Neurotransistor-based Memristive Crossbar Memcomputing (NeuroMCross)

受大脑启发的尖峰神经网络(SNN)通过相互连接的神经元的时空尖峰来并行和异步工作。与执行大规模矩阵-矢量乘法的传统人工神经网络(ANN)结构相比，对电容性元件进行充电和放电、导致从信号源移除能量或将能量返回到信号源的电容神经网络可以导致相当低的能量消耗。在这方面，记忆晶体管神经元将记忆/记忆电容器件和晶体管结构相结合，可以实现神经元的基本特征，如泄漏-积分-火灾(LIF)行为。这可以被构建成神经形态的记忆计算构建块，用于通用的SNN架构。在这个项目中，我们的目标是开发一种新型的神经交叉开关结构，它结合并扩展了记忆/记忆电容交叉开关和基于先前BioMcross项目中开发的Al_2O_3/Nb_2O_5器件的神经晶体管的各个组件的功能。与以往的记忆晶体管方法不同，我们在硅基n沟道场效应晶体管(NFET)的栅极上制作了一个集成的多输入记忆电容交叉开关结构，它能够沿着晶体管沟道形成一维和二维可控的渗流路径。LIF神经交叉开关集成了高复杂性，例如用于多个输入的完全连接层、针对1D渗流或OR链接的2D渗流路径的全有或全无特征提取以及在单个器件结构中的循环连接。为此，拟议的工作将涵盖几个基本方面：(I)使用3D TCAD模拟的神经交叉开关器件建模，(Ii)通过在晶体管结构上集成多个Al_2O_3/Nb_2O_5记忆/记忆电容器件来制造神经交叉开关和参数优化，(Iii)神经交叉开关电路系统的建模和读/写操作的分析，以及(Iv)用于循环SNN的时间编码信号的设计、硬件感知的离线训练、性能优化以及本地学习规则的实施。最后，我们将研究不同的时间序列数据集，以演示递归SNN互连神经交叉开关结构的实现，并与传统的基于CMOS的SNN和纯记忆ANN加速器系统相比，评估系统的性能。这项研究旨在通过一种新的神经交叉杆构建块来扩展技术水平，并基于高度集成的器件概念提供对实际实现记忆电容式SNN系统的可行性的全面了解。

Brain-inspired Spiking Neural Networks (SNNs) operate in parallel and asynchronously via spatio-temporal spikings of interconnected neurons. Capacitive neural networks that charge and discharge capacitive mem-elements, resulting in the removal or return of energy from/to the signal sources, can lead to considerably lower energy consumption in comparison to conventional artificial neural network (ANN) structures performing massively matrix-vector multiplications. In this regards memtransistor neurons combining memristive/memcapacitive devices and transistor structures can enable fundamental features of neurons like the leaky-integrate-and-fire (LIF) behavior. This can be built into a neuromorphic memcomputing building block for versatile SNN architectures. In this project we aim at the development of a novel neuro-crossbar structure, which combines and extends the functionalities of the individual components of the memristive/memcapacitive crossbar and neurotransistor based on Al2O3/Nb2O5 devices developed in the preceding BioMCross project. Unlike previous memtransistor approaches, we fabricate an integrated multi-input memcapacitive crossbar structure at the gate of a silicon-based n-channel field effect transistor (nFET), which is able to create a controlled 1D and 2D percolation path along the transistor channel. The LIF neuro-crossbar integrates high complexity, such as a fully connected layer for multiple inputs, an all-or-nothing feature extraction for 1D percolation or OR-linked 2D percolation paths as well as recurrent connection in a single device structure. For this purpose, the proposed work will cover several essential aspects: (i) the neuro-crossbar device modeling using 3D TCAD simulation, (ii) the neuro-crossbar fabrication and parameter optimization by integration of multiple Al2O3/Nb2O5 memristive/memcapacitive devices on top of the transistor structure, (iii) the neuro-crossbar circuit system modeling and analysis of read/write operation and (iv) the design of temporal coded signals for recurrent SNNs, hardware-aware offline training, performance optimization as well as implementation of local learning rules. Finally, we will investigate various time series datasets to demonstrate the realization of recurrent SNNs interconnected neuro-crossbar structures and evaluate the system's performance compared to conventional CMOS-based SNN and purely memristive ANN accelerator systems. This research aims to extend the state of the art by a new neuro-crossbar building block and to provide a comprehensive understanding of the feasibility of practical realization of memcapacitive SNN systems based on highly integrated device concepts.