National AI competition: 10 teams for the final
Ten teams at to present their machine learning projects in TĂĽbingen on November 15: The young finalists demonstrated their inventiveness and tinkering skills.
(Image: sdecoret / shutterstock.com)
In March, the TĂĽbingen AI Center announced this year's Federal AI Competition (BWKI). Participants had until mid-September to turn their projects into working code and present them in a pitch video. 101 teams (a total of 227 individuals, including 50 girls) ultimately submitted their work in the hope of securing one of the coveted places in the final. The finalists have now been chosen: ten teams, or 11 boys and 5 girls, will travel to the final in TĂĽbingen on November 15 to present their work. It will still be a while before specific names and detailed project descriptions are published. However, we had the opportunity to get an initial overview and talk to one of the project assessors.
Alexander Braun is a research associate at the Max Planck Institute for Intelligent Systems in TĂĽbingen and one of the nine BWKI project reviewers. He told us what criteria are used to evaluate the projects and how the latest developments in machine learning are reflected in the submissions. In order to avoid unconscious preferences, the projects were anonymized: He and the other reviewers did not get to see the name, age, gender or place of residence.
The range, wealth of ideas and pragmatic implementation are inspiring even when reading the short texts: training, programming and tinkering were carried out; the code and sometimes the associated equipment solve concrete problems, both in everyday life and in society, the environment and science.
Even if LLMs and image generators are currently dominating the discourse and triggering many a controversial discussion, they did not play the main role in the competition and when they were used, they only served as aids. A central technique is still classic pattern recognition using neural networks, whether in image or other data: The young people used it to design and build efficient, resource-saving systems for all kinds of purposes, ranging from agriculture and private households to traffic control and event technology. Other teams transformed simple devices, for example for medical diagnostics, into high-tech devices with the help of AI, developed robotic assistants for sports or solved unsolved scientific puzzles. The young people also found solutions on how to process personal – including medical – data for the benefit of the individual while preserving privacy.
Mr. Braun, you had the exciting and certainly not easy task of judging the submitted projects. What range of topics did you expect?
In terms of subject matter, the projects submitted this year were once again extremely varied and no two projects are the same. Some deal with very big challenges such as climate change or the detection of diseases, others solve small problems that everyone knows from everyday life. All in all, there were a lot of creative and innovative ideas this year that have never been seen before.
Do you see a particular thematic trend compared to previous years?
Last year's submissions were also very diverse, I couldn't see a clear trend or specific developments in the themes. However, there were a few projects that were already in the running last year but didn't quite make it to the final. It is always very exciting to see how an idea has been pursued and how the implementation has improved.
How are the current developments in machine learning, particularly generative AI and especially the large language models, affecting the projects submitted?
We have noticed that more teams are building on existing AI applications and fewer are developing their own ML models from scratch. As a result, language models such as GPT or Llama and object recognition models such as Yolo were used much more frequently. These existing models were then adapted to a specific use case through prompting or fine-tuning. This clearly shows that certain AI applications have become much more accessible to developers in recent years. Even if less is being built "from scratch" as a result, I think this is a very positive development. When programming, you regularly use existing libraries and don't always reinvent the wheel. The barrier to entry drops enormously and as a hobby programmer you suddenly have completely new possibilities because you can build on existing solutions. After all, you can't just build a large language model yourself.
Has this made it more difficult to evaluate the participants' performance?
Of course, we had to take a closer look during the evaluation and differentiate between the individual contribution and innovation in a project and which functions were already available. Here, however, I was rather positively surprised, because in the vast majority of cases there was still a lot of personal work in the projects. Existing models were often very creatively combined or further developed, resulting in new and innovative solutions.
Are the projects becoming more ambitious?
The greater availability of existing AI applications to build on has definitely made it somewhat easier to achieve presentable results quickly. Large language models in particular have been used much more frequently and are creating completely new possibilities. However, the projects that develop and assemble their own hardware in addition to the software are always very impressive. There were some very impressive and ambitious submissions both this year and last year. All in all, the standard is consistently very high.
How mature does a project have to be, what standards do you apply?
We assess how relevant and innovative the idea is and how well it has been implemented, with a focus on the use of machine learning methods, of course. In addition, there is the presentation in the video pitch and the critical examination of your own project regarding weaknesses or possible risks.
What would lead directly to rejection?
What tends not to go down so well is if a project lacks innovation and your own ideas. Of course, it is not a problem to use public and known data sets as long as you create something new from them. Simply reprogramming instructions is a good exercise, but not enough for the final.
On the other hand, there are also some submissions with very great and innovative ideas that don't really work yet or are only at the very beginning. We can only ever evaluate what is actually submitted as code. However, we always try to give a lot of feedback and motivate the teams to continue working on the project. Hopefully, we'll see some projects again next year, and then we'll be even happier when the idea has been implemented.
And what excites you?
Personally, I am always most enthusiastic about projects when they are a "well-rounded thing". It doesn't matter whether the idea changes the world or simply solves a small everyday problem. It's better for the project to be a little smaller, but well implemented and thought through to the end, than to set yourself big goals but only implement a small part of them. For me, projects with a great idea that solve a real problem and end up with something that you can actually look at and try out have a good chance of reaching the final.
Videos by heise
Main, special and audience prizes
Four projects will be chosen at the final on November 15: The winning team that impresses the jury overall will receive the main prize of 1500 euros as well as an internship at robotics and automation specialist FANUC.
The prize in the special category AI for Good goes to a particularly sustainable AI project; the team receives 1000 euros. The special category No risk, no fun! rewards the team with the most innovative idea or a particularly unusual hardware solution with 750 euros.
You can also vote, as there is an audience award: from November 1, everyone can find out about the projects on the BWKI website and vote for their personal favorite; the audience favorite will receive 500 euros.
The Anne Frank School from Molbergen in Lower Saxony has already been named AI School of the Year. Its pupils were particularly active in the online AI course, which was also initiated by the organizers.
heise Medien is a cooperation partner of the BWKI in 2024. c't editor Andrea Trinkwalder will be part of the jury this year.
(atr)
