摩尔定律失效下，类脑计算为什么成为下一代关键技术？

雷锋网AI科技评论按：当前计算机技术面临着两个重要瓶颈：（1）摩尔定律失效；（2）「冯诺依曼」架构导致的能效低下。

随着集成电路的规模越来越接近物理极限，人类若想进一步提升计算机的性能，必然要考虑新的计算机架构。而另一方面，「冯诺依曼」架构中，运算单元和存储单元分离，使得大部分能量和时间都消耗在数据的读取和存储过程中；并且数据处理是基于串行结构，即同一时刻只能执行一个任务。

这与人脑处理信息的方式差别巨大。在进行学习和认知等复杂计算时，人脑的功耗只有 20 瓦；而目前最先进的计算机模拟人脑功能，功耗也高达 800 万瓦以上，速度比人脑要慢 1000 倍以上。究其原因，是因为现代计算机一般使用固定的数字化的程序模型，同步、串行、集中、快速、具有通用性地处理问题，数据存储与计算过程在不同地址空间完成。而与之形成鲜明对比的是，人的大脑会重复利用神经元，并突触、异步、并行、分布式、缓慢、不具通用性地处理问题，是可重构的、专门的、容错的生物基质，并且人脑记忆数据与进行计算的边界是模糊的。

鉴于此，当前借鉴人脑发展的类脑计算技术，被认为是应对当前挑战的重要方案。“类脑计算”本质来说，即利用神经计算来模拟人类大脑处理信息的过程，被认为“下一代人工智能”的重要方向，也是当前人工智能领域的热点方向。

与经典人工智能符号主义、连接主义、行为主义以及机器学习的统计主义这些技术路线不同，类脑计算采取的是仿真主义：

• 结构层次模仿脑（非冯·诺依曼体系结构）

• 器件层次逼近脑（神经形态器件替代晶体管）

• 智能层次超越脑（主要靠自主学习训练而不是人工编程）

因此，类脑计算试图建立一个类似的架构，使得计算机也能保持类脑的复杂性，达到可处理小数据 & 小标注问题、适用于弱监督和无监督问题、关联分析能力强、鲁棒性强、计算资源消耗较少、具备认知推理能力、时序相关性好、可能解决通用场景问题的目的，最终实现强人工智能和通用智能。

目前，国际上类脑计算研究已经取得显著进展，技术探索阶段已经过去，技术预研已经开始，一些关键技术获得突破，相关的技术原型和系统原型已开发成功。

总的来说，类脑智能技术体系分四层：基础理论层、硬件层、软件层、产品层。

基础理论层基于脑认知与神经计算，主要从生物医学角度研究大脑可塑性机制、脑功能结构、脑图谱等大脑信息处理机制研究；

硬件层主要是实现类脑功能的神经形态芯片，也就是非冯诺依曼架构的类脑芯片，如脉冲神经网络芯片、忆阻器、忆容器、忆感器等；

软件层包含核心算法和通用技术，核心算法主要是弱监督学习和无监督学习机器学习机制，如脉冲神经网络、增强学习、对抗神经网络等；

通用技术主要是包含视觉感知、听觉感知、多模态融合感知、自然语言理解、推理决策等；产品层主要包含交互产品和整机产品，交互产品包含脑机接口、脑控设备、神经接口、智能假体等，整机产品主要有类脑计算机、类脑机器人等。

类脑计算，被视为未来信息技术最具有发展前景的重要领域之一，正如欧盟人脑旗舰研究计划所指出的：「在未来 20 到 30 年内，谁要想主导世界经济，谁必须在类脑计算这个领域领先」。

中国学术领域近年来也越来越关注类脑计算的研究。这表现在多个方面。其一，国家层面上，从2016年起制定了为期15年的「脑计划」，类脑计算正是其核心研究领域之一。另一方面，在近些年来，有越来越多的会议开始设置「类脑计算」专场或论坛。例如前不久刚结束的由中国计算机学会、雷锋网、香港中文大学共同举办的 CCF-GAIR 2019大会便设置了此专场。

近期将举办的国际图像图形学学术会议也将举办「类脑智能论坛」。

国际图象图形学学术会议（ICIG）是中国图象图形学学会主办的最高级别的系列国际会议，创建于2000年，每两年举办一届，迄今已经成功举办九届。

第十届国际图象图形学学术会议（ICIG2019）将于2019年8月23-25日在北京友谊宾馆召开，主题为“人工智能时代的图像图形前沿研究”，由清华大学、北京大学和中国科学院自动化研究所承办，得到了国际模式识别协会（IAPR）的支持。本次的会议共包含了3个特邀报告、2个讲习班、3个workshops，多个论坛。其中之一便为「类脑智能论坛」。

类脑智能论坛由CSIG机器视觉专委会承办，由中科院自动化所何晖光研究员和深圳职业技术学院人工智能学院院长杨金峰教授共同组织，邀请到6位专家从不同的角度来介绍类脑智能的研究进展、类脑研究中的难点问题，并对今后的研究进行展望。

各专家报告内容可参考如下：

 

Si Wu

Peking University

Title：Push-pull Feedback Implements Rough-to-fine Information Processing

Abstract: Experimental data has revealed that in addition to feedforward connections, there exist abundant feedback connections in a hierarchical neural pathway. Although the importance of feedback in neural information processing has been widely recognized in the field, the detailed mechanism of how it works remains largely unknown. Here, we investigate the role of feedback in hierarchical memory retrieval. Specifically, we consider a multi-layer network which stores hierarchical memory patterns, and each layer of the network behaves as an associative memory of the corresponding hierarchy. We find that to achieve good retrieval performance, the feedback needs to be dynamical: at the early phase, the feedback is positive (push), which suppresses inter-class noises between memory patterns; at the late phase, the feedback is negative (pull), which suppresses intra-class noises between memory patterns. Overall, memory retrieval in the network progresses from rough to fine. Our model agrees with the push-pull phenomenon observed in neural data and sheds light on our understanding of the role of feedback in neural information processing.

Biography: Dr. Si Wu is Professor at School of Electronics Engineering & Computer Science, Principle Investigator at IDG/McGovern Institute for Brain Research, and Principle Investigator at PKU-Tsinghua Center for Life Science in Peking University. He was originally trained as a theoretical physicist and received his BSc, MSc, and PhD degrees all from Beijing Normal University (87-95). His research interests have turned to Artificial Intelligence and Computational Neuroscience since graduation. He worked as Postdocs at Hong Kong University of Science & Technology (95-97), Limburg University of Belgium (97-98), and Riken Brain Science Institute of Japan (98-00), and as Lecturer/Senior Lecturers at Sheffield University (00-02) and Sussex University (03-08) of UK. He came back to China in 2008, and worked as PI at Institute of Neuroscience in Chinese Academy of Sciences (08-11) and Professor in Beijing Normal University (11-18). His research interests focus on Computational Neuroscience and Brain-inspired Computing. He has published more than 100 papers, including top journals in neuroscience, such as Neuron, Nature Neuroscience, PNAS, J. Neurosci., and top conferences in AI, such as NIPS. He is now Co-editor-in-chief of Frontiers in Computational Neuroscience.      

 

Sen Song

Tsinghua University

Title: Recent progress in brain research and inspirations for neurocomputing

Abstract: Recently, big scale neuronal recordings are starting to reveal the way information is represented in the nervous system. At the same time, analysis of artificial neural networks trained by deep learning is also starting to reveal its representations. In this talk, I will try to summarize and compare representations in deep neural networks and the brain, regarding objects, object features, object relations, tree like structures and graph-like structures, and start to build a mathematical framework to describe them.

 Biography: Dr. Sen Song is an principal investigator at Tsinghua Laboratory for Brain and Intelligence and Department of Biomedical Engineering at Tsinghua University. He received his Ph.D. degree in Neuroscience from Brandeis University in 2002. Before joining Tsinghua in 2010, he did post-doctoral research at Cold Spring Harbor Laboratory and Massachusetts Institute of Technology. His work in computational neuroscience on spike-timing dependent plasticity and motif analysis of cortical connectivity have been widely cited and form some of the theoretical foundations of brain-inspired computing. His current work involves computational neuroscience, neural circuits underlying emotions and motivations, and the interface between neuroscience and artificial intelligence.

 

Wenming Zheng

Southeast University

Title: Action Intention Understanding and Emotion Recognition for Human Computer Interface

 Abstract: Action intention understanding and emotion recognition play an important role in human computer interface. In this talk, I will address the methods of action intention understanding and emotion recognition from psychophysiological signals, such as EEG or audiovisual signals. Then, I will also briefly address the applications of this research in medical treatment and education.

Biography: Wenming Zheng received his PhD degree in signal and information processing from the Department of Radio Engineering, Southeast University, Nanjing, China, in 2004. He is currently a Professor and the Director of the Key Laboratory of Child Development and Learning Science, Southeast University. He ever worked as a visiting scholar or visiting professor at Microsoft Research Asia (MSRA), Chinese University of Hong Kong (CUHK), University of Illinois at Urbana-Champaign (UIUC), and Cambridge University, respectively. His current research interests include affective information processing for multi-modal signals, e.g., facial expression, speech, and EEG signals, and their applications in education and medical care. Dr. Zheng was an Awardee the Microsoft Young Professor Professorship. He won the Second Prize of the National Technological Invention in 2018, the Second Prize of the Natural Science of Ministry of Education in 2008 and 2015, the Second Prize of the Jiangsu Provincial Science and Technology Progress in 2009. He served as an Associated Editor of several peer reviewed journals, such as IEEE TRANSACTIONS ON AFFECTIVE COMPUTING, Neurocomputing, and The Visual Computer. He is a Council Member of the Chinese Society of Cognitive Science.

Yijun Wang

Chinese Academy of Sciences

Title: Recent Progress in Brain-Machine Integration Technology

Abstract: The brain-computer interface (BCI) technology establishes a direct communication channel between the brain and external devices, which can replace, restore or enhance human’s perception, cognition and motor functions. In recent years, as a new form of hybrid intelligence, the BCI-based brain-machine integration technology has shown great potential in the fields of healthcare, human-computer interaction, and national defense. In this talk, I will introduce recent progress in the development of the brain-machine integration technology. I will first review the history, current status, methodology, and challenges in this field. I will then present examples of progress of the brain-machine integration technology in communication and control, human augmentation, multi-modal integration, and biometrics.

Biography: Yijun Wang is a Research Fellow at the Institute of Semiconductors, Chinese Academy of Sciences, and a member of CAS Center for Excellence in Brain Science and Intelligence Technology. He was selected by the Thousand Youth Talents Plan of China in 2015. He received a B.E. degree and a Ph.D. degree in biomedical engineering from Tsinghua University in 2001 and 2007, respectively. From 2008 to 2015, he was first a Postdoctoral Fellow and later an Assistant Project Scientist at the Institute for Neural Computation, University of California San Diego, USA. His research mainly focuses on neural engineering and neural computation. His research interests include brain-computer interface (BCI), biomedical signal processing, and machine learning. He has published more than 100 papers in scientific journals and conferences such as PNAS, Journal of Neuroscience, IEEE Transactions on Biomedical Engineering. His papers have been cited more than 4500 times according to Google Scholar.

Xiaolin Hu

Tsinghua University

Title: Deep Learning Predicts Correlation between a Functional Signature of Higher Visual Areas and Sparse Firing of Neurons

Abstract: Visual information in the visual cortex is processed in a hierarchical manner. Recent studies show that higher visual areas, such as V2, V3, and V4, respond more vigorously to images with naturalistic higher-order statistics than to images lacking them. This property is a functional signature of higher areas, as it is much weaker or even absent in the primary visual cortex (V1). However, the mechanism underlying this signature remains elusive. We studied this problem using computational models. In several typical hierarchical visual models including the AlexNet, VggNet and SHMAX, this signature was found to be prominent in higher layers but much weaker in lower layers. By changing both the model structure and experimental settings, we found that the signature strongly correlated with sparse firing of units in higher layers but not with any other factors, including model structure, training algorithm (supervised or unsupervised), receptive field size, and property of training stimuli. The results suggest an important role of sparse neuronal activity underlying this special feature of higher visual areas.

Biography: Xiaolin Hu is an associate professor in the Department of Computer Science and Technology, Tsinghua University, Beijing, China. He got his PhD degree in Automation and Computer-Aided Engineering at The Chinese University of Hong Kong in 2017. He was a postdoc at Tsinghua University during 2017-2019. His research areas include artificial neural networks and computational neuroscience. His main research interests include developing brain-inspired computational models and revealing the visual and auditory information processing mechanism in the brain. He has published over 70 research papers in journals include IEEE Transactions on Neural Networks and Learning Systems, IEEE Transactions on Image Processing, IEEE Transactions on Cybernetics, PLoS Computational Biology, Neural Computation, and conferences include CVPR, NIPS, AAAI. He serve as an Associate Editor of IEEE Transactions on Neural Networks and Learning Systems and Associate Editor of Cognitive Neurodynamics.

Jian Liu

University of Leicester

Title: Towards the next generation of computer vision: visual computation with spikes

Abstract: Neuromorphic computing has been suggested as the next generation of computational strategy.  In terms of vision, the retina is the first stage of visual processing in the brain. The retinal coding is for understanding how the brain processes stimulus from the environment, moreover, it is also a cornerstone for designing algorithms of visual coding, where encoding and decoding of incoming stimulus are needed for better performance of physical devices. Here, by using the retina as a model system, we develop some data-driven approaches, spike-triggered non-negative matrix factorization and deep learning nets for characterizing the encoding and decoding of natural scenes by retinal neuronal spikes. I further demonstrate how these computational principles of neuroscience can be transferred to neuromorphic chips for the next generation of the artificial retina. As a proof of concept, the revealed mechanisms and proposed algorithms here for the retinal visual processing can provide new insights into neuromorphic computing with the signal of events or neural spikes.

Biography: Dr. Jian Liu received the Ph.D. in mathematics from UCLA, then worked as Postdoc Fellow at CNRS, France, and University of Goettingen, Germany. He is currently a Lecturer of Computational Neuroscience at University of Leicester, UK. His area of research includes computational neuroscience and brain-inspired computation for artificial intelligence. His work was published in Nature communications, eLife, Journal of neuroscience, PLoS computational biology, IEEE Transactions on Neural Networks and Learning Systems, IEEE Transactions on cybernetics, etc.

雷锋网(公众号：雷锋网)报道。

。