An implementation of the classic Artificial Neural Net with back propagation learning.

When Artificial Neural Nets were first invented and picked up by mainstream popular science many people believed it was going to usher in a new era of thinking machines. That was many decades ago and it turns out that intelligence is more than a few hundred artificial neurons. Way back when memory was measured in MB and CPU was measured in MHz (1,000x slower than today) intelligent operations could more effectively be performed with very different solutions, such as AlphaBeta. Fast forward to today and Artificial Neural Nets have solved a host of modern problems and are an important component in classification, heuristics, and content generation.

For more serious implementations try Tensorflow, Torch, or other such frameworks.

Section Links : Construction , Execution , Examples , Misc , Related , and References

Construction

NeuralNet constructor

var NeuralNetConstructor = require('neuralnet')
var neuralnet = NeuralNetConstructor( config )

Super-minimalistic configuration which works in most cases:

var config = {
	inputs 				: numberOfInputs , 
	outputs				: numberOfOutputs
}
var NeuralNetConstructor = require('neuralnet')
var neuralnet = NeuralNetConstructor( config )

That is all the configuration you need to get started. Jump to the Execution section for the prediction and training functions.

Full Config

If you need more control there are a few more configuration options you can set like so:

var config = {
	inputs 				: numberOfInputs , 		// defaults to 2
	hiddenLayers 			: numberOfHiddenLayers ,	// defaults to 1
	hiddenNodesPerLayer 		: numberOfNodesPerLayer ,	// defaults to inputs * 5
	outputs				: numberOfOutputs ,		// defaults to inputs
	learningRate			: learningRate ,		// defaults to 1
	filterTraining			: functionToLimitTraining	// defaults to no filter
}
var neuralnet = require('neuralnet')( config );

That creates one instance of a NeuralNet calculator which uses the initial configuration you supply. Other than 'inputs' all configuration options are optional.

Implementation Note : There is actually an additional input beyond what you specify in the configuration which is always set to -1. No need to add one if you were thinking of doing so, and no need to remove it because it will only help your NeuralNet.

The filterTraining is used to limit training. The default works fine, but suppose you are training a binary decision and you don't want to over train. The desired output is 0 and 1, but you really only care if the output is less than .5 or greater than .5 because you round. Here is an example filterTraining to prevent training a neural net that has already answered correctly enough:

function ( output , expected ) {
	var equalEnough = true
	for( var i = 0 ; i < expected.length && equalEnough ; i++ ) {
		equalEnough = ( Math.round( expected[i] ) == Math.round( output[i] ) )
	}
	// only train if the output and expected are not equal enough
	return ! equalEnough
}

If filterTraining returns true then the data will pass the filter and go on to train the neural net. If filterTraining returns false then the data will be blocked. NOTE There are much better ways of preventing over-fitting which are not covered in this basic npm module!

neuralnet.clone( )

Use .clone if you want another NeuralNet based on the configuration and learning of an existing NeuralNet. This allows you to take a snapshot of a NeuralNet after some training and explore what would happen if it received different training without modifying the original.

var aTotalCloneNeuralNet = existingNeuralNet.clone()

neuralnet.new( )

use .new if you want a new untrained NeuralNet with the same configuration as another NeuralNet. This is different than clone which copies over all the training. .new leaves the original NeuralNet unchanged.

var aFreshUntrainedNeuralNet = existingNeuralNet.new()

neuralnet.new( config )

A short cut for require('neuralnet')( config ) once you have a NeuralNet.

var aFreshUntrainedNeuralNet = existingNeuralNet.new( config )

neuralnet.learningRate( )

Retrieve the learning rate of this NeuralNet.

var whatsMyLearningRate = neuralnet.learningRate()

A good initial value to try is 1.

neuralnet.learningRate( newLearningRate )

Change the learning rate of this NeuralNet. It returns the value set if it is a legal value otherwise it returns the previously set configuration.

var shouldEqualFive = neuralnet.learningRate(5)

neuralnet.config( )

Retrieve a copy of the complete configuration that this NeuralNet uses. Perhaps you want see what happens with an additional hidden layer. No problem, you can do that like so:

var config = neuralnet.config()
config.numberOfHiddenLayers++
var thickerNeuralNet = neuralnet.new(config).reset()

neuralnet.reset( )

Randomize the weights connecting the layers in the neural net and return the neuralnet.

var aCopyWithRandomizedWeights = neuralnet.clone().reset();

Execution

NeuralNet takes an array of decimal numbers, expecting it to be the size of inputs in length and returns an array of decimal numbers in length equal to outputs. The numbers the NeuralNet takes as input is any decimal number from 0.0 to 1.0. For output the reasonable range is more like 0.1 to 0.9. If you want to input and output numbers outside of this range you'll need to do the mathmatical mapping. Typically though the input is used in a binary way, either exactly 0 or exactly 1, as is the expected output for training purposes.

neuralnet.predict( inputArray )

Predicts the output with grayscale (floating point) numbers. Call the prediction function like so:

var inputArray = [ /* your data */ ]
var predictionArray = neuralnet.predict( inputArray )

An example grayscale predictionArray is : [ 0.223 , 0.118 , 0.734 ] Think of the grayscale prediction as a confidence score where 0.5 represents "I really don't know between 'no' (0.0) and 'yes' (1.0)". The closer the score is to either 0.0 or 1.0 the more confident it is. This could allow you to use the grayscale for more than a binary, yes or no, response. For example, you could use values of 0.0 to 0.25 as 'no', 0.75 to 1.00 as 'yes', and 0.25 to 0.75 as 'I am uncertain'.

neuralnet.predictBoolean( inputArray )

Same as neuralnet.predict but changes grayscale values into booleans

var inputArray = [ /* your data */ ]
var predictionArray = neuralnet.predictBoolean( inputArray )

An example boolean predictionArray is : [ 0 , 0 , 1 ] The boolean prediction is simply the greyscale prediction passed through Math.round().

neuralnet.train( inputArray , expectedOutputArray )

Use backpropogation to train the neural net. It also returns a prediction. Use it like so:

var inputArray = [ /* your data */ ]
var expectedOutputArray = [ /* the expected output */ ]
var actualPredictionArray = neuralnet.train(inputArray,expectedOutputArray)

You can optionally override the learningRate for one training session like so:

neuralnet.train( inputArray , expectedOutputArray , learningRate )

Use this to change the learning rate on a case by case basis if you have some indicator of how much weight a specific learnt item should have. Think of it as shocking the student. It can seriously speed up learning but at the cost of overfitting. Use like so:

var inputArray = [ /* your data */ ]
var expectedOutputArray = [ /* the expected output */ ]
var learningRate = 5 // set the learning rate to 5 times default.
var actualPredictionArray = neuralnet.train(
								inputArray,
								expectedOutputArray,
								learningRate
							)

Example

If you have included NeuralNet as a node package then first change directory to node_packages/neuralnet

Template

A simple template which has one input array, one training output array, and trains the NeuralNet once.

node example/template.js

Rock Paper Sissors (Boolean example)

Boolean operations is the classic use of Neural Nets, either fire or don't fire. In this example the Neural Net learns the rules of the game of Rock Paper Sissors.

node example/rock_paper_sissors.js

Multiplication (Grayscale example)

Grayscale means using the actual output as it is and not converting it to a true/false. This example shows that the numeric value of the output can be fairly precise as an estimator, not just a boolean operation.

node example/multiply.js

Rotation of an image

Learn to rotate a 4x4 black and white image by 90 degrees. The 4x4 is printed to the screen so you can see how the NeuralNet is doing. I was quite amazed at how few learning sessions it needed to solve this problem and how well it does. Honestly I was expecting it to kinda get it but have some trouble and instead it caught on very quickly and accurately.

# the default 4x4 image
node example/rotate.js

# A more difficult 5x5
node example/rotate.js 5

# Still succeeds with 8x8
node example/rotate.js 8

Projectile (Grayscale example)

Estimate the motion of a rock (or anything really) being thrown up and then coming back down.

example/projectile.js

Misc

neuralnet.layers

You can inspect and modify the data in the layers by accessing the layers, although there probably isn't a need.

var layers = neuralnet.layers

Related AI Projects

This is part of a set of related projects.

• AlphaBeta
• Boosting
• GeneticAlgorithm
• NearestNeighbour
• NeuralNet

References

• Instructor: Patrick Winston from MIT
• Wikipedia entry for Artificial Neural Networks

Alternate Neural Net Project

synaptic is a feature-rich but complicated-to-use Neural Net package. It includes LSTM (Long Short Term Memory) and other enhancements.