Task Leader: Luc de Raedt (KU Leuven) - luc.deraedt@kuleuven.be

Goal: develop integrated representations for learning and reasoning and accompanying scalable algorithms for learning and inference, and providing explanations

• Study how to integrate various types of representations, in particular (subsets of) logic, probabilistic graphical models, constraints, neural models, and embeddings. Develop the necessary foundations, semantics, loss functions, and inference and learning techniques. Of particular interest is the level to which different representations can be unified, that is, combined in a way that preserves the properties of the original representations. Special attention will be given to frameworks that not only allow for end-to-end differentiation but that also support symbolic reasoning.

• Design and implement scalable algorithms for learning and reasoning. T4.1 will build upon existing streams of research, in particular NeSy and StarAI, and more generally, combinations for knowledge representation and reasoning with machine learning.

• Develop methods to explain the results of inference and learning for the integrated representations. To this end, T4.1 will realize synergies with symbolic reasoning and differentiable programming. We will extend relevance propagation, attention mechanisms and similar explanation approaches to integrated representations. Likewise, we will explore uncertainty quantification to understand the learned models, also in connection with uncertainty tolerant reasoning. In particular, we will consider estimation and propagation of epistemic uncertainty in modular deep learning systems.

Task Leader: Michela Milano (University of Bologna) - michela.milano@unibo.it

Goal: obtain combinatorial optimisation models and solvers that learn from experience

Study foundations, techniques, algorithms and tools for integrating learning into combinatorial optimisation techniques such as constraint programming. At least two settings are relevant here. First, how elements of the combinatorial optimisation models (like parameters, constraints and optimisation functions) can be learned from past data or experience. A key challenge here is how to learn the optimisation function and the constraints together. A second question is how to continuously improve the performance of the optimisers, e.g., by using ML to choose amongst different formulations of problems, or to aid with reformulation steps to help with a reasoning or explanation task. This question of course also applies to continuous optimisation and work on optimisation in nonconvex spaces, and this subtask will profit from interactions with WP 7 on AutoAI.

Task Leader: Marco Gori (University of Siena) - marco.gori@unisi.it

Goal: provide dedicated integrated learning and reasoning approaches for working with embeddings, knowledge graphs, & ontologies.

Building upon and extending the results of Task 4.1, this task will develop dedicated methods for this type of application given the prominence of this domain. In particular, it will develop methods to learn semantically rich embeddings that preserve as much relational information as possible of the underlying data, and to perform efficient reasoning tasks (also multi-step) using embedded space. It will generalize the concept of semantic spaces to arbitrary domains, data types and relations, and study the feasibility of using embeddings as the core internal representation bridging low-level data processing and high-level reasoning, learning and optimization. Based on this, Task 4.3 will then investigate learning methods that combine discriminative models with generative models that are respecting expert knowledge in the form of, for instance, probabilistic rules and programs. The goal is to understand how expert-driven generative models can boost pure data driven models. Finally, links to geometric and graph-based deep learning will be explored.

Task Leader: Bastian Leibe (RWTH Aachen Center for Artificial Intelligence) - leibe@vision.rwth-aachen.de

Goal: provide dedicated integrated learning and reasoning approaches for scene understanding.

Recent deep learning methods in vision can provide an account of what is visible in a scene at unprecedented levels of accuracy, but they are still very limited in their interpretation why something is happening. This is a classic reasoning task, but common representations need to be found in order to connect learning and reasoning. In this task, we will explore the connection between learning and reasoning in the context of perception and intention estimation tasks in vision and robotics, for which large-scale datasets and benchmarks exist.

Task Leader: Jens Lehmann (Fraunhofer IAIS) - jens.lehmann@iais.fraunhofer.de

Task Leader: Kristian Kersting (TU Darmstadt) - kersting@cs.tu-darmstadt.de