The new Cluster of Excellence you are involved in is called “Bilateral AI”. What does that mean?
 
This project aims to combine the two most important methods of AI: symbolic and sub-symbolic AI. This is a unique opportunity to join forces in fundamental research and to develop new methods. In the future, this could benefit many different applications of AI.
 
What is the difference between symbolic and sub-symbolic AI?
 
Sub-symbolic AI learns from the statistical distributions in the numerical representation of data. The advantage of these models is that they can process vast amounts of data. However, the resulting patterns cannot be easily explained. This is why such an AI also “unintentionally” learns things from the data that do not correspond to facts, laws and norms and, for example, adopts prejudices from human language – which is of course a problem. The famous hallucinations known from ChatGPT are based on this weakness. Symbolic AI, on the other hand, can be used for rule-based modeling. This kind of AI can learn what a fair selection process looks like or, more generally, what fairness is. This way, it becomes possible to model social norms, rules, and laws.

How can these two different approaches be combined?
 
That is the crux of the matter. The methods for combining these two approaches are still in their infancy, partly because the research communities that deal with these two directions of AI have historically not pulled in the same direction. The special characteristic of this project is that top researchers from both fields have been brought together; a total of 35 excellent key researchers from six Austrian universities will work together, in addition to new doctoral students and postdoc positions funded by the FWF. This constellation is unique internationally. Such large-scale projects are only made possible by the FWF's “Cluster of Excellence” initiative.

What is the biggest challenge in this endeavor?
 
The two approaches represent completely different ways of thinking about how knowledge is represented. The big challenge is how to find common representations. One possibility is to create so-called knowledge graphs: These combine factual knowledge with source knowledge, as well as knowledge of the existing rules – i.e. the necessary context – in a graph. Our team hopes to gain new insights into how such knowledge graphs can be created in a scalable way and how these graphs can contribute to making machine learning more comprehensible.
 
What would an AI that combines both approaches be capable of?
 
The combination of symbolic and sub-symbolic AI would potentially allow the development of a “broad AI”: in other words, an AI that can be applied to various complex problems. This includes questions such as how to optimize complex systems – for example the energy supply system. Another example is scenarios in which complex systems are partly based on external environmental influences, but are also partly controlled by human-made rules. 
 
Which other experts from WU are involved in this project?
 
The WU team includes Kurt Hornik, who has worked extensively in the field of machine learning and neural networks and is a strong representative of subsymbolic AI. Nils Wlömert from the Department of Marketing complements us with his research on causal relationships with a focus on practical applications in the field of marketing. Sabrina Kirrane is researching the representation of policies and legal frameworks – in other words, the question of how an AI can be made to behave in accordance with the rules. My own research group brings in the aspect of knowledge graphs, where we try to automatically create or improve such knowledge graphs and thus make machine learning more explainable and robust.

When can we expect the first results?
 
This is a very ambitious basic research project that is scheduled to run for five years. Everyone involved is already constantly delivering new, exciting findings: Since the application was submitted, for example, our colleagues at JKU have already presented a completely new machine learning architecture. We ourselves have just completed studies that analyze how dynamically generated knowledge graphs, which are already used in practice as background knowledge for AI applications, can be repaired by the “crowd”. all of this work provides us with valuable foundations: however, we need to develop solutions that actually combine all of these exciting individual results into a larger whole, and this is what we need this cluster for. The Cluster of Excellence is scheduled to run for five years, which is a very realistic timeframe for basic research, but I am very optimistic that we can expect to see integrated results and completely new approaches within the next one to two years.