GRK 2853:  Neuroexplicit models of language, vision, and action

• Subject Area: Computer Science ; Linguistics
• Term: since 2023
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 471607914

The goal of this RTG is to develop novel neuroexplicit models that accurately solve tasks in natural language processing, computer vision, and action-decision making, and to investigate the theoretical and practical principles of designing effective neuroexplicit models. We will train three cohorts of PhD students in carrying out research on these models at the highest international level, producing graduates who are highly qualified for both the academic and the industrial job market. Over the past ten years, deep neural models have revolutionized computer science and set new standards for what artificial intelligence can achieve. They are in contrast to explicit models, which are designed to capture knowledge about a domain or a task with representations that can be understood and authored by human experts. Neuroexplicit models combine neural and explicit components, with the goal of inheriting the strengths of both model types. Explicit models include symbolic models, e.g. in computational linguistics and action planning, but also approaches to computer vision that model the physics of the world with differential equations. Thus neuroexplicit models encompass neurosymbolic models, which have recently been gaining in popularity across many areas of AI. Despite widespread and growing international interest, this RTG will be the first research center and, in particular, the first PhD training program for neuroexplicit methods in Europe. Neural models greatly outperform purely explicit models in terms of accuracy and scalability. At the same time, they have their own limitations, in particular with respect to generalization (learning abstractions which will transfer to related tasks), robustness to perturbations of the input, and interpretability. As the participating researchers of this RTG and others have shown, neuroexplicit models can help overcome these limitations. However, the design challenges of neuroexplicit models currently must be carefully addressed from scratch for each new task, and the fundamental design principles of neuroexplicit models are not yet well understood. This makes the development of novel neuroexplicit models cumbersome, requiring extensive experimentation to get the model right. In this RTG, we will work out design principles for effective neuroexplicit models through the interdisciplinary collaboration of PhD students from multiple fields of AI, and thus accelerate the development of efficient and accurate neuroexplicit models in the future.

• DFG Programme: Research Training Groups
• Applicant Institution: Universität des Saarlandes
• Participating Institution: CISPA - Helmholtz-Zentrum für Informationssicherheit; Deutsches Forschungszentrum für Künstliche Intelligenz GmbH (DFKI); Max-Planck-Institut für Informatik; Max-Planck-Institut für Softwaresysteme
• Spokesperson: Professor Dr. Alexander Koller
• Participating Researchers: Professorin Dr. Vera Demberg; Professor Dr. Jörg Hoffmann; Professor Dr.-Ing. Eddy Ilg; Professor Dr. Dietrich Klakow; Professor Dr. Bernt Schiele; Professor Dr.-Ing. Philipp Slusallek; Professor Dr. Christian Theobalt; Professorin Dr. Mariya Toneva, Ph.D.; Professorin Dr. Isabel Valera; Professor Dr. Jilles Vreeken; Professor Dr. Joachim Weickert; Professorin Dr. Verena Wolf