Der Vortrag wird auf Deutsch gehalten!
How we build scientific theories and the extent to which AI can support us in doing so
Scientists in all fields are constantly engaged in developing and expanding scientific theories: They recognise relationships, develop new hypotheses, and try to test them. But how exactly do we develop new hypotheses? This is a central question in the philosophy of science and one that I am investigating in my PhD. More specifically, I am investigating abduction.
Besides deduction - certain reasoning - and induction - the generalisation of statements - abduction is the third, and most powerful, type of reasoning: abduction is the only type of reasoning that allows us to introduce new concepts into a theory. For example, Newton deduced from the fall of an apple that gravity exists, Fleming deduced from spots in a Petri dish that an antibiotic is at work there, and Darwin deduced evolution from the formation of certain characteristics in plants and animals. Abduction thus plays an enormously important role in science, since it is the only way to introduce new theoretical concepts, i.e. to infer from observations to theoretical concepts that describe underlying causes.
Precisely because of its power and complexity, abduction is controversially discussed in the philosophy of science and there are many different manifestations and models, for example a psychologically oriented one by Charles Peirce and an explanation-based one by Peter Lipton. Furthermore, it is controversial whether abduction can be formalised, i.e. whether artificial intelligence can at least theoretically be capable of introducing new concepts in a theory.
In my PhD, I developed a new theory of abduction based on conditionals that builds on existing theories while avoiding their problems such as being able to find only certain concepts or having inherent contradictions. Unlike many other theories, it also allows abductive inference to be formalised. Now that the theory has been developed and successfully published as an article, I would like to prove that it can indeed be formalised and profitably used in science.
To do this, I apply it to a previously unsolved problem in the field of artificial intelligence, the Abstraction and Reasoning Corpus (ARC). The ARC is a kind of intelligence test intended for both human and artificial intelligence approaches. While humans can solve 80% of the tasks on average, the best artificial intelligence approach is around 30%. This is because the intelligence test consists of 1000 individual tasks, all of which are different in nature, and existing artificial intelligence approaches are not able to successfully abstract the solution of one task and transfer it to a new, different task.
Abduction can make a valuable contribution here, as each task involves finding the underlying law that describes the transformation from the problem state to the target state. In other words, an underlying cause must be inferred from observations. The abduction theory developed so far is to be implemented accordingly in the form of a computer programme and applied to the ARC test.
If it is possible to solve the ARC test by means of abduction, it can be shown that the philosophical theory of abduction can be formalised and practically applied, which is an important step in demonstrating its feasibility. Furthermore, the successful implementation would show that the generation of scientific hypotheses can be formalised and that it is in principle possible to use artificial intelligence to develop hypotheses and form new scientific theories or extend existing theories. This would be a result that has a major impact on all areas of science and should therefore be discussed interdisciplinarily from the outset and developed further in a joint exchange.
The aim of my 10–12-minute presentation would be to introduce the core ideas and present the main results of the research and their implications. The aim is not to discuss individual components of my work in detail, but to show the audience - especially scientists but also the public - how we develop theories further and to what extent artificial intelligence can support us in this.