CAREER: DeepMatter: A Scalable and Programmable Embedded Deep Neural Network

职业：DeepMatter：可扩展且可编程的嵌入式深度神经网络

• 批准号: 2348983
• 负责人: Tinoosh Mohsenin
• 金额: $ 47.51万
• 依托单位: Johns Hopkins University
• 依托单位国家: 美国
• 项目类别: Continuing Grant
• 财政年份: 2023
• 资助国家: 美国
• 起止时间: 2023-10-01 至 2024-08-31
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=2348983&HistoricalAwards=false
• 关键词: CAREER; DeepMatter; Scalable; Programmable; Embedded

Deep neural networks (DNNs), modeled loosely after the human brain, have shown tremendous success to accurately interpret sensory data and recognize patterns. However, they have not been explored for current and future low power multi-sensor applications, such as Internet of Things (IoT), wearable health and mobile smart devices. The fundamental problem with embedded exploration is that current DNN models are very complex, making them challenging to deploy in embedded systems with limited hardware resources and power budgets. The project investigates novel and transformative methodologies for DNN network modeling, sparsification, and approximation techniques in software, termed DeepMatter. The research develops new architectures to design a programmable domain-specific many-core platform that implements the optimized network and provides performance, scalability, programmability, and power efficiency requirements necessary for embedded DNN implementations. An application program interface (API) will be designed to allow designers to rapidly prototype and deploy the next generation of sophisticated and intelligent applications. For demonstration, five applications including multi-physiological processing for seizure and distress detection, multi-modal assistive device, air quality monitoring and vision-based situational awareness will be evaluated on DeepMatter. The success of this research project will result in small and energy efficient wearable/mobile computing devices which can perform knowledge extraction and classification on raw data at the sensor without sending massive raw data to the cloud for processing. This can revolutionize several fields including healthcare, transportation, ecology, surveillance, public utilities. Software models, hardware and tools will be available for the research community to prototype and evaluate different applications. This research provides a multidisciplinary platform for educational objectives of developing embedded smart processors and involves middle and high school students and teachers as well as undergraduate and graduate students.

深度神经网络（DNN），松散地模仿人类大脑，在准确地解释感官数据和识别模式方面取得了巨大的成功。然而，它们还没有被探索用于当前和未来的低功耗多传感器应用，例如物联网（IoT）、可穿戴健康和移动的智能设备。嵌入式探索的根本问题是，当前的DNN模型非常复杂，这使得它们在硬件资源和功耗预算有限的嵌入式系统中部署具有挑战性。 该项目研究了DNN网络建模，稀疏化和软件近似技术的新颖和变革性方法，称为DeepMatter。该研究开发了新的架构，以设计可编程的特定于域的众核平台，该平台实现了优化的网络，并提供了嵌入式DNN实现所需的性能，可扩展性，可编程性和能效要求。将设计一个应用程序接口（API），使设计人员能够快速原型化和部署下一代复杂和智能的应用程序。为了进行演示，将在DeepMatter上评估五种应用，包括癫痫发作和遇险检测的多生理处理，多模式辅助设备，空气质量监测和基于视觉的态势感知。该研究项目的成功将产生小型和节能的可穿戴/移动的计算设备，这些设备可以在传感器处对原始数据进行知识提取和分类，而无需将大量原始数据发送到云端进行处理。这可以彻底改变几个领域，包括医疗保健，交通，生态，监控，公用事业。软件模型、硬件和工具将提供给研究界，以制作原型和评估不同的应用。本研究提供了一个多学科的平台，开发嵌入式智能处理器的教育目标，涉及初中和高中的学生和教师，以及本科生和研究生。