alert = s => document.querySelector ("pre").innerText += s + "\n" Math.sigmoid = x => 1/(1 + Math.exp(-x)) class Matrix { constructor() { this.data = new Array } get dim() { return [this.data.length, this.data[0].length] } get T() { const [n, m] = this.dim let res = Matrix.zero(m, n) for(let i = 0; i < n; i++) for(let j = 0; j < m; j++) res.data[j][i] = this.data[i][j] return res } static array(arr) { let matrix = new Matrix matrix.data = arr return matrix } static zero(n, m) { let matrix = new Matrix for(let i = 0; i < n; i++) matrix.data.push(new Array(m).fill(0)) return matrix } static random(n, m) { let matrix = new Matrix for(let i = 0; i < n; i++) { matrix.data.push(new Array) for(let j = 0; j < m; j++) matrix.data[i].push(2*Math.random()-1) } return matrix } static diagonal(mat) { const n = mat.dim[1] let matrix = new Matrix for(let i = 0; i < n; i++) { matrix.data.push(new Array) for(let j = 0; j < n; j++) matrix.data[i].push(i==j?mat.data[0][i]:0) } return matrix } row(r) { const n = this.dim[1] return Matrix.each(Matrix.zero(1, n), (_, i, j) => this.data[r][j]) } add(mat) { const [n, m] = this.dim let res = Matrix.zero(n, m) for(let i = 0; i < n; i++) for(let j = 0; j < m; j++) res.data[i][j] = this.data[i][j] + mat.data[i][j] return res } sub(mat) { const [n, m] = this.dim let res = Matrix.zero(n, m) for(let i = 0; i < n; i++) for(let j = 0; j < m; j++) res.data[i][j] = this.data[i][j] - mat.data[i][j] return res } dot(mat) { const n = this.dim[0], m = mat.dim[1] let matrix = Matrix.zero(n, m) for(let j = 0; j < m; j++) for(let i = 0; i < n; i++) for(let k = 0; k < this.dim[1]; k++) matrix.data[i][j] += this.data[i][k]*mat.data[k][j] return matrix } mult(num) { const [n, m] = this.dim let res = Matrix.zero(n, m) for(let i = 0; i < n; i++) for(let j = 0; j < m; j++) res.data[i][j] = num*this.data[i][j] return res } multMat(mat) { const [n, m] = this.dim let res = Matrix.zero(n, m) for(let i = 0; i < n; i++) for(let j = 0; j < m; j++) res.data[i][j] = this.data[i][j]*mat.data[i][j] return res } static each(mat, func) { const [n, m] = mat.dim let res = Matrix.zero(n, m) for(let i = 0; i < n; i++) for(let j = 0; j < m; j++) res.data[i][j] = func(mat.data[i][j], i, j) return res } toString() { return "[" + this.data.map(row => "[" + row.map(d => d.toFixed(3)).join(",") + "]" ).join(",\n ") + "]" } } class Layer { constructor(inp, out) { this.weights = Matrix.random(out, inp) this.dim = this.weights.dim } forward(data) { return Matrix.each (this.weights.dot(data), Math.sigmoid) } backprop(data, output, error, rate) { const ksy = pred => pred*(1 - pred) // current del matrix let del = Matrix.each(output, ksy) del = error.multMat(del) let adjustment = del.dot(data.T).mult(rate) this.weights = this.weights.add(adjustment) } } onload = function() { let l = new Layer(2, 1) // training operation f(x,y) = x let x = Matrix.array([[0,0],[0,1],[1,1],[1,0]]) let y = Matrix.array([[0,0,1,1]]).T let samples = x.dim[0] let iterations = 2000 alert("data sample:") alert(x) alert("output sample:") alert(y) alert("initial weights:") alert(l.weights) alert("initial predictions:") let p = Matrix.each(Matrix.zero(4, 1), (_, i) => l.forward(x.row(i).T).data[0][0]) alert(p) alert("running training process...") alert("samples amount: " + samples) alert("iterations: " + iterations) for(let i = 0; i < iterations; i++) for(let s = 0; s < samples; s++) { let data = x.row(s).T, out = y.row(s).T // make prediction let pred = l.forward(data) // count error value (derivative actually) let err = out.sub(pred) // perform backpropagation... l.backprop(data, pred, err, 1) } alert("weights after training:") alert(l.weights) alert("predictions after training:") p = Matrix.each(Matrix.zero(4, 1), (_, i) => l.forward(x.row(i).T).data[0][0]) alert(p) }

my rnn test