idnits 2.17.1 draft-choi-icnrg-aiot-06.txt: Checking boilerplate required by RFC 5378 and the IETF Trust (see https://trustee.ietf.org/license-info): ---------------------------------------------------------------------------- No issues found here. Checking nits according to https://www.ietf.org/id-info/1id-guidelines.txt: ---------------------------------------------------------------------------- == The page length should not exceed 58 lines per page, but there was 1 longer page, the longest (page 1) being 60 lines Checking nits according to https://www.ietf.org/id-info/checklist : ---------------------------------------------------------------------------- ** The document seems to lack a Security Considerations section. ** There are 534 instances of too long lines in the document, the longest one being 5 characters in excess of 72. Miscellaneous warnings: ---------------------------------------------------------------------------- == The copyright year in the IETF Trust and authors Copyright Line does not match the current year == Line 815 has weird spacing: '... of unsup...' -- The document date (May 23, 2021) is 1068 days in the past. Is this intentional? Checking references for intended status: Informational ---------------------------------------------------------------------------- == Missing Reference: 'Kriegal' is mentioned on line 725, but not defined == Missing Reference: 'Hochreiter' is mentioned on line 792, but not defined == Missing Reference: 'Krizhevsky' is mentioned on line 811, but not defined == Missing Reference: 'Goodfellow' is mentioned on line 830, but not defined == Missing Reference: 'Venkataramani' is mentioned on line 870, but not defined == Unused Reference: 'Kriegel' is defined on line 1059, but no explicit reference was found in the text Summary: 2 errors (**), 0 flaws (~~), 9 warnings (==), 1 comment (--). Run idnits with the --verbose option for more detailed information about the items above. -------------------------------------------------------------------------------- 2 ICN Research Group J.K. Choi 3 Internet Draft J.S. Han 4 Intended status: Informational G.H. Lee 5 Expires: November 23, 2021 N.K. Kim 6 KAIST 7 May 23, 2021 9 Requirements and Challenges for User-level Service Managements of 10 IoT Network by utilizing Artificial Intelligence 11 draft-choi-icnrg-aiot-06 13 Abstract 15 This document describes the requirements and challenges to employ 16 artificial intelligence (AI) into the constraint Internet of Things 17 (IoT) service environment for embedding intelligence and increasing 18 efficiency. 20 The IoT service environment includes heterogeneous and multiple IoT 21 devices and systems that work together in a cooperative and 22 intelligent way to manage homes, buildings, and complex autonomous 23 systems. Therefore, it is becoming very essential to integrate IoT 24 and AI technologies to increase the synergy between them. However, 25 there are several limitations to achieve AI enabled IoT as the 26 availability of IoT devices is not always high, and IoT networks 27 cannot guarantee a certain level of performance in real-time 28 applications due to resource constraints. 30 This document intends to present a right direction to empower AI in 31 IoT for learning and analyzing the usage behaviors of IoT 32 devices/systems and human behaviors based on previous records and 33 experiences. With AI enabled IoT, the IoT service environment can be 34 intelligently managed in order to compensate for the unexpected 35 performance degradation often caused by abnormal situations. 37 Status of this Memo 39 This Internet-Draft is submitted in full conformance with the 40 provisions of BCP 78 and BCP 79. 42 Internet-Drafts are working documents of the Internet Engineering 43 Task Force (IETF), its areas, and its working groups. Note that 44 other groups may also distribute working documents as Internet-Drafts. 46 Internet-Drafts are draft documents valid for a maximum of six months 47 and may be updated, replaced, or obsoleted by other documents at any 48 time. It is inappropriate to use Internet-Drafts as reference 49 material or to cite them other than as "work in progress." 51 The list of current Internet-Drafts can be accessed at 52 http://www.ietf.org/ietf/1id-abstracts.txt 54 The list of Internet-Draft Shadow Directories can be accessed at 55 http://www.ietf.org/shadow.html 57 This Internet-Draft will expire on November 23, 2021. 59 Copyright Notice 61 Copyright (c) 2021 IETF Trust and the persons identified as the 62 document authors. All rights reserved. 64 This document is subject to BCP 78 and the IETF Trust's Legal 65 Provisions Relating to IETF Documents 66 (http://trustee.ietf.org/license-info) in effect on the date of 67 publication of this document. Please review these documents carefully, 68 as they describe your rights and restrictions with respect to this 69 document. Code Components extracted from this document must include 70 Simplified BSD License text as described in Section 4.e of the Trust 71 Legal Provisions and are provided without warranty as described in 72 the Simplified BSD License. 74 Table of Contents 76 1. Introduction ................................................ 3 77 2. Challenging Issues of IoT network 78 ............................ 6 79 2.1. Untrusted and incorrect IoT devices ..................... 6 80 2.2. Traffic burstiness of IoT network ....................... 6 81 2.3. Management overheads of heterogeneous IoT sensors 82 ........ 7 83 3. Overview of AI/ML-based IoT services ......................... 9 84 4. Requirements for AI/ML-based IoT services ................... 11 85 4.1. Requirements for AI/ML-based IoT data collection and delivery 87 ........................................................... 12 88 4.2. Requirements for intelligent and context-aware IoT services13 89 4.3. Requirements for applying AI/ML to IoT data ............ 15 90 4.3.1. Training AI/ML algorithm 91 .......................... 15 92 4.3.2. AI/ML inference in IoT application ................ 15 93 4.3.3. AI/ML models update in IoT application ............ 16 95 5. State of arts of the artificial intelligence/machine learning 96 technologies for IoT services 97 .................................. 16 98 5.1. Machine learning and artificial intelligence technologies 99 review ..................................................... 16 100 5.1.1. Supervised learning for IoT ....................... 17 101 5.1.2. Unsupervised learning for IoT ..................... 18 102 5.1.3. Reinforcement learning for IoT .................... 19 103 5.1.4. Neural Network based algorithms for IoT ........... 20 104 5.2. Technologies for lightweight and real-time intelligence 105 . 21 106 6. Use cases of AI/ML into IoT service ......................... 23 107 6.1. Surveillance and Security in Smart Home ................ 23 108 6.1.1. Characteristics of Smart Home for AI/ML processing 109 . 23 110 6.1.2. Use case ......................................... 24 111 6.2. Smart Building Management 112 .............................. 24 113 6.2.1. Characteristics of Smart Building for AI/ML processing24 114 6.2.2. Use case ......................................... 25 115 6.3. Manufacturing optimization in Smart factory ............ 25 116 6.3.1. Characteristics of Smart Building for AI/ML processing25 117 6.3.2. Use case ......................................... 25 118 7. IANA Considerations ........................................ 25 119 8. Acknowledgements ........................................... 26 120 9. Contributors ............................................... 26 121 10. Informative References 122 ..................................... 26 124 1. Introduction 126 The document explains the effects of applying artificial intelligence 127 /machine learning (AI/ML) algorithms in the Internet of Thing (IoT) s 128 ervice environments. 130 IoT applications will be deployed in heterogeneous and different area 131 s such as the energy, transportation, automation and manufacturing in 132 dustries as well as the information and communication technology (IC 133 T) industry. Many IoT sensors and devices can connect to an IoT servi 134 ce environment where IoT objects cannot interoperate with each other 135 and can interact with different applications. The IoT service may not 136 run in a single administrative domain. If market demand exists, the 137 cross-domain service scenarios for IoT applications could be widely d 138 eployed. Future IoT applications occur at multiple domains of heterog 139 eneity with various time scales. 141 The IoT service requirements for common architectures and public APIs 142 poses some challenges to the underlying service environment and netw 144 orking technologies. Some IoT applications require significant securi 145 ty and privacy as well as significant resource and time constraints. 146 These mission-critical applications can be separated from many common 147 IoT applications that current technology may not provide. It means t 148 hat IoT service requirements are difficult to classify common require 149 ments and functional requirements depending on IoT service scenario. 151 Recently, artificial intelligence technologies can help the context-a 152 ware IoT service scenarios apply rule-based knowledge accumulation. T 153 he IoT service assumes that many sensing devices are connected to sin 154 gle or multiple IoT network domains. Each sensor sends small packets 155 to the IoT servers periodically or non-periodically. Detection data c 156 ontains periodic status information that monitors whether the system 157 is in a normal state or not. In some cases, alert information is incl 158 uded for quick processing. Most IoT applications can operate in two m 159 odes. One is a simple monitoring mode and the other is an abnormal mo 160 de for rapid processing. In a simple monitoring phase, the IoT device 161 periodically sends sensing data to the server. If the measured data 162 is outside the normal range, the IoT service can change the operating 163 mode to an abnormal phase and activate future probes. Alarm conditio 164 ns should be promptly notified to responsible persons. For mission-cr 165 itical applications, reliable communication with robust QoS requireme 166 nts in terms of error and latency is required. 168 Periodic data accumulation from IoT devices is cumbersome. Under norm 169 al conditions, the IoT data is simply accumulated without further act 170 ion. In an unusual situation, incoming IoT data can cause an urgent a 171 ction to notify the administrator of the problem. Streaming data traf 172 fic from thousands of IoT devices is annoying to store in the databas 173 e because it is not easy to extract unidentified or future incidents. 174 Only a significant portion of the incoming data stream can be stored 175 in a real-time database that is time-sensitive and capable of rapid 176 query processing. A combination of different IoT detection data, incl 177 uding location, time, and status, allows you to sort and categorize a 178 portion of streaming data when an additional inspection is required, 179 and perform real-time processing. One of the missions of the IoT dat 180 abase is to be able to extract preliminary symptoms of unexpected acc 181 idents from a large amount of streaming data. 183 If some transmitted data is important to invoke the corresponding act 184 ion, there are some questions about whether the incoming data is corr 185 ect. If the incoming data contains accurate and time-critical events, 186 appropriate real-time control and management can be performed. Howev 187 er, if the incoming data is inaccurate or intentionally corrupted, ad 188 ditional accidents may occur. In these cases, incoming data can trigg 189 er to initiate additional inspections to protect against future unacc 190 eptable situations. But, if time-critical data is missed due to error 191 s in the sensing devices and the delivery protocol, there is no reaso 192 n to configure IoT networks and devices at a high cost. 194 It is not easy to analyze data collected through IoT devices installe 195 d to monitor complex IoT service environments. If the sensor malfunct 196 ions, the data of the sensor cannot be trusted. Additional investigat 197 ion should be done if abnormal status from specific sensors is collec 198 ted. The data of the redundant sensor installed in the same area shou 199 ld be received or combined with other sensor information adjacent to 200 the sensor to determine the abnormal state. 202 For sensors installed in a specific area, sensing records will remain 203 for a certain period of time. IoT service operators can look at the 204 operational history of the sensor for a period of time to determine w 205 hat problems were encountered when data was collected. When an abnorm 206 al situation occurs, IoT sensor should investigate whether it noticed 207 normal operations and notified the IoT service operator. If the abno 208 rmal situation is not properly detected, the operator should analyze 209 whether it was caused by malfunction of the IoT sensor or other reaso 210 ns. 212 In the IoT service environment, it is possible to analyze the situati 213 on accurately by applying recent artificial intelligence and machine 214 learning technologies. If there is an operational record of the past, 215 it is possible to determine when an abnormal situation arises. Most 216 problems are likely to be repeated, so if the past learning experienc 217 e is accumulated, the anomaly of IoT services can be easily and immed 218 iately identified. In addition, when information gathered from variou 219 s sensors is synthesized, it is possible to accurately determine whet 220 her abnormal situations have occurred. 222 Various types of IoT sensors are installed with certain purposes. It 223 expects that all the IoT sensors intend to monitor the occurrence of 224 special abnormal situations in advance. Therefore, it should be set i 225 n advance what actions are required when a specific anomaly occurs. T 226 he appropriate work is performed on the abnormal situation according 227 to the procedure, predefined by the human. By using artificial intell 228 igence and machine learning algorithms, the appropriate actions are t 229 aken when an abnormal situation is detected from various IoT sensors. 231 2. Challenging Issues of IoT network 233 This section describes the challenging issues of data sensing, 234 collection, transfer, and intelligent decision from untrusted data 235 quality and unexpected situations of IoT service environments. 237 2.1. Untrusted and incorrect IoT devices 239 IoT traffic is similar to traditional Internet traffic with small 240 packet sizes. Mobile IoT traffic can cause some errors and delays 241 because wireless links are unstable and signal strength may be 242 degraded with device mobility. If the signal strength of the IoT 243 device with a power limit is not so strong, the reception quality of 244 the IoT server may not be sufficient to obtain the measurement data. 246 For mission-critical applications, such as smart-grid and factory- 247 automation, expensive IoT sensors with self-rechargeable batteries 248 and redundant hardware logic may be required. However, unexpected 249 abnormal situations may occur due to sensor malfunctions. There are 250 trade-offs between implementation cost and efficiency for cost- 251 effective IoT services. When smart-grid and factory-automation 252 applications are equipped with IoT devices, the acceptable quality 253 from IoT solutions can be required. Sometimes, expensive and 254 duplicated IoT solutions may be needed. 256 2.2. Traffic burstiness of IoT network 258 IoT traffic includes two types of traffic characteristic: periodic 259 with small packet sizes and bursty with high bandwidth. Under normal 260 conditions, the IoT traffic periodically transmits status information 261 with a small bandwidth, several kilobits/sec. However, in an abnormal 262 state, IoT devices need a high bandwidth, up to several tens of 263 megabits/sec, in order to identify actual events and investigate 264 accurate status information. In addition, traffic volume can 265 explosively increase in response to emergencies. For example, in the 266 case of smart-grid application, the bandwidth of several kilobits/sec 267 is usually used, and when an urgent situation occurs, a broadband 268 channel is required up to several tens of megabits/sec. 270 The other traffic can be integrated at an IoT network to increase 271 bandwidth efficiency. If an emergency situation occurs in the IoT 272 service, IoT traffic volumes suddenly increase, in which case network 273 processing capacity may be not sufficient. If the IoT service is 274 integrated with voice and video applications, the problem can become 275 more complex. As time goes by, traffic congestion and bottlenecks are 276 frequent in some areas. In addition, if an existing service policy 277 changes (for example, prioritizing certain traffic or suddenly 278 changing the route), other unexpected problems may be encountered. 279 Various congestion control and load balancing algorithms with the 280 help of artificial intelligence can be applied to handle time-varying 281 traffic on a network. 283 Until now, much research has been done on traffic variability in an 284 integrated network service environment. All networks have their own 285 traffic characteristics, depending on geographical area, number of 286 subscribers, subscribers' preferences, and types of applications used. 287 In the case of IoT traffic, the normal bandwidth is very small. If 288 the IoT traffic volume increases abruptly in an abnormal situation, 289 the network may suffer unacceptable delay and loss. If emergency 290 situations detected by IoT networks occur in a smart grid or 291 intelligent transportation system, the processing power of the IoT 292 network alone cannot solve the problem and the help of existing 293 network resources is inevitable. 295 2.3. Management overheads of heterogeneous IoT sensors 297 Traffic management in an integrated network environment is not easy. 298 In order to operate the network steadily, a network operator has its 299 own know-hows and experiences. If there are plenty of network 300 resources, it is easy to set up a bypass route even if network 301 failure or congestion occurs in a specific area. For operating 302 network steadily, network resources may be designed to be over- 303 provisioned in order to cope with various possible outages. A network 304 operator predicts the amount of traffic generated by the 305 corresponding equipment and grasps to what extent a transmission 306 bandwidth is required. If traffic fluctuation is very severe, the 307 network operator can allocate network resources in advance. In case 308 of frequent failures or severe traffic fluctuation, some network 309 resources are separated in order not to affect normal traffic. 311 More than a billion IoT devices are expected to connect to 312 smartphones, tablets, wearables, and vehicles. Therefore, IoT 313 services are targeted at mobile applications. In particular, 314 intelligent transportation systems need the help of IoT technology to 315 provide traffic monitoring and prevent public or private traffic 316 accidents. IoT technology can play an important role in reducing 317 traffic congestion, saving people's travel time and costs, and 318 providing a pleasant journey. 320 The IoT service has troublesome administrative problems to configure 321 an IoT network which consists of IoT servers, gateways, and many 322 sensing devices. The small-sized but large-numbered IoT devices may 323 incur administrative overhead since all the IoT devices should be 324 initialized and the bootstrapping information of IoT resources should 325 be loaded into the IoT service environments. Whenever some IoT 326 devices are newly added and some devices have to be removed, the 327 dynamic reconfiguration of IoT resources is essential. In addition, 328 the IoT device's preinstalled software should be regularly inspected 329 and upgraded according to its version. Frequent upgrades and changes 330 to some IoT devices may require autonomic management and 331 bootstrapping techniques. 333 Network management generally assumes that all network resources 334 operate reliably with acceptable quality. In most failure situations, 335 the network operator decides to switch to a redundant backup device 336 or bypass the failed communication path. If some IoT devices are not 337 stable, duplicate IoT devices can be installed for the same purpose. 338 If IoT resources are not duplicated, various mechanisms are needed to 339 reduce the damage. Therefore, it is necessary to prioritize the 340 management tasks to be performed first when an abnormality occurs in 341 the IoT service environment. However, managing duplicate networks can 342 cause another problem. If two IoT devices are running at the same 343 time, the recipient can get redundant information. If two or more 344 unusual situations occur at the same time, it is difficult to solve 345 the problem since tasks for urgent processing should be distinguished 346 from tasks that can be performed over time. 348 In addition, the operations manager's mistakes or misunderstanding of 349 problem situations can lead to other unexpected complications. 350 Therefore, artificial intelligence technologies can help what kind of 351 network management work is required when an unexpected complicated 352 situation occurs even though a procedure for an abnormal situation is 353 already prepared. 355 3. Overview of AI/ML-based IoT services 357 In this section, successful applications of artificial intelligence i 358 n IoT domains are provided. The common property of IoT applications a 359 nd services is that they require fast analytics rather than later ana 360 lytics with piled data. Recently, neural-network-based artificial int 361 elligence technologies are widely used across many IoT applications. 363 Simple IoT applications include dynamic contexts that share common fe 364 atures among social relations at the same administration domain. IoT 365 devices in the same domain can provide their service contexts to the 366 IoT server. When a dynamic change occurs in an IoT service context, t 367 he IoT device needs real-time processing to activate urgent events, a 368 lert notifications, update, and reconnect contexts. The IoT service m 369 ust support real-time interactions between the IoT device and the sys 370 tem in the same domain. The IoT service contexts must be shared betwe 371 en physical objects and social members in the same domain as well. 373 Artificial intelligence technologies have been shown promising in man 374 y areas, including IoT. For example, contextual information for a car 375 -sharing business must interact with customers, car owners, and car s 376 haring providers. All entities in the value chain of a car sharing bu 377 siness must share the corresponding situation to pick up, board, and 378 return shared cars. Communication networks and interactive informatio 379 n, including registration and payment, can be shared tightly among th 380 e entities. Home IoT service environment can be equipped with sensors 381 for theft detection, door lock, temperature, fire detection, gas det 382 ection, short circuit, air condition to name a few. Office IoT servic 383 e environments, including buildings such as shopping centers and bus/ 384 airport terminals, have their own sensors, including alarm sensors. W 385 hen an alarm signal is detected by the sensor, the physical position 386 and occurrence time of the sensor is determined in advance. All signa 387 ls from various sensors are analyzed comprehensively to make the righ 388 t decision. If some sensors frequently malfunction, the situation can 389 be grasped more accurately by analyzing the information of the adjac 390 ent sensor. In particular, when installing multiple sensors in a part 391 icular building (e.g., surveillance camera, location monitoring, temp 392 erature, etc.), a much wider range of sensors can be used when utiliz 393 ing artificial intelligence and machine learning technologies. 395 (Smart home) Smart home concept span over multiple IoT applications, 396 health, energy, entertainment, education, etc. It involves voice reco 397 gnition, natural language processing, image-based object recognition, 398 appliance management, and many more artificial intelligence technolo 399 gies integrated with IoT. Smart connected-devices monitor the house t 400 o provide better control over home supplies and expenses. The energy 401 consumption and efficiency of home appliances are monitored and analy 402 zed with deep learning based technologies, such as artificial neural 403 network, long-short-term-memory, etc. 405 (Smart city) Smart city, as well, contains multiple IoT domains, tran 406 sportation, infrastructure, energy, agriculture, etc. Since heterogen 407 eous data from different domains are gathered in smart cities, variou 408 s artificial intelligence approaches are studied in smart-city applic 409 ation. Public transportation behaviors and crowd movements patterns a 410 re important issues, and they are often dealt with neural network bas 411 ed methods, long-short-term-memory and convolutional neural network. 413 (Smart energy) As two-way communication energy infrastructure is depl 414 oyed, smart grid has become a big IoT application, which requires int 415 elligent data processing. The traditional energy providers are highly 416 interested in recognizing local energy consumption patterns and fore 417 casting the needs in order to make appropriate decisions on real-time. 419 Moreover, the energy consumers, as well, want analyzed information o 420 n their own energy consumption behaviors. Recently, many works on ene 421 rgy consumption prediction, energy flexibility analysis, etc. are act 422 ively ongoing. Most works are based on the latest deep learning techn 423 ologies, such as multi-layered-perceptron, recurrent neural network, 424 long-short-term-memory, autoencoder, etc. 426 (Smart transportation) The intelligent transportation system is anoth 427 er source of big data in IoT domains. Many use cases, such as traffic 428 flow and congestion prediction, traffic sign recognition, vehicle in 429 trusion detection, etc., have been studied. Moreover, a lot of advanc 430 ed artificial intelligence technologies are required in autonomous an 431 d smart vehicles, which require many intelligent sub-tasks, such as p 432 edestrian's detection, obstacle avoidance, etc. 434 (Smart healthcare) IoT and artificial intelligence are integrated int 435 o the healthcare and wellbeing domain as well. By analyzing food imag 436 es with convolutional neural network on mobile devices, dietary intak 437 es can be measured. With voice signal captured from sensor devices, v 438 oice pathologies can be detected. Moreover, recurrent neural network 439 and long-short-term-memory technologies are actively being studied fo 440 r early diagnosis and prediction of diseases with time series medical 441 data. 443 (Smart agriculture) To manage a vast area of land, IoT and artificial 444 intelligence technologies are recently used in agriculture domains. 445 Deep neural network and convolutional neural network are utilized for 446 crop detection or classification and disease recognition in the plan 447 ts. Moreover, for automatic farming with autonomous machine operation, 448 obstacle avoidance, fruit location, and many more sub-tasks are hand 449 led with advanced artificial intelligence technologies. 451 4. Requirements for AI/ML-based IoT services 453 In this section, the requirements for AI/ML-based IoT data collection 454 and delivery, intelligent and context-aware IoT services, and applyi 455 ng AI/ML to IoT data will be described. 457 4.1. Requirements for AI/ML-based IoT data collection and delivery 459 IoT services store a vast amount of data that IoT devices 460 periodically generate, and the refining and analyzing are costly. 461 Effective analysis of IoT data has been considered to be the most 462 important factor in data processing, but the analysis of efficient 463 data collection and delivery methods are becoming other significant 464 factors as the amount of the data collected is explosively increasing. 466 In particular, as a number of IoT devices have been deployed within 467 the IoT network, controlling data collection and delivery for each of 468 them has become impossible. The introduction of AI/ML techniques for 469 simultaneous and efficient management of the IoT devices should be 470 considered as a countermeasure. For IoT data collection and delivery, 471 the following two factors will need to be considered, IoT devices 472 energy and data quality. 474 (IoT Device Energy) As many IoT devices have begun to be deployed 475 within the IoT network, it is impossible to deliver energy to many 476 IoT devices simultaneously. Consequently, the efficient battery use 477 has become an important issue. 479 If IoT data collection and delivery periods are too short, a lifetime 480 of the IoT device will be shortened through the reckless use of IoT 481 device energy. Thereby, it increases the cost required to provide IoT 482 service. On the other hand, if IoT data collection and delivery 483 period are too long, the quality of the IoT services provided will be 484 reduced due to the lack of details in the data for situation 485 recognition and real-time processing. Therefore, taking into account 486 the energy consumption of the IoT devices, research on proper IoT 487 data collection and delivery period is necessary. 489 (Data Quality) Since the data collected from the majority of IoT 490 devices usually contain redundant information, it causes additional 491 costs for the data collection and refinement processes. Therefore, it 492 will be necessary to select and deliver meaningful information from 493 redundant IoT data to reduce unnecessary cost on the IoT network. To 494 do so, it will need the research to identify the relationships among 495 the data collected various devices and interpret the information that 496 the data contains. 498 (Optimal IoT Device Operation) Each IoT service may have its own 499 defined information requirements and it is imperative that the data 500 with certain quality level should be collected to extract the core 501 information for each situation. As scheme of data collection 502 considerably affects the data quality, the operation of IoT devices 503 should be effectively adjusted to ensure the data quality. In general, 504 high resolution of data collected from IoT devices as well as data 505 preprocessing are required for high quality data. However, these 506 methods leading tremendous operation cost of IoT devices causes a 507 lifetime of IoT networks, degrading the QoS of IoT services. 508 Eventually, an optimal IoT device operation technique for each IoT 509 service must be considered. 511 4.2. Requirements for intelligent and context-aware IoT services 513 In a context-aware IoT service environment, it is important to 514 establish a context to be aware of in advance since IoT devices will 515 be deployed according to a pre-designed architecture and to check how 516 characteristics of IoT data and data-to-data characteristics are 517 expressed under these circumstances. For the data produced by IoT 518 devices, since it contains the device's relative location information, 519 sensing value over time, event information, it should be reviewed to 520 provide the target context-aware service using this information. Some 521 of the necessary technologies will be described in the following. 523 (Physical Clustering) To increase the accuracy of context-awareness, 524 the provision of context-aware services should be considered in a 525 situation where the relationship between IoT devices with respect to 526 physical layout or physical environment is taken into account. 527 Setting a rule using the service provider's domain knowledge may be 528 possible, but introducing the physical clustering into a diverse IoT 529 environment (e.g., in bedroom, kitchen, balcony, or a space connected 530 through an open door) will require identifying the physical 531 relationship between the devices using data generated from IoT 532 devices. 534 (Extra Data Processing) In order to prevent degradation of service 535 quality from errors in data values or device malfunctions, extra 536 sensors should be placed in the majority of IoT environments. In a 537 context-aware service, they contain the same information, so the 538 technologies filtering the data that contains only essential part 539 among the same information while preventing data errors would be 540 required. 542 (Unreported data handling) If an event is detected on a particular 543 IoT device, it will transmit data regardless of the device's sensing 544 and delivery interval. At this time, the data of IoT devices which 545 are physically clustered are needed to accurately detect events that 546 occurred, and it is difficult to expect that these devices will 547 provide data at the same time. A gateway can request data from 548 clustered devices, but it has a problem for real-time processing for 549 emergency situations. Therefore, handling unreported data will be 550 required based on previously collected data. 552 (Abnormal data in AI/ML) In the case of context-aware services that 553 operates based on the predetermined rule, the flexibility to cope 554 with emergency situations that have not been considered is low, and 555 thus AI/ML algorithms are required to intelligently cope with a 556 myriad of situations. However, many abnormal data are generated 557 depending on environmental conditions such as device status, so AI/ML 558 algorithms that can operate in that imperfect environment should be 559 considered. 561 (Edge computing in IoT) There are two necessary prerequisites 562 required in context-aware IoT services: IoT devices real-time 563 management and IoT network architecture supporting the high-volume 564 data transmission. When an abnormal situation is discovered, high- 565 volume data should be utilized to adequately to monitor the situation 566 through the IoT device's real-time management. As contrasted with 567 conventional cloud computing structures, an edge computing structure, 568 where IoT data processing servers are located in closer proximity to 569 IoT devices, provides higher energy efficiency and lesser 570 communication delay. 572 4.3. Requirements for applying AI/ML to IoT data 574 In this subsection, the requirements for applying AI/ML to IoT data 575 are described. 577 4.3.1. Training AI/ML algorithm 579 To use AI/ML algorithm, two elements are required, AI/ML model and 580 training data. The presence of training dataset in good quality is an 581 important factor of the AI/ML model performance since the model is 582 iteratively trained with the training data. However, for anomaly 583 detection, there is not enough training data since not only the 584 probability of anomaly occurrence is very low but also it is almost 585 impossible to retrieve the ground truth value even when the situation 586 has occurred. Therefore, using domain knowledge, AI/ML learning based 587 on abnormal situation data generation or simulation should be 588 considered. For example, for an external intrusion detection 589 application within a smart home, when a camera and a motion sensor 590 detect an intruder, a light sensor checks the measuring value. If the 591 light does not turn on, then the IoT application recognizes it as an 592 abnormal situation. In this way, by using the domain knowledge, the 593 rule regarding the operational scenario of the IoT application is 594 generated as the training data, and the generated training data can 595 be used for model learning. This will not only enable learning the 596 anomaly detection algorithm in IoT application but also improving the 597 accuracy. Therefore, IoT application, in which it is difficult to 598 acquire dataset in good quality, will require data generation based 599 on domain knowledge for AI/ML. 601 4.3.2. AI/ML inference in IoT application 603 In order for AI ML technology to be applied to IoT applications, the 604 training data and the input data for model testing and inferencing 605 must have the same characteristics such as dimension, time interval, 606 types of features, etc. However, due to the volatile IoT data 607 characteristics that vary from situations in many IoT applications, 608 it is difficult to directly apply the AI/ML algorithms. For example, 609 in a simple monitoring phase, the IoT devices periodically send 610 sensing data, and AI/ML have no difficulty in operating. However, in 611 an abnormal mode, the IoT applications require a fast response, and 612 IoT devices transmit data at shorter intervals than normal, which 613 changes the characteristics of the data being input to the AI/ML 614 algorithm. Therefore, data preprocessing technology handling the 615 abnormal data will be required in advance, such as data imputation, 616 correcting data anomalies, and Interpolation of unreported data. 618 4.3.3. AI/ML models update in IoT application 620 While IoT devices deployed in an IoT environment continuously collect 621 and transmit the necessary data to an IoT server, the IoT server 622 delivers intelligent IoT services using AI/ML models, based on these 623 collected data. As IoT devices monitor a wide variety of IoT 624 environmental information such as time-relevant and place-relevant 625 information, the data gathered from IoT devices changes significantly 626 depending on the events occurred in an IoT environment. For this 627 reason, complete information is no longer reflected during AI/ML 628 training and this, sooner or later, affects the quality of IoT 629 services provided. Therefore, AI/ML models must be updated 630 periodically upon the data subsequently collected. However, if the 631 models are updated in a too short period, not only the high 632 management (or labor) cost for updating is required, but also it may 633 detrimentally affect the overall performance of the models. Therefore, 634 models should be updated by fairly considering between the cost 635 required for updating AI/ML models and the merits gained from the 636 usage of updated AI/ML models. 638 5. State of arts of the artificial intelligence/machine learning 639 technologies for IoT services 641 In this section, well-known machine learning and artificial 642 intelligence technologies applicable to IoT applications are reviewed. 644 5.1. Machine learning and artificial intelligence technologies review 646 The classical machine learning models can be divided into three types, 647 supervised, unsupervised, and reinforcement learnings. Therefore, in 648 this subsection, machine learning and artificial intelligence 649 technology reviews are done in four different categories: supervised, 650 unsupervised, reinforcement, and neural-network-based. 652 5.1.1. Supervised learning for IoT 654 Supervised learning is a task-based type of machine learning, which 655 approximates function describing the relationship and causality 656 between input and output data. Therefore, the input data needs to be 657 clearly defined with proper output data since supervised learning 658 models learn explicitly from direct feedback. 660 (K-Nearest Neighbor) Given a new data point in K-Nearest Neighbor 661 (KNN) classifier, it is classified according to its K number of the 662 closest data points in the training set. To find the K nearest 663 neighbors of the new data point, it needs to use a distance metric 664 which can affect classifier performance, such as Euclidean, 665 Mahalanobis or Hamming. One limitation of KNN in applying for IoT 666 network is that it is unscalable to large datasets because it 667 requires the entire training dataset to classify a newly incoming 668 data. However, KNN required less processing power capability compared 669 to other complex learning methods. 671 (Naive Bayes) Given a new data point in Naive Bayes classifiers, 672 it is classified based on Bayes' theorem with the "naive" assumption 673 of independence between the features. Since Naive Bayes classifiers 674 don't need a large number of data points to be trained, they can deal 675 with high-dimensional data points. Therefore, they are fast and 676 highly scalable. However, since its "naive" assumptions are somewhat 677 strong, a certain level of prior knowledge on the dataset is required. 679 (Support Vector Machine) Support Vector Machine (SVM) is a binary and 680 non-probabilistic classifier which finds the hyperplane maximizing 681 the margin between the classes of the training dataset. SVM has been 682 the most pervasive machine learning technology until the study on 683 neural network technologies are advanced recently. However, SVM still 684 has advantages over neural network based and probabilistic approaches 685 in terms of memory usage and capability to deal with high-dimensional 686 data. In this manner, SVM can be used for IoT applications with 687 severe data storage constraint. 689 (Regression) Regression is a method for approximating the 690 relationships of the dependent variable, which is being estimated, 691 with the independent variables, which are used for the estimation. 692 Therefore, this method is widely used for forecasting and inferring 693 causal relationships between input data and output data in time- 694 sensitive IoT application. 696 (Random Forests) In random forests, instead of training a single 697 decision tree, a group of trees is trained. Each tree is trained on a 698 subset of the training set using a randomly chosen subset of M input 699 variables. Random forests considering various tree structures have 700 very high accuracy, so it can be utilized in the accuracy-critical 701 IoT applications. 703 5.1.2. Unsupervised learning for IoT 705 Unsupervised learning is a data-driven type of machine learning which 706 finds hidden structure in unlabeled dataset without feedback during 707 the learning process. Unlike supervised learning, unsupervised 708 learning focuses on discovering patterns in the data distributions 709 and gaining insights from them. 711 (K-means clustering) K-means clustering aims to assign observations 712 into K number of clusters in which each observation belongs to the 713 cluster having the most similarities. The measure of similarity is 714 the distance between K cluster centers and each observation. K-means 715 is a very fast and highly scalable clustering algorithm, so it can be 716 used for IoT applications with real-time processing requirements such 717 as smart transportation. 719 (Density-based spatial clustering of applications with noise) 720 Density-Based approach to Spatial Clustering of Applications with 721 Noise (DBSCAN) is a method that clusters dataset based on the density 722 of its data samples. In this model, dense regions which include data 723 samples with many close neighbors are considered as clusters, and 724 data samples in low-density regions are classified as outliers 725 [Kriegal]. Since this method is robust to outliers, DBSCAN is 726 efficient data clustering method for IoT network environments with 727 untrusted big datasets in practice. 729 5.1.3. Reinforcement learning for IoT 731 Reinforcement learning is a reactive type of machine learning that 732 learn a series of actions in a given set of possible states, actions, 733 and rewards or penalties. It can be seen as the exploring decision- 734 making process and choosing the action series with the most reward or 735 the least penalty which can be cost, priority, time to name a few. 736 Reinforcement learning can be helpful for selecting action of IoT 737 device by providing a guideline. 739 (Q-learning) Q-Learning is a model-free, off-policy reinforcement 740 learning algorithm based on the well-known Bellman Equation. The goal 741 is to learn an action-selection policy maximizing the Q-value, which 742 tells an agent what action to take. It can be used for IoT device to 743 determine which action it should take according to conditions. 745 (State-Action-Reward-State-Action) Though State-Action-Reward-State- 746 Action (SARSA) is a much similar algorithm to Q-learning, the main 747 difference is that it is an on-policy algorithm in which agent 748 interacts with the environment and updates the policy based on 749 actions taken. It means that the Q-value is updated by an action 750 performed by the current policy instead of the greed policy that 751 maximizes Q-value. In this perspective, it is relevant when an action 752 of one IoT device will greatly influence the condition of the 753 environment. 755 (Deep Q Network) Deep Q network (DQN) is developed to solve the 756 exploration problem for unseen states. In the case of Q-learning, the 757 agent is not capable of estimating value for unseen states. To handle 758 this generality problem, DQN leverages neural network technology. As 759 a variation of the classic Q-Learning algorithm, DQN utilizes a deep 760 convolutional neural net architecture for Q-function approximation. 761 In real environments not all possible states and conditions are not 762 able to be observed. Therefore, DQN is more relevant than Q-learning 763 or SARSA in real applications such as IoT. Since DQN could be used 764 within only discrete action space, it can be utilized for traffic 765 routing in the IoT network. 767 (Deep Deterministic Policy Gradient) DQN has solved generality and 768 exploration problem of the unseen or rare states. Deep Deterministic 769 Policy Gradient (DDPG) takes DQN into the continuous action domain. 770 DDPG is a deterministic policy gradient based actor-critic, model- 771 free algorithm. The actor decides the best action for each state and 772 critic is used to evaluate the policy, the chosen action set. In IoT 773 applications, DDPG can be utilized for the tasks that require 774 controlled in continuous action spaces, such as energy-efficient 775 temperature control, computation offloading, network traffic 776 scheduling, etc. 778 5.1.4. Neural Network based algorithms for IoT 780 (Recurrent Neural Network) Recurrent Neural Network (RNN) is a 781 discriminative type of supervised learning model that takes serial or 782 time-series input data. RNN is specifically developed to address 783 issue of time dependency of sequential time-series input data. It 784 processes sequences of data through internal memory, and it is useful 785 in IoT applications with time-dependent data, such as identifying 786 time-dependent patterns of sensor data, estimating consumption 787 behavior over time, etc. 789 (Long Short Term Memory) As an extension of RNN, Long Short Term 790 Memory (LSTM) is a discriminative type of supervised learning model 791 that is specialized for serial or time-series input data as well 792 [Hochreiter]. The main difference of LSTM from RNN is that it 793 utilizes the concept of gates. It actively controls forget gates to 794 prevent the long term time dependency from waning. Therefore, 795 compared to RNN, it is more suitable for data with long time 796 relationship and IoT applications requiring analysis on the long lag 797 of dependency, such as activity recognition, disaster prediction, to 798 name a few [Chung]. 800 (Convolutional Neural Network) Convolutional neural network (CNN) is 801 a discriminative type of supervised learning model. It is developed 802 specifically for processing 2-dimensional image data by considering 803 local connectivity, but now generally used for multidimensional data 804 such as multi-channel sound signals, IoT sensor values, etc. As in 805 CNN neurons are connected only to a small subset of the input and 806 share weight parameters, CNN is much more sparse compared to fully 807 connected network. However, it needs a large training dataset, 808 especially for visual tasks. In CNN, a new activation function for 809 neural network, Rectified Linear Unit (ReLU), was proposed, which 810 accelerates training time without affecting the generalization of the 811 network [Krizhevsky]. In IoT domains, it is often used for detection 812 tasks that require some visual analysis. 814 (Variational Autoencoder) Autoencoder (AE) is a generative type 815 of unsupervised learning model. AE is trained to generate output to 816 reconstruct input data, thus it has the same number of input and 817 output units. It is suitable for feature extraction and 818 dimensionality reduction. Because of its behavior to reconstructing 819 the input data at the output layer, it is often used for machinery 820 fault diagnosis in IoT applications. The most popular type of AE, 821 Variational Autoencoder (VAE) is a generative type of semi-supervised 822 learning model. Its assumptions on the structure of the data are weak 823 enough for real applications and its training process through 824 backpropagation is fast [Doersch]. Therefore, VAE is suitable in IoT 825 applications where data tends to be diverse and scarce. 827 (Generative Adversarial Network) Generative Adversarial Network (GAN) 828 is a hybrid type of semi-supervised learning model which contain two 829 neural networks, namely the generative and discriminative networks 830 [Goodfellow]. The generator is trained to learn the data distribution 831 from a training dataset in order to generate new data which can 832 deceive the latter network, so-called the discriminator. Then, the 833 discriminator learns to discriminate the generated data from the real 834 data. In IoT applications, GAN can be used in situations when 835 something needs to be generated from the available data, such as 836 localization, way-finding, and data type conversion. 838 5.2. Technologies for lightweight and real-time intelligence 840 As the era of IoT has come, some sort of light-weight intelligence is 841 needed to support smart objects. Prior to the era of IoT, most of the 842 works on learning did not consider resource-constrained environments. 843 Especially, deep learning models require many resources such as 844 processing power, memory, stable power source, etc. However, it has 845 been recently shown that the parameters of the deep learning models 846 contain redundant information, so that some parts of them can be 847 delicately removed to reduce complexity without much degradation of 848 performance [Ba], [Denil]. In this section, the technologies to 849 achieve real-time and serverless learning in IoT environments are 850 introduced. 852 (network compression) Network compression is a method to convert a 853 dense network into a sparse one. With this technology the network can 854 be reduced in its size and complexity. By pruning irrelevant parts or 855 sharing redundant parameters, the storage and computational 856 requirements can be decreased [Han]. After pruning, the performance 857 of the network is examined and the pruning process is repeated until 858 the performance reaches the minimum requirements for the specific 859 applications and use cases. As many parameters are removed or shared, 860 the memory required is reduced, as well as computational burden and 861 energy. Especially as most energy in neural network is used to access 862 memory, the consumed energy dramatically drops. Although its main 863 limitation is that there is not a general solution to compress all 864 kinds of network, but it rather depends on the characteristics of 865 each network. However, network compression is still the most 866 widespread method to make deep learning technologies to be 867 lightweight and IoT-friendly. 869 (approximate computing) Approximate computing is an approach to 870 support deep learning in smart devices [Venkataramani], [Moons]. It 871 is based on the facts that the results of deep learning do not need 872 to be exact in many IoT applications but still valid if the results 873 are in an acceptable range. By integrating approximate computing into 874 deep learning, not only the execution time but also the energy 875 consumption is reduced [Mohammadi]. Based on the optimal trade-off 876 between accuracy and run-time or energy consumption, the network can 877 be adjustably approximated. The network approximate technology can be 878 well-used in such situations when the response time is more important 879 than sophisticatedly analyzed results. Although it is a technology to 880 facilitate real-time and lightweight intelligence, the process of 881 training models and converting it to approximate network require some 882 amount of resource. Therefore, the approximated model can be deployed 883 on smart devices but the learning and approximation processes still 884 need to take places on resource rich platforms. 886 6. Use cases of AI/ML into IoT service 888 Many IoT service environments are equipped with camera, door lock, 889 temperature sensor, fire detector, gas detector, alarm, and so on. 890 Each sensor is deployed with particular purposes of each own to 891 provide a specific service. However, there is a problem that the 892 sensor utilization is not high enough due to the provision of the 893 service using only a single sensor rather than multiple sensors and 894 their mutual relations. Therefore, the quality of the service 895 provided is not high as well. To enhance the sensor utilization and 896 the service quality, all signals from various sensors should be 897 analyzed comprehensively to make the right decision. This section 898 describes the use cases for introducing AI / ML techniques in actual 899 IoT service, utilizing multiple sensors. In advance of each use case 900 description of various IoT service domains, characteristics of each 901 domain to adopt AI/ML techniques are investigated. 903 6.1. Surveillance and Security in Smart Home 905 To minimize users' inconvenience and ensure their safety, 906 surveillance and safety IoT applications provided within smart homes 907 require fast notification with good level of precision IoT service 908 quality for abnormal conditions detection. To do this, both data 909 preprocessing techniques and AI/ML technologies for analysis of 910 anomalies with high accuracy will be required. 912 6.1.1. Characteristics of Smart Home for AI/ML processing 914 (Training Data Generation) For Surveillance and Security, the 915 processed data is necessary because there is little data for 916 anomalies and the ground truth values are hardly available. Therefore, 917 first, the steps to detect and calibrate the abnormal data are 918 essential before the anomaly data should be generated using domain 919 knowledge. First, constructing simulators about targeted smart home 920 and generating events against external intrusions and then collecting 921 the anomaly data can be considered. Furthermore, based on the data 922 collected in the actual environment, anomaly data generation can 923 proceed by breaking the relationship between sensors considering 924 possible links between them within any intrusive environment. 926 (AI/ML Algorithm) One of the characteristics of the IoT environment 927 for surveillance and safety is that a massive amount of data is 928 collected and real-time responses are required. For the kNN algorithm, 929 since the more data sets, the stronger against the noise and the 930 higher the accuracy. If the appropriate dataset is used, the fast 931 response can be expected. It makes suitable for the service 932 environment to be considered. In addition, considering real-time data 933 forecasting and analysis via LSTM, it is believed that improved 934 accuracy for real-time anomalies detection can be expected. 936 6.1.2. Use case 938 (To be continued) 940 6.2. Smart Building Management 942 Smart buildings often consist of heterogeneous IoT devices. These 943 devices cooperate and their data is integrated for efficient 944 autonomous building management. Many of the events in a large 945 building may not require deep, complicated learning or processing. 946 Some of them may require a fast response than an accurate analysis. 947 Above all, a lot of events simultaneously occur and can arise heavy 948 loads on the main server. The edge-computing techniques can be used 949 to offload the main server's tasks. 951 6.2.1. Characteristics of Smart Building for AI/ML processing 953 (Training Data Generation) In smart buildings, heterogeneous IoT 954 devices are deployed. They are diverse in their types, functions, 955 performances, etc. To utilize the data from diverse devices, data 956 needs to be able to well-integrated. Therefore, it is better for data 957 to be in a common data format, or it needs to be able to transform 958 into one another. The other characteristic is that the IoT devices 959 may interact in local and global environments of the building. 960 Therefore, the scope of the dataset used in training can play a 961 critical role in developing AL/ML model for building management. 963 (AI/ML Algorithm) To offload and reduce the burden of the main server 964 and to provide fast, efficient decision makings, the IoT and the 965 other network-related devices can use their computing resources. 966 Various edge-computing techniques can be applied to do so, such as 967 developing light-weighted AI/ML models that can be easily deployed in 968 the edge devices or balancing the learning and processing computation 969 load from the server to the edge devices. 971 6.2.2. Use case 973 (To be continued) 975 6.3. Manufacturing optimization in Smart factory 977 When IoT meets smart factory Industry, Industrial Internet of Things 978 (IIoT) technology integrating various technologies such as AI/ML, big 979 data analytics, massive sensor connection, Machine to Machine 980 communication (M2M) and automation are generally utilized. Its IIoT 981 applications has initiated a profound change within companies making 982 them increasingly connected and allowing them to exploit data to 983 optimize their production processes, since IIoT is able to support 984 real-time monitoring of each process line, thanks to different types 985 of sensors positioned in critical points. 987 However, to anticipate and prevent possible failures, the health 988 state of assets is always monitored, which generating tremendous 989 volume of data, and instantons and intelligent responses to 990 unexpected behavior are required. 992 6.3.1. Characteristics of Smart Building for AI/ML processing 994 (To be continued) 996 6.3.2. Use case 998 (To be continued) 1000 7. IANA Considerations 1002 This document requests no action by IANA. 1004 8. Acknowledgements 1006 9. Contributors 1008 10. Informative References 1010 [Hochreiter]S. Hochreiter and J. Schmidhuber, "Long short-term 1011 memory," Neural Comput., vol. 9, no. 8, pp. 1735-1780, Nov. 1012 1997. 1014 [Chung] J. Chung, C. Gulcehre, K. Cho, and Y. Bengio, "Empirical 1015 evaluation of gated recurrent neural networks on sequence 1016 modeling," arXiv preprint arXiv:1412.3555v1 [cs.NE], 2014. 1018 [Krizhevsky]A. Krizhevsky, I. Sutskever, and G. E. Hinton, "ImageNet 1019 classification with deep convolutional neural networks," in 1020 Proc. Adv. Neural Inf. Process. Syst., 2012, pp. 1097-1105.. 1021 Hochreiter and J. Schmidhuber, "Long short-term memory," 1022 Neural Comput., vol. 9, no. 8, pp. 1735-1780, Nov. 1997. 1024 [Doersch] C. Doersch, "Tutorial on variational autoencoders," arXiv 1025 preprint arXiv:1606.05908v2 [stat.ML], 2016. 1027 [Goodfellow]I. Goodfellow et al., "Generative adversarial nets," in 1028 Proc. Adv. Neural Inf. Process. Syst., 2014, pp. 2672-2680. 1030 [Ba] J. Ba and R. Caruana, "Do deep nets really need to be deep?" 1031 in Proc. Adv. Neural Inf. Process. Syst., Montreal, QC, 1032 Canada, 2014, pp. 2654-2662. 1034 [Denil] M. Denil, B. Shakibi, L. Dinh, N. de Freitas, and M. Ranzato, 1035 "Predicting parameters in deep learning," in Proc. Adv. Neural Inf. 1036 Process. Syst., 2013, pp. 2148-2156. 1038 [Han] S. Han, J. Pool, J. Tran, and W. Dally, "Learning both 1039 weights and connections for efficient neural network," in 1040 Proc. Adv. Neural Inf. Process. Syst., Montreal, QC, Canada, 1041 2015, pp. 1135-1143. 1043 [Venkataramani]S. Venkataramani, A. Ranjan, K. Roy, and A. 1044 Raghunathan, "AxNN: Energy-efficient neuromorphic systems 1045 using approximate computing," in Proc. Int. Symp. Low Power 1046 Electron. Design, ACM, 2014, pp. 27-32. [Moons]S. 1047 Hochreiter and J. Schmidhuber, "Long short-term memory," 1048 Neural Comput., vol. 9, no. 8, pp. 1735-1780, Nov. 1997. 1050 [Moons] B. Moons, B. De Brabandere, L. Van Gool, and M. Verhelst, 1051 "Energy- efficient ConvNets through approximate computing," 1052 in Proc. IEEE Winter Conf. Appl. Comput. Vis. (WACV), Lake 1053 Placid, NY, USA: IEEE, 2016, pp. 1-8. 1055 [Mohammadi] Mohammadi, Mehdi, et al. "Deep learning for IoT big data 1056 and streaming analytics: A survey," IEEE Communications 1057 Surveys & Tutorials, 2018, pp. 2923-2960. 1059 [Kriegel] Kriegel, HansPeter, et al. "Densitybased clustering," 1060 Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery, 1061 2011, pp. 231-240. 1063 Authors' Addresses 1065 Jun Kyun Choi (editor) 1066 Korea Advanced Institute of Science and Technology (KAIST) 1067 193 Munji Ro, Yuseong-gu, Daejeon, Korea 1069 Email: jkchoi59@kaist.ac.kr 1071 Jae Seob Han 1072 Korea Advanced Institute of Science and Technology (KAIST) 1073 193 Munji Ro, Yuseong-gu, Daejeon, Korea 1075 Email: j89449@kaist.ac.kr 1077 Gyeong Ho Lee 1078 Korea Advanced Institute of Science and Technology (KAIST) 1079 193 Munji Ro, Yuseong-gu, Daejeon, Korea 1081 Email: gyeongho@kaist.ac.kr 1083 Na Kyoung Kim 1084 Korea Advanced Institute of Science and Technology (KAIST) 1085 193 Munji Ro, Yuseong-gu, Daejeon, Korea 1087 Email: nkim71@kaist.ac.kr