High Performance Computing in Memristive Crossbars by Thermal-Heat Coupling (MemCouple)

尽管机器学习取得了令人印象深刻的成功，但人工智能（AI）仍然没有超越自然智能。由于他们不理解他们所学习的内容，他们在认知任务中的工作效率更低。今天的机器学习运行在串行冯-诺依曼计算机上，使得训练过程不灵活且能量密集，而人类大脑是一个巨大的并行神经元网络，分为功能专业化的区域，作为一个依赖于上下文的，自组织的和瞬态的子网络参与，每0.5到2秒就会随着注意力的变化而改变。然而，这导致了认知神经科学中最令人困惑的问题之一，即众所周知的绑定问题，这与对网络动态及其连接性的相关性的理解密切相关，即，对时空模式的理解在这里，神经元可塑性是导致突触权重的结构连接矩阵的关键因素，该结构连接矩阵决定了网络的整体功能。为了给认知电子学铺平道路，该项目旨在模拟忆阻交叉阵列中时空模式的形成，并使其可用于非常规计算。为此，将利用热串扰的影响。这应该通过定制忆阻器件特性和交叉开关结构属性来实现，使得它们满足局部学习规则。实验研究将通过为制造的忆阻器件和阵列级模型开发可变性感知的紧凑模型来补充，这些模型包括器件之间在空间和时间上的热耦合效应。基于模拟和实验验证，我们希望推导出忆阻交叉杆阵列的设计规则，以最有效的方式利用热耦合效应。为了确保忆阻神经网络的最佳训练特性，我们希望开发忆阻交叉开关，通过合适的可塑性机制实现局部学习机制，以生成时空模式并在混沌边缘稳定系统动力学。在这个临界状态下，神经网络在灵敏度、动态范围、相关长度、信息传递和敏感性方面具有最大的计算特性。因此，一个特定的目标是使用热串扰和泄漏电流来启动关联学习机制。最终的目标是一个神经形态系统，工作在尽可能低的电路开销，并具有最小数量的输入和输出神经元。

Despite the impressive successes in machine learning, artificial intelligence (AI) still not outperform natural ones. As they do not understand what they learn, they work more inefficiently on cognitive tasks. Today’s machine learning runs on serial von-Neumann computers making the training process inflexible and energy-intensive, whereas the human brain is a massive parallel network of neurons divided into functionally specialized regions participating as a context-dependent, self-organized, and transient subnetwork which is shifted by changes in attention every 0.5 to 2 s. This, however, leads to one of the most puzzling issues in cognitive neuroscience, well known as the binding problem, which is strongly interwoven with an understanding of the correlation of the network dynamics and its connectivity, i.e., the understanding of spatial -temporal patterns. Here, neuronal plasticity is a crucial factor leading to a structural connectivity matrix of synaptic weights that determines the overall function of the network. To pave the way to cognitive electronics this project aims to emulate the formation of temporal-spatial patterns in memristive crossbar arrays and to make them usable for unconventional computing. For this purpose, the effect of thermal crosstalk will be exploited. This should be achieved by tailoring the memristive device characteristics and crossbar structures properties so that they satisfy local learning rules. The experimental investigation will be complemented by developing variability-aware compact models for the fabricated memristive devices and array-level models that include thermal coupling effects between devices in space and time. Based on the simulations and the experimental validation, we want to deduce design rules for memristive crossbar arrays that enable the exploitation of thermal coupling effects in the most efficient way. To ensure optimal training properties for memristive neural networks we want to develop memristive crossbars that implement local learning mechanisms via suitable plasticity mechanisms to generate spatial-temporal patterns and stabilize the system dynamics at the edge of chaos. At this critical state neural networks have maximum computational properties in terms of sensitivity, dynamic range, correlation length, information transfer, and susceptibility. Therefore, a particular goal is to use thermal crosstalk and leakage currents to initiate associative learning mechanisms. The final objective is a neuromorphic system that works with the lowest possible circuit overhead and has a minimum number of input and output neurons.