Extracting State Machines from Feedforward Neural Networks

Artificial neural networks (ANNs) are a category of AI models based on the principles of connectionism that have recently been successfully used to solve problems that were not solvable with classic AI approaches in the past. The aim of this paper is to introduce a method to extract state machines from Feedforward Neural Networks (FNN) at a specific point of training to provide an approach in Inter-cognitive communication for better understanding the logical decision processes that are simulated by a FNN's calculations. This method is directly fitted to be used with FNNs that calculate decisions over continuous input data, where the future situation is based on the recent situation and the decisions that are made in the present. Based on the extraction from FNN's, themselves being a basis for many other neural network types, our method is supposed to deepen the link between connectionist's models and symbolic models, thus improving the usability of artificial neural networks data processing in the environment of symbolic human concepts.