The architecture of OpenANN is very general because it combines different learning approaches.

These are

supervised learning
unsupervised learning
reinforcement learning
evolutionary optimization

The most general concepts in the library are Optimizer and Optimizable. Optimizer is the common base class for optimization algorithms like IPOPCMAES, CG, LBFGS, LMA and MBSGD. An Optimizable provides at least an error function that has to be optimized during the learning process and could additionally provide the first and second derivative of the error function with respect to the weights to speed up the learning process. Most optimizers even require the first derivative. In supervised or unsupervised learning, we usually want to optimize a Learner. Each Learner combines a model with a DataSet. The dataset is called training set in this context. For each Learner we can compute an error and a gradient on the training set and it derives from Optimizable. However, we cannot only optimize learners with optimization algorithms but we can also optimize e.g. reinforcement learning agents that implement the Optimizable interface. The most important Learner in the library is the Net. It represents a feedforward multilayer neural network that can be trained supervised. Another example is the RBM, which is usually trained unsupervised and can then be used as a layer of a Net. A Net can consist of many different types of layers. They only have to implement the Layer interface.