A lightweight Neural Network library in Javascript

This library is inspired by Toy-Neural-Network, which works for one hidden layer. mini-ANN-js provides basic ANN functionalities which includes ability to create multilayer architecture, feed forward, training through backpropagation and some functions for genetic algorithm.

Documentation

• Initialize Neural Network

// ANN with 4 inputs, 3 neurons in hidden layer and 2 outputs
const my_ann = new NeuralNetwork([4, 3, 2]);
// initializes ANN with random weights and biases

• Changing Activation function

// set ReLU function as activation function
my_ann.setActivation(NeuralNetwork.ReLU);

// set Sigmoid function as activation function
my_ann.setActivation(NeuralNetwork.SIGMOID);

// By default Sigmoid is the activation function

• Performing feed forward

// passing 4 inputs as follows...
const output = my_ann.feedforward([0, 2, 1, 2]);
// returns an array with outputs of ANN

// pass 2nd arg for feedforward as true to get all layers instead of just output
const all_layers = my_ann.feedforward([0, 2, 1, 2], true);

• Training ANN

let input = [0, 2, 1, 2];
let expected_output = [1,0];
my_ann.train(input, expected_output);

• Functions for Genetic algorithms

// mutate weights and biases of ANN
my_ann.mutate(0.2); //mutation rate = 0.2 (min-0 & max-1)

// creates a copy of ann
new_ann = my_ann.copy()

// crossover
const offspring = my_ann.crossover(other_ann);

• Visualization function

const canvas = document.getElementById("canvas")
const ctx = canvas.getContext("2d")
const nn = new NeuralNetwork([3,5,3,2])
nn.draw(
  ctx, // canvas context
  100, // top left x coordinate to start drawing from
  100, // top left y coordinate to start drawing from
  [-0.5, 0.75, 0.2], // OPTIONAL network inputs. When given also visualizes activation states
  20 // OPTIONAL neuron radius. Default = 20
)

Example output: