Synthetic neural networks for neuromorphic applications

人类大脑包含数十亿个神经元，它们通过突触交换信号。我们拥有超越最快的超级计算机的独特智力；例如，高水平的模式识别、能源效率，以及从经验中学习的终极技能。这些属性激发了所谓的神经形态（类脑）设备的创造，这种高度连接的电子电路试图模仿大脑中的结构。神经元和突触构成了生物神经系统的基本组成部分。在神经启发技术中，最大的挑战之一是找到合适的构建模块，可以模拟合成领域的大脑突触。此外，我们需要知道如何整合和控制这些构建块，以满足神经形态计算的基本要求。这包括区块之间的分散通信，记忆和处理的协同定位，以及支持学习和适应的多状态/模拟记忆响应。我的目标是创建一个研究项目，为开发尖端神经形态设备揭开新材料概念和构建模块的面纱。我将研究认知特征是如何从纳米级材料中产生的，在纳米级材料中，它们的导电性不是静态的，而是随着其终端中设置的电流/电压的大小而变化。这通常在记忆系统中看到，非易失性存储材料的电阻表现为一个动态量。我已经构建了一个创新的计算平台，该平台将模拟纳米级认知材料自组装网络的记忆特征，这些材料被视为神经形态的有希望的候选者。这种网络的例子是由随机分散的纳米线制成的意大利面状结构，其中复杂的记忆现象发生在它们的导线接触点上。我的目标是理论描述紧急电阻机制控制传播和记忆的电信号在其无序框架。建立与电子控制系统集成的认知网络材料虚拟电路模型，模拟典型的大脑功能，如数据识别、记忆和容错处理。我们将揭示最佳的材料特性、电路设计和训练方案，这些将在长期合作伙伴的实验室中进行测试，这些合作伙伴将在模拟结果的启发下设想制造概念验证的神经形态设备。该计划将使年轻科学家能够参与一个真正的跨学科环境，连接众多领域，如纳米技术，计算机科学和神经科学；它还将使加拿大的一个机构（和加拿大）成为一项颠覆性的大脑激发技术的知识交流中心，这项技术将极大地影响人工智能，预计人工智能将在未来几十年成为全球主要的经济驱动力之一。

The human brain contains billions of neurons that exchange signals through synapses. We have unique intellectual abilities that outstrip the fastest supercomputers; e.g. high-level pattern recognition, energy-efficiency, and the ultimate skill of learning from experience. Such attributes have inspired the creation of so-called neuromorphic (brain-like) devices, highly connected electronic circuits that attempt to mimic the architectures present in the brain. Neurons and synapses form the primitive building blocks in biological neural systems. One of the biggest challenges in neural-inspired technologies is to find suitable building blocks that can emulate brain synapses in the synthetic realm. Moreover, we need to know how to integrate and to control these building blocks in order to meet the basic requirements of neuromorphic computing. These involve decentralized communication between the blocks, co-location of memory and processing, and multistate/analog memory response that supports learning and adaptation. My goal is to create a research program that will unveil new material concepts and building blocks for the development of cutting-edge neuromorphic devices. I will investigate how cognitive features emerge from nanoscale materials in which their electric conductivity is not static but changes with the amount of current/voltage set in their terminals. This is typically seen in memristive systems, nonvolatile memory materials whose resistance behaves as a dynamical quantity. I have structured an innovative computational platform that will model the memristive characteristics of self-assembled networks of nanoscale cognitive materials seen as promising candidates for neuromorphics. Examples of such a network are spaghetti-like structures made by randomly dispersed nanowires in which complex memristive phenomena take place in their wire-wire contact points. I target the theoretical description of emergent resistive mechanisms controlling the propagation and memorization of electrical signals throughout their disordered frame. Virtual circuit models of cognitive network materials integrated with electronic control systems will be built to simulate typical brain-functions, e.g. data recognition, memorization, and fault-tolerant processing. We will reveal optimal materials properties, circuit designs, and training protocols that will be tested in the laboratory of long-term collaborators that envision the fabrication of a proof-of-concept neuromorphic device inspired by the outcomes of our simulations. This program will enable young scientists to engage in a truly interdisciplinary environment connecting numerous fields, e.g. nanotechnology, computer science, and neuroscience; it will also place a Canadian institution (and Canada) as the knowledge exchange hub of a disruptive brain-inspired technology that will greatly impact artificial intelligence, a sector expected to be one of the leading economic drivers world-wide in the next decades.