liquidcarrot / carrot

🥕 Evolutionary Neural Networks in JavaScript

ℹ️ The new TypeScript version is coming! If you would like to try the expiremental version please clone the repository and checkout the typescript branch of the project. Docs for this new version can temporarily be found here

Carrot is an architecture-free neural network library built around neuroevolution

Why / when should I use this?

Whenever you have a problem that you:

• Don't know how-to solve
• Don't want to design a custom network for
• Want to discover the ideal neural-network structure for

You can use Carrot's ability to design networks of arbitrary complexity by itself to solve whatever problem you have. If you want to see Carrot designing a neural-network to play flappy-bird check here

For Documentation, visit here

Key Features

• Simple docs & interactive examples
• Neuro-evolution & population based training
• Multi-threading & GPU (coming soon)
• Preconfigured GRU, LSTM, NARX Networks
• Mutable Neurons, Layers, Groups, and Networks
• SVG Network Visualizations using D3.js

Demos

Flappy bird neuro-evolution

Install

$ npm i @liquid-carrot/carrot

Carrot files are hosted by JSDelivr

For prototyping or learning, use the latest version here:

<script src="https://cdn.jsdelivr.net/npm/@liquid-carrot/carrot/dist/carrot.umd2.min.js"></script>

For production, link to a specific version number to avoid unexpected breakage from newer versions:

<script src="https://cdn.jsdelivr.net/npm/@liquid-carrot/[email protected]/dist/carrot.umd2.min.js"></script>

Getting Started

💡 Want to be super knowledgeable about neuro-evolution in a few minutes?

Check out this article by the creator of NEAT, Kenneth Stanley

💡 Curious about how neural-networks can understand speech and video?

Check out this video on Recurrent Neural Networks, from @LearnedVector, on YouTube

This is a simple perceptron:

.

How to build it with Carrot:

let { architect } = require('@liquid-carrot/carrot');

// The example Perceptron you see above with 4 inputs, 5 hidden, and 1 output neuron
let simplePerceptron = new architect.Perceptron(4, 5, 1);

Building networks is easy with 6 built-in networks

let { architect } = require('@liquid-carrot/carrot');

let LSTM = new architect.LSTM(4, 5, 1);

// Add as many hidden layers as needed
let Perceptron = new architect.Perceptron(4, 5, 20, 5, 10, 1);

Building custom network architectures

let architect = require('@liquid-carrot/carrot').architect
let Layer = require('@liquid-carrot/carrot').Layer

let input = new Layer.Dense(1);
let hidden1 = new Layer.LSTM(5);
let hidden2 = new Layer.GRU(1);
let output = new Layer.Dense(1);

// connect however you want
input.connect(hidden1);
hidden1.connect(hidden2);
hidden2.connect(output);

let network = architect.Construct([input, hidden1, hidden2, output]);

Networks also shape themselves with neuro-evolution

let { Network, methods } = require('@liquid-carrot/carrot');

// this network learns the XOR gate (through neuro-evolution)
async function execute () {
  // no hidden layers...
   var network = new Network(2,1);

   // XOR dataset
   var trainingSet = [
       { input: [0,0], output: [0] },
       { input: [0,1], output: [1] },
       { input: [1,0], output: [1] },
       { input: [1,1], output: [0] }
   ];

   await network.evolve(trainingSet, {
       mutation: methods.mutation.FFW,
       equal: true,
       error: 0.05,
       elitism: 5,
       mutation_rate: 0.5
   });

   // and it works!
   network.activate([0,0]); // 0.2413
   network.activate([0,1]); // 1.0000
   network.activate([1,0]); // 0.7663
   network.activate([1,1]); // 0.008
}

execute();

Build vanilla neural networks

let Network = require('@liquid-carrot/carrot').Network

let network = new Network([2, 2, 1]) // Builds a neural network with 5 neurons: 2 + 2 + 1

Or implement custom algorithms with neuron-level control

let Node = require('@liquid-carrot/carrot').Node

let A = new Node() // neuron
let B = new Node() // neuron

A.connect(B)
A.activate(0.5)
console.log(B.activate())

Try with

Data Sets

• MNIST

Contributors ✨

This project exists thanks to all the people who contribute. We can't do it without you! 🙇

Thanks goes to these wonderful people (emoji key):

Luis Carbonell 💻 🤔 👀 📖

Christian Echevarria 💻 📖 🚇

Daniel Ryan 🐛 👀

IviieMtz ⚠️

Nicholas Szerman 💻

tracy collins 🐛

Manuel Raimann 🐛 💻 🤔

This project follows the all-contributors specification. Contributions of any kind welcome!

💬 Contributing

Your contributions are always welcome! Please have a look at the contribution guidelines first. 🎉

To build a community welcome to all, Carrot follows the Contributor Covenant Code of Conduct.

And finally, a big thank you to all of you for supporting! 🤗

Planned Features

* [ ] Performance Enhancements * [ ] GPU Acceleration * [ ] Tests * [ ] Benchmarks * [ ] Matrix Multiplications * [ ] Tests * [ ] Benchmarks * [ ] Clustering | Multi-Threading * [ ] Tests * [ ] Benchmarks * [ ] Syntax Support * [ ] Callbacks * [ ] Promises * [ ] Streaming * [ ] Async/Await * [ ] Math Support * [ ] Big Numbers * [ ] Small Numbers

Patrons

Acknowledgements

A special thanks to:

@wagenaartje for Neataptic which was the starting point for this project

@cazala for Synaptic which pioneered architecture free neural networks in javascript and was the starting point for Neataptic

@robertleeplummerjr for GPU.js which makes using GPU in JS easy and Brain.js which has inspired Carrot's development