Data Flow and Distributed Deep Neural Network based low latency IoT-Edge computation model for big data environment

Highlights

• Data Flow and Distributed Deep Learning based Edge computation model for Big Data.

• Data Flow based data gathering at the Edge with reduced bandwidth.

• Intelligent decisions at various levels of Edge in a distributed manner.

• Aggregation schemes for improving the overall performance of the system.

• AI-assisted Edge services for resource-constrained heterogeneous devices.

Abstract

The trillion-fold increase in computing power brings the accessibility of deep learning to everyone. Deep learning offers precise information almost all the time when compared to other learning algorithms. On the other hand, the popularity of Internet of Things (IoT) has increased in various areas such as Smart City, Oil Mining, and Transportation. Edge/Fog computing environment helps to handle significant challenges faced by the IoT, viz. latency, bandwidth consumption, and everlasting network connectivity. For analytics in Edge computing, which is distributed in nature, the trend is more towards distributed machine learning. This research work is focused on the integration of data flow and distributed deep learning in the IoT-Edge environment to bring down the latency and increase accuracy starting from the data generation phase. To this end, a novel Data Flow and Distributed Deep Neural Network (DF-DDNN) based IoT-Edge model for big data environment has been proposed. Our proposed method has resulted in latency reduction of up to 33% when compared to the existing traditional IoT-Cloud model.

Introduction

Computing has evolved in multiple phases in the past 30 years. Information and Communication Technology (ICT) has contributed greatly to the betterment of human life in all aspects which have a market value of US $3.5 trillion and growing at the speed of 5% per annum. The word Internet of Thing (IoT) was coined by Kevin Ashton in 1999 (Kevin, 2009, Kevin, 2016). IoT may be a network of physical devices, together with things like smartphones, home appliances, vehicles, and more, that connect with an exchange of information on computers. IoT represents a general idea for the flexibility of network devices to sense and collect information from around the globe and share that information across a network where it will be processed and used for numerous exciting commitments. In the last five-year span from 2015 to 2020, IoT has become quicker than some other category of connected gadgets that drive worldwide IP traffic as indicated in a recent Cisco report. The number of machine-to-machine (m2m) connections ought to raise firmly 2.5 times, i.e., from 4.9 billion in 2015 to 12.2 billion in 2020. Cisco IBSG predicts (Kevin, 2016, Jazib et al., 2019) that there will be 50 billion devices connected to the Internet by 2020. This growth will impact the Globe more in the ICT global market value. The IoT devices are connected to the cloud data center via the backhaul network. This rapid growth gives a new challenge to the ICT industries. In IoT, data needs to be analyzed rapidly. This requires a minimal delay in sending and receiving the data, which can be very problematic because data needs to reach the cloud data centers before processing takes place. This brings the need for Edge computing that can reduce latency. This travel time is crucial for applications like monitoring of in-flight jet engines or the operation of driverless cars where even millisecond delay is unacceptable.

The main requirements of IoT devices are privacy, latency, and connectivity.

• Privacy: sensitive data can be processed on-site. Sending data to an external cloud introduces new privacy risks regardless of how well-secured the connection is.

• Latency: Transporting the whole data directly to the cloud will increase the traffic in the network and also increase the load in the cloud server.

• Connectivity issues: IoT devices need to remain connected with the network always. Practically this has many challenges.

Edge computing answers these challenges by processing the sensor data locally, reduce data traffic, bandwidth usage, and latency (Papageorgiou et al., 2016, Teerapittayanon et al., 2017, Savaglio et al., 2019). The main idea behind Edge computing is to process data as close to the source as possible. Rather than sending all the data collected by IoT sensors directly to the cloud, it processes this data within the network, and only relevant data or information conveniently bundled, is sent. After a wave of digitalization and centralization, Edge computing fulfills the need for decentralization, because the centralized model does not fit all IoT infrastructures (Papageorgiou et al., 2016; Teerapittayanon et al., 2017; Savaglio et al., 2019; MD Golam et al., 2019; Pei Yun et al., 2018; Roberto et al., 2019). The development of 5G networks by the telecom industry makes it much easier for the enterprise to access Edge computing (Ridhawi et al., 2017). The 5G providers are likely to build locally based microdata centers into their wireless network signal towers, and this enables the access of micro datacenters for Edge computing and makes use of the 5G network to transport the data to their cloud.

Similarly, Artificial Intelligence (AI) is having an immediate effect on human lives. AI is characterized by the capacity of a machine to replicate human behavior. Machine Learning (ML) is a subset of AI which has got similar characteristics of AI. The machines are trained towards performing certain activities like classification, prediction, or clustering based on machine learning algorithms similar to a human being. Deep learning is a subset of ML.

Usually, an individual’s utilization of Deep Learning is alluding to Deep Artificial Neural Networks, less frequently to deep reinforcement learning. Deep Artificial Neural Networks (DANN) is a set of algorithms that bring new records of precision for a huge number of discriminating problems such as image recognition, voice recognition, and recommendation frameworks.

The specific research objective of this paper is to develop an approach for the integration of data flow and distributed deep learning in the IoT-Edge environment to bring down the latency and increase accuracy starting from the data generation phase. We propose a data flow and distributed deep neural network based IoT-Edge model. Thus, the main contributions of this paper are as follows:

• Data Flow and Distributed Deep Learning based Edge computation model for Big Data.

• Data Flow based data gathering at the Edge with reduced bandwidth.

• Intelligent decisions at various levels of Edge in a distributed manner.

• Aggregation schemes for improving the overall performance of the system.

• AI-assisted Edge services for resource-constrained heterogeneous devices.

The proposed model has been compared with the traditional IoT-Cloud model and not with any other IoT-Edge model driven by AI. The reason behind this is that Edge/Fog computing is a fairly new technology introduced in the world of IoT. Edge/Fog computing is been introduced for bringing down the latency, cloud workload, and network congestion. The proposed work is to validate as how AI assisted Edge could reduce the latency, network congestion and cloud workload as compared to performing AI enabled decision at the cloud. Comparing our proposed AI assisted Edge with other Edge computing model could be part of another research paper in future with many AI assisted Edge based models developed as prototype or simulator

The rest of the paper is organized as follows. Section 2 discusses the literature review on existing works in fog computing and distributed deep neural network in fog computing. Section 3 explains the proposed data flow-based distributed deep learning in Edge computing. Section 4 describes the experimental setup and results. Section 5 analyzes the experimental results, while Section 6 gives the concluding remarks and the possibility of future work.