Tinyneuron NPM

tinyneuron

An implementation of various neural networks concepts in JavaScript

Perceptrons

Creation/config

Create a perceptron like so:

const config = { inputs: [0.25, 0.5, -0.75, 1.0, -1.25] };
const perceptron = new Perceptron(config);

A Perceptron has inputs, weights, a learningRate, and an activation function activation, which can all be passed into the constructor config. The activation function defaults to the sign function, const sign = (x) => x >= 0 ? 1 : -1. The number of weights defaults to inputs length. You can also optionally disable the bias neuron with the config flag addBias: false, or tweak the bias output with bias: value. The learning rate defaults to 5%, which still allows for dramatic improvements between cycles.

To generate perceptron outputs:

const perceptron = new Perceptron({ numWeights: 2, inputs: [1.0, -1.0] });
console.log(perceptron.getOutput()) // -> activation( (Math.random() * 2 - 1) * 1.0 + (Math.random() * 2 - 1) * -1.0)

weights are assigned randomly as values between -1 and 1.

To reassign weights and inputs, call percepton.assignWeights(array) or perceptron.assignInputs(array)

Training

To train the perceptron against a target value, run:

perceptron.train(targetOutput);

This function calculates the predicted output based on current weights and inputs and an error between the target and predicted outputs tempered by the learning rate. It then adjusts its weights to compensate for error. For a large learning rate generally only a few rounds of training at most are necessary.

Activation

To use another activation function (e.g. sigmoid):

import { activation, Perceptron } from 'tinyneuron';
const perceptron = new Perceptron({ activation: activation.sigmoid });

Example: AND gate

const andNode = new Perceptron({
  addBias: true,
  weights: [1, 1, -1.5]
});

andNode.assignInputs([0, 0]);
andNode.getOutput() === -1;

andNode.assignInputs([0, 1]);
andNode.getOutput() === -1;

andNode.assignInputs([1, 0]);
andNode.getOutput() === -1;

andNode.assignInputs([1, 1]);
andNode.getOutput() === 1;

Neural Networks

Creation/config

Create a neural network like so:

import { NeuralNetwork } from 'tinyneuron';

const config = {
  inputNodes: 2,
  hiddenNodes: 2,
  outputNodes: 1,
  learningRate: 0.25
}
const network = new NeuralNetwork(config);

Training

To train a network, call train(inputs, targets) whereinputsis an array with dimensions matching the number of input nodes, andtargetsis an array with dimensions matching the number of output nodes.train` returns an array of outputs for the training round. For example:

Example

The following network can learn XOR:

const network = new NeuralNetwork({
  inputNodes: 2,
  hiddenNodes: 10,
  outputNodes: 1,
  learningRate: 0.25
});
const inputs = [[1, 0], [0, 1], [1, 1], [0, 0]];
const targets = [1, 1, 0, 0];

for (let i = 0; i < 50000; i++) {
  const idx = Math.floor(Math.random() * 4);
  network.train(inputs[idx], [targets[idx]]);
}

network.feedForward([1, 0]) // => approximately [1]
network.feedForward([0, 1]) // => approximately [1]
network.feedForward([0, 0]) // => approximately [0]
network.feedForward([1, 1]) // => approximately [0]