gdg-tangier / Cervello

Cervello

A simple machine learning lib written in Typescript.

Features

• Linear Regression with Gradient Descent.

• Feed forward Neural Network with backpropagation.

Try it out: Demo

Installation

Via CDN

<script src="https://unpkg.com/cervello/dist/cervello.js"></script>

<script>
    // Note that we should install NeuralNetwork feature in the global mode by calling:
    // Cervello.Cervello.NeuralNetwork
    // [LibName].[Namespace].[Feature]
    var nn = new Cervello.Cervello.NeuralNetwork()
    var lr = new Cervello.Cervello.LinearRegression()
    
    //Check
    console.log(nn, lr)
</script>

Via NPM

npm: npm install cervello --save

Linear Regression

import {Cervello} from "cervello"

let data = {inputs: [1,2,3,4,5,6], outputs: [1,2,3,4,5,6]}

let lr = new Cervello.LinearRegression().train(data)

let result = lr.predict(7)

console.log(result) // => 7

Import and export

You can serialize Linear model and export it by export(), or you can load (a,b) of a trained linear model and use them as well.

// Export
let seralize = lr.export() // [a,b] => (ax + b) 

// Import/Load
lr.import([1.5, 2]) // which means => (y = 1.5x + 2)

Neural Network

import {Cervello} from "cervello"

// XOR Problem
let data = [{input: [0, 0], output: [0]},
            {input: [0, 1], output: [1]},
            {input: [1, 0], output: [1]},
            {input: [1, 1], output: [0]}]

let nn = new Cervello.NeuralNetwork().train(data)

let result = nn.predict([0,1])

console.log(result) // => [0.99]

You can specify which options you want to use such like hiddenLayers, iterations, neurons (units), Learning Rate and activation method.

Available options:

• hiddenLayer: (numebr) : The number of hiddenLayer.
• units: (number) : The number of neurons foreach hiddenLayer.
• iteration: (number) : The number of iteration of (forward/back) probagation.
• learningRate: (number) : The number of how faster could our net learn.
• activator : (string) : activation method weather sigmoid or htan.

Other example (Digit recognition)

import {Cervello} from "cervello"

let one = number(
  '.....$.' +
  '..$$$$.' +
  '.....$.' +
  '.....$.' +
  '.....$.' +
  '.....$.' +
  '...$$$$'
)

let two = number(
  '$$$$$$$' +
  '......$' +
  '......$' +
  '$$$$$$$' +
  '$......' +
  '$......' +
  '$$$$$$$'
)

let three = number(
  '$$$$$$$' +
  '......$' +
  '......$' +
  '$$$$$$$' +
  '......$' +
  '......$' +
  '$$$$$$$'
)

let four = number(
  '$......' +
  '$.....$' +
  '$.....$' +
  '$$$$$$$' +
  '......$' +
  '......$' +
  '......$'
)

// Neural network options
let options = {hiddenLayer: 2, units: 4, iteration: 100000, activator: "sigmoid"}

// Init Neural Network.
let nn = new Cervello.NeuralNetwork(options)

// Train
nn.train([{input: one, output: results("1")},
          {input: two, output: results("2")},
          {input: three, output: results("3")},
          {input: four, output: results("4")}])

// Predict
let result = nn.predict(number(
  '.....$.' +
  '.$$$$$.' +
  '.....$.' +
  '.....$.' +
  '.....$.' +
  '.....$.' +
  '...$$$$'
))

console.log(result) // ~ [0.1]

// helpers
function number(string) {
  return string
    .trim()
    .split('')
    .map(function(symbol){
        if ('$' === symbol) return 1
        if ('.' === symbol) return 0
    })
}

function results(number) {
  if (number === '1') return [ 0.1 ]
  if (number === '2') return [ 0.3 ]
  if (number === '3') return [ 0.5 ]
  if (number === '4') return [ 0.7 ]
  return 0
}

Import and export

It's the same as Linear model, you can serialize the model wedges by export() that will return a serlized object of NN wedges.

// Export model
let serialized = nn.export() // => seralized object (string)

// Import model
nn.import(seralized)

Activation methods

• Sigmoid
• hyperbolic tangent

Contributors

LICENSE

MIT