Exploring Compositional Neural Networks for Real-Time Systems

Sobhan Chatterjee, Nathan Allen, Nitish D. Patel, Partha S. Roop

Abstract: Real-time CPSs using Artificial Neural Networks (ANNs) are traditionally developed as monolithic black-boxes. This results in designs that are often difficult to formally verify against safety specifications and implement on hardware for formal timing analysis. Consequently, their implementation as a composition of smaller ANNs has received recent interest. These are easier to implement, parallelise and validate. Despite this, the question of how to produce hardware-implementable compositional designs from existing monolithic ones remains largely unanswered. This work develops a novel procedure to replace large ANN monolithic designs with smaller compositional designs and implement them on a Field Programmable Gate Array (FPGA) for timing analysis using synchronous compositional semantics. To illustrate our approach, we develop regression and classification ANN designs for multiple real-life datasets. Using various design and model architecture variations, we show that using a compositional design instead of a monolithic design can achieve an $\mathbf{8 5 \%}$ reduction in WCET, around a $\mathbf{53 \%}$ reduction in hardware resources and around a 40% reduction in computations and neuron connections for a minor reduction in performance.