Memristive Electronics for Neuromorphic Engineering and Computing

这项工作的目的是开发记忆/电阻RAM(RRAM)器件在新型神经形态(神经启发)计算体系结构和系统中的应用。这样的系统在提供在实现机器学习(ML)和机器智能方面远远优于现有CMOS硬件的解决方案方面具有巨大的潜力。其主要优点之一是显著降低了电路复杂性，并极大地提高了电源效率。这对于将ML引入自适应嵌入式系统是至关重要的。此外，这一点还可以在“物联网/智能物联网”时代得到利用，ML算法直接在物联网上实现，从而促进高效的本地数据处理并使设备能够在本地做出决策，而不是依赖于数据流和易于延迟的云计算。RRAM(RRAM)技术是记忆系统的一个子类，它基于简单的两个终端纳米器件，这些纳米器件可以重复地改变它们的电阻，具有低操作能量和非常高的集成水平。值得注意的是，它们类似于不同的神经元功能--最重要的是，通过逐渐改变它们的阻力(“突触重量”)，具有突触样的可塑性。通过利用“内存中计算”，有可能解决“冯·诺伊曼瓶颈”这一长期存在的问题：需要在处理核心和内存之间不断地转移数据。此外，硬件神经网络的实现可能是高密度、低功率神经形态系统的关键推动因素。该项目将是建模和实验工作的结合，将包括以下组件/目标：基于氧化物的神经形态设备的制造和结构特征；设备的内在自适应(“突触样”)能力的演示和优化；模拟和构建神经形态电路系统；在博士期间，学生将获得半导体加工和表征技术、新型半导体物理、神经形态结构和神经形态计算算法的广泛知识。该项目与ICT中的人工智能主题相关。

The objective of the work is to exploit the use of memristive/resistive RAM (RRAM) devices for novel neuromorphic (neuro-inspired) computing architectures and systems. Such systems have massive potential in delivering solutions that are far superior to existing CMOS hardware in implementation of machine learning (ML) and machine intelligence. One of the main benefits is a significant reduction in circuit complexity and vast improvements in power efficiency. This is crucial for bringing ML in adaptive embedded systems. Furthermore, this could be leveraged in "Internet/Intelligence of things" era with ML algorithms implemented directly on board, facilitating efficient local data processing and enabling devices to make decisions locally, rather than to rely on data streaming and latency-prone cloud computing.Resistive RAM (RRAM) technology, a subclass of memristive systems, is based on simple two terminal nanodevices that can repeatedly vary their resistance, with low operational energy and very high levels of integration. Remarkably, they resemble different neuronal functions - most importantly a synaptic-like plasticity by gradually changing their resistance ("synaptic weights"). By utilising the "computing-in-memory", it is possible to solve the long-lasting problem of the "von Neumann bottleneck": the need to continually shuffle data between processing cores and memory. Furthermore, implementation of hardware neural networks can be a key enabling factor for high-density, low-power neuromorphic systems.The project will be the mix of modelling and experimental work and will include the following components/objectives:Fabrication and structural characterisation of oxide-based neuromorphic devices;Demonstration and optimisation of devices' intrinsic adaptive ("synaptic-like") capabilities;Simulation and building neuromorphic circuit systems;During the PhD, the student will gain an extensive knowledge of semiconductor processing and characterisation techniques, novel semiconductor physics, neuromorphic architectures and neuromorphic computing algorithms.The project is relevant for the Artificial Intelligence theme within ICT.