The challenge

• Objectives:

Designing, implementation, and operation of integral solutions for building “Predictive and Learning Approaches based on Distributed-to-Centralized Machine Learning (ML) and Artificial Intelligence (AI) in Management of Large-Scale Internet of Things (IoT) networks in Smart Cities.”

• Challenges:

Why is it necessary to consider and extend ML and AI techniques at the edge of IoT networks?

1. Data Growth in terms of “Size,” “Time,” and “Scale” (Large-Scale ICT and its Data Management): Exponential growth of city-data in widely distributed storage media of Smart Cities, as data is the main ingredient for ML and AI techniques and algorithms;
2. Data Privacy: Citizens and data stakeholders may not be willing to share their information in a public data storage (e.g., Cloud technologies platform);
3. General Data Protection Regulation (GDPR): GDPR is a regulation in the EU law on data protection and privacy. GDPR also addresses the transfer of personal data outside the EU and EEA areas;
4. Cybersecurity Concerns: In case of an attack at the centralized platform place (Cloud technologies) or multi-attacks at the IoT devices network will occur, how can we collect and send data/datasets to a centralized platform place to extract our knowledge requirements through data analysis/analytic and ML and AI techniques?
5. Complexities of running and managing complex ML and AI techniques at the edge of networks: Due to computational and memory limitations of IoT devices and low processing abilities at the edge of the Smart City networks, IoT devices are often not capable of running and managing complex ML and AI techniques and algorithms at the edge of the Smart City networks.
6. Cost of Cloud storage service: In most cases having Cloud storage can be quite expensive. Being charged year after year for a monthly or annual subscription can add up. Cloud storage providers know that customers will pay those high fees to have their data backed up and ultimately to have peace of mind.

• Introduction:

Smart Cities are known to be complex, large scale, and distributed environments that require growing use of IoT devices in various domains, such as transportation and healthcare. A Smart City is the result of the production of massive volumes of city-data from sensor networks and other physical data sources to non-physical data sources in the cities. This produced city-data has been stored in local data storage media in a city. Based on Cloud computing technologies, the produced city-data may shift from distributed local data storage media in a city to centralized data storage media (e.g., Cloud data storage media) somewhere outside the city in most cases. This data movement will be organized through particular update mechanism policies by the city manager. With regard to focus on designing “Predictive and Learning Approaches based on D2C-ML and AI in Management of Large-Scale IoT networks in Smart Cities,” data is the main ingredient for ML and AI techniques and algorithms. So, in the Large-Scale distributed IoT system in Smart Cities, before focusing on ML and AI techniques, it is essential to effectively and securely organize widely distributed IoT devices networks through a comprehensive ICT and its data management architecture (e.g., Distributed-to-Centralized Data Management: A New Sense of Large-Scale ICT Management of Smart City IoT Networks).

ML and AI algorithms and techniques are overgrowing and have been continuously changing and improving smart predictive approaches through different use cases, domains, and scenarios, such as Smart Cities. However, one of the main limiting factors which prevent many ML tasks is the need for huge and diverse training datasets. In Smart Cities, widely distributed IoT devices networks are continuously capturing various environmental city events and producing many data/datasets. Collecting many datasets may provide several facilities for the data stakeholders in the Smart Cities, but several challenges raise when the exponential growth of available city-data is stored in the local and centralized data storage media. Also, due to the computational and memory limitations of IoT devices and low processing abilities at the edge of the Smart City networks, IoT devices are often not capable of running and managing complex ML and AI techniques and algorithms at the edge of the Smart City networks. Finally, sometimes citizens and data stakeholders may not like to share their personal information with others in a public platform (e.g., Cloud technologies).

Recently new calls have been made to design large-scale IoT networks management in the Smart Cities from a very small scale (distributed) to a large scale (centralized) of the city. Edge-to-Cloud computing orchestration may offer a splendid solution to manage Large-Scale IoT networks in Smart Cities, such as data, resources, and software/services, as well as making it possible to apply complex ML and AL techniques and algorithms at the edge of networks. This solution also is useful to create and train datasets from Edge to Cloud concerning the privacy of city-data and other related issues.

Edge-to-Cloud orchestration may be put forward with different architectural layers in Smart Cities, such as Fog and cloudlet. Through different architectural layers from Edge-to-Cloud orchestration, there are several opportunities for building predictive model approaches based on distributed-to-centralized learning approaches to train the datasets through different ML and AI techniques. Therefore, this makes it possible to predict and manage different requirements of Smart Cities, such as resource allocation and consumption, and intelligent cyber-attacks predication.

In the case of the combined and hierarchical computing platform and architecture (i.e., Fog-to-Cloud or Fog-to-cloudlet-to-Cloud) from the distributed-to-centralized schema, it is necessary to define the architectural framework for predicting the Smart Cities requirements based on various ML and AI techniques. This framework makes numerous facilities to analyze varieties of datasets from different business domains. Therefore, it is essential to define and design that kind of architectural framework for combined Edge to Cloud technologies systems to manage and predict Smart Cities’ requirements, which is a yet demanding specific effort from the research community. Currently, two main approaches can be applied through different architectural layers of Edge-to-Cloud orchestration to train varieties of datasets as well as using different ML and AI techniques, as described below:

1. 1. Replicated Learning approach: Each IoT node/device produces data – which must be sent to the centralized node for training and prediction. Therefore, the centralized node receives and aggregates all data from IoT nodes and builds (a) training model(s), and only that model(s) is(are) shared from the centralized server to all distributed IoT nodes. In summary, replicated learning enables Edge devices to share their local data to the centralized node collaboratively, and the centralized node learns a machine learning model but keeps all the data on the storage media of the centralized node, instead of storing data on the local distributed nodes. Replicated Learning may come with the following advantages if the effective middleware platform/controller will be designed in between Edge-to-Cloud computing technologies:
      • • It may make it possible to build more accurate ML models that are closer to the city-data sources in comparison with building ML models in the Cloud computing platform;
        • It may improve the predication level of resources closer to the city-data sources;
        • It may be possible to store city-data in the middleware platform instead of sending to the Cloud data storage;
        • It improves data privacy efficiency on the middleware platform;
        • This approach optimizes system bandwidth and data availability between geographically distributed data centers in the middleware platform.

       

   2. Federated Learning approach: Federated learning is a family of ML algorithms that has the core idea: a connected network exists in which there is a central server node. Each of the IoT nodes/devices creates data – that must be used for training and prediction. Each node trains a local model, and only that model is shared with the server, not the data. In summary, Federated learning enables Edge devices to learn an ML model collaboratively but keeping all the data on the IoT devices itself, instead of moving data to the Cloud computing technologies. Federated Learning has the following advantages:

• • • Ability to build more accurate models faster;
    • Low latency during inference;
    • Privacy-preserving;
    • Improved energy efficiency of the devices.

Organization

Submission

Program

Start the video-workshop at 18:00 – Japan Standard Time (Standard Time), 08 January 2021 (Friday)

18:00 – 18.10 Welcome message by the Program Chair (Amir Sinaeepourfard, Norwegian University of Science and Technology (NTNU), Norway)

18.10 – 18.30 Keynote Talk

Presenter: Vitor Barbosa Souza, Ph.D., Informatics Department (DPI), Universidade Federal de Vicosa – UFV, Vicosa-MG, Brazil

Title: “Improving the availability of fog resources by means of a smart handover of mobile devices across distinct fog domains in a smart city”

Abstract: Recently, computer network architectures have presented a great evolution largely influenced by the introduction of millions of edge devices acting both as data producers and consumers, the deployment of applications demanding real-time broadband communication, and the popularization of mobile computing requiring strict QoS provisioning and reduced service disruption. Fog computing is a novel architecture focusing on addressing several related issues through the employment of computing resources at the edge to deploy highly virtualized micro data centers located close to the end-user. In such architecture, the effective distribution of computing, storage, and control, aiming at the optimal utilization of available resources, is mandatory. Thus, since locality is an important fog characteristic, frequent inter-fog domain handovers is an open challenge since it may result in frequent task migrations besides preventing the continuous availability of shared edge resources. This talk discusses the employment of machine learning strategies to maximize the availability of resources at fog domains by minimizing the frequency of handovers in vehicular scenarios by means of smart placement of resources within fog domains.

18.30 – 18.45 Accepted Paper Presentation

Title: “Macroscopic Traffic Stream Variables Prediction with Weather Impact Using Hybrid CNN-LSTM model”

Authors: Archana Nigam (Dhirubhai Ambani Institute of Information and Communication Technology Gandhinagar, India), Sanjay Srivastava (Dhirubhai Ambani Institute of Information and Communication Technology Gandhinagar, India)

Abstract: Accurate prediction of the macroscopic traffic stream variables such as speed and flow is important for traffic operation and management in an intelligent transportation system. Adverse weather conditions like fog, snow, and rainfall affect the driver’s visibility, road capacity, and mobility. The accurate prediction of the traffic stream variables in adverse weather conditions is challenging because of the non-linear and complex characteristics of the traffic stream and spatiotemporal correlation between traffic and weather variables. Prolonged heavy rain causes massive waterlogging in developing countries due to weak drainage systems, narrow streets, and encroachment, further affecting these traffic stream variables. Snow reduces the road capacity as much as waterlogging does. Prolonged snowfall creates a thick layer on the road, which affects the traffic stream variables. Traffic data has a high spatial and temporal
resolution compared to weather data, which makes the problem more challenging. In this paper, we define a soft temporal threshold to capture the prolonged impact of weather variables. To capture the traffic and weather data’s spatiotemporal and temporal features, we propose a hybrid CNN-LSTM model. To validate model performance, data from San Diego and Minneapolis Minnesota Twin city are used. The test experiments show that the hybrid CNN-LSTM model learns spatiotemporal and temporal features accurately compared to other deep learning models.

Keywords—Traffic stream variables prediction, CNN, LSTM, Rainfall, Snowfall

18.45 – 19.00 Accepted Paper Presentation

Title: “ VDA: Deep Learning based Visual Data Analysis in Integrated Edge to Cloud Computing Environment”

Authors: Atanu Mandal (Department of Computer Science and Engineering Jadavpur University Kolkata, India), Amir Sinaeepourfard (Norwegian University of Science and Technology (NTNU), Norway)), Sudip Kumar Naskar (Department of Computer Science and Engineering Jadavpur University Kolkata, India)

Abstract: In recent years, video surveillance technology has become pervasive in every sphere. The manual generation of videos’ descriptions requires enormous time and labor, and sometimes essential aspects of videos are overlooked in human summaries. The present work is an attempt towards the automated description generation of Surveillance Video. The proposed method consists of the extraction of key-frames from a surveillance video, objects detection in the key-frames, natural language (English) description generation of the key-frames, and summarizing the descriptions. The key-frames are identified based on a structural similarity index measure. Object detection in a key-frame is performed using the architecture of Single Shot Detection. We used Long Short Term Memory (LSTM) to generate captions from frames. Translation Error Rate (TER) is used to identify and remove duplicate event descriptions. Term frequency-inverse document frequency (Tf-idf) is used to rank the event descriptions generated from a video, and the top-ranked the description is returned as the system generated a summary of the video. We evaluated the Microsoft Video Description Corpus (MSVD) data set to validate our proposed approach, and the system produces a Bilingual Evaluation Understudy (BLEU) score of 46.83.
Keywords—Smart City, Smart Surveillance, Video Summarization, Content-based Video Retrieval, Internet of Things

19.00 – 19.15 Final Technical Discussion Panel by the committee members and keynote speakers, including “Main outcomes and open challenges of the workshop discussion”

19.15  End of Workshop

Venue

Please visit this link for further information.

Contact

For further information, please contact us at a.sinaee@ntnu.no. Please put the title of your email “3SCity-E2C 2021 Workshop-Question”.