let width = 1200;

let height = 600;

let training

let output

//let b = 0.1; //Bias

let n = 16; //Number of pixels

let N = 1000 //Training iterations

let η = 0.5; //Learning rate

function setup() {

  createCanvas(width, height);

  for (let i = 0; i < n; i++) {

    let temp = [];

    for (let j = 0; j < n; j++) {

      temp.push(1 / n);

    }

    weights.push(temp);

  }

  calc = createButton("Train");

  calc.mousePressed(train);

  //Training for N iterations

  for (let i = 0; i < N; i++) {

    train();

  }

}

​

function draw() {

  background(220);

  //This creates the image of the machine output

  for (let i = 0; i < sqrt(n); i++) {

    for (let j = 0; j < sqrt(n); j++) {

      fill(output[i + 4 * j] * 255);

      rect(

        (i * width) / sqrt(n) / 2,

        (j * height) / sqrt(n),

        ((i + 1) * width) / sqrt(n) / 2,

        ((j + 1) * height) / sqrt(n)

      );

    }

  }

  //This draws the training image

  for (let i = 0; i < sqrt(n); i++) {

    for (let j = 0; j < sqrt(n); j++) {

      fill(training[i + 4 * j] * 255);

      rect(

        ((i + 4) * width) / sqrt(n) / 2,

        (j * height) / sqrt(n),

        ((i + 5) * width) / sqrt(n) / 2,

        ((j + 1) * height) / sqrt(n)

      );

    }

  }

}

//This trains the machine and adjusts the weights on a random training image

function train() {

  //Here I create a new random training image

  training = [];

  for (let i = 0; i < n; i++) {

    training.push(round(random()));

  }

  output = matMul(weights, training);

  //Here I adjust the weights according to the output and the expected output(training)

  for (let i = 0; i < weights.length; i++) {

    for (let j = 0; j < weights[i].length; j++) {

      weights[i][j] +=

        (training[i] - output[i])*output[j]*(1-output[j])*training[j]* η;

    }

  }

  print(weights)

  //print(output)

}

//This is the scalar product to calculate each pixel of the output picture

function dot(v, w) {

  let S = 0;

  for (let i = 0; i < v.length; i++) {

    S += v[i] * w[i];

  }

  return S;

}

//This performs all the scalar products, takes the sigmoid on the result and produces the total image

function matMul(A, x) {

  let S = [];

  for (let i = 0; i < A.length; i++) {

    S.push(sigmoid(dot(A[i], x)));

  }

  return S;

}

//The sigmoid function makes sure our values are between 0 and 1

function sigmoid(x) {

  return 1 / (1 + exp(-x));

}

​