Doctoral Schools WUT

Red teamming AI foundation models. Foundation models (for language or image data) will play an increasingly important role in AI systems. They will also provide more and more opportunities for a diverse range of applications, often not foreseen at the model training stage. An in-depth analysis and diagnostics of foundation models will become crucial in ensuring the security of AI systems. Today, however, there are no good methods or tools to conduct such analysis and diagnostics. This research direction will create methods, processes and tools to support the search for vulnerabilities in AI systems using foundation models.

• Keywords: Wyjaśnialne uczenie maszynowe |

Polyrepresentation learning for image or text models. AI systems use numerous approaches to train data representations. Often, however, a single approach is based on one selected perspective related to the loss function used to train the representation. When looking for fundamental models, it may be beneficial to combine perspectives into a broader representation - a polyrepresentation.
This line of research will focus on analyzing what representations provide complementary perspectives and how to build and analyze such perspectives. The ongoing research may concern one modality (text or image) or different modalities.

• Keywords: Sztuczna inteligencja | Modele podstawowe | Red teamming |