Collaborative Research: Spintronics Enabled Stochastic Spiking Neural Networks with Temporal Information Encoding

合作研究：自旋电子学支持具有时间信息编码的随机尖峰神经网络

• 批准号: 2333881
• 负责人: Abhronil Sengupta
• 金额: $ 25万
• 依托单位: Pennsylvania State Univ University Park
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2024
• 资助国家: 美国
• 起止时间: 2024-06-01 至 2027-05-31
• 项目状态: 未结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=2333881&HistoricalAwards=false
• 关键词: Collaborative; Research; Spintronics; Enabled; Stochastic

Neuromorphic computing architectures attempt to bridge the computational efficiency gap of Artificial Intelligence platforms by emulating certain facets of the computational units and in-situ synaptic storage of the brain in the underlying algorithms and hardware substrate. This research addresses one of the main challenges facing neuromorphic computing today -- How to make bio-plausible spiking neural networks (SNNs) scalable and efficient for large-scale machine learning tasks while persevering the benefits of sparse, event-driven computation and learning? Currently, SNNs remain very similar to non-spiking networks with the temporal aspect remaining largely unexploited. The project is driven by the motivation that the current gap in SNN efficiency metrics (recognition accuracy, hardware power, energy and area efficiency) will be bridged by a transformative rethinking of spike information encoding in the temporal domain along with exploring nanoelectronic devices amenable for such alternate spike encoding schemes that leverage its inherent stochastic physics for brain-like probabilistic inference. Combining these two perspectives, stochastic biomimetic hardware, encoding information in the temporal domain, has the potential of enabling a new generation of brain-inspired computing platforms that leverages the associated advantages of two complementary insights from computational neuroscience -- how information is encoded in the brain and how computing occurs in the brain. The cross-layer nature of the project ranging from device design, circuit, system and algorithm explorations will serve as an ideal platform to enable interdisciplinary training and education of graduate and undergraduate students including women and underrepresented minority communities.The research involves a transformative research agenda, at the intersection of hardware and software, that develops a cross-layer design effort from devices to algorithms and underlying learning methodologies. The project spans cross-cutting explorations across the following thrust areas: (i) Thrust 1 investigates spin device physics and proposes device-circuit primitives suitable for temporal information encoding and learning in stochastic neuromorphic computing platforms. (ii) Thrust 2 considers system development that inherently exploits the temporal encoding of information in stochastic magnetic devices. (iii) Hardware-algorithm co-design resulting from Thrusts 1 and 2 will culminate in Thrust 3 that will consider large-scale system level simulations and performance evaluation across a benchmark application suite. Such an end-to-end framework can enable the fusion of appropriate neuromorphic computing paradigms with the intrinsic operation of the underlying hardware to improve its performance (classification accuracy) and efficiency for complex machine learning tasks. Successful completion of the project offers the basis for a significant leap in the quest to implement machine intelligence with brain-scale efficiency by pursuing a multi-disciplinary perspective spanning devices, circuits, systems, machine learning and computational neuroscience.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.

神经形态计算架构试图通过在底层算法和硬件基板中模拟计算单元的某些方面以及大脑的原位突触存储来弥补人工智能平台的计算效率差距。这项研究解决了当今神经拟态计算面临的主要挑战之一——如何使生物合理的尖峰神经网络 (SNN) 可扩展且高效地完成大规模机器学习任务，同时保持稀疏、事件驱动计算和学习的优势？目前，SNN 仍然与非尖峰网络非常相似，但时间方面在很大程度上仍未得到利用。该项目的动机是，当前 SNN 效率指标（识别精度、硬件功率、能源和​​面积效率）的差距将通过对时域尖峰信息编码的变革性重新思考以及探索适合这种情况的纳米电子设备来弥补。替代尖峰编码方案利用其固有的随机物理原理进行类脑概率推理。将这两个观点结合起来，在时域中编码信息的随机仿生硬件，有可能实现新一代类脑计算平台，该平台利用计算神经科学的两个互补见解的相关优势——信息如何在时间域中编码。大脑以及计算如何在大脑中发生。该项目的跨层性质涵盖设备设计、电路、系统和算法探索，将成为一个理想的平台，为研究生和本科生（包括女性和少数族裔群体）提供跨学科培训和教育。该研究涉及变革性研究议程，在硬件和软件的交叉点，开发从设备到算法和底层学习方法的跨层设计工作。该项目跨越以下主旨领域的交叉探索：（i）主旨 1 研究自旋器件物理并提出适合随机神经形态计算平台中的时间信息编码和学习的器件电路原语。 (ii) 推力 2 考虑了本质上利用随机磁性设备中信息的时间编码的系统开发。 (iii) 由主旨 1 和 2 产生的硬件算法协同设计将在主旨 3 中达到顶峰，该主旨 3 将考虑跨基准应用套件的大规模系统级仿真和性能评估。这样的端到端框架可以实现适当的神经拟态计算范式与底层硬件的固有操作的融合，以提高复杂机器学习任务的性能（分类精度）和效率。该项目的成功完成为通过追求跨越设备、电路、系统、机器学习和计算神经科学的多学科视角，实现具有大脑规模效率的机器智能的重大飞跃奠定了基础。该奖项反映了 NSF 的法定使命通过使用基金会的智力价值和更广泛的影响审查标准进行评估，并被认为值得支持。