【热点热聊】神经形态计算·区间值决策·混合仿生智能·大鼠机器人·玻尔兹曼机

本专栏特邀编委

Nenggan Zheng, Qiushi Academy for Advanced Studies, Zhejiang University, China

Cristian Axenie, Neuroscientific System Theory Group, Department of Electrical and Computer Engineering, Technical University of Munich, Germany

Dakai Zhu, Department of Computer Science, University of Texas at San Antonio, U.S.A

Man Lin, Department of Mathematics, Statistics and Computer Science, St. Francis Xavier University, Canada

       生物脑被认为是富有创造力且高效的计算系统，能够完成各类不同难度的感知、认知及运动控制等复杂任务。这背后所蕴含的神经机制和信息处理过程，受到越来越多的科研人员的高度重视。人们试图研究各类新颖的技术方案来探索这些基础的神经计算原理，研制模拟大脑处理机制的神经认知计算模型。近年来，脑机接口技术的飞速发展，为连接生物脑和计算机提供了直接的信息通道，从而形成了生物智能与机器智能融合的混合智能系统。这类革命性的新型智能系统能够实现了两种智能形式的优势互补。

       从硬件系统设计方面，神经形态计算系统模拟生物神经元工作机制，用于实现两个目标，一方面加深我们对大规模神经回路工作机制的理解；另一方面可以设计出非冯诺依曼架构的新型高效计算系统。近年来各类神经拟态芯片推陈出新，不仅有助于我们以多学科的视角来验证大脑的认知计算过程，还可以构建出具有工程应用价值的认知计算系统。

       本专栏旨在概括认知神经计算和混合智能的最新发展，从国内外众多投稿中收录5篇文章，反映了近年来在神经计算和混合智能系统领域的部分研究成果。

A Survey of Neuromorphic Computing Based on Spiking Neural Networks 主要介绍了神经形态计算方法，用于建立能够仿真生物脑的数字及模拟计算系统。它具有高性能和低功耗的优势。本文对基于脉冲神经网络的神经形态计算的发展历史、常用神经元模型、主要的研究项目、神经形态传感器以及在脑机接口领域的应用进行了系统性的介绍。

      DOI: cje.2018.05.006

Rough Set Model for Cognitive Expectation Embedded Interval-Valued Decision Systems 提出了带认知期望的区间值决策系统的概念，基于认知期望与区间值之间的距离构建新的优势关系，并构造了相应的粗糙集模型，用以提高智能系统的描述和分析能力。

      DOI: cje.2017.09.024

An HSR Data Model for Cyborg Insect Research Experiments 提出了一种新型数据模型，采用元数据集成的方法同步采集分布式传感器的异构实时数据。基于该数据模型的昆虫-机器实验平台在数据同步、实时处理和混合数据集成等方面，满足混合智能昆虫机器人的研究需求。

      DOI: cje.2017.08.010

A Posture Recognition System for Rat Cyborg Automated Navigation 提出了一种能给出大鼠机器人具体运动参数和精确身体姿态的监控系统。利用形状信息识别大鼠姿态，提出了计算性能和计算准确率优化的形状描述特征。该监控系统主要用于大鼠机器人的自动导航，提高控制的精确性。

     DOI: cje.2018.04.003    

Training Restricted Boltzmann Machine Using Gradient Fixing Based Algorithm 分析由采样算法所引起的近似梯度与真实梯度之间的数值误差与方位误差，并提出了梯度修正模型来减少此类梯度误差。基于此模型设计出了两个用于受限玻尔兹曼机训练的新算法GFGS和GFPT。实验证明新算法以合理的计算运行时间代价获得更高的训练准确性。

     DOI: cje.2018.05.007

SPECIAL ISSUE ON NEURAL COGNITIVE COMPUTATION AND CYBORG INTELLIGENT SYSTEMS

Guest Editors:

Nenggan Zheng, Qiushi Academy for Advanced

Studies, Zhejiang University, China

Cristian Axenie, Neuroscientific System Theory Group, Department of Electrical and Computer Engineering, Technical University of Munich, Germany

Dakai Zhu, Department of Computer Science, University of Texas at San Antonio, U.S.A

Man Lin, Department of Mathematics, Statistics and Computer Science, St. Francis Xavier University, Canada

It is believed that brain is one of the most creative and powerful computation systems, exhibiting diverse cognitive functionalities. Scientific problems related to the neural mechanisms and computation processes underlying the brains have attracted researchers to propose solutions to discover the fundamental principles, design research methods and present novel algorithms. Besides, emergent engineering approaches also extend the area explored largely and broaden our vision dramatically in recent decades. One of the typical examples may be the neural machine interface technology, which provides new ways to directly connect man-made devices with biological brains. Nowadays, information is becoming to flow smoothly between intelligent machines and biological neural substrates, which results in the revolutionary hybrid system of Cyborg intelligence. By achieving the idealized complementary integration, cyborg intelligent systems can solve complicated problems in general environments, which neither biological nor artificial intelligence system can tackle alone.                                                                   

Besides, another direction of exploration is the field of Neuromorphic Computing which investigates principles of computation in the nervous system and methods for implementing them in full-custom hardware/software co-designed technical systems. The neuromorphic research goal is two-fold: (a) advance our understanding of how large scale neuronal circuits perform coherent computation, and (b) push current engineering technology beyond today’s von-Neumann computing architectures. Various implementations of neuromorphic hardware systems, ranging from single synapses to large scale models, have been developed in recent decades. Examples include sensory devices (e.g. silicon retinas or cochleae), analog VLSI chips and massively distributed digital systems for real- or accelerated-time spiking neural networks. Enroute to future applications of neuromorphic computing the current challenge lies in the identification and implementation of functional brain algorithms in a multi-disciplinary approach.

This timely special issue aims to summarize recent developments in cognitive neural computation and cyborg intelligence in a broad scope, and thus make a better understanding of the underlying mechanisms of biological neural substrates and the novel approaches for cyborg intelligent systems. The 5 articles included in this special issue reflect exciting achievements in recent years.

A Survey of Neuromorphic Computing Based on Spiking Neural Networks (Zhang, et al.) mainly introduces the Neuromorphic Computing which builds digital or analog computer systems that emulate or simulate the biological brain, in order to achieve high performance and low power consumption for intelligent information processing applications. And this article reviews on neuromorphic computing based on Spiking neural networks (SNNs), including its history of development, common neuron models, major research projects, neuromorphic sensors, and applications in brain-computer Interfaces.

DOI: cje.2018.05.006

Rough Set Model for Cognitive Expectation Embedded Interval-Valued Decision Systems (Dai, et al.) proposes the concept of Interval-valued decision system with expectations (IDSE), a new dominance relation based on the distances between expectations and interval values is constructed. And rough set model is investigated to improve the perception ability of biological intelligence.

DOI: cje.2017.09.024

An HSR Data Model for Cyborg Insect Research Experiments (Hong, et al.) presents a novel data model, naming as the Hybrid-synchronized-real time (HSR) data model, which can synchronize the hybrid raw data channels by a meta data integration method and process them with a fixed priority scheduling algorithm. Their work satisfies the requirement of the cyborg intelligent insect research, collecting data from the distributed sensing models and processing the multi-modal signals in real-time.

DOI: cje.2017.08.010

A Posture Recognition System for Rat Cyborg Automated Navigation (Zhang, et al.) develops a monitoring system that is capable of giving detailed motion parameters and accurate body postures of the cyborg rats. The authors explore the methods to recognize rat postures using shape information and propose several simple shape descriptors fast to compute with acceptable performance. The monitoring system is mainly used to extract detailed motion state parameters to improve the automatic navigation performance of cyborg rats.

DOI: cje.2018.04.003

Training Restricted Boltzmann Machine Using Gradient Fixing Based Algorithm (Li, et al.) analyses the approximation error between the approximate gradient and the true gradient and proposes a gradient fixing model to reduce this error. Two novel algorithms are suggested to limit the gradient error and improve the training accuracy.

DOI: cje.2018.05.007

点击下方“阅读原文”前往本刊官网浏览并下载文章