Empowering things with intelligence: a survey of the progress, challenges, and opportunities in artificial intelligence of things

May combine two claims.

The architecture of the network matters: even with fewer model parameters and computational complexity, recently proposed networks such as ResNet and DenseNet outperform previous ones such as VGGNet.

“Besides, the architecture of the network matters. Even with fewer model parameters and computational complexity, the recently proposed networks such as ResNet and DenseNet outperform previous ones such as VGGNet.”

Probably not a substantial claim: may not pass S3 (load-bearing).

With the increase of network depth and the number of parameters, the representation capacity also increases, leading to lower top1 classification error on the ImageNet dataset.

“As can be seen, with the increase of network depth and the number of parameters, the representation capacity also increases, leading to lower top1 classification error on the ImageNet dataset.”

May combine two claims.

Transmitting huge amounts of IoT data requires massive bandwidth, and cloud processing and sending results back to end devices leads to high latency.

“Transmitting this huge amount of data requires massive bandwidth, and cloud processing and sending the results back to end devices leads to high latency.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The IoT represents a key infrastructure for supporting various applications such as smart homes, smart transportation, smart grids, and smart healthcare.

“The IoT represents a key infrastructure for supporting various applications [4], e.g., smart homes [5], [6], smart transportation [7], [8], smart grids [9], and smart healthcare [10], [11].”

• cites [4] [4] L. Da Xu, W. He, and S. Li, 'Internet of things in industries: A survey,' IEEE Transactions on Industrial Informatics , vol. 10, no. 4, pp. 22332243, 2014. -Not checked yet

  Not checked in the free run.

• cites [5] [5] X. Li, R. Lu, X. Liang, X. Shen, J. Chen, and X. Lin, 'Smart community: An internet of things application,' IEEE Communications Magazine , vol. 49, no. 11, pp. 68-75, 2011. -Not checked yet

  Not checked in the free run.

• cites [6] [6] B. L. R. Stojkoska and K. V. Trivodaliev, 'A review of internet of things for smart home: Challenges and solutions,' Journal of Cleaner Production , vol. 140, pp. 1454-1464, 2017. -Not checked yet

  Not checked in the free run.

May combine two claims.

50 billion devices will be IoT connected by 2025, generating 79.4 zettabytes of data.

“Based on reports from Cisco [14] and IDC [15], 50 billion devices will be IoT connected by 2025, generating 79.4 zettabytes of data.”

• cites [14] [14] J. Macaulay, L. Buckalew, and G. Chung, 'Internet of things in logistics: A collaborative report by dhl and cisco on implications and use cases for the logistics industry,' DHL Trend Research and Cisco Consulting Services , 2015. -Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [15] [15] M. Shirer and C. MacGillivray, 'The growth in connected iot devices is expected to generate 79.4 zb of data in 2025, according to a new idc forecast,' 2019. -Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing).

Fog computing aims to bring storage, computation, and networking capacity to the edge of the network in proximity to devices.

“To address this issue, 'fog computing', coined by Cisco [16], aims to bring storage, computation, and networking capacity to the edge of the network (e.g., to distributed fog nodes such as routers) in proximity to the devices.”

• cites [16] [16] F. Bonomi, 'Connected vehicles, the internet of things, and fog computing,' in The Eighth ACM International Workshop on Vehicular Inter-Networking , 2011, pp. 13-15. -Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing).

A typical IoT architecture has three layers: a perception layer, a network layer, and an application layer.

“A typical IoT architecture has three layers [13]: a perception layer, a network layer, and an application layer.”

• cites [13] [13] J. Lin, W. Yu, N. Zhang, X. Yang, H. Zhang, and W. Zhao, 'A survey on internet of things: Architecture, enabling technologies, security and privacy, and applications,' IEEE Internet of Things Journal , vol. 4, no. 5, pp. 1125-1142, 2017. -Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Classified as probably not a substantial claim.

In conventional IoT solutions, data collected from sensors are transmitted to the cloud computing platform through networks for further processing and analysis before delivering results or commands to end devices or actuators.

“In conventional IoT solutions, data collected from sensors are transmitted to the cloud computing platform through the networks for further processing and analysis before delivering the results/commands to end devices/actuators.”

Probably not a substantial claim: may not pass S4 (based).

Edge cases exist in the real world that may not be seen in the training set nor defined in the label set, resulting in degeneration of a pre-trained model.

“For example, edge cases exist in the real world that may not be seen in the training set nor defined in the label set, resulting in degeneration of a pre-trained model.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing). May combine two claims.

This heralds the era of the artificial intelligence of things (AIoT), where AI encounters IoT.

“This heralds the era of the artificial intelligence of things (AIoT), where AI encounters IoT.”

• AssumptionRests on an assumption: Advanced communication technologies such as Wi-Fi 6 and 5G and AI technologies will enable mass connection.

May combine two claims.

Fog computing offers the advantages of low latency and high computational capacity for IoT applications.

“Fog computing offers the advantages of low latency and high computational capacity for IoT applications [17], [18].”

• cites [17] [17] M. Chiang and T. Zhang, 'Fog and iot: An overview of research opportunities,' IEEE Internet of Things Journal , vol. 3, no. 6, pp. 854864, 2016. - sourceNot checked yet

  Not checked in the free run.

• cites [18] [18] J. Ni, K. Zhang, X. Lin, and X. S. Shen, 'Securing fog computing for internet of things applications: Challenges and solutions,' IEEE Communications Surveys & Tutorials , vol. 20, no. 1, pp. 601-628, 2017. - sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

IoT applications based on specific things, e.g., the Internet of Vehicles and Internet of Video Things, have also been rapidly developed.

“Furthermore, IoT applications based on specific things, e.g., the Internet of Vehicles (IoV) [7] and Internet of Video Things (IoVT) [23] have also been rapidly developed.”

• cites [7] [7] J. Contreras-Castillo, S. Zeadally, and J. A. Guerrero-Iba˜ nez, 'Internet of vehicles: Architecture, protocols, and security,' IEEE Internet of Things Journal , vol. 5, no. 5, pp. 3701-3709, 2017. - sourceNot checked yet

  Not checked in the free run.

• cites [23] [23] C. W. Chen, 'Internet of video things: Next generation iot with visual sensors,' IEEE Internet of Things Journal , 2020. - sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Edge computing has been proposed by further deploying computing capacity on edge devices in proximity to sensors and actuators.

“'Edge computing' has also recently been proposed by further deploying computing capacity on edge devices in proximity to sensors and actuators [19], [20].”

• cites [19] [19] N. Abbas, Y. Zhang, A. Taherkordi, and T. Skeie, 'Mobile edge computing: A survey,' IEEE Internet of Things Journal , vol. 5, no. 1, pp. 450-465, 2017. -Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [20] [20] J. Pan and J. McElhannon, 'Future edge cloud and edge computing TABLE IX LIST OF ABBREVIATIONS. | Abbreviation | Description | |----------------|------------------------------------------| | 5G | 5th Generation Mobile Networks | | AI | Artificial Intelligence | | AIoT | Artificial Intelligence of Things | | AR/VR | Augmented Reality / Virtual Reality | | ASIC | Application-Specific Integrated Circuit | | ASR | Automatic Speech Recognition | | CNN | Convolutional Neural Networks | | CTC | Connectionist Temporal Classification | | DA | Domain Adaptation | | DNN | Deep Neural Network | | DRLNot matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing).

Massive sensor data must be processed to extract useful information before being used for further analysis and decision-making.

“Massive sensor data must be processed to extract useful information before being used for further analysis and decision-making.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Deep learning has facilitated progress in these areas, suggesting that deep learning can be similarly leveraged to improve IoT system learning.

“Deep learning has facilitated progress in these areas, suggesting that deep learning can be similarly leveraged to improve IoT system learning.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The context of IoT sensors can provide auxiliary information to help understand sensor data.

“Moreover, the context of IoT sensors can provide auxiliary information to help understand sensor data. Therefore, various context-aware computing methods have been proposed for IoT [54].”

• cites [54] [54] C. Perera, A. Zaslavsky, P. Christen, and D. Georgakopoulos, 'Context aware computing for the internet of things: A survey,' IEEE Communications Surveys & Tutorials , vol. 16, no. 1, pp. 414-454, 2013. - sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

In industry, operating modes of machines may drift or change due to fatigue or wear and tear, and models trained for the initial mode cannot adapt to this variation, leading to performance loss.

“Another example is in industry, where the operating modes of machines may drift or change due to fatigue or wear and tear. Consequently, models trained for the initial mode cannot adapt to this variation, leading to a performance loss.”

May combine two claims.

Multimedia sensor data are high-dimensional and unstructured, so it is difficult to design handcrafted features for them without domain knowledge.

“First, multimedia sensor data are high-dimensional and unstructured (semantics are unavailable without additional processing), so it is difficult to design handcrafted features for them without domain knowledge.”

Probably not a substantial claim: may not pass S4 (based).

Perceiving and understanding dynamic and complex environments from sensor data is fundamental to IoT applications providing useful services to users.

“Perceiving and understanding dynamic and complex environments from sensor data is fundamental to IoT applications providing useful services to users.”

Probably not a substantial claim: may not pass S3 (load-bearing).

According to Cisco's white paper, 99.4% of physical objects are still unconnected.

“According to Cisco's white paper [43], 99.4% of physical objects are still unconnected.”

• cites [43] [43] J. Bradley, J. Barbier, and D. Handler, 'Embracing the internet of everything to capture your share of $14.4 trillion,' White Paper, Cisco , vol. 318, 2013. -Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

To interact with the environment and humans, an IoT system should be able to reason and behave.

“Furthermore, to interact with the environment and humans, an IoT system should be able to reason and behave.”

Probably not a substantial claim: may not pass S3 (load-bearing).

These issues are related to machine learning topics including few-shot learning, zero-shot learning, meta-learning, unsupervised learning, semi-supervised learning, transfer learning, and domain adaptation.

“These issues are related to some well-known machine learning research topics including fewshot learning [31], zero-shot learning [32], meta-learning [33], unsupervised learning [34], semi-supervised learning [35], transfer learning [36], and domain adaptation [37], [38].”

• cites [31] [31] Y. Wang, Q. Yao, J. T. Kwok, and L. M. Ni, 'Generalizing from a few examples: A survey on few-shot learning,' ACM Computing Surveys , vol. 53, no. 3, pp. 1-34, 2020. -Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [32] [32] W. Wang, V. W. Zheng, H. Yu, and C. Miao, 'A survey of zero-shot learning: Settings, methods, and applications,' ACM Transactions on Intelligent Systems and Technology , vol. 10, no. 2, pp. 1-37, 2019. - sourceNot checked yet

  Not checked in the free run.

• cites [33] [33] C. Finn, P. Abbeel, and S. Levine, 'Model-agnostic meta-learning for fast adaptation of deep networks,' in Proceedings of the 34th International Conference on Machine Learning , 2017. - sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

These data may contain biometric information such as faces, voice, or fingerprints.

“These data may contain some biometric information such as faces, voice, or fingerprints.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Edge computing can process and analyze data on premises and make decisions instantly, thereby benefitting latency-sensitive IoT applications.

“Edge computing can process and analyze data on premises and make decisions instantly, thereby benefitting latency-sensitive IoT applications.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Deep neural networks have been widely used and achieved excellent performance in many applications including speech recognition, face recognition, image classification, object detection, semantic segmentation, and natural language processing.

“Deep neural networks (DNNs) have been widely used in many areas and have achieved excellent performance in many applications including speech recognition [24], face recognition [25], image classification [26], object detection [27], semantic segmentation [28], natural language processing [29]”

• cites [24] [24] H. Sak, A. Senior, and F. Beaufays, 'Long short-term memory recurrent neural network architectures for large scale acoustic modeling,' in Proceedings of the Fifteenth Annual Conference of the International Speech Communication Association , 2014. - sourceNot checked yet

  Not checked in the free run.

• cites [25] [25] Y. Taigman, M. Yang, M. Ranzato, and L. Wolf, 'Deepface: Closing the gap to human-level performance in face verification,' in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , 2014, pp. 1701-1708. - sourceNot checked yet

  Not checked in the free run.

• cites [26] [26] A. Krizhevsky, I. Sutskever, and G. E. Hinton, 'Imagenet classification with deep convolutional neural networks,' in Advances in Neural Information Processing Systems , 2012, pp. 1097-1105. -Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Cyberattacks on IoT systems may result in data leakage, so data security and privacy have become a critical concern in IoT applications.

“Cyberattacks on IoT systems may result in data leakage, so data security and privacy have become a critical concern in IoT applications [18].”

• cites [18] [18] J. Ni, K. Zhang, X. Lin, and X. S. Shen, 'Securing fog computing for internet of things applications: Challenges and solutions,' IEEE Communications Surveys & Tutorials , vol. 20, no. 1, pp. 601-628, 2017. -Not checked yet

  Not checked in the free run.

May combine two claims.

An IoT perception system alone is insufficient; simply adapting to and interacting with the dynamic and complex world is insufficient.

“Although an IoT's perception system is a critical component of the architecture, simply adapting to and interacting with the dynamic and complex world is insufficient.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

There has recently been rapid progress in deep learning, with positive effects impacting IoT data processing, e.g., streaming data analysis, mobile multimedia processing, manufacturing inspection, and health monitoring.

“There has recently been rapid progress in deep learning, with these positive effects also impacting IoT data processing, e.g., streaming data analysis [55], mobile multimedia processing [56], manufacturing inspection [57], and health monitoring.”

• cites [55] [55] M. Mohammadi, A. Al-Fuqaha, S. Sorour, and M. Guizani, 'Deep learning for iot big data and streaming analytics: A survey,' IEEE Communications Surveys & Tutorials , vol. 20, no. 4, pp. 2923-2960, 2018. -Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [56] [56] K. Ota, M. S. Dao, V. Mezaris, and F. G. D. Natale, 'Deep learning for mobile multimedia: A survey,' ACM Transactions on Multimedia Computing, Communications, and Applications , vol. 13, no. 3s, pp. 1-22, 2017. - sourceNot checked yet

  Not checked in the free run.

• cites [57] [57] L. Li, K. Ota, and M. Dong, 'Deep learning for smart industry: Efficient manufacture inspection system with fog computing,' IEEE Transactions on Industrial Informatics , vol. 14, no. 10, pp. 4665-4673, 2018. - sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Data mining and machine learning approaches have been used for IoT data processing and analysis.

“Data mining and machine learning approaches have been used for IoT data processing and analysis [52], [53].”

• cites [52] [52] C.-W. Tsai, C.-F. Lai, M.-C. Chiang, and L. T. Yang, 'Data mining for internet of things: A survey,' IEEE Communications Surveys & Tutorials , vol. 16, no. 1, pp. 77-97, 2013. - sourceNot checked yet

  Not checked in the free run.

• cites [53] [53] M. S. Mahdavinejad, M. Rezvan, M. Barekatain, P. Adibi, P. Barnaghi, and A. P. Sheth, 'Machine learning for internet of things data analysis: A survey,' Digital Communications and Networks , vol. 4, no. 3, pp. 161-175, 2018. -Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Massive user data are collected via ubiquitous connected sensors, which may be transmitted and stored in the cloud through IoT networks.

“Massive user data are collected via ubiquitous connected sensors, which may be transmitted and stored in the cloud through IoT networks.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Application scenarios such as smart parking could benefit from recent advances in causal inference and discovery, graph-based reasoning, reinforcement learning, and speech recognition and synthesis.

“These application scenarios could benefit from recent advances in causal inference and discovery [39], graphbased reasoning [40], reinforcement learning [41], and speech recognition and synthesis [24], [42].”

• cites [39] [39] Y. Zhang, S. Qian, Q. Fang, and C. Xu, 'Multi-modal knowledgeaware hierarchical attention network for explainable medical question answering,' in Proceedings of the 27th ACM International Conference on Multimedia , 2019, pp. 1089-1097. -Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [40] [40] X. Wang, D. Wang, C. Xu, X. He, Y. Cao, and T.-S. Chua, 'Explainable reasoning over knowledge graphs for recommendation,' in Proceedings of the AAAI Conference on Artificial Intelligence , vol. 33, 2019, pp. 5329-5336. - sourceNot checked yet

  Not checked in the free run.

• cites [41] [41] J. Zhang, B. Hao, B. Chen, C. Li, H. Chen, and J. Sun, 'Hierarchical reinforcement learning for course recommendation in moocs,' in Proceedings of the AAAI Conference on Artificial Intelligence , vol. 33, 2019, pp. 435-442. - sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Access control and trust management approaches have been reviewed to protect the security and privacy of IoT.

“Recently, access control [58] and trust management [59] approaches have been reviewed to protect the security and privacy of IoT.”

• cites [58] [58] J. Qiu, Z. Tian, C. Du, Q. Zuo, S. Su, and B. Fang, 'A survey on access control in the age of internet of things,' IEEE Internet of Things Journal , 2020. - sourceNot checked yet

  Not checked in the free run.

• cites [59] [59] Z. Yan, P. Zhang, and A. V. Vasilakos, 'A survey on trust management for internet of things,' Journal of Network and Computer Applications , vol. 42, pp. 120-134, 2014. - sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

The last few years have witnessed a renaissance in artificial intelligence assisted by deep learning.

“The last few years has witnessed a renaissance in artificial intelligence (AI) assisted by deep learning.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

ONNX offers porting models among different frameworks by defining an open format built to represent machine learning models, which has been supported by TensorFlow and PyTorch.

“The Open Neural Network Exchange (ONNX) 5 offers this feature by defining an open format built to represent machine learning models, which has been supported by TensorFlow and Pytorch.”

Probably not a substantial claim: may not pass S3 (load-bearing).

State-of-the-art object detection methods can be categorized into two groups: two-stage methods and one-stage methods.

“State-of-the-art methods can be categorized into two groups: two-stage methods and one-stage methods.”

Probably not a substantial claim: may not pass S3 (load-bearing).

FPGAs have also been used for DNN acceleration due to their low power consumption and high throughput.

“Besides, Field-Programmable Gate Arrays (FPGA) have also been used for DNN acceleration due to their low power consumption and high throughput.”

May combine two claims.

Fog computing can provide continuity of service without the need for the Internet, which is important for specific AIoT applications with an unstable Internet connection, e.g., in agriculture, mining, and shipping domains.

“Besides, fog computing can provide continuity of service without the need for the Internet, which is important for specific AIoT applications with an unstable Internet connection, e.g., in agriculture, mining, and shipping domains.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Mask R-CNN performance can be improved by exploiting and enhancing the feature hierarchy of deep convolutional networks, employing non-local attention, and leveraging the reciprocal relationship between detection and segmentation via hybrid task cascade.

“Its performance can be improved further by exploiting and enhancing the feature hierarchy of deep convolutional networks [80], employing non-local attention [81], and leveraging the reciprocal relationship between detection and segmentation via hybrid task cascade [82].”

• cites [80] [80] T.-Y. Lin, P. Doll´ ar, R. Girshick, K. He, B. Hariharan, and S. Belongie, 'Feature pyramid networks for object detection,' in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , 2017, pp. 2117-2125. - sourceNot checked yet

  Not checked in the free run.

• cites [81] [81] X. Wang, R. Girshick, A. Gupta, and K. He, 'Non-local neural networks,' in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , 2018, pp. 7794-7803. - sourceNot checked yet

  Not checked in the free run.

• cites [82] [82] K. Chen, J. Pang, J. Wang, Y. Xiong, X. Li, S. Sun, W. Feng, Z. Liu, J. Shi, W. Ouyang et al. , 'Hybrid task cascade for instance segmentation,' in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , 2019, pp. 4974-4983. - sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Some popular applications in smartphones also benefit from advances in image recognition for recognizing flowers and birds, and food items and calories.

“Besides, some popular applications in smartphones also benefit from the advances in this area for recognizing flowers and birds, and food items and calories.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Lightweight networks are appealing to AIoT applications where DNNs are deployed on edge devices.

“Lightweight networks are appealing to AIoT applications where DNNs are deployed on edge devices.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Fog computing offers a key advantage of low latency since it is closer to devices.

“Although functioning similarly to cloud computing, fog computing offers a key advantage, i.e., low latency, since it is closer to devices.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Edge things and fog nodes are always distributed while the cloud is centralized in the AIoT network topology.

“It is noteworthy that the edge things and fog nodes are always distributed while the cloud is centralized in the AIoT network topology.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Recently, computationally efficient networks like MobileNet have been proposed by leveraging depth-wise convolutions, point-wise convolutions, or binary operations.

“Recently, some computationally efficient networks like MobileNet have been proposed by leveraging depth-wise convolutions [65], point-wise convolutions [66], or binary operations [67].”

• cites [65] [65] A. G. Howard, M. Zhu, B. Chen, D. Kalenichenko, W. Wang, T. Weyand, M. Andreetto, and H. Adam, 'Mobilenets: Efficient convolutional neural networks for mobile vision applications,' arXiv preprint arXiv:1704.04861 , 2017. - sourceNot checked yet

  Not checked in the free run.

• cites [66] [66] J. Zhang, Y. Cao, Y. Wang, C. Wen, and C. W. Chen, 'Fully pointwise convolutional neural network for modeling statistical regularities in natural images,' in Proceedings of the 26th ACM international conference on Multimedia , 2018, pp. 984-992. - sourceNot checked yet

  Not checked in the free run.

• cites [67] [67] M. Rastegari, V. Ordonez, J. Redmon, and A. Farhadi, 'Xnor-net: Imagenet classification using binary convolutional neural networks,' in Proceedings of the European Conference on Computer Vision . Springer, 2016, pp. 525-542. - sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

Empowering things with the perceiving ability, i.e., understanding the environment using various sensors, is fundamental for AIoT systems.

“Empowering things with the perceiving ability, i.e., understanding the environment using various sensors, is fundamental for AIoT systems.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

With off-the-shelf deep learning tools and scalable computing hardware, it is easy to set up the production environment on the cloud, where deep neural networks are trained and deployed to process massive amounts of data.

“With the off-the-shelf deep learning tools and scalable computing hardware, it is easy to set up the production environment on the cloud, where deep neural networks are trained and deployed to process the massive amounts of data.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Different open-source deep learning frameworks have been developed, from Caffe and MatConvNet to TensorFlow and PyTorch.

“To this end, different open-source deep learning frameworks have been developed, from the beginners like Caffe 1 and MatConvNet 2 to the popular TensorFlow 3 and PyTorch 4 .”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Caffe2 improves Caffe accordingly, and Caffe2 has been later merged into PyTorch.

“Caffe2 improves it accordingly, which has been later merged into PyTorch.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Caffe is implemented in C++ with Python and Matlab interfaces and is well-known for its speed but does not support distributed computation and mobile deployment.

“Caffe is implemented in C++ with Python and Matlab interfaces and well-known for its speed but does not support distributed computation and mobile deployment.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The cloud enables AIoT enterprises to use computing resources virtually via the Internet instead of building their physical infrastructure on premises.

“The cloud enables AIoT enterprises to use computing resources virtually via the Internet instead of building their physical infrastructure on premises.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Several machine learning processors have been developed for fog and edge computing, e.g., Google Edge TPU and NVIDIA Jetson Nano.

“Several machine learning processors have also been developed for fog and edge computing, e.g., Google Edge TPU and NVIDIA Jetson Nano.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Detection of specific categories of objects such as pedestrian, car, traffic-sign, and license plate has been widely studied, which is useful for improving the perceiving ability of AIoT systems for traffic and public safety surveillance and autonomous driving.

“Detection of specific category of objects such as pedestrian, car, traffic-sign, and the license plate has been widely studied, which are useful for improving the perceiving ability of AIoT systems for traffic and public safety surveillance and autonomous driving [70].”

• cites [70] [70] C. Chen, A. Seff, A. Kornhauser, and J. Xiao, 'Deepdriving: Learning affordance for direct perception in autonomous driving,' in Proceedings of the IEEE International Conference on Computer Vision , 2015, pp. 2722-2730. -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Mask R-CNN is a well-known baseline model for instance segmentation that adopts an extra branch for object mask prediction in parallel with the existing one for bounding box regression.

“A well-known baseline model is Mask R-CNN which adopts an extra branch for object mask prediction in parallel with the existing one for bounding box regression [79].”

• cites [79] [79] K. He, G. Gkioxari, P. Doll´ ar, and R. Girshick, 'Mask r-cnn,' in Proceedings of the IEEE International Conference on Computer Vision , 2017, pp. 2961-2969. - sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

TensorFlow and PyTorch support deploying models into mobile devices by enabling model compression/quantization and hardware acceleration.

“They also support for deploying models into mobile devices by enabling model compression/quantization and hardware acceleration.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Object detection is a crucial technique for video data structuring in many AIoT systems using visual sensors, aiming to extract and organize compact structured semantic information from video data for further retrieval, verification, statistics, and analysis at low transmission, storage, and computation cost.

“Besides, object detection is a crucial technique for video data structuring in many AIoT systems using visual sensors, which aims to extract and organize compact structured semantic information from video data for further retrieval, verification, statistics, and analysis at low transmission, storage, and computation cost.”

May combine two claims.

The edge computing layer may function like the perception layer and supports control and execution over sensors and actuators, aiming to empower AIoT systems with the ability to perceive and behave.

“The edge computing layer may function like the perception layer in [13] and smart visual sensing block in [23]. It also supports control and execution over sensors and actuators. Thereby, this layer aims to empower AIoT systems with the ability to perceive and behave.”

• cites [13] [13] J. Lin, W. Yu, N. Zhang, X. Yang, H. Zhang, and W. Zhao, 'A survey on internet of things: Architecture, enabling technologies, security and privacy, and applications,' IEEE Internet of Things Journal , vol. 4, no. 5, pp. 1125-1142, 2017. - sourceNot checked yet

  Not checked in the free run.

• cites [23] [23] C. W. Chen, 'Internet of video things: Next generation iot with visual sensors,' IEEE Internet of Things Journal , 2020. - sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

In practice, it is common to deploy multiple different models into cloud platforms, fog nodes, and edge devices in an AIoT system to build an intelligent hybrid computing architecture.

“In practice, it is common to deploy multiple different models into cloud platforms, fog nodes, and edge devices in an AIoT system to build an intelligent hybrid computing architecture.”

Probably not a substantial claim: may not pass S3 (load-bearing).

General-purpose GPU turns the massive computational power of its shader pipeline into general-purpose computing power, which has sparked the deep learning revolution along with DNN and big data.

“While GPU is initially developed for accelerating image rendering on display devices, the generalpurpose GPU turns the massive computational power of its shader pipeline into general-purpose computing power (e.g., for massive vector operations), which has sparked the deep learning revolution along with DNN and big data.”

May combine two claims.

Edge computing reduces latency and network bandwidth since it can process data into compact structured information on-site before transmission, which is especially useful for AIoT applications using multimedia sensors.

“A great advantage of edge computing over fog and cloud computing is the reduction of latency and network bandwidth since it can process data into compact structured information on-site before transmission, which is especially useful for AIoT applications using multimedia sensors.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Fog computing protects data security and privacy since data can be held within the LAN.

“The other advantage of fog computing is the protection of data security and privacy since data can be held within the LAN.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The fully CNN with an encoder-decoder structure has become the de facto paradigm for semantic segmentation because it can learn discriminative and multi-resolution features through cascaded convolution blocks while preserving spatial correspondence.

“The fully CNN with an encoder-decoder structure has become the de facto paradigm for semantic segmentation [75], [76], since it can learn discriminative and multi-resolution features through cascaded convolution blocks while preserving spatial correspondence.”

• cites [75] [75] O. Ronneberger, P. Fischer, and T. Brox, 'U-net: Convolutional networks for biomedical image segmentation,' in Proceedings of the International Conference on Medical Image Computing and ComputerAssisted Intervention , 2015, pp. 234-241. - sourceNot checked yet

  Not checked in the free run.

• cites [76] [76] J. Long, E. Shelhamer, and T. Darrell, 'Fully convolutional networks for semantic segmentation,' in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , 2015, pp. 3431-3440. - sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Boundary refinement in semantic segmentation can be achieved by using conditional random field as the postprocessing technique on the predicted probability maps.

“Boundary refinement aims to obtain sharp boundaries between different categories in the segmentation map, which can be achieved by using conditional random field as the postprocessing technique on the predicted probability maps [28].”

• cites [28] [28] L.-C. Chen, G. Papandreou, I. Kokkinos, K. Murphy, and A. L. Yuille, 'Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs,' IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 40, no. 4, pp. 834848, 2017. - sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Cloud computing can provide flexible, scalable, and reliable resources including computation, storage, and network for enabling various AIoT applications.

“It can provide flexible, scalable, and reliable resources including computation, storage, and network for enabling various AIoT applications.”

Classified as probably not a substantial claim.

Due to limited computation capacity, only lightweight DNNs can run on edge devices.

“However, due to its limited computation capacity, only lightweight DNNs can run on edge devices.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Features like dynamic computation graphs and automatic computation of gradients in TensorFlow and PyTorch have made them easy to use and popular.

“The features like dynamic computation graphs and automatic computation of gradients in TensorFlow and PyTorch have made them easy to use and popular.”

Probably not a substantial claim: may not pass S4 (based).

By intelligently offloading part of the computation workload from edge devices to fog nodes and cloud, it is expected to achieve low latency while leveraging deep learning capacities for processing massive amounts of data.

“By intelligently offloading part of the computation workload from edge devices to the fog nodes and cloud, it is expected to achieve low latency while leveraging deep learning capacities for processing massive amounts of data.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Research topics including neural network architecture design or search for mobile setting and network pruning/compression/quantization have attracted increasing attention recently.

“Therefore, research topics including neural network architecture design or search for mobile setting and network pruning/compression/quantization have attracted increasing attention recently.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Classical machine learning methods based on hand-crafted features have been surpassed by DNNs on large-scale benchmark datasets like ImageNet, sparking a wave of research on the architecture of DNNs.

“Classical machine learning methods based on hand-crafted features have been surpassed by DNNs [26] on large-scale benchmark datasets like ImageNet [30], sparking a wave of research on the architecture of DNNs.”

• cites [26] [26] A. Krizhevsky, I. Sutskever, and G. E. Hinton, 'Imagenet classification with deep convolutional neural networks,' in Advances in Neural Information Processing Systems , 2012, pp. 1097-1105. -Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [30] [30] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, 'Imagenet: A large-scale hierarchical image database,' in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , 2009, pp. 248-255. - sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

An ASIC named TPU is designed by Google specifically for neural network machine learning.

“Recently, an application-specific integrated circuit (ASIC) named TPU is designed by Google specifically for neural network machine learning.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The fog and cloud computing layers mainly aim to empower AIoT systems with the ability of learning and reasoning since they can access massive amounts of data and have vast computation resources.

“The fog and cloud computing layers mainly aim to empower AIoT systems with the ability of learning and reasoning since they can access massive amounts of data and have vast computation resources.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Cloud computing can leverage all the data from registered devices in an AIoT application, which is useful for training deep models with better representation and generalization ability.

“Another feature is that it can leverage all the data from the registered devices in an AIoT application, which is useful for training deep models with better representation and generalization ability.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

Image recognition can be very useful in many AIoT applications, such as smart education tools or toys, which help and teach children to explore the world with cameras.

“Image recognition can be very useful in many AIoT applications, such as smart education tools or toys, which help and teach children to explore the world with cameras.”

Probably not a substantial claim: may not pass S3 (load-bearing).

There are other deep learning frameworks like MXNet, Theano, PaddlePaddle, and ncnn.

“There are other deep learning frameworks like MXNet 6 , Theano 7 , PaddlePaddle 8 , and neural network inference computing framework for mobile devices like ncnn 9 .”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

A simple but strong baseline model for panoptic segmentation adds a semantic segmentation branch into the Mask R-CNN framework and uses a shared feature pyramid network backbone.

“A simple but strong baseline model is proposed in [84], which adds a semantic segmentation branch into the Mask R-CNN framework and uses a shared feature pyramid network backbone [80].”

• cites [84] [84] A. Kirillov, R. Girshick, K. He, and P. Doll´ ar, 'Panoptic feature pyramid networks,' in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , 2019, pp. 6399-6408. -Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [80] [80] T.-Y. Lin, P. Doll´ ar, R. Girshick, K. He, B. Hariharan, and S. Belongie, 'Feature pyramid networks for object detection,' in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , 2017, pp. 2117-2125. - sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

The methods for object detection from images have been revolutionized by DNNs.

“The methods for object detection from images have been revolutionized by DNNs.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Lots of operations in neural networks such as convolution can be computed in parallel on GPU, significantly reducing training and inference time.

“Lots of operations in the neural network such as convolution can be computed in parallel on GPU, significantly reducing the training and inference time.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Network compression such as pruning and quantization can be used to obtain lightweight models from heavy models.

“Besides, network compression such as pruning and quantization can be used to obtain lightweight models from heavy models, which will be reviewed in Section III-A17.”

Probably not a substantial claim: may not pass S3 (load-bearing).

An image can be decomposed into the reflectance map and illumination map based on the Retinex theory.

“An image can be decomposed into the reflectance map and illumination map based on the Retinex theory [145].”

• cites [145] [145] E. H. Land, 'The retinex theory of color vision,' Scientific american , vol. 237, no. 6, pp. 108-129, 1977. - sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

A differentiable architecture named neural graph optimizer is proposed for global pose graph optimization, and together with a local pose estimation model it achieves a complete end-to-end neural network solution for SLAM.

“Recently, a differentiable architecture named neural graph optimizer is proposed for global pose graph optimization [136]. Together with a local pose estimation model, it achieves a complete end-to-end neural network solution for SLAM.”

• cites [136] [136] E. Parisotto, D. Singh Chaplot, J. Zhang, and R. Salakhutdinov, 'Global pose estimation with an attention-based recurrent network,' in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops , 2018, pp. 237-246. - sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Due to the prevalence of cameras and great progress in deep learning and computer vision, visual hand gesture recognition has vast potential.

“Nevertheless, due to the prevalence of cameras and great progress in deep learning and computer vision, visual hand gesture recognition has the vast potential, which can be categorized into two groups, i.e., static ones and dynamic ones.”

• AssumptionRests on an assumption: Cameras are prevalent.

Probably not a substantial claim: may not pass S3 (load-bearing).

Estimation of human pose and recognition of action can be very useful in many real-world AIoT scenarios, such as rehabilitation exercises monitoring and assessment, dangerous behavior monitoring, and human-machine interaction.

“Estimation of human pose and recognition of action can be very useful in many real-world AIoT scenarios, such as rehabilitation exercises monitoring and assessment [120], dangerous behavior monitoring [121], and human-machine interaction (HMI).”

• cites [120] [120] D. Gonz´ alez-Ortega, F. D´ ıaz-Pernas, M. Mart´ ınez-Zarzuela, and M. Ant´ on-Rodr´ ıguez, 'A kinect-based system for cognitive rehabilitation exercises monitoring,' Computer Methods and Programs in Biomedicine , vol. 113, no. 2, pp. 620-631, 2014. -Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [121] [121] A. Kashevnik, I. Lashkov, and A. Gurtov, 'Methodology and mobile application for driver behavior analysis and accident prevention,' IEEE Transactions on Intelligent Transportation Systems , vol. 21, no. 6, pp. 2427-2436, 2019. - sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Face detection benefits from the recent success of deep learning in generic object detection.

“Face detection, as a specific subarea of object detection, benefits from the recent success of deep learning in generic object detection.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Different geometric cues such as lines and vanishing points have been exploited for formulating optimization constraints in optimization-based methods or loss functions in learning-based methods [157].

“Different geometric cues have been exploited for formulating the optimization constraints in optimization-based methods or loss functions in learning-based methods [157], such as lines and vanishing points.”

• cites [157] [157] Z. Xue, N. Xue, G.-S. Xia, and W. Shen, 'Learning to calibrate straight lines for fisheye image rectification,' in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , 2019, pp. 1643-1651. -Not checked yet

  Not checked in the free run.

May combine two claims.

CNN-based depth estimation must address challenges including occlusions, dynamic objects, and the scale issue of per-frame ambiguity and temporal inconsistency.

“Although CNN has powerful representation capacity, special effort has to be made to address the challenges including occlusions and dynamic objects, as well as the scale issue (per-frame ambiguity and temporally inconsistent).”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Hand gesture recognition can be very useful for interactions with things in AIoT systems, such as non-contact control of television and car infotainment system, and communication with the speech and hearing impaired.

“Hand gesture recognition can be very useful for interactions with things in AIoT systems, e.g., non-contact control of television and car infotainment system, and communication with the speech and hearing impaired [125].”

• cites [125] [125] T. Starner, J. Weaver, and A. Pentland, 'Real-time american sign language recognition using desk and wearable computer based video,' IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 20, no. 12, pp. 1371-1375, 1998. - sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Generative adversarial networks have been used to generate style-transferred images for bridging the domain gap between different datasets in person re-identification.

“Recently, generative adversarial networks (GAN) have been used to generate style-transferred images for bridging the domain gap between different datasets [111].”

• cites [111] [111] L. Wei, S. Zhang, W. Gao, and Q. Tian, 'Person transfer gan to bridge domain gap for person re-identification,' in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , 2018, pp. 79-88. - sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

A DCNN-based palmprint verification system named DeepMPV is proposed for mobile payment.

“Besides, a DCNN-based palmprint verification system named DeepMPV is proposed for mobile payment in [104].”

• cites [104] [104] Y. Zhang, L. Zhang, X. Liu, S. Zhao, Y. Shen, and Y. Yang, 'Pay by showing your palm: A study of palmprint verification on mobile platforms,' in Proceedings of the IEEE International Conference on Multimedia and Expo . IEEE, 2019, pp. 862-867. - sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Most CNN-based crowd counting methods directly regress the crowd density map, with ground truth constructed by placing Gaussian density maps at the head regions.

“Therefore, most CNN-based methods directly regress the crowd density map, in which the ground truth is constructed by placing Gaussian density maps at the head regions.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Human-in-the-loop artificial intelligence can be leveraged to achieve high performance with low labor effort in practical person re-identification systems.

“Although extra efforts are needed to build practical person re-identification systems, one can leverage the idea of humanin-the-loop artificial intelligence to achieve high performance with low labor effort.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Compared with fingerprint, palmprint has abundant features and can be captured using common built-in cameras of mobile phones rather than sensitive sensors.

“Compared with fingerprint, palmprint has abundant features and can be captured using common built-in cameras of mobile phones rather than sensitive sensors.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

The choice of a specific biometric recognition solution depends on sensors, usage scenarios, latency, and power consumption.

“In practice, the choice of a specific biometric recognition solution depends on sensors, usage scenarios, latency, and power consumption.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The widely used calibration method is proposed by Z. Zhang in [156] based on planar patterns at a few different orientations, where radial lens distortion is modelled.

“The widely used calibration method is proposed by Z. Zhang in [156] based on planar patterns at a few different orientations, where radial lens distortion is modelled.”

• cites [156] [156] Z. Zhang, 'A flexible new technique for camera calibration,' IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 22, no. 11, pp. 1330-1334, 2000. - sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Nighttime image dehazing, when images are captured in low-light and haze environments, is a more challenging case.

“When images are captured in the low-light and haze environment, it comes to the more challenging case, i.e., nighttime image dehazing.”

May combine two claims.

Depth estimation, pose estimation, VO/VIO, and SLAM constitute the important 3D perceiving ability of AIoT and could be very useful in smart transportation, smart industry, smart agriculture, smart cities, and smart homes.

“Depth estimation, pose estimation, VO/VIO, and SLAM constitute the important 3D perceiving ability of AIoT, which could be very useful in smart transportation [137], smart industry [138], smart agriculture [139], [140], [141], smart cities and homes [142], [143], [144].”

• cites [137] [137] X. Shao, L. Zhang, T. Zhang, Y. Shen, H. Li, and Y. Zhou, 'A tightlycoupled semantic slam system with visual, inertial and surround-view sensors for autonomous indoor parking,' in Proceedings of the 28th ACM International Conference on Multimedia , 2020. - sourceNot checked yet

  Not checked in the free run.

• cites [138] [138] R. Krug, T. Stoyanov, V. Tincani, H. Andreasson, R. Mosberger, G. Fantoni, and A. J. Lilienthal, 'The next step in robot commissioning: Autonomous picking and palletizing,' IEEE Robotics and Automation Letters , vol. 1, no. 1, pp. 546-553, 2016. - sourceNot checked yet

  Not checked in the free run.

• cites [139] [139] F. A. Cheein, G. Steiner, G. P. Paina, and R. Carelli, 'Optimized eif-slam algorithm for precision agriculture mapping based on stems detection,' Computers and Electronics in Agriculture , vol. 78, no. 2, pp. 195-207, 2011. - sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Images captured in a haze environment are in low contrast due to the haze attenuation and scattering effects.

“Images captured in a haze environment are in low contrast due to the haze attenuation and scattering effects.”

• AssumptionRests on an assumption: Haze attenuation and scattering cause low contrast in captured images.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Recovering the clear image from a single hazy input is an ill-posed problem that can be addressed by both prior-based and learning-based methods.

“Recovering the clear image from a single hazy input is also an illposed problem, which can be addressed by both prior-based and learning-based methods [149], [150], [151].”

• cites [149] [149] K. He, J. Sun, and X. Tang, 'Single image haze removal using dark channel prior,' IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 33, no. 12, pp. 2341-2353, 2010. - sourceNot checked yet

  Not checked in the free run.

• cites [150] [150] B. Cai, X. Xu, K. Jia, C. Qing, and D. Tao, 'Dehazenet: An end-to-end system for single image haze removal,' IEEE Transactions on Image Processing , vol. 25, no. 11, pp. 5187-5198, 2016. - sourceNot checked yet

  Not checked in the free run.

• cites [151] [151] J. Zhang and D. Tao, 'Famed-net: A fast and accurate multi-scale endto-end dehazing network,' IEEE Transactions on Image Processing , vol. 29, pp. 72-84, 2019. - sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Deep learning has advanced text spotting by learning more representative features, devising better representation of text proposals, and using large-scale synthetic datasets.

“Deep learning has advanced this area by learning more representative feature [87], devising better representation of text proposals [88], and using large-scale synthetic dataset [89].”

• cites [87] [87] M. Jaderberg, A. Vedaldi, and A. Zisserman, 'Deep features for text spotting,' in Proceedings of the European Conference on Computer Vision , 2014, pp. 512-528. - sourceNot checked yet

  Not checked in the free run.

• cites [88] [88] X. Liu, D. Liang, S. Yan, D. Chen, Y. Qiao, and J. Yan, 'Fots: Fast oriented text spotting with a unified network,' in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , 2018, pp. 5676-5685. - sourceNot checked yet

  Not checked in the free run.

• cites [89] [89] A. Gupta, A. Vedaldi, and A. Zisserman, 'Synthetic data for text localisation in natural images,' in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , 2016, pp. 2315-2324. -Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Cai et al. propose the first deep CNN model for image dehazing, which outperforms traditional prior-based methods by leveraging the powerful representation capacity of CNNs.

“Cai et al. propose the first deep CNN model for image dehazing, which outperforms traditional prior-based methods by leveraging the powerful representation capacity of CNNs [150].”

• cites [150] [150] B. Cai, X. Xu, K. Jia, C. Qing, and D. Tao, 'Dehazenet: An end-to-end system for single image haze removal,' IEEE Transactions on Image Processing , vol. 25, no. 11, pp. 5187-5198, 2016. - sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Low-light images usually contain noises that will be amplified after enhancement, motivating robust Retinex models that account for noise and estimate reflectance, illumination, and noise simultaneously.

“Since lowlight images usually contain noises that will be amplified after enhancement, some robust Retinex models have been proposed to account for noise and estimate reflectance, illumination, and noise simultaneously [147], [148].”

• cites [147] [147] M. Li, J. Liu, W. Yang, X. Sun, and Z. Guo, 'Structure-revealing lowlight image enhancement via robust retinex model,' IEEE Transactions on Image Processing , vol. 27, no. 6, pp. 2828-2841, 2018. - sourceNot checked yet

  Not checked in the free run.

• cites [148] [148] A. Zhu, L. Zhang, Y. Shen, Y. Ma, S. Zhao, and Y. Zhou, 'Zero-shot restoration of underexposed images via robust retinex decomposition,' - in Proceedings of 2020 IEEE International Conference on Multimedia and Expo . IEEE, 2020, pp. 1-6. - sourceNot checked yet

  Not checked in the free run.

• AssumptionRests on an assumption: Low-light images usually contain noise that will be amplified after enhancement.

Probably not a substantial claim: may not pass S3 (load-bearing).

Significant progress in face recognition has been achieved in the last decade mainly owing to deep representation learning and metric learning.

“For face recognition, significant progress has been achieved in the last decade, mainly owing to deep representation learning and metric learning.”

Probably not a substantial claim: may not pass S3 (load-bearing).

It is a typical ill-posed problem to obtain the reflectance and illumination from a single image.

“However, it is a typical ill-posed problem to obtain the reflectance and illumination from a single image.”

Probably not a substantial claim: may not pass S3 (load-bearing).

A representation learning method using triplet loss directly learns discriminative and compact face embedding.

“To address this issue, a representation learning method using triplet loss is proposed to directly learn discriminative and compact face embedding [99].”

• cites [99] [99] F. Schroff, D. Kalenichenko, and J. Philbin, 'Facenet: A unified embedding for face recognition and clustering,' in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , 2015, pp. 815-823. - sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

Person re-identification has vast potential for AIoT applications such as smart security in an uncontrolled and non-contact environment, where other biometric recognition techniques are not applicable.

“Person re-identification has vast potential for AIoT applications such as smart security in an uncontrolled and non-contact environment, where other biometric recognition techniques are not applicable.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Images captured in a low-light environment are in low visibility and difficult to see details due to insufficient incident light or underexposure.

“Images captured in a low-light environment are in low visibility and difficult to see details due to insufficient incident light or underexposure.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Rectification methods can be categorized as camera calibration-based methods and distortion model-based methods.

“The rectification methods can be categorized as camera calibration-based methods and distortion model-based methods.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

In video surveillance, it is necessary to count the crowd in both indoor and outdoor areas and prevent crowd congestion and accident.

“In the video surveillance scenario, it is necessary to count the crowd in both indoor and outdoor areas and prevent crowd congestion and accident.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Face recognition is one of the most widely used perceiving techniques for identity verification and access control in various AIoT applications, such as smart cities and smart homes.

“Face recognition is one of the most widely used perceiving techniques for identity verification and access control in various AIoT applications, e.g., smart cities and smart homes.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Depth estimation from monocular video together with camera pose estimation has recently attracted a lot of attention.

“Recently, depth estimation from monocular video together with camera pose estimation has attracted a lot of attention.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Face landmark detectors are usually lightweight and run very fast, making them useful for latency-sensitive AIoT applications.

“Face landmark detectors are usually lightweight and run very fast, which are very useful for latency-sensitive AIoT applications.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Liu et al. propose an online camera pose optimization method for the surround view system [158], which is composed of several fisheye cameras around the vehicle.

“For example, Liu et al. propose an online camera pose optimization method for the surround view system [158], which is composed of several fisheye cameras around the vehicle.”

• cites [158] [158] X. Liu, L. Zhang, Y. Shen, S. Zhang, and S. Zhao, 'Online camera pose optimization for the surround-view system,' in Proceedings of the 27th ACM International Conference on Multimedia , 2019, pp. 383-391. - sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Top-down human pose estimation methods still dominate the leaderboard of public benchmark datasets like MS COCO, but they are usually slower than bottom-up methods.

“Although top-down methods still dominate the leaderboard of public benchmark datasets like MS COCO 10 , they are usually slower than bottom-up methods [112].”

• cites [112] [112] Z. Cao, T. Simon, S.-E. Wei, and Y. Sheikh, 'Realtime multi-person 2d pose estimation using part affinity fields,' in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , 2017, pp. 7291-7299. - sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Text detection is a different and challenging problem from generic object detection because text may be in variable length and shape depending on the number of characters and their orientation, and the appearance of the same text may change significantly due to fonts, styles, and background context.

“Although text detection is related to generic object detection, it is a different and challenging problem: 1) while generic objects have regular shapes, text may be in variable length and shape depending on the number of characters and their orientation; 2) the appearance of same text may change significantly due to fonts, styles as well as background context.”

May combine two claims.

Synthetic datasets can be used for crowd counting and have demonstrated value for the task, either in a pretraining-finetuning scheme or by domain adaptation.

“Since it is costly to collect and annotate crowd images, synthetic datasets can be used and have demonstrated its value for this task, i.e., either being used in the pretraining-finetuning scheme or by domain adaptation [128].”

• cites [128] [128] Q. Wang, J. Gao, W. Lin, and Y. Yuan, 'Learning from synthetic data for crowd counting in the wild,' in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , 2019, pp. 8198-8207. -Not checked yet

  Not checked in the free run.

May combine two claims.

SLAM can be used for lane departure warning, lane-keeping, and high-precision map construction and update.

“Besides, SLAM can be used for lane departure warning, lanekeeping, high-precision map construction and update [137].”

• cites [137] [137] X. Shao, L. Zhang, T. Zhang, Y. Shen, H. Li, and Y. Zhou, 'A tightlycoupled semantic slam system with visual, inertial and surround-view sensors for autonomous indoor parking,' in Proceedings of the 28th ACM International Conference on Multimedia , 2020. - sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Given two or more fisheye cameras with calibrated parameters, a panorama image can be obtained from their images by image stitching.

“Given two or more fisheye cameras with calibrated parameters, a panorama image can be obtained from their images by image stitching.”

Probably not a substantial claim: may not pass S3 (load-bearing).

End-to-end modeling of text detection and recognition has achieved impressive performance.

“Recently, end-toend modeling of text detection and recognition has achieved impressive performance [90], [91].”

• cites [90] [90] H. Li, P. Wang, and C. Shen, 'Towards end-to-end text spotting with convolutional recurrent neural networks,' in Proceedings of the IEEE International Conference on Computer Vision , 2017, pp. 5238-5246. - sourceNot checked yet

  Not checked in the free run.

• cites [91] [91] J. Liu, Z. Chen, B. Du, and D. Tao, 'Asts: A unified framework for arbitrary shape text spotting,' IEEE Transactions on Image Processing , vol. 29, pp. 5924-5936, 2020. - sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Visual odometry and visual-inertial odometry are always used as the front-end in a SLAM system, where the back-end refers to nonlinear optimization of the pose graph aiming to obtain globally consistent and drift-free pose estimation results.

“VO and VIO are always used ad the front-end in a SLAM system, where the back-end refers to the nonlinear optimization of the pose graph, aiming to obtain globally consistent and drift-free pose estimation results.”

May combine two claims.

Fusing multi-modal data at the feature level is most prevalent, which can be further categorized into early fusion [186], late fusion [185], and fusion at multiple levels [85], [182].

“Among them, fusing multi-modal data at the feature level is most prevalent, which can be further categorized into three groups, i.e., early fusion [186], late fusion [185], and fusion at multiple levels [85], [182].”

• cites [186] [186] Z. Xu, S. Liu, J. Shi, and C. Lu, 'Outdoor rgbd instance segmentation with residual regretting learning,' IEEE Transactions on Image Processing , vol. 29, pp. 5301-5309, 2020. - sourceNot checked yet

  Not checked in the free run.

• cites [185] [185] A. Wang, J. Lu, J. Cai, T.-J. Cham, and G. Wang, 'Large-margin multimodal deep learning for rgb-d object recognition,' IEEE Transactions on Multimedia , vol. 17, no. 11, pp. 1887-1898, 2015. - sourceNot checked yet

  Not checked in the free run.

• cites [85] [85] Z. Chen, J. Zhang, and D. Tao, 'Progressive lidar adaptation for road detection,' IEEE/CAA Journal of Automatica Sinica , vol. 6, no. 3, pp. 693-702, 2019. - sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

RNN, especially LSTM, is used to model long-range dependencies in speech sequence and decode text sequentially [24].

“For example, RNN, especially Long Short-Term Memory (LSTM), is used to model the long-range dependencies in the speech sequence and decode the text sequentially [24].”

• cites [24] [24] H. Sak, A. Senior, and F. Beaufays, 'Long short-term memory recurrent neural network architectures for large scale acoustic modeling,' in Proceedings of the Fifteenth Annual Conference of the International Speech Communication Association , 2014. -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Neural machine translation based on deep learning has made rapid progress in recent years, outperforming traditional statistical MT methods or example-based MT methods by leveraging powerful representation capacity and large-scale training data.

“Neural machine translation (NMT) based on deep learning has made rapid progress in recent years, outperforming the traditional statistical MT methods or example-based MT methods by leveraging the powerful representation capacity and largescale training data.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

RNN processes data in a sequential manner, which is parallel-unfriendly.

“for another, RNN processes data in a sequential manner, which is parallel-unfriendly.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Recent studies on cross-media matching and retrieval try to align both domains semantically by leveraging deep learning, especially adversarial learning [177].

“Recent studies on cross-media matching and retrieval try to align both domains semantically by leveraging deep learning, especially adversarial learning [177].”

• cites [177] [177] B. Wang, Y. Yang, X. Xu, A. Hanjalic, and H. T. Shen, 'Adversarial cross-modal retrieval,' in Proceedings of the 25th ACM International Conference on Multimedia , 2017, pp. 154-162. - sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Traditional ASR models are based on hand-crafted features like cepstral coefficient and HMM [159], which have been revolutionized by deep neural networks for end-to-end modeling without the need of domain knowledge for feature engineering, HMM design, as well as explicit dependency assumption.

“Traditional ASR models are based on hand-crafted features like cepstral coefficient and Hidden Markov Model (HMM) [159], which have been revolutionized by the deep neural network for end-to-end modeling without the need of domain knowledge for feature engineering, HMM design, as well as explicit dependency assumption.”

• cites [159] [159] L. Bahl, P. Brown, P. De Souza, and R. Mercer, 'Maximum mutual information estimation of hidden markov model parameters for speech recognition,' in Proceedings of the IEEE International Conference on Acoustics, Speech, and Signal Processing , vol. 11, 1986, pp. 49-52. -Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

May combine two claims.

The surround view system can capture a 360° view around the vehicle, which is useful in IoV for the advanced driver assistant system and crowd-sourcing high-precision map update.

“The surround view system can capture a 360 ◦ view around the vehicle, which is useful in IoV for the advanced driver assistant system and crowd-sourcing high-precision map update.”

• cites [158] [158] X. Liu, L. Zhang, Y. Shen, S. Zhang, and S. Zhao, 'Online camera pose optimization for the surround-view system,' in Proceedings of the 27th ACM International Conference on Multimedia , 2019, pp. 383-391. - sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

There are three ways of fusing multi-modal data: at the input level [181], at the feature level [185], [186], [85], [182], [183], and at the output level [184].

“There are three ways of fusing multi-modal data, i.e., at the input level [181], at the feature level [185], [186], [85], [182], [183], and at the output level [184], respectively.”

• cites [181] [181] L. Li, B. Qian, J. Lian, W. Zheng, and Y. Zhou, 'Traffic scene segmentation based on rgb-d image and deep learning,' IEEE Transactions on Intelligent Transportation Systems , vol. 19, no. 5, pp. 1664-1669, 2017. - sourceNot checked yet

  Not checked in the free run.

• cites [185] [185] A. Wang, J. Lu, J. Cai, T.-J. Cham, and G. Wang, 'Large-margin multimodal deep learning for rgb-d object recognition,' IEEE Transactions on Multimedia , vol. 17, no. 11, pp. 1887-1898, 2015. -Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [184] [184] J. Liu, Y. Li, S. Song, J. Xing, C. Lan, and W. Zeng, 'Multi-modality multi-task recurrent neural network for online action detection,' IEEE Transactions on Circuits and Systems for Video Technology , vol. 29, no. 9, pp. 2667-2682, 2018. - sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

ASR can be used in smart input system [163], automatic transcription system, smart voice assistant [164], [165], computer-assisted speech rehabilitation and language teaching [166].

“For example, ASR can be used in the smart input system [163], automatic transcription system, smart voice assistant [164], [165], computer-assisted speech rehabilitation and language teaching [166].”

• cites [163] [163] F. Longo, L. Nicoletti, and A. Padovano, 'Ubiquitous knowledge empowers the smart factory: The impacts of a service-oriented digital twin on enterprises' performance,' Annual Reviews in Control , vol. 47, pp. 221-236, 2019. - sourceNot checked yet

  Not checked in the free run.

• cites [164] [164] J. M. Owens, S. B. McLaughlin, and J. Sudweeks, 'On-road comparison of driving performance measures when using handheld and voicecontrol interfaces for mobile phones and portable music players,' SAE International Journal of Passenger Cars-Mechanical Systems , vol. 3, no. 2010-01-1036, pp. 734-743, 2010. - sourceNot checked yet

  Not checked in the free run.

• cites [166] [166] K. Angelov, B. Bringert, and A. Ranta, 'Speech-enabled hybrid multilingual translation for mobile devices,' in Proceedings of the Demonstrations at the 14th Conference of the European Chapter of the Association for Computational Linguistics , 2014, pp. 41-44. - sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Network compression is an effective technique to improve the efficiency of DNNs for AIoT applications with limited computational budgets.

“Network compression is an effective technique to improve the efficiency of DNNs for AIoT applications with limited computational budgets.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Attention mechanism is used to attend to all source words (global attention) or only part of them (local attention) when decoding at each step of RNN [173], [174], [175].

“Later, attention mechanism is used to attend to all source words (i.e., global attention) or only part of them (i.e., local attention) when decoding at each step of RNN [173], [174], [175].”

• cites [173] [173] D. Bahdanau, K. Cho, and Y. Bengio, 'Neural machine translation by jointly learning to align and translate,' in Proceedings of the 3rd International Conference on Learning Representations , 2015. - sourceNot checked yet

  Not checked in the free run.

• cites [174] [174] Y. Wu, M. Schuster, Z. Chen, Q. V. Le, M. Norouzi, W. Macherey, M. Krikun, Y. Cao, Q. Gao, K. Macherey et al. , 'Google's neural machine translation system: Bridging the gap between human and machine translation,' arXiv preprint arXiv:1609.08144 , 2016. - sourceNot checked yet

  Not checked in the free run.

• cites [175] [175] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, 'Attention is all you need,' in Advances in Neural Information Processing Systems , 2017, pp. 59986008. -Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing).

Connectionist temporal classification is proposed to address the need for presegmentation by directly maximizing probabilities of correct label sequence in a differentiable way [92].

“To address the first issue, the connectionist temporal classification is proposed by directly maximizing the probabilities of the correct label sequence in a differentiable way [92].”

• cites [92] [92] A. Graves, S. Fern´ andez, F. Gomez, and J. Schmidhuber, 'Connectionist temporal classification: Labelling unsegmented sequence data with recurrent neural networks,' in Proceedings of the 23rd International Conference on Machine Learning , 2006, pp. 369-376. - sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Speaker recognition is one of the important means for identity identification, which has many applications in various AIoT domains, for example, automatic transcription system for multi-person meetings, personalized recommendation by smart voice assistants [170], and audio forensics [171].

“Speaker recognition is one of the important means for identity identification, which has many applications in various AIoT domains, for example, automatic transcription system for multi-person meetings, personalized recommendation by smart voice assistants [170], and audio forensics [171].”

• cites [170] [170] H. Feng, K. Fawaz, and K. G. Shin, 'Continuous authentication for voice assistants,' in Proceedings of the 23rd Annual International Conference on Mobile Computing and Networking , 2017, pp. 343-355. - sourceNot checked yet

  Not checked in the free run.

• cites [171] [171] A. Ross, S. Banerjee, and A. Chowdhury, 'Security in smart cities: A brief review of digital forensic schemes for biometric data,' Pattern Recognition Letters , 2020. - sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Video coding for machines can facilitate down-stream tasks given compact coded features as well as support human-in-the-loop inspection and intervention, therefore having vast potential for supporting many AIoT applications.

“It can facilitate down-stream tasks given the compact coded features as well as support human-in-the-loop inspection and intervention, therefore having vast potential for supporting many AIoT applications.”

• cites [188] [188] L.-Y. Duan, J. Liu, W. Yang, T. Huang, and W. Gao, 'Video coding for machines: A paradigm of collaborative compression and intelligent analytics,' arXiv preprint arXiv:2001.03569 , 2020. -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Speaker recognition can be integrated with face recognition for access control.

“Besides, speaker recognition can be integrated with face recognition for access control.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Unsupervised NMT has been studied, which could be trained on monolingual corpora.

“Recently, unsupervised NMT has also been studied, which could be trained on monolingual corpora.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Transformer architecture is devised using scaled dot-product attention and multi-head attention to mitigate the parallel-unfriendly issue of RNN [160].

“To mitigate the other issue, the Transformer architecture is devised using scaled dot-product attention and multi-head attention [160].”

• cites [160] [160] L. Dong, S. Xu, and B. Xu, 'Speech-transformer: A no-recurrence sequence-to-sequence model for speech recognition,' in Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing , 2018, pp. 5884-5888. - sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

A multi-branch group fusion module is proposed to fuse features from RGB and thermal infrared images at different levels in [182], since semantic information and visual details differ at different levels.

“For example, a multi-branch group fusion module is proposed to fuse features from RGB and thermal infrared images at different levels in [182], since the semantic information and visual details differ at different levels.”

• cites [182] [182] Q. Zhang, N. Huang, L. Yao, D. Zhang, C. Shan, and J. Han, 'Rgbt salient object detection via fusing multi-level cnn features,' IEEE Transactions on Image Processing , vol. 29, pp. 3321-3335, 2019. - sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

A novel idea named video coding for machines is proposed recently [188], which attempts to bridge the gap between feature coding for machine vision and video coding for human vision.

“It is noteworthy that a novel idea named video coding for machines is proposed recently [188], which attempts to bridge the gap between feature coding for machine vision and video coding for human vision.”

• cites [188] [188] L.-Y. Duan, J. Liu, W. Yang, T. Huang, and W. Gao, 'Video coding for machines: A paradigm of collaborative compression and intelligent analytics,' arXiv preprint arXiv:2001.03569 , 2020. -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Unsupervised representation learning has shown promising performance for many down-stream language tasks by learning context-aware and informative embeddings, e.g., BERT [29].

“Unsupervised representation learning has shown promising performance for many down-stream language tasks by learning context-aware and informative embeddings, e.g., BERT [29].”

• cites [29] [29] J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, 'Bert: Pre-training of deep bidirectional transformers for language understanding,' in Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , 2019, pp. 4171-4186. -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Previously speaker recognition methods are dominated by the i-vector representation and probabilistic linear discriminant analysis framework [168], where i-vector refers to extracting low-dimensional speaker embeddings from sufficient statistics.

“Previously speaker recognition methods are dominated by the i-vector representation and probabilistic linear discriminant analysis framework [168], where i-vector refers to extracting low-dimensional speaker embeddings from sufficient statistics.”

• cites [168] [168] N. Dehak, P. J. Kenny, R. Dehak, P. Dumouchel, and P. Ouellet, 'Frontend factor analysis for speaker verification,' IEEE Transactions on Audio, Speech, and Language Processing , vol. 19, no. 4, pp. 788-798, 2010. - sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Recently, several end-to-end deep speaker recognition models have been devised [169], achieving better performance than i-vector baselines.

“Recently, several end-to-end deep speaker recognition models have been devised [169], achieving better performance than i-vector baselines.”

• cites [169] [169] J. S. Chung, A. Nagrani, and A. Zisserman, 'Voxceleb2: Deep speaker recognition,' Proc. Interspeech 2018 , pp. 1086-1090, 2018. - sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

The authors in [85] leverage residual learning to fuse multilevel RGB image features and Lidar features via a residual structure in a cascaded manner.

“Besides, the authors in [85] leverage the residual learning idea to fuse the multilevel RGB image features and Lidar features via a residual structure in a cascaded manner.”

• cites [85] [85] Z. Chen, J. Zhang, and D. Tao, 'Progressive lidar adaptation for road detection,' IEEE/CAA Journal of Automatica Sinica , vol. 6, no. 3, pp. 693-702, 2019. - sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The computing paradigm for ASR could be on-device edge computing, fog computing with a powerful computing device and sound pickup system, or cloud computing with acceptable latency.

“The computing paradigm could be on-device edge computing (e.g., off-line mode of smart voice assistant), fog computing with a powerful computing device and sound pickup system (e.g., automatic transcription system for conferences), as well as the cloud computing with acceptable latency (e.g., on-line mode of smart voice assistant).”

Probably not a substantial claim: may not pass S3 (load-bearing).

Leveraging BERT as contextual embedding has been proved useful for NMT by borrowing informative context from the pre-trained model [176].

“For example, leveraging BERT as contextual embedding has been proved useful for NMT by borrowing informative context from the pre-trained model [176].”

• cites [176] [176] J. Zhu, Y. Xia, L. Wu, D. He, T. Qin, W. Zhou, H. Li, and T. Liu, 'Incorporating bert into neural machine translation,' in Proceedings of the International Conference on Learning Representations , 2019. - sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Multimedia coding has been advanced by deep learning [187].

“Another research topic that is close to AIoT is multimedia coding, which has also been advanced by deep learning [187].”

• cites [187] [187] D. Liu, Y. Li, J. Lin, H. Li, and F. Wu, 'Deep learning-based video coding: A review and a case study,' ACM Computing Surveys , vol. 53, no. 1, pp. 1-35, 2020. - sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Attention can be useful for learning context features related to the target and achieve joint alignment and translation, showing better performance for long sentences.

“Attention can be useful for learning context features related to the target and achieve joint alignment and translation, showing better performance for long sentences.”

• cites [173] [173] D. Bahdanau, K. Cho, and Y. Bengio, 'Neural machine translation by jointly learning to align and translate,' in Proceedings of the 3rd International Conference on Learning Representations , 2015. - sourceNot checked yet

  Not checked in the free run.

• cites [174] [174] Y. Wu, M. Schuster, Z. Chen, Q. V. Le, M. Norouzi, W. Macherey, M. Krikun, Y. Cao, Q. Gao, K. Macherey et al. , 'Google's neural machine translation system: Bridging the gap between human and machine translation,' arXiv preprint arXiv:1609.08144 , 2016. - sourceNot checked yet

  Not checked in the free run.

• cites [175] [175] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, 'Attention is all you need,' in Advances in Neural Information Processing Systems , 2017, pp. 59986008. -Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing).

Disentangled representation learning has been proposed to address modality-exclusive information by maximizing mutual information between feature embeddings from different modalities and separating modality-exclusive features [178].

“To address this issue, disentangled representation learning has been proposed [178], which tries to maximize the mutual information between feature embeddings from different modalities and separate modality-exclusive features from them.”

• cites [178] [178] W. Guo, H. Huang, X. Kong, and R. He, 'Learning disentangled representation for cross-modal retrieval with deep mutual information estimation,' in Proceedings of the 27th ACM International Conference on Multimedia , 2019, pp. 1712-1720. - sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Presegmenting training sequences is needed so that classification loss can be calculated at each point in the sequence independently.

“However, for one thing, extra effort is needed to presegment training sequences so that the classification loss can be calculated at each point in the sequence independently”

Probably not a substantial claim: may not pass S3 (load-bearing).

The prevalent architecture for NMT is the encoder-decoder [172].

“The prevalent architecture for NMT is the encoder-decoder [172].”

• cites [172] [172] K. Cho, B. van Merrienboer, D. Bahdanau, and Y. Bengio, 'On the properties of neural machine translation: Encoder-decoder approaches,' in Proceedings of the 8th Workshop on Syntax, Semantics and Structure in Statistical Translation , 2014. -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Real-time ASR systems have been developed either using on-device computing or in a cloud-assisted manner [161], [162].

“Recently, real-time ASR systems have been developed either using on-device computing or in a cloud-assisted manner [161], [162].”

• cites [161] [161] Y.-S. Chang, S.-H. Hung, N. J. Wang, and B.-S. Lin, 'Csr: A cloudassisted speech recognition service for personal mobile device,' in Proceedings of the International Conference on Parallel Processing , 2011, pp. 305-314. - sourceNot checked yet

  Not checked in the free run.

• cites [162] [162] Y. He, T. N. Sainath, R. Prabhavalkar, I. McGraw, R. Alvarez, D. Zhao, D. Rybach, A. Kannan, Y. Wu, R. Pang et al. , 'Streaming end-to-end speech recognition for mobile devices,' in Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing , 2019, pp. 6381-6385. - sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

By using them with RGB images as input, cross-modal perceiving has attracted increasing attention in real-world applications, e.g., scene parsing for autonomous driving [181], [85], object detection and tracking in low-light scenarios [182], [183], and action recognition [184].

“By using them with RGB images as input, cross-modal perceiving has attracted increasing attention in real-world applications, e.g., scene parsing for autonomous driving [181], [85], object detection and tracking in low-light scenarios [182], [183], and action recognition [184].”

• cites [181] [181] L. Li, B. Qian, J. Lian, W. Zheng, and Y. Zhou, 'Traffic scene segmentation based on rgb-d image and deep learning,' IEEE Transactions on Intelligent Transportation Systems , vol. 19, no. 5, pp. 1664-1669, 2017. - sourceNot checked yet

  Not checked in the free run.

• cites [182] [182] Q. Zhang, N. Huang, L. Yao, D. Zhang, C. Shan, and J. Han, 'Rgbt salient object detection via fusing multi-level cnn features,' IEEE Transactions on Image Processing , vol. 29, pp. 3321-3335, 2019. - sourceNot checked yet

  Not checked in the free run.

• cites [184] [184] J. Liu, Y. Li, S. Song, J. Xing, C. Lan, and W. Zeng, 'Multi-modality multi-task recurrent neural network for online action detection,' IEEE Transactions on Circuits and Systems for Video Technology , vol. 29, no. 9, pp. 2667-2682, 2018. - sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

ASR is useful in many AIoT applications because speech is one of the most important non-contact interaction modes.

“ASR is very useful in many AIoT applications since speech is one of the most important non-contact interaction modes.”

• AssumptionRests on an assumption: Speech is one of the most important non-contact interaction modes.

Probably not a substantial claim: may not pass S3 (load-bearing).

Network compression mainly includes four kinds of techniques: network pruning, network quantization, low-rank factorization, and knowledge distillation.

“It mainly included four kinds of techniques, i.e., network pruning, network quantization, low-rank factorization, and knowledge distillation.”

Probably not a substantial claim: may not pass S3 (load-bearing).

A speaker recognition system is composed of three modules: speech acquisition and production, feature representation and selection, and pattern matching and classification [167].

“A speaker recognition system is composed of three modules, i.e., speech acquisition and production, feature representation and selection, and pattern matching and classification [167].”

• cites [167] [167] J. P. Campbell, 'Speaker recognition: A tutorial,' Proceedings of the IEEE , vol. 85, no. 9, pp. 1437-1462, 1997. -Not checked yet

  Not checked in the free run.

May combine two claims.

FL can be used to improve the perceiving and learning ability of AIoT devices such as connected vehicles for autonomous driving [199] and wearable devices for health monitoring [222], [197].

“For example, FL can be used to improve the perceiving and learning ability of AIoT devices such as connected vehicles for autonomous driving [199] and wearable devices for health monitoring [222], [197].”

• cites [199] [199] X. Liang, Y. Liu, T. Chen, M. Liu, and Q. Yang, 'Federated transfer reinforcement learning for autonomous driving,' arXiv preprint arXiv:1910.06001 , 2019. - sourceNot checked yet

  Not checked in the free run.

• cites [222] [222] T. S. Brisimi, R. Chen, T. Mela, A. Olshevsky, I. C. Paschalidis, and W. Shi, 'Federated learning of predictive models from federated electronic health records,' International Journal of Medical Informatics , vol. 112, pp. 59-67, 2018. - sourceNot checked yet

  Not checked in the free run.

• cites [197] [197] A. Saeed, F. D. Salim, T. Ozcelebi, and J. Lukkien, 'Federated self-supervised learning of multi-sensor representations for embedded intelligence,' IEEE Internet of Things Journal , 2020. - sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Domain adaptation has been used in many computer vision tasks including classification, object detection, and especially semantic segmentation [37], where obtaining dense pixel-level annotations in target domain is costly and time-consuming.

“Domain adaptation has been used in many computer vision tasks including classification, object detection, and especially semantic segmentation [37], where obtaining the dense pixellevel annotations in the target domain is costly and timeconsuming.”

• cites [37] [37] Q. Zhang, J. Zhang, W. Liu, and D. Tao, 'Category anchor-guided unsupervised domain adaptation for semantic segmentation,' in Advances in Neural Information Processing Systems , 2019, pp. 435-445. - sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

The architecture of FL usually consists of a central server (or collaborator) and many distributed client devices.

“The architecture of FL usually consists of a central server (or collaborator) and many distributed client devices.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Most NAS methods fall into three categories: evolutionary methods, reinforcement learning-based methods, and gradient-based methods.

“Most NAS methods fall into three categories, i.e., evolutionary methods, reinforcement learning-based methods, and gradient-based methods.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

It has been almost the de facto practice to fine-tune models pre-trained on ImageNet in different down-stream tasks, e.g., object detection and semantic segmentation, for faster convergence and better generalization.

“For example, it has been almost the de facto practice to fine-tune the models pre-trained on ImageNet in different down-stream tasks, e.g., object detection and semantic segmentation, for faster convergence and better generalization.”

Probably not a substantial claim: may not pass S3 (load-bearing).

A mobile domain adaptation framework is proposed for edge computing in AIoT [38] by knowledge distillation from teacher model on server to student model on edge device.

“Recently, a mobile domain adaptation framework is proposed for edge computing in AIoT [38] by knowledge distillation from the teacher model on the server to the student model on the edge device.”

• cites [38] [38] J. Yang, H. Zou, S. Cao, Z. Chen, and L. Xie, 'Mobileda: Towards edge domain adaptation,' IEEE Internet of Things Journal , 2020. -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Network quantization compresses the original network by reducing the number of bits required for each weight, which significantly reduces memory use and float point operations with a slight loss of accuracy.

“Network quantization compresses the original network by reducing the number of bits required for each weight, which significantly reduces memory use and float point operations with a slight loss of accuracy.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

DNN has a strong representation capacity for learning compact and discriminative feature representation from high-dimensional image and video data, which enables RL to deal with previously intractable problems by learning better policy and value function.

“DNN has a strong representation capacity for learning compact and discriminative feature representation from the high-dimensional image and video data, which enables RL to deal with previously intractable problems by learning better policy and value function.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Federated learning was initially proposed to address the learning problem in which datasets are distributed across multiple devices and not allowed to be leaked to others [220].

“Federated learning was initially proposed to address the learning problem in which datasets are distributed across multiple devices and not allowed to be leaked to others [220].”

• cites [220] [220] J. Koneˇ cn` y, H. B. McMahan, F. X. Yu, P. Richt´ arik, A. T. Suresh, and D. Bacon, 'Federated learning: Strategies for improving communication efficiency,' arXiv preprint arXiv:1610.05492 , 2016. - sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Self-supervised learning learns discriminative visual representation by solving predefined pretext tasks [194].

“Another hot research sub-area is self-supervised learning, which learns discriminative visual representation by solving predefined pretext tasks [194].”

• cites [194] [194] L. Jing and Y. Tian, 'Self-supervised visual feature learning with deep neural networks: A survey,' IEEE Transactions on Pattern Analysis and Machine Intelligence , 2020. - sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

SimCLR defines a context-based contrasting task for self-supervised learning [34], obtaining comparable performance as fully supervised models.

“For example, the recently proposed SimCLR method defines a context-based contrasting task for self-supervised learning [34], obtaining comparable performance as fully supervised models.”

• cites [34] [34] T. Chen, S. Kornblith, M. Norouzi, and G. Hinton, 'A simple framework for contrastive learning of visual representations,' arXiv preprint arXiv:2002.05709 , 2020. - sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Gradient-based NAS methods adopt a continuous relaxation of architecture representation, so optimization can be conducted in continuous space using gradient descent, which is orders of magnitude faster.

“By contrast, gradient-based methods adopt a continuous relaxation of the architecture representation. Therefore, the optimization of neural architecture can be conducted in a continuous space using gradient descent, which is orders of magnitude faster.”

May combine two claims.

FL offers a general learning framework to use massive distributed and isolated data while preserving data security and privacy, which is particularly appealing to AIoT applications in the context of edge computing [221], [197].

“FL offers a general learning framework to use massive distributed and isolated data while preserving data security and privacy, which is particularly appealing to AIoT applications in the context of edge computing [221], [197].”

• cites [221] [221] Q. Wu, K. He, and X. Chen, 'Personalized federated learning for intelligent iot applications: A cloud-edge based framework,' IEEE Computer Graphics and Applications , 2020. - sourceNot checked yet

  Not checked in the free run.

• cites [197] [197] A. Saeed, F. D. Salim, T. Ozcelebi, and J. Lukkien, 'Federated self-supervised learning of multi-sensor representations for embedded intelligence,' IEEE Internet of Things Journal , 2020. - sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

There are two main groups of DRL methods including deep Q-network [210] and policy gradient-based methods such as asynchronous advantage actor-critic [211].

“There are two main groups of DRL methods including deep Q-network [210] and policy gradient-based methods such as asynchronous advantage actor-critic [211].”

• cites [210] [210] V. Mnih, K. Kavukcuoglu, D. Silver, A. A. Rusu, J. Veness, M. G. Bellemare, A. Graves, M. Riedmiller, A. K. Fidjeland, G. Ostrovski et al. , 'Human-level control through deep reinforcement learning,' Nature , vol. 518, no. 7540, pp. 529-533, 2015. - sourceNot checked yet

  Not checked in the free run.

• cites [211] [211] V. Mnih, A. P. Badia, M. Mirza, A. Graves, T. Lillicrap, T. Harley, D. Silver, and K. Kavukcuoglu, 'Asynchronous methods for deep reinforcement learning,' in Proceedings of the International Conference on Machine Learning , 2016, pp. 1928-1937. -Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing).

Transfer learning does not require the data distributions of the source and target domains to be identical.

“Note that different from the aforementioned semi-supervised learning where labeled and unlabeled data are usually drawn from the same distribution, transfer learning does not require the data distributions of the source and the target domains to be identical.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

In a recent study [36], a computational taxonomic map is discovered for transfer learning between twenty-six visual tasks, providing valuable empirical insights, e.g., what tasks transfer well to other target tasks, and how to reuse supervision among related tasks to reduce demand for labeled data while achieving same performance.

“In a recent study [36], a computational taxonomic map is discovered for transfer learning between twenty-six visual tasks, providing valuable empirical insights, e.g., what tasks transfer well to other target tasks, and how to reuse supervision among related tasks to reduce the demand for labeled data while achieving same performance.”

• cites [36] [36] A. R. Zamir, A. Sax, W. Shen, L. J. Guibas, J. Malik, and S. Savarese, 'Taskonomy: Disentangling task transfer learning,' in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , 2018, pp. 3712-3722. -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Network pruning can be carried out at different levels of granularity, e.g., weight pruning, neuron pruning, filter pruning, and channel pruning based on the magnitude of weights or responses calculated by L1/L2 norm [189].

“Network pruning can be carried out at different levels of granularity, e.g., weight pruning, neuron pruning, filter pruning, and channel pruning based on the magnitude of weights or responses calculated by L 1 /L 2 norm [189].”

• cites [189] [189] Y. He, X. Zhang, and J. Sun, 'Channel pruning for accelerating very deep neural networks,' in Proceedings of the IEEE International Conference on Computer Vision , 2017, pp. 1389-1397. -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Rapid progress has been made by incorporating deep learning in RL (i.e., DRL) [209].

“Recently, rapid progress has been made by incorporating deep learning in RL (i.e., DRL) [209].”

• cites [209] [209] K. Arulkumaran, M. P. Deisenroth, M. Brundage, and A. A. Bharath, 'Deep reinforcement learning: A brief survey,' IEEE Signal Processing Magazine , vol. 34, no. 6, pp. 26-38, 2017. -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Various GAN models have been proposed which can generate high-resolution and visually realistic images from random vectors.

“Recently, various GAN models have been proposed which can generate high-resolution and visually realistic images from random vectors.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

While actor-critic based methods directly optimize cumulative reward, a surrogate objective function can be used to address distribution shift problem in policy gradient based RL methods.

“While actor-critic based methods directly optimize the cumulative reward, a surrogate objective function can be used to address the distribution shift problem in policy gradient based RL methods.”

Probably not a substantial claim: may not pass S2 (checkable).

Empowering things with learning ability is important for AIoT so that it can update and evolve in response to variations.

“Thereby, empowering things with learning ability is important for AIoT so that it can update and evolve in response to the variations.”

May combine two claims.

DRL can empower things with the ability to interact with and adapt to the dynamic world, making it useful in many AIoT applications, such as autonomous driving in smart transportation [203], [199], 3D-landmark detection of CT scans [212] and robot control [213] in smart healthcare, course recommendation in smart education [41], real-time scheduling for smart factory [214], load scheduling in smart

“DRL can empower things with the ability to interact with and adapt to the dynamic world, making it useful in many AIoT applications (Figure 9), such as autonomous driving in smart transportation [203], [199], 3D-landmark detection of CT scans [212] and robot control [213] in smart healthcare, course recommendation in smart education [41], real-time scheduling for smart factory [214], load”

• cites [203] [203] X. Pan, Y. You, Z. Wang, and C. Lu, 'Virtual to real reinforcement learning for autonomous driving,' arXiv preprint arXiv:1704.03952 , 2017. - sourceNot checked yet

  Not checked in the free run.

• cites [212] [212] F.-C. Ghesu, B. Georgescu, Y. Zheng, S. Grbic, A. Maier, J. Hornegger, and D. Comaniciu, 'Multi-scale deep reinforcement learning for realtime 3d-landmark detection in ct scans,' IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 41, no. 1, pp. 176-189, 2017. -Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [41] [41] J. Zhang, B. Hao, B. Chen, C. Li, H. Chen, and J. Sun, 'Hierarchical reinforcement learning for course recommendation in moocs,' in Proceedings of the AAAI Conference on Artificial Intelligence , vol. 33, 2019, pp. 435-442. - sourceNot checked yet

  Not checked in the free run.

May combine two claims.

In real-world AIoT systems, there are always many related tasks involved, e.g., object detection and tracking, and semantic segmentation in video surveillance. Therefore, finding transfer learning dependencies across these tasks and leveraging such prior knowledge to learn better models are of practical value for AIoT [199], [200], [201], [202].

“In real-world AIoT systems, there are always many related tasks involved, e.g., object detection and tracking, and semantic segmentation in video surveillance. Therefore, finding the transfer learning dependencies across these tasks and leveraging such prior knowledge to learn better models, are of practical value for AIoT [199], [200], [201], [202].”

• cites [199] [199] X. Liang, Y. Liu, T. Chen, M. Liu, and Q. Yang, 'Federated transfer reinforcement learning for autonomous driving,' arXiv preprint arXiv:1910.06001 , 2019. - sourceNot checked yet

  Not checked in the free run.

• cites [200] [200] M. D'Incecco, S. Squartini, and M. Zhong, 'Transfer learning for nonintrusive load monitoring,' IEEE Transactions on Smart Grid , vol. 11, no. 2, pp. 1419-1429, 2019. - sourceNot checked yet

  Not checked in the free run.

• cites [201] [201] S. Bargoti and J. Underwood, 'Deep fruit detection in orchards,' in Proceedings of the IEEE International Conference on Robotics and Automation , 2017, pp. 3626-3633. - sourceNot checked yet

  Not checked in the free run.

• AssumptionRests on an assumption: In real-world AIoT systems, there are always many related tasks involved.