Neuromorphic memristive circuits to simulate inhibitory and excitatory dynamics of neuron networks: from physiological similarities to deep learning

为什么人类的大脑不能像计算机一样运作？两者在运算时都使用逻辑和数字。然而，与局部化的计算机比特记忆相比，人类的记忆更具分布性(类似于模式)，具有时间衰减(堵塞)，在训练时会改变“连接”，并被设计为处理信息信号(激发波)，而不是像计算机那样将比特存储在不同的存储位置。尽管现代计算机在数字方面的表现远远超过人脑，但它们仍然无法处理需要猜测和模糊逻辑的任务。这解释了人们对人工智能系统的浓厚兴趣，这些系统被训练来执行某些在数值模拟中不可行的任务。目前，人工智能技术由两个不同的目标驱动：(I)技术需求(自动驾驶汽车、在线交易、用于医学成像分析的人工智能等)。以及(Ii)试图创造一个具有思考、感觉和与人类互动能力的电子大脑。使用不同的学习算法，人工智能在几个战略和决策任务(例如围棋)中展示了与人脑相当的能力，甚至超过了人脑。然而，目前尚不清楚人工智能算法如何/是否与信息处理和大脑中相应的心理生理过程有关。回答这个大问题不仅有助于制造具有人类能力的机器，还有助于阐明大脑是可以简化为生物连接的电路，还是具有简单的电气/化学功能以外的东西。为了真正模拟大脑中的信息处理(神经形态计算)，应该开发新一代计算机体系结构。神经形态计算和信号处理最有前景的技术之一是基于忆阻器，其中电阻通过通过系统的总电荷来调节(例如，在两种状态之间切换)(例如，电阻取决于所施加的电脉冲序列/历史，其可以编码通过大脑传播的信息信号)。利用拉夫堡在固态物理、功能材料、薄膜、建模和人工智能方面的专业知识，在世界领先的神经形态研究中心(马萨诸塞大学阿默斯特分校)和神经科学/生理专业知识(索尔克生物研究所)的协同下，在英国工业合作伙伴的需求驱动下，我们打算开发一种记忆神经形态芯片组的原型，能够通过模拟大脑皮层中的神经过程来分析图像流并做出决策和选择。受生物神经元操作和深度学习AI范式的启发，我们建议开发一种通过两个相互竞争的过程工作的电路：(1)混合或干扰在不同时刻(例如，随后的视频帧)由视觉刺激产生的电信号，以及；(2)将信号从一个电路层传输到另一个电路层，以提取主要的视觉特征/概念。对于神经形态系统，这些过程的组合用于图像流分析以前从未被考虑过，这是本研究的主要创新之处。这使我们能够比较视频中的图像帧，并降低信息的复杂性，从而做出二元决策(选择)。这种神经形态的两个过程概念有一个明确的大脑功能类比：感官刺激激发无数产生叠加信号的感受器，其最终效果可以通过一句简短的识别(“这是我妈妈”)或辨别(“那是一辆汽车而不是一辆自行车”)来表达。在二元选择之后的干扰和卷积信息的循环似乎特别适合于忆阻器分层结构，其中初始复电压图案编码图像流减少到结构的更深层中的某个忆阻器的开关或不开关(发出做出哪个决定/选择的信号)。开发的原型将是视觉皮质的第一个记忆实现。

Why does the human brain not operate as a computer? Both use logic and numbers when operating. Nevertheless, the human memory is much more distributed (pattern-like) in contrast to localised computer bit-memory, has time decay (clogging), changes "wiring" when trained and is designed to process information signals (excitation waves) rather than just store bits in different memory locations as a computer does. Although modern computers significantly numerically outperform the human brain, they still cannot handle tasks requiring guessing and fuzzy logic.This explains a booming interest in AI systems trained to perform certain tasks infeasible for numerical simulations. Currently AI technology is driven by two distinct goals: (i) technological demand (autonomous vehicles, online trading, AI for medical imaging analysis etc.) and (ii) an attempt to create an electronic brain with the ability to think, feel and interact with humans. Using different learning algorithms, AI has demonstrated abilities comparable to, or even outperforming, the human brain in several strategic and decision-making tasks (e.g., the game of Go). However, it is unclear how/if AI algorithms relate to information processing and the corresponding psycho-physiological processes in the brain. Answering this big question would help in not only making machines with human abilities but also elucidating whether a brain can be reduced to a biologically-wired electric circuit only or it has something beyond simple electric/chemical functionalities.To truly emulate information processing in the brain (neuromorphic computing), a new generation of computer architecture should be developed. One of the most promising technologies for neuromorphic computing and signal processing is based on memristors, where resistance is tuned (e.g., switching between two states) by total charge passed through the system (e.g., resistance depends on applied electric pulse sequence/history which can encode an information signal propagating through the brain). Using Loughborough's expertise in solid state physics, functional materials, thin films, modelling, and AI, in synergy with a world-leading centre of neuromorphic research (the University of Massachusetts, Amherst) and neuroscience/physiological expertise (Salk Institute for Biological studies), and driven by the demand of UK industrial partners, we intend to develop a prototype of a memristive neuromorphic chipset able to analyse image-streams and to make decisions and choices by mimicking neural process in a brain cortex.Inspired by biological neuron operation, and the deep learning AI paradigm, we propose to develop an electric circuit operating via two competing processes:(1) Intermixing or interfering electric signals generated by visual stimuli at different time moments (e.g., subsequent video frames) and;(2) Transmitting signals from one circuit layer to another in order to extract the main visual features/concepts.A combination of these processes for image-stream analysis has never been considered before for neuromorphic systems and is the main novelty of the proposed research. This allows us to compare image frames in a video and to reduce the complexity of the information towards a binary decision (choice). This neuromorphic two-process concept has a clear brain-functioning analogy: sensory stimuli excite a myriad of receptors generating superimposing signals, an end effect of which can be expressed by a short statement of recognition ("It's my Mom") or discrimination ("It was a car not a bike"). Cycles of interfering and convolving information followed by binary choice seems particularly well fit to memristor layered architecture where initial complex voltage-pattern encoding image stream reduces to switching or not of a certain memristor (signalling which decision/choice is made) in a deeper layer of the structure. The developed prototype will be the first memristive realisation of a visual cortex.