Gated Synaptic Memory Devices with Adaptive Short-Term States for Neuromorphic Computing

用于大数据分析的人工智能（AI）技术变得非常重要。然而，人工智能软件算法是计算资源密集型的，这对其实际应用造成了限制，因为目前可用的数据处理器并不适合这些需求。例如，人工智能的训练算法可能需要几个小时到几天才能完成训练过程。此外，其他一些挑战，例如对庞大训练数据集的需求，缺乏实时训练和多模态数据融合功能，以及系统在有限输入数据下可靠决策的局限性，也是众所周知的。如果能够开发出受大脑启发的神经形态数据处理器，这些问题中的许多都可以得到解决。然而，由于两个主要原因，这是一项重要的任务。首先，大脑中的皮层回路尚未完全理解，仍然是神经科学界的研究课题。其次，用于整合大脑启发架构的人工神经模拟组件尚未开发出与生物大脑的计算效率和多样性相匹配的组件。 从目前对生物脑皮层回路的认识中可以看出，突触作为神经元之间的可重构连接，在学习记忆的形成中起着关键作用。这个项目的重点是开发可以集成在神经形态架构中的人工纳米电子突触器件。该项目为培训研究生和本科生理解和开发人工智能神经形态处理器提供了重要机会。将在研究生一级开设一门关于“人工智能神经形态计算机”的新课程。将通过利用“从一开始就培养教育准备和发展”方案和通过当地计算机协会分会，努力增加代表性不足的群体在STEM中的参与。 纳米电子突触器件将开发利用时间依赖的陷阱动力学的氧化物结合运输的内在或外在的掺杂剂在一个新的门控突触记忆器件（门控SMD）配置。这些动力学将导致模拟增强（电导增加）和抑制（电导减少）作为门上的电压脉冲的时间序列的函数，其可以被探索用于实现生物启发学习算法。本论文的主要研究内容包括：（1）制作栅控SMD器件，研究器件特性，包括电阻态在不同时间尺度上的增强和抑制随栅偏压的变化规律，并对其进行建模;（2）了解这些设备的可扩展性-缩放布局设计，并将单元尺寸与其他候选存储器技术进行比较和基准测试;以及（3）开发新颖的实时学习算法并使用用于神经形态架构的门控SMD实现生物启发学习方案。该研究的知识意义在于为神经形态电路提供知识库和器件平台，为神经形态电路提供纳米电子突触解决方案。如果成功，该项目将产生以下成果：（i）对门控SMD和以实验数据为基准的器件模型的基本理解，（ii）通过设计器件参数来控制门控SMD中电阻状态的增强和抑制率的策略，（iii）为门控SMD量身定制的实时学习算法，以及（iv）门控SMD的大规模集成路线和可扩展性数据。这些成果的实现将对人工智能神经形态数据处理器的开发产生变革性影响。该奖项反映了NSF的法定使命，并被认为值得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来支持。

Artificial Intelligence (AI) techniques for big data analytics are becoming very important. However, AI software algorithms are computation resource intensive which imposes limitations on their practical applications as the currently available data processors are not well-suited for these needs. For example, training algorithms for AI can take several hours to days for completing the training process. Additionally, some of the other challenges, such as the requirement for huge training datasets, lack of real-time training and multi-modal data fusion capabilities, and limitations for the system to make decisions reliably with limited input data are well-recognized. Many of these problems can be addressed if brain-inspired neuromorphic data processors can be developed. However, it is a non-trivial task because of two primary reasons. First, the cortical circuits in brain is not fully understood and still is a topic of research in the neuroscience community. Second, artificial neuromimetic components for integration in brain-inspired architectures are not yet developed to match the computational efficiency and diversity of biological-brains. It has been identified from current understanding of cortical circuits in biological-brain that a synapse which is a reconfigurable connection between neurons, play pivotal a role in learning and memory formation. The focus of this project is to develop artificial nanoelectronic synaptic devices that can be integrated in neuromorphic architectures. The project provides significant opportunities for training graduate and undergraduate students in understanding and developing neuromorphic processors for AI. A new course on "Neuromorphic Computers for AI" at the graduate level will be developed. Efforts will be made to increase participation of underrepresented groups in STEM by leveraging the program on "Nurturing Educational Readiness and Development from the Start (NERDS)" and through local Association for Computing Machinery (ACM) chapter. The nanoelectronic synaptic device will be developed by exploiting time-dependent trap dynamics in oxides in conjunction with the transport of intrinsic or extrinsic dopants in a novel gated-Synaptic Memory Device (gated-SMD) configuration. These dynamics will result in an analog potentiation (increase in conductance) and depression (decrease in conductance) as a function of the temporal sequences of voltage-pulses on gate that can be explored for implementing bio-inspired learning algorithms. The objective of the proposed research will be achieved by executing the following specific aims: (1) fabricating gated-SMDs and studying the device characteristics, including potentiation and depression of resistive states on different time-scales as a function of gate-bias and modeling it; (2) understanding the scalability of these devices by large-scale layout designs and comparing and benchmarking the cell sizes against other candidate memory technologies; and (3) developing novel real-time learning algorithms and implementing bio-inspired learning schemes using gated-SMDs for neuromorphic architectures. The intellectual significance of the proposed research lies in knowledge base and a device platform to provide a solution of nanoelectronic synapses for neuromorphic circuits. If successful, the project will yield the following outcomes: (i) a fundamental understanding of gated-SMDs and device models benchmarked against experimental data, (ii) strategies to control potentiation and depression rates of resistive states in gated-SMD by engineering the device parameters, (iii) real-time learning algorithms tailored for gated-SMDs, and (iv) large-scale integration routes for gated-SMDs and scalability data. The achievement of these outcomes will have transformative impact on developing neuromorphic data processors for AI.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.