神经网络七十年:回顾与展望

发布时间：2018-05-11 11:40

本文选题：类脑智能 + 神经网络　；参考：《计算机学报》2016年08期

【摘要】：作为联接主义智能实现的典范,神经网络采用广泛互联的结构与有效的学习机制来模拟人脑信息处理的过程,是人工智能发展中的重要方法,也是当前类脑智能研究中的有效工具.在七十年的发展历程中,神经网络曾历经质疑、批判与冷落,同时也几度繁荣并取得了许多瞩目的成就.从20世纪40年代的M-P神经元和Hebb学习规则,到50年代的Hodykin-Huxley方程、感知器模型与自适应滤波器,再到60年代的自组织映射网络、神经认知机、自适应共振网络,许多神经计算模型都发展成为信号处理、计算机视觉、自然语言处理与优化计算等领域的经典方法,为该领域带来了里程碑式的影响.目前,模拟人脑复杂的层次化认知特点的深度学习已经成为类脑智能中的一个重要研究方向.通过增加网络层数所构造的"深层神经网络"使机器能够获得"抽象概念"能力,在诸多领域都取得了巨大的成功,又掀起了神经网络研究的一个新高潮.文中回顾了神经网络的发展历程,综述了其当前研究进展以及存在的问题,展望了未来神经网络的发展方向.
[Abstract]:As an example of intelligent realization of connectionism, neural network is an important method in the development of artificial intelligence, which uses extensive interconnected structure and effective learning mechanism to simulate the process of human brain information processing. It is also an effective tool in the research of brain-like intelligence. In the development of 70 years, neural networks have been questioned, criticized and ignored, but also several times prosperity and made a lot of achievements. From M-P neurons and Hebb learning rules in the 1940s to Hodykin-Huxley equations in the 1950s, perceptron models and adaptive filters, to self-organizing mapping networks, neural cognitive machines, adaptive resonance networks in the 1960s, Many neural computing models have developed into classical methods in the fields of signal processing, computer vision, natural language processing and optimization computing, which have brought about a milestone impact in this field. At present, deep learning, which simulates the complex and hierarchical cognitive characteristics of the human brain, has become an important research direction in brain-like intelligence. The "deep neural network", which is constructed by increasing the number of network layers, enables the machine to acquire the "abstract concept" capability. It has achieved great success in many fields and set off a new climax in the research of neural networks. This paper reviews the development of neural networks, summarizes their current research progress and existing problems, and looks forward to the future development of neural networks.
【作者单位】：西安电子科技大学智能感知与图像理解教育部重点实验室智能感知与计算国际联合研究中心智能感知与计算国际合作联合实验室;西安电子科技大学计算机学院;
【基金】：国家“九七三”重点基础研究发展计划项目基金(2013CB329402) 国家自然科学基金重大研究计划(91438201,91438103) 教育部“长江学者和创新团队发展计划”(IRT_15R53)资助
【分类号】：TP183

【参考文献】