Rule extraction from recurrent neural networks

This thesis investigates rule extraction from recurrent neural networks, which takes the form of automated construction of models of an underlying network. Typically the models are expressed as finite state machines and they should mimic the network while being more intelligible. It is argued that rule extraction allows a deeper and more general form of analysis than other, more or less ad hoc, methods which are typically applied after the training of the recurrent networks. The first part of this thesis reviews and analyses the development of related techniques. The second part presents a novel algorithm, the Crystallizing Substochastic Sequential Machine Extractor (CrySSMEx), which efficiently generates a sequence of increasingly refined stochastic finite state models of an underlying system. Novel features of CrySSMEx include, for example, freedom from parameters, deterministic extraction, a hierarchical vector quantizer, and a stochastic finite state model which can be constructed also when some data is missing. Experiments show that CrySSMEx is, compared to other methods, applicable to a wider range of problems (such as high-dimensional or chaotic dynamic systems). Finally, the field is discussed from a more theoretical perspective in terms of scientific methodology targeted at simulated systems. It is suggested that a rule extractor (or Empirical Machine) can actively select data from the system it is set to model by continuously targeting the weakest point of its currently strongest model. These automated experimenters can, in turn, be made part of a framework (or Popperian Machine) in which theories about populations of systems are generated and tested in order to establish falsifiable statements. These statements should have a high empirical content and thus concisely describe emergent, and previously unknown, properties of the systems.