceptron

Coming soon..

1.3 The network (neuronal network)

1.3.1 Create a neuronal network with given weight matrices

var weightMatrices =  [
    new Matrix([[0.9, 0.3, 0.4], [0.2, 0.8, 0.2], [0.1, 0.5, 0.6]]),
    new Matrix([[0.3, 0.7, 0.5], [0.6, 0.5, 0.2], [0.8, 0.1, 0.9]])
];

var neuronalNetwork = new NeuronalNetwork(weightMatrices);

var input  = new Vector([0.9, 0.1, 0.8]);
var output = neuronalNetwork.calculateOutput(input);

console.log(JSON.stringify(output.array));
// prints [0.7263033450139793,0.7085980724248232,0.778097059561142]

1.3.2 Let the neuronal network calculate the weight matrices

var planes = [3, 3, 3];

var neuronalNetwork = new NeuronalNetwork(planes);

var input  = new Vector([0.9, 0.1, 0.8]);
var output = neuronalNetwork.calculateOutput(input);

console.log(JSON.stringify(output.array));
// prints (example) [0.5601474395121488,0.6247316906773285,0.6346952211353676]

1.3.3 Train and test the network

1.3.3.1 I

var planes = [2, 2, 2];

var neuronalNetwork = new NeuronalNetwork(planes, true);

var input    = new Vector([0.7, 0.6]);
var expected = new Vector([0.9, 0.2]);

/* train the network */
for (var i = 0; i < 100; i++) {
    neuronalNetwork.train(input, expected);
}

/* test the network (the output should be the expected value */
var output = neuronalNetwork.calculateOutput(input);

console.log(Math.round(output.array[0] * 10) / 10, Math.round(output.array[1] * 10) / 10);

It prints

0.9 0.2

1.3.3.2 II

We want to have the double output value of the given input value:

var planes = [1, 1, 1];
var neuronalNetwork = new NeuronalNetwork(planes, true);

/* the in- and outputs */
var inputs    = new Array(
    new Vector([0.1]),
    new Vector([0.2]),
    new Vector([0.3]),
    new Vector([0.4]),
    new Vector([0.5]),
    new Vector([0.11])
);
var outputs = new Array(
    new Vector([0.2]),
    new Vector([0.4]),
    new Vector([0.6]),
    new Vector([0.8]),
    new Vector([1.0]),
    new Vector([0.22])
);

/* train the network */
for (var i = 0; i < 1000; i++) {
    for (var j = 0; j < inputs.length; j++) {
        neuronalNetwork.train(inputs[j], outputs[j]);
    }
}

/* the check values */
var inputsTest = new Array(
    new Vector([0.33]),
    new Vector([0.12]),
    new Vector([0.11]),
    new Vector([0.44]),
    new Vector([0.49]),
    new Vector([0.39]),
);

/* check the trained network */
for (var j = 0; j < inputs.length; j++) {
    console.log(inputsTest[j].vector[0], neuronalNetwork.calculateOutput(inputsTest[j]).vector[0]);
}

It prints

0.33 0.6874777236350406
0.12 0.23842579104867254
0.11 0.2269773560587364
0.44 0.882471971205491
0.49 0.9254848257237483
0.39 0.8127789003178846

1.3.3.3 III

XOR Function

var planes = [2, 2, 1];
var neuronalNetwork = new NeuronalNetwork(planes, true);

var inputs    = new Array(
    new Vector([0, 0]),
    new Vector([0, 1]),
    new Vector([1, 0]),
    new Vector([1, 1])
);
var outputs = new Array(
    new Vector([0]),
    new Vector([1]),
    new Vector([1]),
    new Vector([0])
);

for (var i = 0; i < 3000; i++) {
    for (var j = 0; j < inputs.length; j++) {
        neuronalNetwork.train(inputs[j], outputs[j]);
    }
}

var inputsTest = new Array(
    new Vector([0, 0]),
    new Vector([0, 1]),
    new Vector([1, 0]),
    new Vector([1, 1])
);

for (var j = 0; j < inputs.length; j++) {
    console.log(inputsTest[j].vector[0], inputsTest[j].vector[1], neuronalNetwork.calculateOutput(inputsTest[j]).vector[0]);
}

It prints

0 0 0.02324451873204845
0 1 0.9750358589087668
1 0 0.9748872859116073
1 1 0.03127867654707072

2. Test the libraries

2.1 Neuronal vector library

Call tests/neuronalNetwork.html in your browser. It adds a div element with id testResult to your body and returns for example something like this:

-------------------------------------
Start test "Neuronal Network - Tests"
-------------------------------------
 
  1) NeuronalNetwork: Running error test "Given parameter is not an array" (Code: 101).
     Test succeeded (0.5 ms).
  2) NeuronalNetwork: Running error test "Array must be longer than one" (Code: 103).
     Test succeeded (0.2 ms).
  3) NeuronalNetwork: Running error test "The given element is no number" (Code: 104).
     Test succeeded (0.3 ms).
  4) NeuronalNetwork: Running success test "Init neuronal network with planes." (Code: 201).
     Test succeeded (0.8 ms).
  5) NeuronalNetwork: Running success test "Init neuronal network with planes and bias." (Code: 202).
     Test succeeded (0.1 ms).
  6) NeuronalNetwork: Running success test "Init neuronal network with weight matrices." (Code: 203).
     Test succeeded (0.3 ms).
  7) NeuronalNetwork: Running success test "Init neuronal network with weight matrices and bias." (Code: 204).
     Test succeeded (0.3 ms).
  8) NeuronalNetwork: Running error test "The current vector does not fit as input value" (Code: 103).
     Test succeeded (0.4 ms).
  9) NeuronalNetwork: Running success test "Calculation test." (Code: 205).
     Test succeeded (1.3 ms).
 10) NeuronalNetwork: Running error test "The current vector does not fit as input value" (Code: 103).
     Test succeeded (0.3 ms).
 11) NeuronalNetwork: Running success test "Calculation test with bias." (Code: 206).
     Test succeeded (0.6 ms).
 12) NeuronalNetwork: Running success test "Test the learn method." (Code: 207).
     Test succeeded (38 ms).
 
---------------------------------------------------------------
RESULT
-> All test succeeded (46 ms) [success: 12; error: 0; all: 12].
---------------------------------------------------------------

A. Other Tutorials

• Coming soon..

B. Literature

Recommended for practical use:

• Coming soon..

C. Sources

• Coming soon..

D. Authors

• Björn Hempel bjoern@hempel.li - Initial work - https://github.com/bjoern-hempel

E. Licence

This tutorial is licensed under the MIT License - see the LICENSE.md file for details