IEC-人工智能将赋能哪些行业-2019.3-98页.pdf

Artificial intelligence across industriesWhite Paper3Artificial intelligence is currently attracting considerable interest and attention from industry,researchers,governments as well as investors,who are pouring record amounts of money into the development of new machine learning technologies and applications.Increasingly sophisticated algorithms are being employed to support human activity,not only in forecasting tasks but also in making actual decisions that impact society,businesses and individuals.Whether in the manufacturing sector,where robots are adapting their behaviour to work alongside humans,or in the home environment,where refrigerators order food supplies based on the homeowners preferences,artificial intelligence is continuously making inroads into domains previously reserved to human skills,judgment or decision-making.While artificial intelligence has the potential to help address some of humanitys most pressing challenges,such as the depletion of environmental resources,the growth and aging of the worlds population,or the fight against poverty,the increasing use of machines to help humans make adequate decisions is also generating a number of risks and threats that businesses,governments and policy makers need to understand and tackle carefully.New concerns related to safety,security,privacy,trust,and ethical considerations in general are definitely emerging together with the technological innovations enabled by artificial intelligence.These challenges are common to all societies across the globe and will need to be dealt with at the international level.The present White Paper provides a framework for understanding where artificial intelligence stands today and what could be the outlook for its development in the next 5 to 10 years.Based on an explanation of current technological capabilities,it describes the main systems,techniques and algorithms that are in use today and indicates what kinds of problems they typically help to solve.Adopting an industrial perspective,the White Paper discusses in greater detail four application domains offering extensive opportunities for the deployment of artificial intelligence technologies:smart homes,intelligent manufacturing,smart transportation and self-driving vehicles,and the energy sector.The analysis of various specific use cases pertaining to these four domains provides clear evidence that artificial intelligence can be implemented across and benefit a wide set of industries.This potential is paving the way for artificial intelligence to become an essential part of the equation in resolving issues generated by todays and tomorrows megatrends.Building upon this analysis,the White Paper provides a detailed description of some of the major existing and future challenges that artificial intelligence will have to address.While industry and the research community constitute the principal drivers for developing initiatives to tackle technical challenges related to data,algorithms,hardware and computing infrastructures,governments and regulators urgently need to elaborate new policies to deal with some of the most critical ethical and social issues foreseen to be the by-products of artificial intelligence.Standardization and conformity assessment are expected to play an essential role not only in driving market adoption of artificial intelligence but also in mitigating some of the most pressing challenges related to decision-making by machines.As a leading organization providing a unique mix of standardization and conformity assessment capabilities for industrial and information technology systems,the IEC is ideally positioned to address some of these challenges at the international level.Executive summary4The following specific recommendations targeted at the IEC and its committees are provided in the last part of the White paper:Promote the central role of JTC 1/SC 42 in horizontal artificial intelligence standardization.Coordinate the standardization of data semantics and ontologies.Develop and centralize artificial intelligence-related use cases.Develop an artificial intelligence reference architecture with consistent interfaces.Explore the potential for artificial intelligence conformity assessment needs.Foster a dialogue with various societal stakeholders concerning artificial intelligence.Include artificial intelligence use cases in testbeds involving the IEC.As it is foreseen that artificial intelligence will become a core technology across many different industries and one of the driving forces of the coming fourth industrial revolution,the standardization community will play a critical role in shaping its future.Building upon its long track record in safety and reliability,the IEC can be instrumental in achieving this goal and fulfilling the promise of artificial intelligence as a benefit to humanity.AcknowledgmentsThis White Paper has been prepared by the artificial intelligence project team in the IEC Market Strategy Board(MSB),with major contributions from the project partner,the German Research Centre for Artificial Intelligence(DFKI),and the project leader,Haier Group.The project team was directed by Mr Ye Wang,Vice President of Haier Home Appliance Industry Group and an MSB member.The project team is listed below:Mr Jens Popper,DFKI/SmartFactoryKL,Project Partner LeadMr Jesko Hermann,DFKI/SmartFactoryKL Mr Kai Cui,Haier,Project ManagerMr Simon Bergweiler,DFKIDr Stephan Weyer,DFKIProf.Dr-Ing.Martin Ruskowski,DFKIMr Miao Wang,HaierMr Liang Guang,HuaweiMr Yun Chao Hu,HuaweiDr Victor Kueh,HuaweiMr Di Wang,HuaweiMs Mary Carol Madigan,SAPDr Ian Oppermann,NSW Data AnalyticsMr Sung-Min Ryu,KEPCOMr Lagyoung Kim,LG ElectronicsDr Sha Wei,CESIMr Ruiqi Li,CESIMr Xiaohui Du,ITEIMr Yujia Shang,ITEIMr Xiangqian Ding,OUCMr Guangrui Zhang,OUCDr Gilles Thonet,IECExecutive summary5Table of contentsExecutive summary 3List of abbreviations 9Glossary 11Section 1 Introduction 131.1 Artificial intelligence:miracle or mirage?131.2 From winter to rebirth of artificial intelligence 141.3 Great opportunities come with risks and challenges 151.4 Definitions of artificial intelligence 161.5 Scope of the White Paper 171.6 Outline of the White Paper 18Section 2 Need for artificial intelligence 212.1 Scarcity of natural resources 212.2 Climate change 222.3 Demographic trends 222.4 Economic policy 242.5 Service and product customization 24Section 3 Enablers and drivers of artificial intelligence 273.1 Enablers of artificial intelligence 273.1.1 Increased computational power 283.1.2 Availability of data 293.1.3 Improved algorithms 293.2 Drivers of artificial intelligence 303.2.1 Cloud and edge computing 303.2.2 Internet of Things 313.2.3 Big data 313.2.4 Consumer acceptance 326Table of contentsSection 4 Inside artificial intelligence 354.1 Categories of machine learning 354.1.1 Supervised learning 354.1.2 Unsupervised learning 364.1.3 Reinforcement learning 364.2 Current machine learning systems 364.2.1 Computer vision 364.2.2 Anomaly detection 374.2.3 Time series analysis 374.2.4 Natural language processing 374.2.5 Recommender systems 384.3 Algorithms for machine learning 384.3.1 Decision trees 384.3.2 Support vector machines 384.3.3 Nave Bayes 404.3.4 k-nearest neighbour 404.3.5 k-means 414.3.6 Hidden Markov model 424.3.7 Artificial neural networks 424.3.8 Convolutional neural networks 434.3.9 Recurrent neural networks 43Section 5 Deployment of artificial intelligence 455.1 Artificial intelligence in smart homes 465.1.1 Smart television control system 465.1.2 Bathroom self-service system 475.1.3 Intelligent food identification system 475.1.4 Challenges for smart homes 485.2 Artificial intelligence in smart manufacturing 485.2.1 Predictive maintenance 495.2.2 Collaborative robots 515.2.3 Quality control 515.2.4 Challenges in smart manufacturing 5375.3 Artificial intelligence in smart transportation and the automotive sector 535.3.1 Autonomous driving 535.3.2 Traffic management 545.3.3 Traffic robots 565.3.4 Challenges in smart transportation 565.4 Artificial intelligence in smart energy 565.4.1 Grid management and operations 575.4.2 Consumer engagement and services 575.4.3 Integrated smart energy platforms 575.4.4 Challenges in smart energy 58Section 6 Artificial intelligence challenges 616.1 Social and economic challenges 616.1.1 Changes in decision-making 616.1.2 Advanced supply chain operations 616.2 Data-related challenges 626.2.1 Selection of training data 626.2.2 Standardized data 636.3 Algorithm-related challenges 646.3.1 Robustness 646.3.2 Transfer learning 656.3.3 Interpretability 656.3.4 Objective functions 666.4 Infrastructure-related challenges 666.4.1 Hardware bottlenecks 666.4.2 Lack of platforms and frameworks 676.5 Trustworthiness-related challenges 676.5.1 Trust 676.5.2 Privacy 686.5.3 Security 686.6 Regulatory-related challenges 686.6.1 Liability 686.6.2 Privacy 696.6.3 Ethics 69Table of contents8Section 7 Standardization gaps in artificial intelligence 717.1 Standardization activities in artificial intelligence 717.1.1 ISO/IEC JTC 1 727.1.2 IEC 737.1.3 ISO 737.1.4 ITU 737.1.5 IEEE 747.1.6 ETSI 747.1.7 Standardization activities in China 757.1.8 Standardization activities in the United States 757.1.9 European AI Alliance 767.1.10 Consortia and other organizations 767.2 Standardization gaps 777.2.1 Harmonized data models and semantics 777.2.2 Common ontology based on data models 777.2.3 Verification of artificial intelligence algorithms 777.2.4 Benchmarking and evaluation of artificial intelligence infrastructures 78Section 8 Conclusions and recommendations 798.1 Industry recommendations 798.2 Regulatory recommendations 808.3 Recommendations addressed to the IEC and its committees 81Annex A Future developments 83A.1 Biology-inspired artificial intelligence 83A.2 Human/artificial intelligence interaction 83A.3 Artificial intelligence-enabled digital twin 84A.4 Automated machine learning 84Bibliography 87Table of contents9 AI artificial intelligence AIR automated image recognition AMI advanced metering infrastructure ANN artificial neural network API application programming interface ASIC application-specific integrated circuit CNN convolutional neural network CART classification and regression tree CPU central processing unit DNN deep neural network EBL explanation-based learning FPGA field-programmable gate array GDPR(EU)General Data Protection Regulation GPU graphics processing unit GRU gated recurrent unit HMM hidden Markov model HTM hierarchical temporal memory ICT information and communication technology ID3 Iterative Dichotomiser 3 IoT Internet of Things IT information technology k-NN k-nearest neighbour KPI key performance indicator LSTM long-short-term memory NLP natural language processing NPU neural processing unit RDBMS relational database management systemList of abbreviationsTechnical andscientific terms10 ReLu rectified linear unit RNN recurrent neural network SME small-to-medium enterprise SVM support vector machine TPU tensor processing unit CESI China Electronics Standardization Institute DFKI German Research Center for Artificial Intelligence EC European Commission ENI(ETSI ISG)Experiential Networked Intelligence ETSI European Telecommunications Standards Institute FG-ML5G (ITU-T)Focus Group on machine learning for future networks including 5G IDC International Data Corporation IEEE Institute of Electrical and Electronics Engineers ISG(ETSI)Industry Specification Group ISO International Organization for Standardization ITEI Instrumentation Technology and Economy Institute(China)ITU International Telecommunication Union ITU-T(ITU)Telecommunication Standardization Sector JTC Joint Technical Committee KEPCO Korea Electric Power Corporation NHTSA National Highway Traffic Safety Administration(US)OUC Ocean University of China SAC Standardization Administration of China UN United NationsOrganizations,institutions and companiesList of abbreviations11GlossaryApplication programming interfaceAPIinterface constituted of clearly defined methods of communication between various software componentsApplication-specific integrated circuitASICan electronic circuit specialized to perform a specific set of operations for a specific purpose NOTE The application field cannot be changed since it is defined through its architecture.Artificial intelligenceAIa branch in computer science that simulates intelligent behaviour in computers including problem solving,learning and pattern recognitionArtificial neural networkANNa mathematical construct inspired by biological neural networks that are often used in computer science to perform tasks by giving them training examples without being explicitly programmed to do soCentral processing unitCPUan electronic circuit that performs instructions of a computer programmeConvolutional neural networkCNNa special feed-forward network that is usually applied for tasks such as image recognitionDeep learninga field of machine learning using deep neuronal networksDeep neural networkDNNan artificial neural network that has several consecutive layers of neurons that are connected in order to process an input to an outputExplanation-based learningEBLa form of artificial intelligence that uses domain theory to generalize from training examplesField-programmable gate arrayFPGAan electronic circuit that performs specifically for different applications NOTE In contrast to application-specific integrated circuits,FPGA can be reprogrammed after manufacturing.General Data Protection RegulationGDPRa set of significant regulatory changes to data protection and privacy in the European Union,which also addresses automated decision-making by artificial intelligence systemsGraphics processing unitGPUan electric circuit that is specialized to process images by performing massive amounts of calculations in parallelHidden Markov modelHMMa probabilistic model of linear sequences that can be described using the Markov process NOTE Hidden Markov model is a technique used in machine learning with the assumption that not all states of the described processes can be directly observed and thus are hidden.12Internet of ThingsIoTnetwork of physical devices,embedded electronics or software that enables these components to be connected with a larger network to exchange dataMachine learninga category of algorithms in computer science enabling a device to improve its performance of a specific task with increasing data and without being explicitly programmed to do soNatural language processingNLPan area of computer science dealing with how computers can process natural language for speech recognition,language understanding or language generationNeural processing unitNPUan electric circuit that is not based on a von-Neumann or Harvard architecture,but on the principle of neuromorphingRectified linear unitReLuan activation function of a neuron which consists of two linear partsRecurrent neural networkRNNa class of neural network in which the connections between the neu