Artificial Intelligence has become fashionable in all areas of technological development, including communication networks. How realistic are expectations that AI will significantly improve 5G and beyond networks? When will it happen? And what are the obstacles? Eurescom message editor-in-chief Milon Gupta asked someone about it who has been at the forefront of network development for a long time – professor Hans Dieter Schotten, head of the research department Intelligent Networks at the German Research Center for Artificial Intelligence, DFKI.

How will AI improve 5G and beyond ­networks?

H. D. Schotten: AI is already used in networks today: network management and SON solutions use symbolic AI solutions; anomaly detection and self-learning event classification are part of many network security systems. With the progress in machine learning, algorithms and computing power, many additional application areas in communications become realistic targets. Anomalies of almost any network property can be detected, events can be classified with high reliability, and systems can learn how to best handle these events based on monitoring the success of implementing measures. All these AI concepts can be applied on network management, network optimisation, and even physical layer routines. As a result, complexity and cost of network management can be reduced, energy efficiency can be improved, networks can be used more efficiently by better adapting them to specific scenarios, and physical layers can learn how to best handle challenging scenarios.

What are currently the major challenges for using AI in future networks?

H. D. Schotten: We need data to train AI, and, additionally, we need to improve the learning efficiency. Federated and distributed learning, transfer learning, data synthesis by generative adversarial networks, and other concepts can help to address this challenge. Hybrid model and data-driven AI will improve the learning efficiency.

AI solutions need to be compatible with the telecoms ecosystem. Data is often too valuable to be shared with other parties. Trained-model marketplaces, federated learning, or even AI on homomorphic data could solve this problem.

Security is a big problem, as AI functions can be attacked.

Due to the size of our networks, scalability and overall cost in terms of energy consumption and additional network load need to justify the achievable gains. Dynamically deployable AI functions can help to address this issue.

And, finally, there is always a trust issue when talking about AI.

To what extent can we trust AI functions to run our future networks?

H. D. Schotten: We are using AI already today in many products and even in networks. So, there is some trust in their integrity and performance. The black-box approach is often used to mitigate risks. However, the use of AI functions also allows new, partly AI-enabled attack vectors against the integrity of networks. Some techniques to address this challenge are known, but they need to be adapted and improved continuously.

Certification of AI functionality and implementation will also help to avoid misuse and to create trust.

How critical do you consider data privacy issues for the use of data for training AI?

H. D. Schotten: Data privacy is always a top priority. Some tools to address this issue are known: federated and distributed learning or model-based learning allow to share and use learning results without exposing private data. I assume that efforts to develop and certify these privacy-preserving AI concepts will increase with the growing adoption of AI in public services and infrastructures.

How can 5G networks optimise AI-enabled services and applications?

H. D. Schotten: Whenever we have AI-in-the-loop applications, whenever we consider AI-enabled products that depend on guaranteed and real-time data availability, AI-enabled services and applications will significantly benefit from 5G. Network slicing is a key enabler for many data-driven products and services. And URLLC is a key enabler for applications where AI is processing sensor data collected in real-time for controlling actuators, for example in AI-in-the-loop applications.

What is your vision for the uses and ­benefits of AI in beyond 5G networks?

H. D. Schotten: We will see synergies of 5G and AI that go far beyond what the technologies can achieve separately. Autonomous machines in public infrastructures and new concepts for humans to interact with and control their cyber-physical environment are just two examples. In general, 5G will provide the high-performance networking infrastructure for AI-enabled solutions where the advantages in standardisation and convergence will be additional benefits provided by 5G. 5G networks will allow AI functions to be deployed where and when needed, saving cost and making the use of AI more agile. On the other hand, AI will help to create Smart Networks, providing a holistic connectivity and computing infrastructure that automatically adapts, end-to-end, to the changing needs of services and varying resources.