001 package aima.learning.neural;
002
003 import aima.util.Matrix;
004
005 public class BackPropLearning implements NNTrainingScheme {
006 private final double learningRate;
007 private final double momentum;
008
009 private Layer hiddenLayer;
010 private Layer outputLayer;
011 private LayerSensitivity hiddenSensitivity;
012 private LayerSensitivity outputSensitivity;
013
014 public BackPropLearning(double learningRate, double momentum) {
015
016 this.learningRate = learningRate;
017 this.momentum = momentum;
018
019 }
020
021 public void setNeuralNetwork(FunctionApproximator fapp) {
022 FeedForwardNeuralNetwork ffnn = (FeedForwardNeuralNetwork) fapp;
023 this.hiddenLayer = ffnn.getHiddenLayer();
024 this.outputLayer = ffnn.getOutputLayer();
025 this.hiddenSensitivity = new LayerSensitivity(hiddenLayer);
026 this.outputSensitivity = new LayerSensitivity(outputLayer);
027 }
028
029 public Vector processInput(FeedForwardNeuralNetwork network, Vector input) {
030
031 hiddenLayer.feedForward(input);
032 outputLayer.feedForward(hiddenLayer.getLastActivationValues());
033 return outputLayer.getLastActivationValues();
034 }
035
036 public void processError(FeedForwardNeuralNetwork network, Vector error) {
037 // TODO calculate total error somewhere
038 // create Sensitivity Matrices
039 outputSensitivity.sensitivityMatrixFromErrorMatrix(error);
040
041 hiddenSensitivity
042 .sensitivityMatrixFromSucceedingLayer(outputSensitivity);
043
044 // calculate weight Updates
045 calculateWeightUpdates(outputSensitivity, hiddenLayer
046 .getLastActivationValues(), learningRate, momentum);
047 calculateWeightUpdates(hiddenSensitivity, hiddenLayer
048 .getLastInputValues(), learningRate, momentum);
049
050 // calculate Bias Updates
051 calculateBiasUpdates(outputSensitivity, learningRate, momentum);
052 calculateBiasUpdates(hiddenSensitivity, learningRate, momentum);
053
054 // update weightsAndBiases
055 outputLayer.updateWeights();
056 outputLayer.updateBiases();
057
058 hiddenLayer.updateWeights();
059 hiddenLayer.updateBiases();
060
061 }
062
063 public Matrix calculateWeightUpdates(LayerSensitivity layerSensitivity,
064 Vector previousLayerActivationOrInput, double alpha, double momentum) {
065 Layer layer = layerSensitivity.getLayer();
066 Matrix activationTranspose = previousLayerActivationOrInput.transpose();
067 Matrix momentumLessUpdate = layerSensitivity.getSensitivityMatrix()
068 .times(activationTranspose).times(alpha).times(-1.0);
069 Matrix updateWithMomentum = layer.getLastWeightUpdateMatrix().times(
070 momentum).plus(momentumLessUpdate.times(1.0 - momentum));
071 layer.acceptNewWeightUpdate(updateWithMomentum.copy());
072 return updateWithMomentum;
073 }
074
075 public static Matrix calculateWeightUpdates(
076 LayerSensitivity layerSensitivity,
077 Vector previousLayerActivationOrInput, double alpha) {
078 Layer layer = layerSensitivity.getLayer();
079 Matrix activationTranspose = previousLayerActivationOrInput.transpose();
080 Matrix weightUpdateMatrix = layerSensitivity.getSensitivityMatrix()
081 .times(activationTranspose).times(alpha).times(-1.0);
082 layer.acceptNewWeightUpdate(weightUpdateMatrix.copy());
083 return weightUpdateMatrix;
084 }
085
086 public Vector calculateBiasUpdates(LayerSensitivity layerSensitivity,
087 double alpha, double momentum) {
088 Layer layer = layerSensitivity.getLayer();
089 Matrix biasUpdateMatrixWithoutMomentum = layerSensitivity
090 .getSensitivityMatrix().times(alpha).times(-1.0);
091
092 Matrix biasUpdateMatrixWithMomentum = layer.getLastBiasUpdateVector()
093 .times(momentum).plus(
094 biasUpdateMatrixWithoutMomentum.times(1.0 - momentum));
095 Vector result = new Vector(biasUpdateMatrixWithMomentum
096 .getRowDimension());
097 for (int i = 0; i < biasUpdateMatrixWithMomentum.getRowDimension(); i++) {
098 result.setValue(i, biasUpdateMatrixWithMomentum.get(i, 0));
099 }
100 layer.acceptNewBiasUpdate(result.copyVector());
101 return result;
102 }
103
104 public static Vector calculateBiasUpdates(
105 LayerSensitivity layerSensitivity, double alpha) {
106 Layer layer = layerSensitivity.getLayer();
107 Matrix biasUpdateMatrix = layerSensitivity.getSensitivityMatrix()
108 .times(alpha).times(-1.0);
109
110 Vector result = new Vector(biasUpdateMatrix.getRowDimension());
111 for (int i = 0; i < biasUpdateMatrix.getRowDimension(); i++) {
112 result.setValue(i, biasUpdateMatrix.get(i, 0));
113 }
114 layer.acceptNewBiasUpdate(result.copyVector());
115 return result;
116 }
117 }