Liaison statement
LS/o on the results of the 1st meeting of the ITU-T Focus Group on Machine Learning for Future Networks including 5G (FG ML5G)

We would like to inform you that ITU-T on Machine Learning for Future Networks
including 5G (FG ML5G) held its first meeting in Geneva, 30 January - 2
February 2018, with a workshop taking place on 29 January 2018.

The group was established by ITU-T Study Group 13 at its meeting in Geneva
(6-17 November 2017) with these Terms of Reference. (Terms of Reference
attached)

At its first meeting, FG ML5G established three working groups:

- WG1: Use cases, services & requirements (Terms of reference see Annex A of
this liaison) - WG2: Data formats & ML technologies (Terms of reference see
Annex B of this liaison). - WG3: ML-aware network architecture: (Terms of
reference see Annex C of this liaison).

FG ML5G has a lifetime of one year from its first meeting. The next meeting of
FG ML5G will take 24-27 April 2018, with a workshop on 25 April 2018. The venue
for the focus group meeting and the workshop is Xi’an (China).

ITU Focus Groups are open to all interested parties. Membership in ITU is not
required. Please join the Focus Group’s mailing list, request access to
documents and learn more about its priorities on the group’s homepage.
(https://www.itu.int/en/ITU-T/focusgroups/ml5g/Pages/default.aspx)

We welcome your input to the deliverables and updates on activities related to
ML5G, and look forward to collaboration in this area.

Annexes
Annex A: Terms of reference Working Group 1: Use, cases, services and
requirements Annex B: Terms of reference Working Group 2: Data formats and ML
technology Annex C: Terms of reference Working Group 3: ML-aware network
architecture

                                          Annex A: Terms of reference Working
                                          Group 1: Use, cases, services and
                                          requirements

Motivation
Today, lots of ML-based applications and services for future networks including
5G start to appear in the market. In order to prepare effective standard ways
to build up ML-based applications and services in the networking area, we need
to investigate the way ML-technologies and data are used in them and to derive
the core requirements for them.

Scope
The objective of WG1 is to understand the industry needs and to clarify the
vision about the whole ecosystem in terms of ML for future networks and 5G.
This will be done by collecting use cases from industry and defining the
derived requirements. Expected use cases, and without any limitation, can be
both on the network infrastructure services and applications services (e.g.
self-organized networks, information control networks, networked autonomous
driving). The classification of these use cases will then allow deriving a set
of requirements. These requirements will drive the work on data formats and ML
technologies within WG2 and ML-aware network architectures within WG3.

The activity of the group will also include the exploration of the whole
ecosystem and stakeholders of the market involved or impacting the vision and
the needs for ML in the future networks and 5G. This can go beyond the
classical telecoms operators and would involve new players from vertical
industries (e.g. automotive, manufacturing) and their needs. Aspects related to
specific standardization gaps based on the observation of the whole
standardization environment and in liaison with other SDOs, fora, etc, will be
also addressed.

Since theoutcome of this working group is crucial to the other working groups,
the deliverables should be produced incrementally and quickly. In the beginning
we will start from a small number of important use cases and enlarge the target
use cases later, and then derive the fundamental requirements from the previous
work.

Specific questions to be addressed include:
1. What are the relevant use cases and derived use cases requirements for ML?
2. What are the standardization gaps?
3. What are the liaisons activities?

Tasks include, but are not limited to:
1. Specify important use cases.
2. Derive minimum requirements regarding those use cases to be shared with WG2
and WG3. 3. Analyze technical gaps related to the use cases and its ecosystem

Deliverables
WG1 is to deliver the following documentation:
1. Use cases
2. Ecosystem, terminology and services.
3. Requirements and standardization gap

Relationships
- 3GPP (SA1) outcome
- IEEE
- IETF
- ETSI (ISG ENI, ISG ZSM)
- ONAP (open source project for network orchestration)
- Acumos (open source AI platform supporting network)

                                      Annex B: Terms of reference Working Group
                                      2: Data formats and ML technology

Motivation
Modern communication networks, and in particular mobile networks, generate a
huge amount of data. Powerful machine learning (ML) methods can be used to
extract and leverage this information for various tasks, however, the lack of a
unified data format makes such an analysis a challenging problem. The
application of ML technology to communication networks is further complicated
by constraints and requirements such as limited computation resources,
bandwidth or latency restrictions or distributed data. This working group will
investigate data formats and ML technologies which are tailored for such a
communications scenario.

Scope
This working group will investigate how to collect, prepare, represent and
process data for ML in the context of communication networks. This also
includes the study of privacy and security implications on data formats and ML
techniques. Furthermore, this working group will integrate the inputs received
from the other working groups into its work. These inputs will consist of the
potential use cases, including requirements on the ML technology (what to
compute, what data do we have, how fast to compute, how reliable and
transparent it must be, where to do the computation, how much computational
resources do we have) and requirements on the data (what data to we have, can
we use all the data centrally, can we trust the data, is the data labeled,
where is the data generated), as well as potential network architectures (e.g.,
distributed, centralized, hybrid). Furthermore, this working group will engage
in the categorization of ML algorithms used in communication networks. This
includes the categorization of how different ML methods (e.g., neural networks,
unsupervised methods, reinforcement learning) fit to different communications
problems. Since data is usually distributed in a communication network, this
working group will also investigate how current ML technology can be used in or
extended to a distributed setting. Topics of interest here are the efficient
representation of ML models, efficient at-terminal computation, distributed
learning with reduced overhead and other ML topics such as trustworthiness and
transparency of the algorithms. Finally, this working group will identify
standardization and technology gaps and create liaisons with related activities
in other organizations.

Specific questions to be addressed include:
1. How should data be collected, prepared, represented and processed for ML in
the context of communication networks? 2. What are the privacy and security
implications on data formats and ML? 3. Categorization of ML algorithms in the
context of communication networks, i.e., how do different ML methods fit to
different communications problems? 4. How can current ML technology be used in
a distributed setting (e.g., efficient representation of ML models, efficient
at-terminal computation, distributed learning with reduced overhead)? 5. What
are the standardization and technology gaps?

Tasks include, but are not limited to:
1. Analysis of ML technology and data formats for communication networks, with
special focus on the uses cases of WG1. 2. Providing input to WG3 on data
formats and ML technology, and incorporate output from WG3 on ML-aware network
architectures. 3. Identification of standardization and technology gaps. 4.
Liaisons with other standardization organizations.

Deliverables
1. ML algorithms in communication networks: categorization, terminology &
implications 2. Data formats including privacy and security aspects for ML in
communication networks 3. Standardization and technology gaps

                                      Annex C: Terms of reference Working Group
                                      3: ML-aware network architecture

Motivation
Future networks such as 5G will be highly complex. We expect that ML is a
promising technology to cope with this increased complexity. Moreover, ML
technologies can be used to improve the performance of networks with respect to
OPEX/CAPEX, and enable new use cases, applications and services such as
networked autonomous driving.

Today’s network architectures are not suitable for incorporating ML
technologies. For example, in the case of operation and maintenance, huge
amounts of data must be transferred in order to perform training and prediction
tasks. This would require a large amount of network resources (such as
computational power, energy, storage etc.). Moreover, currently deployed APIs
do not meet the requirements of existing ML technology.

Scope
The WG will study the implications of applying ML technologies to communication
networks. In particular, the focus will be on specification and placement of
functions, interfaces and resources as a result of the integration of ML
technologies. The ultimate objective is to enable efficient use of ML
technologies in future networks.

Specific questions to be addressed include:
1. What are the implications of ML (including distributed ML) on network
architectures? 2. What are the requirements imposed by ML on network
architectures in terms of computational power, energy, storage, interfaces,
communication resources (e.g. which interfaces are needed to support ML-based
network optimization)? 3. What are the standardization gaps? 4. What are the
liaisons activities?

Tasks include, but are not limited to:
1. Analysis of implications of ML (including distributed ML) on network
architectures 2. Incorporate output from WG1 on use cases and requirements and
WG2 on data formats. 3. Analysis of functions, interfaces, resources imposed by
ML on network architecture 4. Gap analysis based on the tasks of different
standard organizations

Other topics can also be studied as appropriate, based on contributions.

Deliverables
1. Analysis of communication network architectures from the viewpoint of ML
2. Description of ML-related functions, interfaces and resources for
communication network architectures 3. Standardization and technology gaps

Relationships
All network architecture related AI and machine learning Standardization
bodies, forums, open source projects: 3GPP IEEE IETF ETSI (ISG ENI, ISG ZSM)
ONAP (open source project for network orchestration) Acumos (open source AI
platform supporting network)