CCF: Small: Real-Number Function Encoding Driven Error Resilient Signal Processing and Control: Application to Nonlinear Systems from Adaptive Filters to DNNs

CCF：小型：实数函数编码驱动的误差弹性信号处理和控制：从自适应滤波器到 DNN 的非线性系统应用

• 批准号: 2128419
• 负责人: Abhijit Chatterjee
• 金额: $ 50万
• 依托单位: Georgia Tech Research Corporation
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2021
• 资助国家: 美国
• 起止时间: 2021-10-01 至 2025-09-30
• 项目状态: 未结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=2128419&HistoricalAwards=false
• 关键词: CCF; Small; Real; Number; Function

Technology scaling, device integration and high circuit speeds have increased the risk of errors in digital processors running computing and control applications. Such errors can jeopardize the operational safety of autonomous systems employing embedded processors in the field, causing severe loss of performance, accidents, or damage to life and property. To put this in perspective, applications such as self-driving cars and drones demand tera-ops of computation throughput and yet must perform with exacting levels of reliability and safety. The latter reliability and safety demands must be met with minimal degrees of hardware and software redundancy without negatively impacting dependability, power consumption, payload, form factor and cost considerations that are critical for commercial success. To alleviate relevant safety threats, this project aims to deploy low-cost and efficient methods for detecting and mitigating the effects of monitored errors in real time and in the field as effectively and rapidly as possible. This will enable a paradigm shift in the way failure-tolerance technologies are applied to autonomous systems while maintaining their reliability, affordability and cost. The project integrates research with education involving development of educational infrastructure for teaching and laboratory work, incorporation of diversity in student participation, exchanges with industry, technology transfer and engagement with undergraduate and high-school students, all with the goal of producing highly qualified trained engineers for the workplace of the future.Today, implementing failure tolerance for nonlinear signal-processing and control algorithms is dependent on some form of computation duplication. The algorithm-based fault-tolerance techniques of the past were designed mostly for linear computations and are not directly applicable to error control in nonlinear computations of modern autonomous systems. To resolve this, the project is built around the concept of algorithmic checks that encode nonlinear computations with linearized or nonlinear checking mechanisms. The checks are created using analytical methods for weakly nonlinear systems or by machine-learning algorithms for generic nonlinear systems and enable real-time error detection in switched capacitor circuits, nonlinear digital filters, adaptive state-estimation filters, nonlinear control algorithms and deep neural networks. Error correction is performed via state restoration in which the system state after error detection is restored to a prior fault-free system state or by using probabilistic error-compensation techniques. The core techniques can be applied to different aspects of the core computations performed in embedded computing infrastructure for autonomous systems, namely perception, intelligence, decision-making and control to make them error-resilient and trustworthy. The underlying science is enabling the design of entirely new classes of self-checking reliable autonomous systems that are beyond the scope of the current state of the art.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.

技术扩展、设备集成和高电路速度增加了运行计算和控制应用的数字处理器出错的风险。 此类错误可能会危及现场使用嵌入式处理器的自主系统的操作安全，导致严重的性能损失、事故或生命和财产损失。从角度来看，自动驾驶汽车和无人机等应用需要万亿次计算吞吐量，但必须以严格的可靠性和安全性水平运行。后者的可靠性和安全性要求必须以最小程度的硬件和软件冗余来满足，而不会对对于商业成功至关重要的可靠性、功耗、有效负载、外形尺寸和成本考虑因素产生负面影响。为了减轻相关的安全威胁，该项目旨在部署低成本、高效的方法，尽可能有效、快速地实时检测和减轻监控错误的影响。这将使容错技术应用于自主系统的方式发生范式转变，同时保持其可靠性、可承受性和成本。 该项目将研究与教育相结合，涉及教学和实验室工作的教育基础设施的开发、学生参与的多样性、与行业的交流、技术转让以及与本科生和高中生的接触，所有这些都旨在培养高素质的训练有素的工程师如今，实现非线性信号处理和控制算法的容错能力依赖于某种形式的计算重复。过去基于算法的容错技术主要是针对线性计算而设计的，不能直接应用于现代自治系统非线性计算中的误差控制。为了解决这个问题，该项目围绕算法检查的概念构建，该算法检查使用线性或非线性检查机制对非线性计算进行编码。 这些检查是使用弱非线性系统的分析方法或通用非线性系统的机器学习算法创建的，并可在开关电容器电路、非线性数字滤波器、自适应状态估计滤波器、非线性控制算法和深度神经网络中进行实时错误检测。纠错是通过状态恢复来执行的，其中将错误检测后的系统状态恢复到先前无故障的系统状态，或者通过使用概率错误补偿技术来执行。 核心技术可应用于自主系统嵌入式计算基础设施中执行的核心计算的不同方面，即感知、智能、决策和控制，使其具有容错性和可信性。基础科学正在实现全新类别的自检可靠自主系统的设计，这些系统超出了当前技术水平的范围。该奖项反映了 NSF 的法定使命，并通过使用基金会的智力评估进行评估，被认为值得支持。优点和更广泛的影响审查标准。

项目成果

A Resilience Framework for Synapse Weight Errors and Firing Threshold Perturbations in RRAM Spiking Neural Networks

RRAM 尖峰神经网络中突触权重误差和激发阈值扰动的弹性框架

• DOI: 10.1109/ets56758.2023.10174229
• 发表时间: 2023-05-22
• 期刊: 2023 IEEE European Test Symposium (ETS)
• 作者: Anurup Saha;C. Amarnath;A. Chatterjee
• 通讯作者: A. Chatterjee

Error Resilient Transformer Networks: A Novel Sensitivity Guided Approach to Error Checking and Suppression

容错变压器网络：一种新颖的灵敏度引导的错误检查和抑制方法

• 发表时间: 2023-05
• 期刊: European Test Symposium
• 作者: Ma, Kwondo.;Amarnath, Chandramouli.;Chatterjee, Abhijit.
• 通讯作者: Chatterjee, Abhijit.

Efficient Low Cost Alternative Testing of Analog Crossbar Arrays for Deep Neural Networks

深度神经网络模拟交叉阵列的高效低成本替代测试

• DOI: 10.1109/itc50671.2022.00060
• 发表时间: 2022-09-01
• 期刊: 2022 IEEE International Test Conference (ITC)
• 作者: Kwondo Ma;Anurup Saha;C. Amarnath;A. Chatterjee
• 通讯作者: A. Chatterjee

Soft Error Resilient Deep Learning Systems Using Neuron Gradient Statistics

使用神经元梯度统计的软错误弹性深度学习系统

• 发表时间: 2022-09
• 期刊: International On-Line Testing Symposium
• 作者: Chandramouli Amarnath; Mohamed Mejri
• 通讯作者: Mohamed Mejri

A Novel Approach to Error Resilience in Online Reinforcement Learning

在线强化学习中的错误恢复新方法

• DOI: 10.1109/iolts59296.2023.10224892
• 发表时间: 2023-07-03
• 期刊: 2023 IEEE 29th International Symposium on On-Line Testing and Robust System Design (IOLTS)
• 作者: C. Amarnath;A. Chatterjee
• 通讯作者: A. Chatterjee