Deep Neural Networks Characterization Framework for Efficient Implementation on Embedded Systems

Bio-inspired machine learning algorithms, such as Convolutional Neural Networks (CNNs), offer interesting solutions to complex real-life problems that cannot be simply modeled. Applications involving image recognition and object detection can greatly benefit from these approaches. Furthermore, their intrinsic and regular parallel structure offer opportunities regarding hardware acceleration. However, moving compute and memory-intensive CNNs to embedded systems while maintaining high energy-efficiency remains challenging. This paper presents the first step of a generic framework targeting the characterization of neural network algorithms to improve their implementation on embedded systems. The presented approach aims at reducing the gap between the fast-changing landscape of applications based on artificial intelligence and the hardware targets. The framework computes different metrics from neural network descriptions (such as computation and memory needs or data locality and reuse) to derive appropriate implementation strategies, or configurations of target architectures. Based on the outputs of the framework, new neural networks topologies can be quickly studied to reduce time-to-market of new systems.