RNCS: Reliable Neuromorphic Computing System Design

深度神经网络（Deep neural networks, dnn）在人工智能和认知处理等领域得到越来越多的关注和应用。由于与使用传统计算架构实现深度神经网络相关的挑战，人们对大脑启发的，又称神经形态计算平台和范式的兴趣日益浓厚，这些平台和范式用于直接和因此更有效地实现深度神经网络。鉴于采用新兴电阻式存储技术实现的神经形态硬件与传统数字电路和架构在电路和功能上的根本差异，在可靠性要求和方法上也存在根本差异。如果NCS的训练和映射明确地意识到预期的错误和不完美，那么NCS可以提供固有的容忍度。这也可以用来容忍硬件故障。本提案的目的是了解和量化与技术相关的神经形态回路功能可靠性故障，并在此基础上，在训练和推理阶段开发各种基于被动和主动冗余的方法，以提高神经形态回路的可靠性。我们还计划提高神经网络在保留和老化故障方面的可用性，并确保系统可以在短时间内修复，以获得可接受的精度。此外，将对系统进行设计和训练，使在运行时出现故障时推理精度的降低降到最低，以实现优雅的降低。

Deep neural networks (DNNs) are gaining increasing attention and usage in many fields related to artificial intelligence and cognitive processing. Due to challenges associated with the implementation of DNNs using traditional computing architectures, there is a growing interest for brain-inspired, aka Neuromorphic computing platforms and paradigms for direct and hence more efficient implementation of DNNs. Given the fundamental differences, in both circuitry and functionality of neuromorphic hardware realized using emerging resistive memory technologies to traditional digital circuits and architectures, there are fundamental differences in reliability requirements and methods. NCS can offer an inherent tolerance to expected errors and imperfections if their training and mapping are made explicitly aware of them. This can also be exploited to tolerate hardware failures.The purpose of this proposal is to understand and quantify the technology-relevant reliability failures on the functionality of the neuromorphic circuits and based on that, to develop various methods based on passive and active redundancy, during both training and inference phases, in order to improve the reliability of neuromorphic circuits. We also plan to improve the availability of neural networks with respect to retention and aging failures, and ensure that the system can be repaired in short amount of time to have acceptable accuracy. In addition, the system will be designed and trained such that the degradation of inference accuracy in presence of runtime failures is minimal to achieve graceful degradation.