panzer / javascript-neuralnet

A neural net written in Javascript. Makes use of the p5.js library for visualizing the net. The net is trained using a genetic algorithm.

Javascript Neural Net

A neural net written in Javascript. The net is trained using a genetic algorithm.

New in the Latest Update

Nets

• New property afunctions. Currently, four functions are in the list (linear, binary, sigmoid, and ReLU) but more are available in AFunctions.js. These are the activation functions available for each Node.

Nodes

• Each Node can mutate to have an activation function. Any function that takes a number and returns a number is eligible, just add it to afunctions in Net.js.
• In the getValue function, Node color is updated based on the activation function.

How it works

1. A population of populationSize nets is created, with each net having n_inputs, n_layers, n_height and n_outputs.
2. A scaled fitness is calculated for each net (the sum of all fitnesses in the population is 1)
   • The fitness of each net is determined by fit_tests provided in the sketch.js file.
   • The net with the least difference between desired output and actual output is the most fit.
3. The best net of that generation is drawn on screen
4. A new generation of Nets is created
   • The best Net is always included
   • The rest of the generation is made up by choosing the most fit Nets and applying slight mutations in the weights
5. Steps 2 - 4 repeat

How to use it

Load up index.html and press the 'Start' button. If you'd like to make modifications to the net, or the training data, you may be able to find what you're looking for in the sketch.js file.

Current Training Data

Currently the nets are training on a 2x2 pixel image.

Inputs

• Top Left pixel [-1, 1]
• Top Right pixel [-1, 1]
• Bottom Left pixel [-1, 1]
• Bottom Right pixel [-1, 1]

Output

• Solid? [0, 1]
• Horizontal? [0, 1]
• Vertical? [0, 1]
• Diagonal? [0, 1]

Try it yourself!

I recommend using Chrome, right clicking, Inspect, and using the console. Use the command bestMember.output(input) where input is an array of your input. You must have at least as many inputs as your net requires!

If you'd like to try the 2x2 pixel image example, keep in mind the expect input range [-1, 1]. Your console input would look something like bestMember.output([1, 1, -0.5, -0.5, 1]). The input of 1 at the end is a constant input factor (which has been included in the training input).

Contributing

Contributions are welcome! If you have some better way of training, if you want to make the user interface better... go ahead! Check out the issues tab and see if you can help out with any of those.

Credits

Makes use of the p5.js library for visualizing the Net and some math functions.

License

This project is licensed under the MIT License - see the LICENSE.md file for details.

javascript-neuralnet's Issues

Allow inputs into the net via some user interface

Make it more pretty and useful so you don't have to go into Chrome console

Have the net size/number of layers genetically mutate

Random mutations of the net size could happen between generations.