/**
* Adds the layers to the architecture such that the architecture represents a Cascade network.
* All layers are fully connected to the input layer Each hidden layer is fully connected to the
* hidden layer added before it.
*
* @param architecture {@inheritDoc }
*/
@Override
public void buildArchitecture(Architecture architecture) {
List<Layer> layers = architecture.getLayers();
layers.clear();
List<LayerConfiguration> layerConfigurations = this.getLayerConfigurations();
int listSize = layerConfigurations.size();
// build the input layer
ForwardingLayer inputLayer = new ForwardingLayer();
inputLayer.setSourceSize(layerConfigurations.get(0).getSize());
if (layerConfigurations.get(0).isBias()) {
inputLayer.setBias(true);
inputLayer.add(new BiasNeuron());
}
layers.add(inputLayer);
// build the hidden layers and output layer
int sumOfPreviousLayerAbsoluteSizes = inputLayer.size();
for (int curLayer = 1; curLayer < layerConfigurations.size(); ++curLayer) {
if (layerConfigurations.get(curLayer).getSize() == 0)
throw new UnsupportedOperationException(
"Hidden layers must have at least one neuron each.");
layerConfigurations
.get(curLayer)
.setBias(false); // All neurons have access to the bias in the input layer.
Layer newLayer =
this.getLayerBuilder()
.buildLayer(layerConfigurations.get(curLayer), sumOfPreviousLayerAbsoluteSizes);
layers.add(newLayer);
sumOfPreviousLayerAbsoluteSizes += newLayer.size();
}
}
/**
* Performs a gradient decent backpropagation given the previous {@link StandardPattern} as input
* as well as the weight updates after the previous execution of a backpropagation. If the
* previous weight updates do not exist, the visitor will create them and initialize them to zero.
*
* @param architecture
*/
@Override
public void visit(Architecture architecture) {
List<Layer> layers = architecture.getLayers();
int numLayers = layers.size();
int currentLayerIdx = numLayers - 1; // skip input layer
Layer currentLayer = layers.get(currentLayerIdx);
int layerSize = currentLayer.size();
Layer nextLayer = null;
int nextLayerSize = -1;
Neuron currentNeuron;
// setup delta storage
layerWeightsDelta = new double[numLayers - 1][]; // not necessary for input layer
// calculate output layer deltas
layerWeightsDelta[currentLayerIdx - 1] = new double[layerSize];
for (int k = 0; k < layerSize; k++) {
currentNeuron = currentLayer.get(k);
double t_k =
layerSize > 1
? ((Vector) previousPattern.getTarget()).doubleValueOf(k)
: ((Real) previousPattern.getTarget()).doubleValue();
double o_k = currentNeuron.getActivation();
layerWeightsDelta[currentLayerIdx - 1][k] =
-1.0 * (t_k - o_k) * currentNeuron.getActivationFunction().getGradient(o_k);
}
// calculate deltas for all hidden layers
for (currentLayerIdx = numLayers - 2; currentLayerIdx > 0; currentLayerIdx--) {
currentLayer = layers.get(currentLayerIdx);
layerSize = currentLayer.size();
layerSize = currentLayer.isBias() ? layerSize - 1 : layerSize;
layerWeightsDelta[currentLayerIdx - 1] = new double[layerSize];
for (int j = 0; j < layerSize; j++) {
layerWeightsDelta[currentLayerIdx - 1][j] = 0.0;
nextLayer = layers.get(currentLayerIdx + 1);
nextLayerSize = nextLayer.size();
nextLayerSize = nextLayer.isBias() ? nextLayerSize - 1 : nextLayerSize;
for (int k = 0; k < nextLayerSize; k++) {
double w_kj = nextLayer.get(k).getWeights().doubleValueOf(j);
layerWeightsDelta[currentLayerIdx - 1][j] += w_kj * layerWeightsDelta[currentLayerIdx][k];
}
currentNeuron = currentLayer.get(j);
layerWeightsDelta[currentLayerIdx - 1][j] *=
currentNeuron.getActivationFunction().getGradient(currentNeuron.getActivation());
}
}
// storage for the weight updates
if (previousWeightUpdates == null) {
previousWeightUpdates = new double[numLayers - 1][];
for (currentLayerIdx = numLayers - 1; currentLayerIdx > 0; currentLayerIdx--) {
for (int k = 0; k < layerSize; k++) {
currentLayer = layers.get(currentLayerIdx);
layerSize = currentLayer.isBias() ? currentLayer.size() - 1 : currentLayer.size();
int previousLayerSize = layers.get(currentLayerIdx - 1).size();
previousWeightUpdates[currentLayerIdx - 1] =
new double[layerSize * previousLayerSize + previousLayerSize + 1];
}
}
}
((ForwardingLayer) layers.get(0)).setSource(new PatternInputSource(previousPattern));
// updates output and all hidden layer weights
for (currentLayerIdx = numLayers - 1;
currentLayerIdx > 0;
currentLayerIdx--) { // loop excludes input layer
currentLayer = layers.get(currentLayerIdx);
layerSize = currentLayer.isBias() ? currentLayer.size() - 1 : currentLayer.size();
int previousLayerSize = -1;
Layer previousLayer = null;
for (int k = 0; k < layerSize; k++) {
currentNeuron = currentLayer.get(k);
previousLayer = layers.get(currentLayerIdx - 1);
previousLayerSize = previousLayer.size();
double tmp = (-1.0 * learningRate) * layerWeightsDelta[currentLayerIdx - 1][k];
for (int j = 0; j < previousLayerSize; j++) {
double weight = currentNeuron.getWeights().doubleValueOf(j);
double newWeightUpdate = tmp * previousLayer.getNeuralInput(j);
double update =
newWeightUpdate
+ momentum
* previousWeightUpdates[currentLayerIdx - 1][k * previousLayerSize + j];
currentNeuron.getWeights().setReal(j, weight + update);
previousWeightUpdates[currentLayerIdx - 1][k * previousLayerSize + j] = newWeightUpdate;
}
}
}
}