EAGER: Exploring the Self-Repair Role of Astrocytes in Neuromorphic Computing

随着深度学习在模式识别任务中取得前所未有的成功，训练和实施这种人工智能（AI）系统所需的计算费用需求也超出了当前的能力。“神经形态计算”通过探索底层计算基元和硬件基板的生物合理性，努力缩小人工智能平台在计算效率上的差距。该项目超越了当前神经形态计算架构对神经元和突触计算模型的关注，以研究生物大脑中可能有助于认知，特别是自我修复的其他计算单元。为此，这个EAGER项目将从计算神经科学中汲取关于神经胶质细胞功能的灵感和见解，并探索它们在新兴硬件支持的神经形态计算平台的容错能力中的作用，从而开辟新的方向。宾夕法尼亚州立大学施赖尔荣誉学院的研究生和本科生将参与该项目。该项目的高度跨学科性质将对培养下一代学生做出重大贡献，他们将获得结合硬件，神经科学和机器学习知识的神经形态计算框架的设计知识。PI计划将这个项目的成果整合到电气工程系K-12夏令营中。先前关于探索星形胶质细胞对自我修复影响的文献主要局限于小规模网络，没有任何机器学习的观点。此外，自我修复主要是从一个简单的软件模拟的角度来研究的，故障卡在零处。星形胶质细胞功能的神经形态硬件实现也仅限于互补金属氧化物半导体（CMOS）技术，由于CMOS晶体管和神经胶质功能之间的功能不匹配，这种技术在能量和面积上效率低下。为了弥补这一差距，提出的研究议程探索了一种硬件软件协同设计方法，通过使用自旋电子技术将神经形态平台中的胶质细胞功能结合起来。EAGER项目侧重于以下研究重点：(i)利用星形细胞计算模型来评估神经形态机器学习平台背景下对自我修复至关重要的神经胶质功能方面；（ii）探索自旋电子设备和电路原语，设计一个能够自我修复的耦合神经元-突触-星形细胞网络，其中底层设备通过其内在物理模拟星形细胞功能；（iii）结合上述自上而下和自下而上的观点，在神经形态人工智能系统中，利用星形胶质细胞自我修复硬件现实故障的背景下，如电阻漂移、寄生效应、设备间的变化等。拟议的研究议程将为新一代高效神经形态平台的开发提供概念验证结果，这些平台能够自主修复非理想的硬件操作。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。

With the unprecedented success of deep learning in pattern recognition tasks, the demands for computational expenses required to train and implement such Artificial Intelligence (AI) systems have also grown beyond current capabilities. “Neuromorphic Computing” strives to reduce the gap in computational efficiency of AI platforms by exploring bio-plausibility in the underlying computational primitives and hardware substrate. This project goes beyond the focus of current neuromorphic computing architectures on computational models for neuron and synapse to examine other computational units of the biological brain that might contribute to cognition and especially self-repair. To this end, this EAGER project will forge new directions by drawing inspiration and insights from computational neuroscience regarding functionalities of glial cells and explores their role in the fault-tolerant capacity of emerging hardware enabled neuromorphic computing platforms. Graduate students and undergraduates from Penn State's Schreyer Honors College will be involved in the project. The highly interdisciplinary nature of the project will contribute significantly to the training of next generation students who will gain knowledge in the design of neuromorphic computing frameworks combining knowledge from hardware, neuroscience and machine learning. The PI plans to integrate the results from this project into the Electrical Engineering departmental K-12 summer camp.Prior literature on exploring impact of astrocytes on self-repair has been primarily confined to small scale networks without any machine learning perspective. Further, self-repair has been studied primarily from a simplistic software simulation standpoint with stuck-at-zero faults. Neuromorphic hardware implementations for astrocyte functionalities have been also limited to Complementary Metal Oxide Semiconductor (CMOS) technology – which is highly energy and area inefficient due to the functional mismatch between CMOS transistors and glial functionality. To bridge this gap, the proposed research agenda explores a hardware-software co-design approach to incorporate glial cell functionality in neuromorphic platforms through the usage of spintronic technologies. The EAGER program focuses on the following research thrusts: (i) Exploiting astrocyte computational models to evaluate the aspects of glial functionality crucial for self-repair in the context of neuromorphic machine learning platforms, (ii) Exploring spintronic device and circuit primitives to design a coupled neuron-synapse-astrocyte network capable of self-repair where the underlying device mimics the astrocyte functionality through their intrinsic physics, and (iii) Combination of the above top-down and bottom-up perspectives to leverage astrocyte self-repair in the context of hardware realistic faults like resistance drift, parasitic effects, device to device variations, among others in neuromorphic AI systems. The proposed research agenda, would provide proof-of-concept results toward the development of a new generation of efficient neuromorphic platforms that are able to autonomously self-repair non-ideal hardware operation.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.