Exploration and design of embedded systems including neural algorithms

The current trend in embedded systems is to make them surrounding the users, providing services thanks to a knowledge of their environment. These self-awareness and context-awareness properties are provided by numerous sensors, from different types. Using the provided information causes at least two problems: the fusion of data from different sources, and the noise induced by sensors which are closer from the processing unit than ever. Additionally, the needed applications that use these information are based on different recognition processings, sometimes not easy to formalize with conventional algorithms. Processing chains using neural-based algorithms are promising approaches for solving these kinds of issues. Unfortunately, embedding bio-inspired algorithms in an embedded system is not so easy since there is no exploration environment for this specific task. Moreover, neural networks often need pre- or postprocessing of data for optimal operation. In fact, there is a balance to find between pre-processing and neural network processing: for example, adding more filtering to clean or to transform data (like convolution filters or FFT) enables to have smaller neural networks, leading to less number of neurons, less learning time and finally more efficient applications. This paper presents early results of a collaboration towards the design of such an exploration environment coming from a joint laboratory between an SME and a Research Institute. The main object coming from the current collaboration is the coupling of a rich exploration environnement of embedded systems (including multi/manycore) with a neural network exploration tool. The combination of the two enables us to have feedbacks concerning both algorithm efficiency and performances and other non-functional metrics regarding the target system for driving the co-design cycle of industrial embedded systems.