Power efficient deep neural networks using analogue and neuromorphic circuits for biomedical applications

深度神经网络（dnn）通常由多层神经元通过加权路由相互耦合组成。dnn最近在不同的应用中表现出了最先进的性能。例如卷积神经网络（CNN）被广泛应用于图像处理，递归神经网络（RNN）被用于自然语言处理。然而，这些应用程序大多是在卷积计算系统上执行的，而卷积计算系统不适合实现这种大规模并行架构。神经形态计算系统显示了一种新的非冯·诺伊曼大规模并行架构，非常适合实现深度神经网络。像大脑一样，模拟神经动力学的神经形态回路更适合于视觉、听觉、嗅觉和其他感觉系统的信号处理。这种实时和节能的设计也符合医疗应用的要求。本项目将从脑科学和神经科学中最前沿的认知和感知机制中寻求灵感，并尝试建立模拟CMOS电路的学习算法。为了将现有的训练算法从数字系统移植到模拟系统中，神经元的电路设计需要实现信号的双向传输。在未来，将引入cmos兼容的记忆器件，以实现计算和内存的共定位。

Deep neural networks (DNNs) typically consist of many layers of neurons coupled among each other by weighted routing. DNNs have recently exhibited state-of-art performance in different applications. For example, convolutional neural network (CNN) is widely applied to image processing, and recurrent neural network (RNN) is used for natural language processing. However, most of these applications are performed on convolutional computing systems, which are ideally unsuited for implementing such massively parallel architectures. Neuromorphic computing systems show a new non-von Neumann massively parallel architecture that is ideally suited to implementing DNNs. Neuromorphic circuits that emulate neural dynamics, like the brain, are more suitable for signal processing of visual, auditory, olfactory and other sensory systems. This real-time and energy-efficient design also matches the requirement of medical applications. This project will seek inspiration from the most cutting-edge cognitive and perceptual mechanisms in brain science and neuroscience and try to establish a learning algorithm for analogue CMOS circuits. The circuit design of the neurons needs to realize the bi-direction transmission of the signal in order to transplant the existing training algorithm from the digital system to the analogue system. In the future, CMOS-compatible memristive devices will be introduced to achieve co-localization of computation and memory.