Synchronous neural networks for cyber-physical systems

Cyber-physical systems (CPS), such as autonomous vehicles or smart power grids, use interactive machine learning modules for decision making. Current design approaches use multiple machine learning modules, often using neural networks, to achieve the desired functionality. Timing validation is performed using measurement-based approaches, which may produce unsound results. To this end, we propose a new approach for designing such systems, by relying on the well known synchronous paradigm. Using this approach, we introduce Synchronous Artificial Neural Networks (SANNs), where we associate logical time to the different operations of the network. This approach provides sound compositional primitives, which enable the composition of interacting neural networks to ensure causality and determinism. We then show that we can embed the generated code on time predictable platforms enabling static analysis. Overall, this paper develops synchronous neural networks implemented in Esterel for the design of time predictable systems. We demonstrate the efficacy of our approach by developing a time predictable implementation of several applications, ranging from 5-100+ neurons, realised using the T-CREST platform. We also implemented a complex Convolutional Neural Network (CNN) application comprising of 1000+ neurons and 16 different layers using Esterel for a soft real-time application. Overall, the developed methodology opens new avenues of research in the direction of time predictable neural networks.