Learning environments of the AI lab

The AI lab's online learning environments are designed to provide teaching approaches to various artificial intelligence methods. This article presents two of the AI lab's learning environments.

Andreas Mühling

Artificial intelligence (AI) is an exciting topic for school lessons - far beyond the subject of computer science - especially in the form of its currently successful variation of machine learning from large amounts of data. Students use the latest generative AI systems such as ChatGPT and our everyday lives have long been permeated by the decisions of "artificial intelligence". This offers numerous possible starting points for contextualized subject lessons, such as the economic use of data, the use of AI in physics or biology, how the systems work and ethical and legal issues.

Fundamental knowledge of the systems is required for such teaching to work, as only then can the dangers and opportunities of AI, for example, be meaningfully assessed. Current systems, however, are the result of many years of cutting-edge research and are therefore difficult to teach at secondary school level. There is also often a lack of appropriate, accessible teaching material.

The AI Lab (https://ki-labor.ddi.leibniz-ipn. de/) offers a solution. The AI Lab was developed by the Computer Science Education working group at Kiel University and the IPN and offers teachers the opportunity to integrate modern AI into their lessons.

The website consists of a collection of entirely web-based online learning environments based on two principles:

1. AI procedures are presented in an application context.

2. Each learning environment contains experimentation opportunities for students.

The learning environments can be flexibly integrated into lessons. The environments are either educationally reduced and augmented or allow experiencing modern methods using large data sets.

"Perzeptron" is an example of the first variant, a highly structured environment. It  uses an interactive comic to describe how simple neural networks work in the context of classification. Students can first experiment with the parameters of a very simple network (the perceptron) and arrive at successful parameter values themselves.

This approach reaches its limits for a more complex network, however, and it becomes apparent that automated optimization ("learning") is necessary.

The learning environment "MNIST Number Recognition", on the other hand, works with a powerful neural network based on a classic machine learning data set - the so-called "MNIST" database. The learning environment focuses on the ability of neural networks to recognize handwriting, a technique that is used, for example, in the automatic sorting of letters.

Students are first instructed in an experimental approach to a fully trained network. The aim is to test the performance of the trained network in comparison to humans. The second step is to train the network itself. The learning parameters (hyperparameters) are made accessible and an untrained network can be trained based on these. It becomes clear that learning itself is a controlled, algorithmic process and that the end result is not always a similarly efficient network. The need for powerful hardware becomes apparent in this context, as this learning stresses the hardware of modern systems and also takes a while. 

Currently, 13 learning environments are already available online for free use. They are continually tested in the classroom, whereby the two learning environments presented here have shown that they can be worked on independently by students in computer science lessons at upper secondary level and that learning success can be measured. Additional offers are being developed on an ongoing basis, for example as part of students' final theses.

About the author:

Prof. Dr. Andreas Mühling has been responsible for the computer science education at Kiel University (CAU) since 2016. He heads the Computer Science Education working group recently established jointly with the CAU at the IPN.