Mapping deep learning algorithms on systems-on chip

该项目旨在定义一种创新方法，用于映射在物联网（IoT）中通信的嵌入式自治系统的深度学习神经网络。自治系统必须能够执行由前一次操作的结果确定的一系列操作，或者参考由系统本身监视和测量的外部环境。今天的自主系统在定义和控制的情况下工作得很好，并且在这一领域有巨大的进一步发展潜力。然而，仍然存在一些关切和制约因素。这主要是因为在未来的物联网中，自治系统将交换持续的海量数据流，每个系统都必须能够快速、安全、可靠地处理数据并激活适当的响应，同时消耗很少的功率。这需要这些系统进行各种复杂的决策、学习和数据处理。 这个拟议的研究项目将为物联网中自主系统的设计、实施和部署方面的关键挑战开发解决方案，即： - 深度学习是未来自治系统中的关键算法，适应边缘计算和自治系统特有的严格性能和功耗约束的新方法; - 新的工具，用于在嵌入式平台上自动优化部署这些极其复杂的学习算法。我们特别关注一组通信算法，这些算法专门用于学习，计算机视觉和语音识别，并映射到集成多核架构和神经网络加速器的专用平台上，例如Synopsys EV 6x。

This project aims to define an innovative approach for mapping deep learning neural networks of embedded autonomous systems communicating in the Internet of Things (IoT). Autonomous systems must be able to perform a series of operations determined by the outcome of the previous operation, or by reference to external circumstances monitored and measured by the system itself. Today's autonomous systems work well within defined and controlled situations, and there is huge potential for further development in this area. There are, however, still concerns and constraints. This is mainly because in the future IoT, autonomous systems will exchange a continuous flow of massive data, and each system must be able to handle the data and activate appropriate responses quickly, safely, and securely, while consuming little power. This requires a variety of complex decision-making, learning and data processing by these systems. This proposed research project will develop solutions to key challenges in the design, implementation, and deployment of autonomous systems in the IoT, namely: - New methods of adapting deep learning, the key algorithms in the autonomous systems of the future, to the tight performance and power constraints specific to edge computing and autonomous systems; - New tools for automatic, optimized, deployment of these extremely complex learning algorithms on embedded platforms. We are particularly concerned with a set of communicating algorithms, specific to learning, computer vision and voice recognition, mapped on dedicated platforms integrating multi-core architectures and neural network accelerators, such as the Synopsys EV6x.