FuSe-TG: Ultra-low-power and Robust Autonomy of Edge Robotics with 2D Semiconductors

FuSe-TG：采用 2D 半导体的边缘机器人的超低功耗和鲁棒自主性

• 批准号: 2235207
• 负责人: Amit Trivedi
• 金额: $ 30万
• 依托单位: University of Illinois at Chicago
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2023
• 资助国家: 美国
• 起止时间: 2023-05-15 至 2025-04-30
• 项目状态: 未结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=2235207&HistoricalAwards=false
• 关键词: FuSe; TG; Ultra; low; power

This project develops foundational semiconductors and co-design methodologies for risk-aware inference at the edge. Risk-aware inference procedures for deep learning such as Bayesian inference of deep neural networks (DNNs) can extract both prediction and prediction risks but also demand overwhelming computations. The inference procedures operate on random numbers and statistical density functions instead of real numbers for their higher expressivity. Meanwhile, their unique workload presents critical complexities for traditional semiconductors and microelectronics designed for digital workloads. To meet the distinct processing challenges, this teaming grant proposal explores the concepts of novel devices based on two-dimensional (2D) materials and co-design methodologies that can process predominant inference operations using an ultra-low-power non-von Neumann execution. A cross-layer simulation tool will be developed to actively bridge the material and device-level novelties to computing model and architecture design space and to explore unconventional system design concepts. Especially our various co-design initiatives will be intersected into exploring autonomous navigation of insect-scale drones as a test platform. Autonomous insect-scale drones can engender unprecedented applications, e.g., our project can potentially enliven a typical internet-of-things into a flying internet-of-things where the sensors riding on insect-scale drones can continually self-organize in space against the changing environmental conditions. The data-driven learning of deep neural networks (DNNs) significantly simplifies the complexity of model abstraction for many decision-making problems. Yet, the generated DNNs act like a black box and do not offer theoretical guarantees on the accuracy of the predictions. Significantly as our reliance on DNN-based predictive models is increasing for mission and safety-critical applications, it has become necessary to assess when DNN’s predictions are likely inaccurate. This teaming grant project explores concepts for future semiconductor technologies where more expressive DNN inference, where both the prediction and prediction risks, can be extracted while operating within the time and energy bounds of edge devices. Additionally, we will develop workshops on emerging co-design methodologies to address the workforce and talent demand for future semiconductor technologies and designs. We will instill co-design skills among undergraduate and graduate students by developing interdisciplinary course lectures and senior design projects. We will also foster a community of researchers on co-design space exploration by creating an online hub of resources such as open-source design kits, device models, and co-design exploration tools.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.

该项目为边缘的风险感知推理开发基础半导体和协同设计方法。深度学习的风险感知推理程序，如深度神经网络的贝叶斯推理（dnn），可以提取预测和预测风险，但也需要大量的计算。由于实数具有更高的表达性，推理过程对随机数和统计密度函数进行操作。同时，它们独特的工作负载为传统半导体和微电子设计的数字工作负载带来了关键的复杂性。为了应对不同的处理挑战，该团队资助提案探索了基于二维（2D）材料和协同设计方法的新型设备的概念，这些设备可以使用超低功耗非冯·诺伊曼执行来处理主要的推理操作。将开发跨层仿真工具，积极地将材料和设备级的新颖性与计算模型和架构设计空间联系起来，并探索非常规的系统设计概念。特别是我们的各种合作设计计划将交叉探索昆虫级无人机的自主导航作为测试平台。自主昆虫级无人机可以产生前所未有的应用，例如，我们的项目可以潜在地将典型的物联网激活为飞行物联网，其中昆虫级无人机上的传感器可以在不断变化的环境条件下不断在空间中自我组织。深度神经网络（dnn）的数据驱动学习显著地简化了许多决策问题的模型抽象复杂性。然而，生成的深度神经网络就像一个黑匣子，不能为预测的准确性提供理论上的保证。重要的是，随着我们对基于DNN的预测模型的依赖越来越多地用于任务和安全关键应用，有必要评估DNN的预测何时可能不准确。该团队资助项目探索了未来半导体技术的概念，其中可以在边缘设备的时间和能量范围内操作时提取更具表现力的DNN推理，其中预测和预测风险都可以。此外，我们将针对新兴的协同设计方法开展研讨会，以解决未来半导体技术和设计对劳动力和人才的需求。我们将通过开发跨学科课程讲座和高级设计项目，向本科生和研究生灌输协同设计技能。我们还将通过创建一个在线资源中心，如开源设计工具包、设备模型和协同设计探索工具，培养一个共同设计空间探索的研究人员社区。该奖项反映了美国国家科学基金会的法定使命，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。