AI on our own terms
Artificial intelligence often feels like a service that takes place somewhere else. The Machine Learning Management Platform (MLMP) at BFH-TI takes a different approach: it gives students, lecturers, mid-level academic staff and specialist staff access to a shared AI infrastructure, keeping knowledge and control close to the university.
SocietyByte: The acronym MLMP sounds rather technical. What lies behind it?
Daniel Reichenpfader: MLMP stands for Machine Learning Management Platform. In concrete terms, it is a shared environment at BFH, Department of Engineering and Computer Science, in which students, lecturers, researchers and specialist staff can develop, train and use AI models. The platform combines high-performance computers with services that make current language models accessible. What is decisive is not the technology itself, but what it makes possible: at BFH, people can experiment, learn, teach and carry out research without having to build their own infrastructure for every project.
What problem does the platform solve?
Peter von Niederhäusern: AI projects need more than a good idea. Training models requires a great deal of computing power, and reliably operating a modern language model also calls for specialist expertise. For an individual course or a single research team, these hurdles can be too high. MLMP pools the necessary resources. Lecturers can prepare an exercise, students can examine how a model behaves, and research teams can test applications on a shared basis. This leaves more time for the actual research question.
| What is a large language model? A large language model (LLM) is an algorithm trained on very large collections of text. It learns statistical patterns of language and generates answers by predicting suitable next pieces of text. This can seem remarkably fluent. Linguistic fluency, however, is no proof of truth or understanding. Important statements need to be verified. |
Many people already use public AI chatbots. Why does a university need a platform of its own?
Daniel Reichenpfader: Public AI chatbots are convenient, almost like takeaway food for text. For universities, however, the question is what should be learned (prepared) and controlled (cooked) in their own kitchen: where data is processed, who understands the technology, and how dependence on individual providers can be reduced. MLMP operates a local high-performance infrastructure and makes open models available to the BFH community. This strengthens data and technology sovereignty: BFH retains more control over infrastructure, models and know-how. At the same time, the entire chain from hardware to application becomes a subject of teaching and research.
Does «local» mean that you can enter any data without a second thought?
Fernando Pareja: No. Responsible use still begins with the people who use the tool. A locally operated service creates more scope for control and learning, but it replaces neither data protection nor critical scrutiny and clear rules. In education this is particularly valuable: students should learn not only what AI can do, but also where its limits and risks lie.
What can you actually do with MLMP?
Peter von Niederhäusern: The range extends from classical machine learning to generative AI. Users can reserve computing capacity in order to train models or run demanding calculations. Language models can be used via the browser or a standardised programming interface. This makes it possible to summarise and analyse texts, trial document searches, build prototypes, transcribe speech or investigate image-text applications. For larger classes, browser-based interactive Jupyter Notebook environments can be prepared so that everyone starts out with the same tools and examples.
| Why do models sometimes invent facts? Language models can formulate very convincingly. Sometimes this sounds like an assured lecture, unfortunately with invented footnotes. An LLM is meant to produce a plausible continuation, not to look something up in a perfect internal encyclopaedia. If training patterns or the context provided are not sufficient for a reliable answer, a convincing-sounding statement can still emerge. Clear instructions, suitable source documents and human oversight reduce this risk, but do not eliminate it entirely. |
How does sharing this high-performance equipment work in everyday practice?
Hanspeter Zimmermann: It takes coordination. Computing resources are limited, which is why intensive work is scheduled and teaching activities are prepared in advance. A class whose students work at different times is easier to serve than one in which everyone sends a demanding request at the same moment. At first this sounds like a restriction, but it conveys an important insight: artificial intelligence does not operate in a vacuum either. Somewhere a computer is humming, electricity is being consumed, heat has to be dissipated, and someone has to make sure that everything keeps running.
Who is the platform intended for, and how should it develop?
Daniel Reichenpfader: The first target group is the BFH community: students, lecturers, researchers and specialists. Beyond that, the aim is to establish a practice-oriented AI competence centre that shares experience and collaborates with other universities and with partners, including small and medium-sized enterprises. The vision is not simply to own fast computers. BFH should use AI responsibly and sovereignly, keep important knowledge within the institution and strengthen the innovation ecosystem.
What should readers remember about MLMP?
Peter von Niederhäusern: AI sovereignty takes concrete shape when people can learn with the technology, examine its limits and design applications themselves. As the MLMP team, we do not make AI magical, but we do make it tangible: as infrastructure, a place of learning and a workshop all at once.
Public MLMP documentation: https://infra.pages.ti.bfh.ch/mlmp/src/
Daniel Reichenpfader is a tenure-track lecturer at the Institute for Patient-Centered Digital Health and is completing a doctorate in digital health at the University of Geneva. His research focuses on the application and optimisation of computational linguistics methods (including chatbots) in healthcare and education.
Peter von Niederhäusern is a lecturer in divisions I and S and teaches data engineering fundamentals, computer graphics and low-level programming languages. He is a doctoral candidate at the Department of Biomedical Engineering at the University of Basel.
Fernando Pareja is a Master's student in data science. His interests lie in particular in data analysis and in applying machine learning methods to generate data-based insights. Professionally, he works on implementing new technologies and innovative concepts in existing platforms.
Hanspeter Zimmermann is a research associate and administers various server environments.
Contributions as RSS
Comments as RSS
Leave a Reply
Want to join the discussion?Feel free to contribute!