What is Marvin.js?

Marvin.js is a fully open-source JavaScript neural network library. You can use it to create neural network implementations with ease! Take a look at the examples to see what's possible.

What is a neural network?

There are plenty good explanations out there, so I'd suggest you take some time to just look it up, but here's a brief explanation:

A neural network is essentially a mathematical function, you give it some variables (numbers!) and it returns some values. The thing with a neural network, though, is that it is generally a huge function, consisting of thousands of variables. When you train a neural network, you give the function some values (from which you know what the result should be) and you compare the result with the desired result. After that you use an algorithm to determine how the function should change in order to return the desired value more often. The beauty is that, generally, the network can be saved to a file of only a couple of kilobytes. That being said, you do need a lot of training data for it to work.

This function can be displayed in a grid of neurons. Layers are formed by the neurons on top of each other. Every neuron has an activation (value) and a relation to all the neurons in the previous layer. The activations of the first layer (the input layer) is directly set as the input. The activations of the last layer (the output layer) are used as an output.

This way, the more layers you have, the greater the chance of the network working as desired, though it needs more training data.

Documentation

Classes

| Class | Function | | ------- | ------- | | network | Create a new neural network with a set amount of layers and neurons | | neuron | It's used by the code of the network, I can't really find a reason to use it on its own. |

Example init calls

Network

  let network = new marvin.network([10, 4, 4, 2], 1);
  // This creates a network consisting of 4 layers
  // Layer 1: (The input layer)         has 10 neurons
  // Layer 2: (The first hidden layer)  has 4 neurons
  // Layer 3: (The last hidden layer)   has 4 neurons
  // Layer 4: (The output layer)        has 2 neurons
  // The learning rate is 1

Neuron

  let neuron = new marvin.neuron(2, 3);
  // This creates a new neuron, keep in mind that it's not added to a network

Network methods

| Method | What it does | | ------- | ------- | | network.log() | Logs the neural network; Useful for debugging | | network.getNeuron(layerNumber, neuronNumber) | Returns the neuron object at neuronNumber in layerNumber | | network.getLayer(layerNumber) | Returns an array of neurons from the specified layer | | network.forAllLayers(function, [backwardsBoolean]) | Executes a function for every layer. Goes back-to-front if backwardsBoolean is true | | network.forAllNeurons(function, [backwardsBoolean]) | Executes a function for every neuron. Goes back-to-front if backwardsBoolean is true | | network.forAllNeuronsInLayer(function, layerNumber) | Executes a function for neuron in the specified layer. | | network.getWeight(neuron1, neuron2) | Gets the weight value for the connection between the neurons | | network.setWeight(neuron1, neuron2, number) | Sets the weight value of the connection between the neurons | | network.train(input, expectedOutput) | Trains the network. The input is an array of the first layer. The expectedOutput is an array of the expected last layer. Returns the cost | | network.feed(input) | Feeds the input into the neural network and returns the last layer | | network.normalize(number) | Normalizes the input using the activation function specified in network.activationFunction | | network.deriveNormalize(number) | Reverts the input back to a normal number using the activation function specified in network.activationFunction | | network.save(path) | Save the network to a file so that you can use it again later | | network.load(path, [callbackFunction]) | Load a network from a file. (Optional): Supply a callback function that gets run after the loading process is done |

neuron Methods

| Method | What it does | | ----- | -------- | | neuron.setWeight(input) | Sets the weights of the neuron equal to the input. Input must be an array of the length of the previous layer | | neuron.setWeight(input, input2) | Sets the weight for previous neuron with the neuronNumber 'input' equal to input2 | | neuron.addWeight() | Adds a weight to the neuron | | neuron.setBias() | Sets the bias of the neuron | | neuron.log() | Logs the variables of the neuron; Useful for debugging |

An example

The goal of this example neural network is to copy the input layer. So, [1, 0, 0] in should mean [1, 0, 0] out.

const marvin = require("./marvin_node.js");

let iterations = 1000;

// Example init call
let n = new marvin.network([3, 4, 4, 3], 1);

console.log(`<<< TRAINING ${iterations} TIMES >>>\n`);

// Training the network with both hardcoded and generated data
for (let i = 1; i <= iterations; i++) {
  n.train([1, 0, 0], [1, 0, 0]);
  n.train([0, 1, 0], [0, 1, 0]);
  n.train([0, 0, 1], [0, 0, 1]);

  let r1 = Math.random();
  n.train([r1, 0, 0], [r1, 0, 0]);
  let r2 = Math.random();
  n.train([0, r2, 0], [0, r2, 0]);
  let r3 = Math.random();
  n.train([0, 0, r3], [0, 0, r3]);
}

console.log(`In: [1, 0, 0]; Expected out: [1, 0, 0]; Out: [${n.feed([1, 0, 0])}]`);
console.log(`In: [0, 1, 0]; Expected out: [0, 1, 0]; Out: [${n.feed([0, 1, 0])}]`);
console.log(`In: [0, 0, 1]; Expected out: [0, 0, 1]; Out: [${n.feed([0, 0, 1])}]`);