From 3a46d27f4ed855dc6181c2f08c08110ec882a8ec Mon Sep 17 00:00:00 2001
From: Makarand Lahane <jayhoylehawkings@gmail.com>
Date: Fri, 9 Feb 2018 14:42:50 +0530
Subject: [PATCH] Multilayer Neural Network

Created a multilayer neural network class  which uses the matrix library in this repository for matrix operations. Some functionality is changed like you need to pass only 2 parameters(number of input nodes, number of output nodes) to the constructor of the neural network and for adding a hidden layer you need to call the function addHiddenLayer(number of neurons) for adding a hiddenLayer to your neural network and before training and after calling all addHiddenLayer function calls you need to call the config() function.
---
 MultiLayerNN.js | 109 ++++++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 109 insertions(+)
 create mode 100644 MultiLayerNN.js

diff --git a/MultiLayerNN.js b/MultiLayerNN.js
new file mode 100644
index 0000000..ed862ba
--- /dev/null
+++ b/MultiLayerNN.js
@@ -0,0 +1,109 @@
+function sigmoid(x) {
+  return 1/(1+Math.exp(-x));
+}
+
+function derivSigmoid(x) {
+  return x*(1-x);
+}
+
+
+function meanError(m) {
+  let sum=0;
+  for(let i=0;i<m.rows;i++) {
+    for(let j=0;j<m.cols;j++) {
+      sum+=m.data[i][j];
+    }
+  }
+  return sum/m.rows;
+}
+
+class NeuralNetwork {
+  constructor(numInput,numOutput) {
+    this.input_nodes=numInput;
+    this.output_nodes=numOutput;
+    this.hiddenLayers=[];
+    this.hiddenLayersConfig=[];
+    this.hiddenLayers.push(numInput);
+  }
+  addHiddenLayer(numOfNeurons) {
+    this.hiddenLayers.push(numOfNeurons);
+    let layerConfig={
+      layerBias:new Matrix(numOfNeurons,1),
+      layerWeight:new Matrix(numOfNeurons,this.hiddenLayers[this.hiddenLayers.length-2]),
+    };
+    layerConfig.layerBias.randomize();
+    layerConfig.layerWeight.randomize();
+    this.hiddenLayersConfig.push(layerConfig);
+  }
+  config() {
+    let outputLayerConfig={
+      layerBias:new Matrix(this.output_nodes,1),
+      layerWeight:new Matrix(this.output_nodes,this.hiddenLayers[this.hiddenLayers.length-1]),
+    };
+    outputLayerConfig.layerBias.randomize();
+    outputLayerConfig.layerWeight.randomize();
+    this.hiddenLayersConfig.push(outputLayerConfig);
+  }
+  predict(input) {
+    let inputLayer=Matrix.fromArray(input);
+    let layers=[];
+    layers.push(inputLayer);
+    for(let i=0;i<this.hiddenLayersConfig.length;i++) {
+      let l=Matrix.multiply(this.hiddenLayersConfig[i].layerWeight,layers[i]);
+      l.add(this.hiddenLayersConfig[i].layerBias);
+      l.map(sigmoid);
+      layers.push(l);
+    }
+    return layers[layers.length-1].toArray();
+  }
+  train(input,output,lr,itr) {
+    let inputLayer=Matrix.fromArray(input);
+    let outputLayer=Matrix.fromArray(output);
+    let layers=[];
+    //feedForward
+    layers.push(inputLayer);
+    for(let i=0;i<this.hiddenLayersConfig.length;i++) {
+      let l=Matrix.multiply(this.hiddenLayersConfig[i].layerWeight,layers[i]);
+      l.add(this.hiddenLayersConfig[i].layerBias);
+      l.map(sigmoid);
+      layers.push(l);
+    }
+    //calculate errors
+    let outputError=Matrix.subtract(outputLayer,layers[layers.length-1]);
+    let meanE=meanError(outputError);
+    if(itr%10000===0) {
+      console.log(abs(meanE));
+    }
+    //calculate errors for each layer except inputLayer
+    let layerErrors=[];
+    layerErrors.push(outputError);
+    for(let i=1;i<this.hiddenLayersConfig.length;i++) {
+      let tW=Matrix.transpose(this.hiddenLayersConfig[this.hiddenLayersConfig.length-i].layerWeight);
+      let layerE=Matrix.multiply(tW,layerErrors[layerErrors.length-1]);
+      layerErrors.push(layerE);
+    }
+    //calculate gradients
+    let gradients=[];
+    for(let i=0;i<layerErrors.length;i++) {
+      let layer=layers[(layers.length-1)-i];
+      let gradientLayer=Matrix.map(layer,derivSigmoid);
+      gradientLayer.multiply(layerErrors[i]);
+      gradientLayer.multiply(lr);
+      gradients.push(gradientLayer);
+    }
+    layers.splice(layers.length-1,1);
+    //calculate deltaWeights
+    let deltaWeights=[];
+    for(let i=0;i<layers.length;i++) {
+      let deltaWeight=Matrix.multiply(gradients[(gradients.length-1)-i],Matrix.transpose(layers[i]));
+      deltaWeights.push(deltaWeight);
+    }
+    gradients.reverse();
+    //change weights and biases
+    for(let i=0;i<this.hiddenLayersConfig.length;i++) {
+      this.hiddenLayersConfig[i].layerBias.add(gradients[i]);
+      this.hiddenLayersConfig[i].layerWeight.add(deltaWeights[i]);
+    }
+
+  }
+}