一个JavaScript神经网络库:Synaptic
Synaptic是一个JavaScript神经网络库,可用于node.js 和浏览器环境。 its generalized algorithm is architecture-free, so you can build and train basically any type of first order or even second order neural network architectures.
This library includes a few built-in architectures like multilayer perceptrons, multilayer long-short term memory networks (LSTM), liquid state machines or Hopfield networks, and a trainer capable of training any given network, which includes built-in training tasks/tests like solving an XOR, completing a Distracted Sequence Recall task or an Embedded Reber Grammar test, so you can easily test and compare the performance of different architectures.
The algorithm implemented by this library has been taken from Derek D. Monner's paper:
A generalized LSTM-like training algorithm for second-order recurrent neural networks
There are references to the equations in that paper commented through the source code.
Introduction
If you have no prior knowledge about Neural Networks, you should start by reading this guide.
Demos
Getting started
Overview
Installation
In node
You can install synaptic with npm:
npm install synaptic --save
In the browser
Just include the file synaptic.js from/distdirectory with a script tag in your HTML:
<script src="synaptic.js"></script>
Usage
var synaptic = require('synaptic'); // this line is not needed in the browser var Neuron = synaptic.Neuron, Layer = synaptic.Layer, Network = synaptic.Network, Trainer = synaptic.Trainer, Architect = synaptic.Architect;
Now you can start to create networks, train them, or use built-in networks from the Architect.
Gulp Tasks
- gulp or gulp build: builds the source code from/srcinto the/distdirectory (bundled and minified).
- gulp debug: builds the source code from/srcinto the/distdirectory (not minifed and with source maps for debugging).
- gulp dev: same as debug but it watches for changes in the source files and rebuilds when any change is detected.
- gulp test: runs all the tests.
Examples
Perceptron
This is how you can create a simple perceptron:
function Perceptron(input, hidden, output) { // create the layers var inputLayer = new Layer(input); var hiddenLayer = new Layer(hidden); var outputLayer = new Layer(output); // connect the layers inputLayer.project(hiddenLayer); hiddenLayer.project(outputLayer); // set the layers this.set({ input: inputLayer, hidden: [hiddenLayer], output: outputLayer }); } // extend the prototype chain Perceptron.prototype = new Network(); Perceptron.prototype.constructor = Perceptron;
Now you can test your new network by creating a trainer and teaching the perceptron to learn an XOR
var myPerceptron = new Perceptron(2,3,1); var myTrainer = new Trainer(myPerceptron); myTrainer.XOR(); // { error: 0.004998819355993572, iterations: 21871, time: 356 } myPerceptron.activate([0,0]); // 0.0268581547421616 myPerceptron.activate([1,0]); // 0.9829673642853368 myPerceptron.activate([0,1]); // 0.9831714267395621 myPerceptron.activate([1,1]); // 0.02128894618097928
Long Short-Term Memory
This is how you can create a simple long short-term memory network with input gate, forget gate, output gate, and peephole connections:
function LSTM(input, blocks, output) { // create the layers var inputLayer = new Layer(input); var inputGate = new Layer(blocks); var forgetGate = new Layer(blocks); var memoryCell = new Layer(blocks); var outputGate = new Layer(blocks); var outputLayer = new Layer(output); // connections from input layer var input = inputLayer.project(memoryCell); inputLayer.project(inputGate); inputLayer.project(forgetGate); inputLayer.project(outputGate); // connections from memory cell var output = memoryCell.project(outputLayer); // self-connection var self = memoryCell.project(memoryCell); // peepholes memoryCell.project(inputGate, Layer.connectionType.ONE_TO_ONE); memoryCell.project(forgetGate, Layer.connectionType.ONE_TO_ONE); memoryCell.project(outputGate, Layer.connectionType.ONE_TO_ONE); // gates inputGate.gate(input, Layer.gateType.INPUT); forgetGate.gate(self, Layer.gateType.ONE_TO_ONE); outputGate.gate(output, Layer.gateType.OUTPUT); // input to output direct connection inputLayer.project(outputLayer); // set the layers of the neural network this.set({ input: inputLayer, hidden: [inputGate, forgetGate, memoryCell, outputGate], output: outputLayer }); } // extend the prototype chain LSTM.prototype = new Network(); LSTM.prototype.constructor = LSTM;
These are examples for explanatory purposes, the Architect already includes Multilayer Perceptrons and Multilayer LSTM network architectures.