On Evolutionary Approximation of Sigmoid Function for HW/SW Embedded Systems

Sigmoid, Linear genetic programming, HW/SW co-design

Providing machine learning capabilities on low cost electronic devices is a challenging goal especially in the context of the Internet of Things paradigm. In order to deliver high performance machine intelligence on low power devices, suitable hardware accelerators have to be introduced. In this paper, we developed a method enabling to evolve a hardware implementation together with a corresponding software controller for key components of smart embedded systems. The proposed approach is based on a multi-objective design space exploration conducted by means of extended linear genetic programming. The approach was evaluated in the task of approximate sigmoid function design which is an important component of hardware implementations of neural networks. During these experiments, we automatically re-discovered some approximate sigmoid functions known from the literature. The method was implemented as an extension of an existing platform supporting concurrent evolution of hardware and software of embedded systems.