物端神经形态类脑芯片设计综述

Review on the edge neuromorphic brain-like chip design

本文简述目前物端神经形态类脑芯片设计的研究现状,首先回顾了目前已发表的神经形态类脑芯片,总结了其特点和局限性,然后简要介绍了脉冲神经网络的基础知识,包括脉冲神经网络的经典神经元模型、网络拓扑结构以及仿生学习算法。接下来重点介绍了目前最新发表的片上实时强化学习物端类脑芯片、片上三重类脑学习物端类脑芯片、视觉压缩感知识别物端类脑芯片以及片上多层脉冲神经网络(spiking neural network,SNN)学习物端类脑芯片4款物端神经形态类脑芯片的算法优化方案、芯片架构和电路设计、以及现场可编程门阵列(field programmable gate array,FPGA)原型或实际制造芯片测试结果。所介绍的神经形态类脑芯片均具备片上实时学习功能,且在各类基准数据集上都实现了较高的识别准确率。同时,提出的芯片架构均为较低成本,能达到相对较高的处理速度,同时还具有较为灵活的可扩展性和可配置性,能够适用于不同的物端智能应用场景,为目前研究领域面临的挑战提供了可行的解决方案。最后指出了目前物端神经形态类脑芯片设计领域发展中的核心瓶颈,并介绍了初步的解决方案。未来将围绕这些方向开展研究,设计新一代高性能物端神经形态类脑芯片。

This paper briefly reviews the status of research on the design of edge neuromorphic brain-like chips.It firstly reviews the published neuromorphic brain-like chips,summarizes their features and limitations,and briefly introduces the basics of spiking neural networks(SNNs),including the classical neuron model of spiking neural networks,network topology and bio-inspired learning algorithms.The review then highlights the algorithm optimization,chip architecture and circuit design,as well as field programmable gnte anrcy(FPGA)prototype or fabricated chip testing results of four recently published typical edge neuromorphic brain-like chips:the spiking Extreme Leaning Machine hardware with on-chip real-time reinforcement learning,the TripleBrain neuromorphic architecture enabling multiple brain-inspired learning schemes,the edge neuromorphic chip with high-speed on-chip learning on spike-domain compressive features,and the MorphBungee neuromorphic chip supporting on-chip deep SNN learning on-chip multilayer SNN learning.The experimental results show that the presented neuromorphic brain-like chips all have on-chip real-time learning capabilities and achieve high recognition accuracies on various benchmark datasets.At the same time,the proposed chip architectures are all low-cost and can achieve relatively high processing speeds,while also being more flexible,scalable and configurable for different edge intelligence application scenarios,providing a feasible solution to the challenges faced by the current research field.Finally,this paper indicates key bottlenecks in the current field of edge neuromorphic brain-like chip design and presents possible solutions.Future research will focus on these directions to foster a new generation of high-performance edge neuromorphic brain-like chips.