The 2nd Artificial Intelligence/Machine Learning (AI/ML) in 5G Challenge is set to conclude in December, capping a successful debut last year.
Organized by the International Telecommunication Union (ITU) as part of its AI for Good initiative, the competition sees participants around the world solving real-world problems by applying machine learning in communications networks. Reinhard Scholl, Deputy Director at ITU’s Telecommunication Standardization Bureau, shares the journey of the Challenge and what’s coming up next.
Reinhard Scholl: When we started the Challenge, we had no idea where it would be going. It was an adventure and turned out to be such a positive experience. Last year, we had participants from 62 countries.
This year, we have participants from 82 countries, with the Grand Challenge Finale scheduled on 14 December.
We never expected such big numbers. We were also surprised by the large number of problem statements – between 15 and 20 each year – that we were able to offer so far. And we are grateful to this year’s sponsors, Xilinx and the Republic of Korea’s Ministry of Science and ICT.
We also published a special issue on AI and machine learning solutions in 5G and future networks in the ITU Journal Future and Evolving Technologies, with a selection of peer-reviewed papers submitted by Challenge participants.
We are on the lookout for new problem statements for the third Challenge.
One thing we are hoping to offer next year is computing resources for participants who might not have the support of a rich university or company. Training machine-learning models can take a lot of time, and several participants informed us that they don’t have the resources to run meaningful models. So, we are working on that.
ITU-T does a lot of technical work related to machine learning in its focus groups – six of which have AI or machine learning in their title – and in its study groups. The specifications of the focus groups are generally turned into ITU standards (“ITU-T Recommendations”).
The most popular standard is the “Architectural framework for machine learning in future networks including IMT-2020” (ITU-T Y.3172), which gives a common nomenclature and primer on how to talk about machine learning in communication networks, so that it can be used by anyone for any network.
Some of the solutions submitted to problem statements in the ITU AI/ML in 5G Challenge reference ITU standards on machine learning. Some have generated contributions to the respective focus groups or study groups – and attracted new ITU-T members. We have run over 50 one-hour, in-depth talks so far – by researchers on machine learning and communication networks – on the AI for Good Discovery Channel, a fabulous resource on what the future of communication networks will look like. We have similar “Discovery Channels” on Trustworthy AI, AI and Health, as well as AI and Climate Science. In January 2022 we are going to launch a Geospatial AI Discovery Channel.
Network operators have used machine learning for some time, but not at the network level. They have used it to analyse the churn rate or to segment their customers. But applying it at the networking level is complicated. Applying machine learning in communication networks is much more difficult than in computer vision or natural language processing, because time scales in a communication network span many orders of magnitude, ranging from parameters which change on an annual basis, like your subscription to a telecom provider, to milliseconds, like resource block allocations in radio access networks) – for which you then have to retrain your machine learning model on a millisecond basis.
As networks get more and more complicated, machine learning will be essential to make sense of the plethora of data being collected. On the other hand, machine learning could also be useful in the standardization process.
For now, standards are produced by people, who meet, make proposals, negotiate, and agree on a certain outcome. But the resulting protocols are often ambiguous and suboptimal, leading to increasing costs in testing and implementation. Part of this process could be taken on by machine learning, where the algorithm proposes a solution. There have been some attempts to do this, but there’s still quite a long way to go.
We are branching into a new Geospatial AI Challenge that draws on location-based data. We have launched a call for problem statements. ITU and the World Health Organization (WHO) meanwhile are working, through their joint focus group, on an incredibly ambitious AI for Health Assessment project. When we take a prescribed medicine or vaccine, there is a sense of trust in the process and in the institutions.
But why would you trust an AI model looking at your X-rays? What does it take to trust a company that has an AI solution to detect skin cancer? The focus group is building a benchmarking framework that allows people to trust in AI health solutions.
The ITU-WHO focus group will come up with a process, guidelines, and best practices to ensure trust in AI solutions.
In addition, it’s developing a platform where a company can submit and test solutions using undisclosed data. A score is generated and published on a leader board, which also allows a regulator to know how good the solution is. You must design a process that allows experts to come to an agreement and then build that into the platform. The prototype will be ready in a matter of weeks. Then it needs to be transformed into a professional platform, which will cost serious money. We are going to start an AI for Good Fund to secure donations for projects like the AI for Health Assessment Platform, along with other work, such as the AI and Road Safety global initiative established in October.
Learn more about the Artificial Intelligence/Machine Learning (AI/ML) in 5G Challenge.