Reconfigurable Neuromorphic Computing to enable Energy-Efficient Edge Intelligence

可重构神经形态计算实现节能边缘智能

• 批准号: 2210804
• 负责人: Sahil Shah
• 金额: $ 33.62万
• 依托单位: University of Maryland, College Park
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2022
• 资助国家: 美国
• 起止时间: 2022-08-01 至 2025-07-31
• 项目状态: 未结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=2210804&HistoricalAwards=false
• 关键词: Reconfigurable; Neuromorphic; Computing; enable; Energy

The goal of this project is to develop an energy-efficient and smart extreme edge device by taking inspiration from the human brain. Extreme edge devices can enable a variety of applications. Specifically, the devices can be used for remote tracking of rapid changes in the arctic, autonomous navigation of aerial and ground robots for deep space exploration, and monitoring and securing critical infrastructure. However, the current technology requires remotely sensed data to be processed in the cloud. This framework has several drawbacks, such as the delay between sensing and decision, shorter battery life due to high power consumption, and privacy concerns related to data transfer. The hardware developed as part of this proposal will seek to address the previously noted challenges and will advance the state-of-the-art in extreme edge devices. Specifically, the grant will enable the development of a reconfigurable brain-inspired processor with the capacity to learn based on input data. The hardware developed as part of this proposal will be used to train and motivate the next generation of students in the areas of microelectronics. Further, the hardware developed will use open-source computer aided design tools to enable broader dissemination of the developed technology. The study proposes to develop a reconfigurable mixed-signal neuromorphic hardware for processing data at the extreme edge. This proposal aims to increase the energy efficiency of neuromorphic hardware by employing mixed-signal circuits to model the neurons and synapses. Further, we propose to incorporate programmable mixed-signal circuit topologies and explore techniques to co-optimize the hardware and software models to learn and adapt the network parameters in the presence of mismatch and variations. In addition, the proposal will also explore circuit topologies to perform learning on-chip. Based on these individual elements, the proposal will investigate a system architecture to develop reconfigurable hardware that can compile a spiking neural network with the capability to learn on-chip for performing extreme edge tasks. The extreme edge task we plan to validate the developed hardware is object detection using openly available datasets and a quadcopter with a limited battery capacity to perform object detection. The proposal will address the knowledge gaps in designing a learning algorithm for a mixed-signal spiking neural network with variation and mismatch, circuit topologies and system architecture for performing on-chip learning with limited resources, and advance the state-of-the-art in energy-efficient neuromorphic hardware.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.

该项目的目的是通过从人脑中汲取灵感来开发节能且智能的极端边缘设备。 Extreme Edge设备可以启用各种应用。具体而言，这些设备可用于远程跟踪北极，空中和地面机器人自动导航的快速变化，以进行深空探索，并监视和确保关键基础设施。但是，当前的技术需要在云中处理远程感知的数据。该框架有几个缺点，例如传感和决策之间的延迟，由于高功耗而导致的电池寿命较短以及与数据传输相关的隐私问题。作为本提案的一部分开发的硬件将寻求应对以前著名的挑战，并将推进极端边缘设备中的最新挑战。具体而言，该赠款将使可重新配置的脑启发的处理器能够根据输入数据进行学习。 作为本提案的一部分开发的硬件将用于训练和激励微电子学领域的下一代学生。此外，开发的硬件将使用开源计算机辅助设计工具来更广泛地传播已发达的技术。 该研究建议开发可重新配置的混合信号神经形态硬件，以处理极端的数据。该建议旨在通过使用混合信号电路来建模神经元和突触来提高神经形态硬件的能效。此外，我们建议合并可编程的混合信号电路拓扑并探索技术以优化硬件和软件模型，以在存在不匹配和变化的情况下学习和调整网络参数。此外，该提案还将探索在片上进行学习的电路拓扑。基于这些个别元素，该提案将研究系统体系结构，以开发可重新配置的硬件，该硬件可以编译尖峰神经网络，并具有学习芯片以执行极端边缘任务的能力。我们计划验证开发的硬件的极端边缘任务是使用公开可用的数据集对象检测，并且是电池可执行对象检测能力有限的四轮驱动器。该提案将解决设计学习算法的知识差距，用于具有变化和不匹配，电路拓扑和系统体系结构的混合信号峰值神经网络，以使用有限的资源进行芯片学习，并提高能源效率高的神经形态硬件的最先进。影响审查标准。