NeuralNetwork.Convolutional.OCR

Prototype for recognition of handwritten digits.

Repo includes following components:

• convolutional neural network (CNN) as a base component;
• back-propagation training algorithms;
• simple GUI app to demonstrate NN working process.

Usage

1. Builder for your neural network

You may create your own builder with specific logic for your NN. Or simply use existing builder named NeuralNetworkBuilder.

  INeuralNetworkBuilder<double> networkBuilder = new NeuralNetworkBuilder<double>();

2. Factory to create specific instance of NN.

By default the the library has already the default implementation of factory for NN.

public class SpecificNeuralNetworkFactory : INeuralNetworkFactory<double>
{
  ...
}

3. Connection schemes.

Additionally, you can configure different schemes of connections between layers in case multi-layer NN. And if you want to connect layers using any specific logic - just create a specific implementation of IConnectionScheme interface.

For example, scheme to connect each neurons with each other (fully-connected layers), specified below:

public class FullyConnectedScheme<TData> 
  : IConnectionScheme<INeuron<TData>, IConnectionFactory<TData, TData>>
{
    private IConnectionFactory<TData, TData> connectionFactory;

    public void SetConnectionFactory(IConnectionFactory<TData, TData> factory)
    {
        connectionFactory = factory;
    }

    public void Apply(
        ILayer<INeuron<TData>, IConnectionFactory<TData, TData>> leftLayer, 
        ILayer<INeuron<TData>, IConnectionFactory<TData, TData>> rightLayer)
    {
        foreach (INeuron<TData> currentRightNeuron in rightLayer.Neurons)
        {
            foreach (INeuron<TData> currentLeftNeuron in leftLayer.Neurons)
            {
                currentRightNeuron.Inputs.Add(connectionFactory.Create(currentLeftNeuron));
                currentLeftNeuron.Outputs.Add(connectionFactory.Create(currentRightNeuron));
            }
        }
    }
}

4. Configure NN using connection schemes and specific factory.

The example, specified below, shows how to create a neural network with three layers which connected using FullyConnectedScheme scheme.

IMultiLayerNeuralNetwork<double> neuralNetwork = networkBuilder
    .AddInputLayer(new Layer(0, 3, true, false))
    .AddLayer<FullyConnectedScheme<double>>(new Layer(1, 5, false, false))
    .AddLayer<FullyConnectedScheme<double>>(new Layer(2, 2, false, true))
    .Build<NeuralNetworkFactory>() as IMultiLayerNeuralNetwork<double>;