Piotr Sowiński
supervisor: Maria Ganzha
Recent years have seen rapid advancements in machine learning, which transformed many industries. Modern AI is becoming an indispensable tool in engineering, finance, marketing, management, and more. There are also many companies trying to broadly implement AI in completely new areas, such as self-driving cars and medicine.
However, these new applications raise some crucial questions. Would I trust a car driver that cannot explain their actions? Do I want to be treated by an AI doctor that does not understand causality? Or do I feel comfortable deploying an AI consultant whose basic language skills were trained on highly biased and toxic online content? All of these questions describe real issues with state-of-the-art AI systems, that were praised for their incredible performance. However, such issues remain largely unsolved and thus may be argued to be the largest obstacles in modern AI research.
We are starting to see the limitations of the modern, neural network-based AI paradigm and its grim implications. In my research I explore ways for combining neural networks and other statistical approaches with explicit knowledge representations, such as knowledge graphs and ontologies. Can an AI look up the necessary knowledge from a database? Can it reason with strict, deductive logic? Yes, it can, however, this is currently extremely hard to implement in practice. Thus, currently, I am working on giving AI researchers new, powerful tools for building such hybrid systems.