CRII: SHF Software and Hardware Architecture Co-Design for Deep Learning on Mobile Device

智能手机已经成为我们生活中不可或缺的一部分，成为许多基本功能的主要工具。智能手机携带着丰富的数据集，包含各种个人信息。在机器学习的支持下，移动的应用程序正在利用这些数据来提高服务质量。当前的实践要求用户将计算任务卸载到云端，这在隐私、性能和用户体验方面带来了挑战。在移动的处理能力急剧增长的推动下，该项目寻求将机器智能带入移动的设备。这些结果有望激发嵌入式机器智能软件基础领域的理论和系统研究。通过该项目获得的经验教训将在下一代移动的操作系统和硬件架构的设计中发挥重要作用。结果将通过出版物和讲座传播。 该项目旨在开发一个高性能、隐私保护和节能的移动平台，并应用行为认证。该研究将研究传感数据的最佳表示，并开发一个紧凑而强大的神经网络架构。包括推理和训练的所有计算将在移动终端上执行。将开发一种协议，使移动的和云之间的功能转移，以减少过度拟合和加快模型收敛，沿着一种新的训练算法，以利用缓存的局部性和减轻内存瓶颈。批量大小和学习率的最佳组合将在内存限制和精度要求范围内进行探索，以最大限度地减少训练时间。何时以及如何在移动终端上安排训练的问题将通过考虑具有高级别用户交互的低级别操作来研究，以实现性能和资源消耗之间的良好平衡。所有这些模块都将在Android中集成和实现，并在各种智能手机型号上进行评估。该奖项反映了NSF的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Smartphones have become an indispensable part of our lives, acting as the primary tool for many essential functions. A smartphone carries a rich set of data with all kinds of personal information. Powered by machine learning, mobile applications are utilizing these data for better quality of service. Current practice requires users to offload computation tasks to the cloud with incumbent challenges on privacy, performance and user experience fronts. Boosted by the dramatic increase in mobile processing power, this project seeks to bring machine intelligence to mobile devices. The results are expected to inspire both theoretical and system research in the areas of software foundations for embedded machine intelligence. Lessons learned through this project will be fundamentally important in the designs of the next generation mobile operating system and hardware architecture. The results will be disseminated through publications and talks. This project seeks to develop a high-performance, privacy-preserving and energy-efficient mobile-based platform, with an application of behavioral authentication. The research will study the optimal representation of sensing data and develop a compact and powerful neural network architecture. All computation including both inference and training will be performed on the mobile device. A protocol to enable feature transfer between the mobile and the cloud will be developed to reduce overfitting and speed up model convergence, along with a new training algorithm to exploit cache locality and mitigate the memory bottleneck. The optimal combinations of batch size and learning rate will be explored within memory constraints and accuracy requirements to minimize training time. The problem of when and how training should be scheduled on a mobile device will be investigated by considering low-level operation with high-level user interaction to achieve a good balance between performance and resource consumption. All these modules will be integrated and implemented in Android and evaluated on various smartphone models.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.