About the RTG "Neuroexplicit Models"

Cutting-edge research

Neural models have revolutionized artificial intelligence, but they also have systematic limitations.

The research in this RTG will help overcome these limitations by combining neural models with human-interpretable explicit models into neuroexplicit models.

Learn More 

World-class faculty

Each PhD student in the RTG is co-supervised by two Principal Investigators, often across research area boundaries.

These are some of the top researchers in their fields in Europe and worldwide.

Learn More 

In-depth qualification program

An RTG is not just a research project - it is a training program for PhD students. Students in the RTG are part of a vibrant community that meets for joint activities and collaborates intensively with each other.

Learn More 

Research Program

The goal of this Research Training Group 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. 24 PhD students, together with 14 PIs and ~20 associated PhD students and postdocs, are carrying out research on neuroexplicit models at the highest international level.

Deep neural models have revolutionized artificial intelligence over the past ten years, by learning from data how to perform a variety of AI tasks at unparalleled accuracy. By contrast, explicit models capture knowledge about a task or a domain in a way that can be understood or authored by human experts, and can therefore be more interpretable and data-efficient. Explicit models can use symbolic representations, or they can capture domain knowledge in other ways, e.g. through differential equations that describe the physics of the world. Neuroexplicit models (such as neurosymbolic ones) combine neural and explicit elements, inheriting the complementary strengths of both.

Neural models have produced quantum leaps in what artificial intelligence can achieve. However, they also have systematic limitations, especially with respect to generalization, robustness, and interpretability. We and others have demonstrated that neuroexplicit models have the potential to overcome these limitations. However, neuroexplicit models need to be designed with care, and the principles of effective design are not well understood. By investigating the rich and varied landscape of neuroexplicit models across multiple areas of artificial intelligence in one coherent, interdisciplinary research group, the RTG will boost our understanding of these models. The RTG is the first research center and in particular the first PhD training program for neuroexplicit methods in Europe.

PhD students in the RTG choose their thesis topic freely, together with their two advisors. We welcome research topics that cut across the boundaries of research fields as much as in-depth advances within a field; some examples are shown below. Each PhD student is funded for four years and has access to travel funds, compute resources, and organizational support. The admission process is selective - we are looking for excellent PhD students in the world who will produce impactful research and become leaders in academia or industry.

Research in the RTG is loosely organized into four research areas: Language, Vision, Action, and Foundations.

Research Areas

L

Language

PIs: Demberg, Hahn, Klakow, Koller, Toneva

Neural models of language, such as GPT-4, have shown groundbreaking performance across even very complex natural language processing tasks. However, these models are data-hungry and slow, intrinsically uninterpretable, struggle with hallucinations and certain kinds of generalization, and are vulnerable to bias and noise in the data. We aim to combine neural and symbolic models that are consistently truthful, can reason robustly, and are linguistically interpretable.

V

Vision

PIs: Ilg, Lenssen, Schiele, Theobalt

Neural models have enabled great improvements across a wide range of computer vision problems, such as estimating correspondences between images and reconstruction of humans and objects in their environment. They are efficient and scale to diverse scenes, but they rely on large quantities of training data and are severely lacking in generalization, robustness, and interpretability. We aim to overcome these limitations by combining them with parametric models, e.g. of the world physics, and set new states of the art in accuracy and runtime efficiency.

A

Action

PIs: Hoffmann, Slusallek, Wolf

Deep reinforcement learning models can beat human world champions in complex games such as chess and Go and are successful across many problems in sequential decision making. However, they do not generalize across related problems, struggle to handle events that were rare in training, and are hard to trust due to lack of interpretability. We aim to learn explicit abstractions that improve generalization and improve the robustness towards rare events and uncertain sensor values through explicit simulations of the environment.

F

Foundations

PIs: Valera, Vreeken, Weickert

The RTG aims not only to improve the state of the art in Language, Vision, and Action, but also to identify the underlying principles of developing effective neuroexplicit models. We aim to develop novel theory and methods in machine learning that directly improve generalization, robustness, and interpretability and cut across the different application areas. This includes training methods that generalize across different domains and objectives, as well as models that can be provably seen as either explicit or neural and thus will help us overcome the distinction between the modeling paradigms.

Qualification Program

A Research Training Group (RTG) is not just a research project - it is a training program for PhD students. Students in the RTG are part of a vibrant community that meets for joint activities and collaborates intensively with each other.

Joint activities

We want our PhD students to be a tight-knit group who work together towards a common goal. You will spend half your time in "RTG offices" near each other, and half your time in your primary advisor's group. You will also meet weekly for brainstorming sessions and reading groups.

Once a year, the whole RTG (students, PIs, and guests) will go on a one-week retreat, e.g. to Schloss Dagstuhl in the Northern Saarland. The program will be designed by the students to fit your interests and could e.g. involve a hackathon and joint research projects. You can also help decide what international guests to invite.

Coursework

We want our PhD students to talk to each other at a technical level, and therefore expect that every student has taken classes in at least two research areas.  If you have focused on a single area so far, you can learn about another one in our MSc-level courses.

Every student in the RTG will also take our award-winning course "Ethics for Nerds". AI increasingly affects our lives at many levels, and as researchers we must live up to our responsibility to having our technology used for good.

People skills

We want to prepare our students for a successful career after the PhD. Each student will typically take two courses per year (one or two days each) on topics such as scientific writing, presenting, self-management, or talking to the media. Students get to select their course program themselves, with support from the RTG office.

You will also have the opportunity to teach classes for advanced students, supervise students assistants or MSc theses, and participate in outreach activities. Of course teaching activities are totally voluntary.