A Compositional Approach for Real-Time Machine Learning

Cyber-Physical Systems are highly safety critical, especially since they have to provide both functional and timing guarantees. Increasingly, Cyber-Physical Systems such as autonomous vehicles are relying on Artificial Neural Networks in their decision making and this has obvious safety implications. While many formal approaches have been recently developed for ensuring functional correctness of machine learning modules involving Artificial Neural Networks, the issue of timing correctness has received scant attention. This paper proposes a new compiler from the well known Keras Neural Network library to hardware to mitigate the above problem. In the developed approach, we compile networks of Artificial Neural Networks, called Meta Neural Networks, to hardware implementations using a new synchronous semantics for their execution. The developed semantics enables compilation of Meta Neural Networks to a parallel hardware implementation involving limited hardware resources. The developed compiler is semantics driven and guarantees that the generated implementation is deterministic and time predictable. The compiler also provides a better alternative for the realisation of non-linear functions in hardware. Overall, we show that the developed approach is significantly more efficient than a software approach, without the burden of complex algorithms needed for software Worst Case Execution Time analysis.