NNet is a (metaprogrammed) neural network. It allows the selection of the network layout (number of layers and activation functions) in a simple way, without runtime costs -- check out the following 2 hidden layers network:

typedef NeuralNetwork<
        float,
        boost::mpl::vector<
                Layer<Tanh, 1>,
                Layer<Tanh, 12>,
                Layer<Tanh, 6>,
                Layer<Tanh, 1>
        > 
> NNet;

Sources include a simple SDL example (where the network is trained to learn the blue points), below a sample execution:

Source are available as follows: tarball download or online browsing.