Page 61 - Proceedings of the 2018 ITU Kaleidoscope
P. 61
Machine learning for a 5G future
data centers) into smaller functions deployed as AI composition at this layer is that some more important
microservices. The microservices can then be deployed decision process with low latency can be executed without
closer to the data sources and users allowing seamless offloading such process and data to the upper layer such as
composition of AI services across the ROOF-Fog-Cloud the Fog and Cloud layers. Note that these microservices are
layer. The AI features and functions are composed from the developed and deployed independently of each other and
distributed microservices AIMS integrated platform, are composable at runtime. The composition of the
allowing AI functionality as intelligence services to be microservices can be realized sequentially based on
implemented and deployed close to the data sources and Network Function Virtualization (NFV) [2] management
users despite the limited resources available at these layers and inter-slice resource brokering process. The dynamic
and the huge computing resources required by AI adopting microservice architecture for the deployment of
algorithms. AI services means that we can now engineer data-driven
IoT based applications that are composed of multiple
hierarchical self-contained, lightweight, portable runtime
and modular components deployed across a federation of
network slices. This means that AI algorithms can be
factored into modular functional entities that can be
implemented as data-driven reusable algorithmic primitives.
For example, the core functionality of a particular AIMS
service could be a service providing regression analysis,
classification, clustering, IoT data pre-processing functions
such as feature extraction, feature reduction, dealing with
missing data values etc. Each microservice is responsible
for the execution of a smaller portion of an AI task with its
own data, processing and notification points accessible to
other microservices.
Figure 1 – Edge-Cloud high-level integrated architecture
over 5G networks for various AI enabled IoT applications
In software engineering, the concept of service-oriented
architecture is not new. This is based on the idea that an
application can be designed such that the functionality it
provides is divided into smaller functions and implemented
as services that can interact via well-defined programming
interfaces and thus allowing scalability, robustness and
interoperability. Network slicing is recognized as a game
changer in the remarkable paradigm shift from 4G to 5G
era because it can maximize the sharing of network
resources and flexibility for dedicated logical networks [15]. Figure 2 – The ecosystem of microservices distributed
AIMS based application involves composing interoperable across the ROOF-Fog-Cloud systems over 5G networks
microservices from the ecosystem of microservices
distributed across the ROOF-Fog-Cloud systems as In Figure 2, the first layer is the physical layer consisting of
illustrated in Figure 2, showing how microservice at each the IoT devices and connectivity protocols. The IoT devices
5G network slice can be composed. At ROOF level, for can be categorized into 3 types. The first type consists of
example, an AIMS service could execute a decision process the edge sensors and actuators. These devices are capable
based on the data obtained from the IoT devices after some of capturing data with little or no processing capability of
other microservices at this layer have executed data operating system. They are equipped with low power 5G
gathering, and pre-processing tasks on the collected data. In radio connectivity with which they can communicate with
fact, the pre-processed data can be temporarily stored on the edge devices. The second type of IoT devices at the
some of the nodes at this layer. One advantage of service edge are the edge devices. These are devices with the
capability to run operating systems such as Android, IOS,
– 45 –