Training Feed-Forward Networks

This tutorial will show you how to construct a feed-forward multi-layer neural network and how to train it efficiently using minibatch training and the Adam Optimization algorithm. It is recommended to read the getting started section, especially the introduction about General Optimization Tasks.

For this tutorial the following includes are needed:

//the model
#include <shark/Models/LinearModel.h>//single dense layer
#include <shark/Models/ConcatenatedModel.h>//for stacking layers, proveides operator>>
//training the  model
#include <shark/ObjectiveFunctions/ErrorFunction.h>//error function, allows for minibatch training
#include <shark/ObjectiveFunctions/Loss/CrossEntropy.h> // loss used for supervised training
#include <shark/ObjectiveFunctions/Loss/ZeroOneLoss.h> // loss used for evaluation of performance
#include <shark/Algorithms/GradientDescent/Adam.h> //optimizer: simple gradient descent.
#include <shark/Data/SparseData.h> //loading the dataset
using namespace shark;

Loading the Data

In this tutorial, we want to learn to recognize handwriten digits from the mnist dataset using a feed-forward network. This is done similar to the previous tutorials by loading a dataset and splitting it into training and test part:

std::size_t batchSize = 256;
LabeledData<RealVector,unsigned int> data;
importSparseData( data, argv[1], 0, batchSize );
data.shuffle(); //shuffle data randomly
auto test = splitAtElement(data, 70 * data.numberOfElements() / 100);//split a test set
std::size_t numClasses = numberOfClasses(data);

Note that we define the batchsize of the dataset during loading of the dataset. This will create the apropriately sized batches which we will later use for minibatch training.

Defining the Network topology

Next, we define our feed forward network. For this, we stack several layers on top of each other. For simplicity we will use only LinearModel with nonlinear activation functions. For this tutorial we decide that the two hidden layers should use Rectified-Linear Units (ReLu). For classification tasks, output neurons should be linear. The Layers are then chained together using operator>>:

//We use a dense linear model with rectifier activations
typedef LinearModel<RealVector, RectifierNeuron> DenseLayer;

//build the network
DenseLayer layer1(data.inputShape(),hidden1);
DenseLayer layer2(layer1.outputShape(),hidden2);
LinearModel<RealVector> output(layer2.outputShape(),numClasses);
auto network = layer1 >> layer2 >> output;

As you can see, we define the number of inputs and outputs for each layer during construction. Note that instead of memorising the number inputs and outputs, we can instead use the inputShape() and outputShape() functions.

Training the Network

After we loaded the dataset and defined the topology of the network, we can train it. Since we use a classification task, we can use the CrossEntropy error to maximize the class probability. The cross entropy assumes the inputs to be the log of the unnormalized probability \(p(y=c|x)\), i.e. the probability of the input to belong to class \(c\). The cross entropy uses an exponential normalisation to transform the inputs into proper normalised probabilities, however this is done in a numerically stable way.

The c-th output neuron of the network encodes in this case the probability of class c. In case of a binary problem, we can omit one output neuron and in this case, it is assumed that the output of the imaginary second neuron is just the negative of the first. The loss function takes care of the normalisation. After training, the most likely class label of an input can be evaluated by picking the class of the neuron with highest activation value. In the case of only one output neuron, the sign decides: negative activation is class 0, positive is class 1.

We will setup our error function to use minibatch training. Every time the error function is evaluated, a random batch in the dataset is evaluated. Thus the batch size defined earlier is an important parameter and a trade-off betwen evaluation speed and noise of the evaluation:

//create the supervised problem.
CrossEntropy loss;
ErrorFunction error(data, &network, &loss, true);//enable minibatch training

//optimize the model
std::cout<<"training network"<<std::endl;
Adam optimizer;
for(std::size_t i = 0; i != iterations; ++i){
        std::cout<<i<<" "<<optimizer.solution().value<<std::endl;

Other network types

Shark offers many different types of neural other neural networks, including radial basis function networks using RBFLayer and Convolutional Neural Networks using Conv2DModel

Full example program

The full example program is FFNNBasicTutorial.cpp.