Thematic Axis: Knowledge Representations and Reasoning
Leader(s)
Description
Here the question of how a system can integrate an inference engine based on explicitly represented knowledge is addressed. Such a system is capable of reasoning, where reasoning is taken in a broad sense: deductive reasoning (i.e. reasoning such that if its premises are considered to be certainly true then their conclusion are necessarily true as well) or hypothetical reasoning (a.k.a non-deductive / non-monotonic / defeasible reasoning, including potentially inductive generalization, abduction, analogy and case-based reasoning). Considering this, representation and reasoning are considered in a broad sense, and therefore also include the study of the capability of deep learning systems (and Large Language Models, LLMs) to learn representations and to exhibit reasoning-like behaviours.
The motivation for focusing on knowledge representation and reasoning as research topics relates to the specific benefits of those methods: they perform with low volumes of data, enable rich and coherent encoding of complex and distributed information, and they are highly interpretable. In that sense, this axis is strongly related to the ones one Hybrid AI (integrating knowledge representations in machine learning methods, including through knowledge graph embedding and graph neural networks), explicability and interpretability (by relying on reasoning traces to justify predictions made by machine learning models), and frugal AI (how knowledge models can represent lighter surrogates for deep learning models). It also, naturally connects to the distributed AI axis since it shares with it the idea of graph-based representation of distributed information/knowledge, and with the uncertainty axis, considering that uncertainty representation can be studied from the point of view of knowledge representation formalisms. Finally, there is an increasing interest in relating research on knowledge representation and reasoning with research on LLMs, not only because LLMs can be used to generate knowledge representations, but also to enable the study and analysis of the capabilities of LLMs with respect to reasoning and processing learned knowledge.