连续吸引子神经网络神经形态芯片实现的研究

Neuromorphic engineering, aims to develop new generation of technologies that carry out computational principles of cognitive brain in real-time while maintaining remarkable energy efficiency. The neuromorphic implementation of the continuous attractor neural network (CANN) has attracted scientists’ attentions in neuromorphic engineering, because CANN involves neural computations of many brain functions, such as sensory perceptions, cognitive functions and motor controls. However, there are still several challenges for electrical implementations of basic units of neural systems, such as silicon synapses and neurons, and also for system integrations of CANN on hardware. This project will focus on the research on the neuromorphic chip of CANN in both circuit and network levels from the point of view of structure and function in the dynamical system approach. In the circuit level, we will study circuit implementations of basic neural units according to expected functions of these circuits and properties of electron devices, and finally give their structures through theoretical analysis and simulations. In the network level, we will derive a partial integro-differential equation characterizing the neural mechanism of on-chip CANN, and study its mismatch improvement through the sensitivity analysis of its performance to parameters in both levels, which guarantee the effective on-chip integration of CANN. It is promising to use this CANN neuromorphic chip in the platform of neuromorphic computing, the integration of the large-scale cognitive brain and the brain-machine interface due to its characteristics of flexible configurations and integrations.

神经形态工程，旨在基于认知大脑工作原理，开发实时低功耗的类脑计算新技术。连续吸引子神经网络，因其动态特性是大脑诸项感知、认知和运动控制功能神经计算的重要基础，其芯片实现是神经形态工程研究的一个热点；但目前，其突触神经元等基本单元电路的功能实现和片上神经网络动态特性的功能整合，尚有挑战。本项目将利用专注结构与功能的动态系统方法，在电路和网络两个层次研究连续吸引子神经网络的芯片实现。在电路层次，根据电子器件属性，改进和设计突触神经元等基本单元电路，并通过理论分析与仿真，确定电路结构。在网络层次，建立宏观动力学方程描述单元电路与网络结构共同作用的动力学机制；分析芯片性能指标对各层次参数的敏感性，研究减小失配恶化网络动态特性的方法；最终确保该神经芯片功能的有效整合。由于该芯片的可配置性和可整合性，本项目对未来类脑计算平台、大规模功能大脑整合以及脑机接口，具有重要的科学和应用价值。

连续吸引子神经网络，其动态特性是大脑诸项感知、认知和运动控制功能神经计算的重要基础。其神经形态芯片实现，因其模块化后的可配置性和可整合性，具有重要的研究与应用价值，如脑机接口、电生理信号处理以及运动控制等。本项目应用动态系统方法，在电路和系统两个层次，研究连续吸引子神经网络模型的物理实现。项目中，我们根据动态系统方法，改进了硅神经元的性能，提出了NMDA型突触电路，设计了一枚含硅神经元与硅突触的神经芯片，并成功流片。我们设计的两变量赢者全拿认知电路，可以实现两选项知觉决策、工作记忆和迟滞现象等大脑中观察到的认知功能。同时，基于数字模式的单芯片设计方案，我们完成了连续吸引子类脑芯片的前端设计，它具有网络动态特性的高度可配置性。这些成果，对未来的大规模功能大脑整合平台，具有重要的应用价值。在该项目的资助下，项目组流片一次，成功测试后获得一枚芯片；项目组已在国内外重要刊物上发表论文2篇，其中SCI检索期刊论文1篇，EI检索论文1 篇，另外还投出1 篇会议论文正在审稿中，完成了预期目标。项目组参加国内外重要学术会议5次，受邀作报告3次。

内容获取失败，请点击重试