Neural networks in JavaScript. Well-documented and object-oriented.
-
In the browser
<script type="text/javascript" src="dist/browser-bundle.min.js"></script>
-
In Node.js
$ npm install @delight-im/neural-network
and
var NeuralNetwork = require("@delight-im/neural-network");
-
Creating a new instance
-
Neural network with 3 input neurons and 1 output neuron
var network = new NeuralNetwork.Type.Feedforward(3, [], 1);
-
Neural network with 4 input neurons, 3 hidden neurons and 2 output neurons
var network = new NeuralNetwork.Type.Feedforward(4, [ 3 ], 2);
-
Neural network with 6 input neurons, two hidden layers with 4 and 2 neurons, and 3 output neurons
var network = new NeuralNetwork.Type.Feedforward(6, [ 4, 2 ], 3);
-
-
Passing any number of additional options to the network
// pass an object containing the desired options as the fourth parameter var network = new NeuralNetwork.Type.Feedforward(3, [ 4 ], 1, { seed: 501935, learningRate: 0.3, hiddenLayerActivationFunction: new NeuralNetwork.Activation.HyperbolicTangent(), outputLayerActivationFunction: new NeuralNetwork.Activation.BinaryStep() });
-
Available activation functions
new NeuralNetwork.Activation.ArcTangent(); new NeuralNetwork.Activation.BinaryStep(); new NeuralNetwork.Activation.GaussianFunction(); new NeuralNetwork.Activation.HyperbolicTangent(); new NeuralNetwork.Activation.Identity(); new NeuralNetwork.Activation.LogisticFunction(); new NeuralNetwork.Activation.RectifiedLinearUnit(); new NeuralNetwork.Activation.RectifiedLinearUnit(0.01); new NeuralNetwork.Activation.SinusoidFunction();
-
Training the network using supervised batch ("all-at-once") learning
// the first parameter is the array of inputs and the second parameter is the array of desired outputs // the third parameter is the optional number of iterations and the fourth parameter is the optional error threshold var error = network.trainBatch( [ [0, 0, 1], [0, 1, 1], [1, 0, 1], [1, 1, 1] ], [ [ 0 ], [ 1 ], [ 1 ], [ 0 ] ], 60000, 0.005 );
-
Training the network using supervised online ("single-pattern") learning
// the first parameter is the input and the second parameter is the desired output var error = network.train([0, 0, 1], [ 0 ]);
-
Asking the network to predict some output from a supplied input pattern
// the single parameter is the input to process network.predict([ 0, 0, 1 ])
-
Saving the network with all its properties to a JSON string
var jsonStr = JSON.stringify(network);
-
Restoring the network with all its properties from a JSON string
var network = NeuralNetwork.Type.Feedforward.fromJson(jsonStr);
-
Prerequisites
$ npm install -g uglify-js $ npm install -g browserify
-
Building the browser bundle
$ browserify src/main.js --standalone NeuralNetwork > dist/browser-bundle.js $ uglifyjs dist/browser-bundle.js --compress --preamble "$(< src/header.js)" > dist/browser-bundle.min.js $ rm dist/browser-bundle.js
-
Running the Node.js examples
$ node examples/node.js
All contributions are welcome! If you wish to contribute, please create an issue first so that your feature, problem or question can be discussed.
This project is licensed under the terms of the MIT License.