Neuralware engineering: develop verifiable ANN-based systems

Abstract

Artificial neural networks (ANN) play an important part in developing intelligent robotic and autonomous systems; it relies on training to formulate the control mechanisms. When such ANN-based components are embedded in a larger system, their interactions become harder to analyze and model. Formal testing of such system for safety properties is extremely hard due to the lack of a complete system model. In this paper we propose the neuralware engineering framework to address the above issues. This framework is based on our experience with verifying and testing complex software systems. It is based on an iterative approach on specification, model checking, and testing. After the ANN-based system is designed and trained using an initial partial system model, a rule extraction algorithm is used to discover what has been learned. The discrepancies between the learned rules and the model is compared to modify the system model. This process is repeated until the behavior of the real system is validated against the model and specification.