CSR: Small: Enabling Deep Neural Networks for Mobile-Cloud Applications

在过去的三年里，深度神经网络(DNN)已经成为解决计算中各种重要问题的主导方法。这包括语音识别、机器翻译、手写识别方面的问题，以及许多计算机视觉问题，如人脸、物体和场景识别。虽然它们以出色的识别性能而闻名，但DNN也是众所周知的计算密集型网络：通常用于语音、视觉和语言理解任务的网络通常会消耗数百MB的内存和Gflop的计算能力，通常是服务器级计算机的专属领域。然而，上述应用与移动环境的相关性和开发新应用的潜力为在移动设备上执行DNN提供了强有力的理由。本项目旨在构建一个移动云平台上的深度神经网络执行框架，以支持广泛的新兴应用，如连续移动视觉。特别是，这项工作将着眼于实现一大套基于DNN的人脸、场景和对象处理算法，这些算法基于将DNN应用于来自可穿戴设备的视频流。在给定任意DNN的情况下，该框架将以适度的精度损失将其编译成资源高效的变体。项目计划包括开发新的技术，以使DNN专门化以适应环境，并在多个同时执行的DNN之间共享资源。最后，它将创建一个运行时系统，用于管理生成的优化模型。以具有挑战性的连续移动视觉领域为例，该计划将展示这些技术在普通移动环境中显著减少了DNN资源的使用，包括内存使用和执行的指令的数量级减少。

Over the past three years, Deep Neural Networks (DNNs) have become the dominant approach to solving a variety of important problems in computing. This includes problems in speech recognition, machine translation, handwriting recognition and many computer vision problems like face, object, and scene recognition. Although they are renowned for their excellent recognition performance, DNNs are also known to be computationally intensive: networks commonly used for speech, visual and language understanding tasks routinely consume hundreds of MB of memory and Gflops of computing power, typically the province of server-class computers. However, the relevance of the above applications to the mobile setting and the potential for developing new applications provides a strong case for executing DNNs on mobile devices.This project is to build an execution framework for deep-neural networks on mobile-cloud platforms so as to enable a broad class of emerging applications such as continuous mobile vision. In particular, this work will look at enabling a large suite of DNN-based face, scene and object processing algorithms based on applying DNNs to video streams from wearable devices. This framework, given an arbitrary DNN, will compile it down to a resource-efficient variant at modest loss in accuracy. The project plans include developing novel techniques to specialize DNNs to contexts and to share resources across multiple simultaneously executing DNNs. Finally, it will create a run-time system for managing the optimized models generated. Using the challenging continuous mobile vision domain as a case-study, the plan is to demonstrate that these techniques yield very significant reductions in DNN resource usage, including orders of magnitude reduction in memory use and instructions executed, in common mobile settings.