面向大数据分析的自学习网络关键技术研究

Big data analytics is a technology that can reveal hidden patterns and internal correlations from a big amount and a large variety of data, which could lead to faster decision making, cost reduction and improved customer experience for companies, and therefore it has become a key technology in the era of big data. This project proposes a novel approach of doing big data analysis using artificial neural networks based on emerging memristors with self-learning capability. This approach allows the analysis of unstructured or semi-structured data, supports online learning, and has the advantages of very compact integration and low power consumption. This project plans to investigate the critical devices, integration technology and algorithms that are involved in this approach. We will adopt a novel heterostructure resistive switching layer and finely tune the ion conduction in the devices via systematically manipulating the stoichiometry of materials, the type and distribution of defects, the barrier of ion migration as well as the ion diffusion coefficient, etc., so as to obtain analog synaptic components with a speed of ≤100 ns, power consumption of ≤50 pJ, endurance of ≥10^6 and other good performance metrics including large on/off ratio as well as high linearity, etc. The devices will be further integrated into ≥1k array. The latest research findings in biological neural networks will be used to guide the design of algorithms and neurons, and the data analysis capability of the self-learning networks will be further implemented in hardware prototype. This project is expected to provide key technologies for the development of novel big data analytics.

大数据分析是从大规模、不同类型的数据中发现隐含信息与内在联系的技术，有助于企业快速决策、降低成本、提高客户满意度等，成为大数据时代的核心技术。本项目提出利用基于新型忆阻器的自学习人工神经网络执行大数据分析任务的新方案。该方法能够用于非结构化数据或半结构化数据的分析，支持在线学习，并且具有集成度高、功耗低等潜在优势。本项目计划对该方案所需的核心器件、集成技术、算法等进行深入研究。通过在阻变层中采用新型异质结复合结构，并系统调控阻变层化学计量比、缺陷类型、缺陷分布、离子迁移势垒、离子扩散系数等精细调控忆阻器中的离子输运，研发响应速度≤100纳秒，操作功耗≤50pJ，循环次数≥10^6，具有良好开关比、线性度等性能的突触单元，并进行≥1k的阵列集成。借鉴生物神经网络最新研究进展进行学习算法、神经元等设计，并通过硬件原型验证实现自学习网络的数据分析功能，为新型大数据分析技术的发展提供关键技术支持。

围绕大数据、人工智能时代众多数据密集型计算任务对于智能化、高能效基础器件及神经网络系统的迫切需求，本项目在纳米尺度神经形态器件微观机理、高精度人工突触与人工神经元、面向高效数据分析的智能系统构建等研究方面取得了一系列创新成果。针对长期困扰氧化物忆阻器研究的机理难题，本项目提出了利用微小静电力表征阻变存储器氧离子输运的创新方法，实现了对氧化铪神经形态器件微观离子输运过程和权值调整机制的解析，为从原理出发设计优化神经形态器件奠定了物理基础；针对已有器件缺乏时空复杂性处理能力、能耗过高等限制，本项目提出并实现了离子栅控型时敏突触，单脉冲能耗为30fJ，达到了与生物突触相当的水平，首次研制了具有非线性增益与时空动力学的人工神经元，为构建神经形态计算系统奠定了器件基础；基于项目研制的神经形态器件实现了突触和神经元完全基于忆阻器、支持在线学习的全忆阻神经网络，并将混沌动力学引入到神经网络中，首次在单个忆阻器阵列中实现了暂态混沌神经网络，实现了函数优化、组合优化等复杂优化问题的高效硬件求解；面向移动端对于在线学习能力的需求，本项目基于4Mb相变存储器芯片实现了大规模卷积神经网络VGG-16的在线训练，并结合优化的直接误差反传算法将训练时间和能耗减少了3倍以上；提出并研制了基于忆阻器的数据高效对称加密解密、汉明距离计算、蒙特卡洛计算、汉明纠错码等一系列高效存算一体数据处理系统，为下一代智能化、高能效存算一体架构和芯片的研发奠定了基础。