Thematic Axis: Hybrid AI

Leader(s)

Descritpion

In this thematic axis, we study and propose hybrid approaches (symbolic, sub-symbolic, numerical) for humans, to make them interactive, and hybrid by design to be interpreted and explained.

More precisely, our research in this axis focuses on the following challenges

Hybrid approaches (symbolic, sub-symbolic, numerical) for human (interactive)
How to inject symbolic expert knowledge into numerical approaches
How numerical approaches can help symbolic approaches
How symbolic approaches can help numerical approaches
Proposing or developing hybrid approaches by design to be interpretable and explicable

The motivation for focusing on hybrid AI as a research topic relates to the recognized limitations of any single approach taken in isolation: purely symbolic systems struggle to scale and to handle noisy, incomplete data, while purely data-driven systems lack transparency and require large volumes of labelled examples. Hybrid approaches offer a way to address both sets of limitations simultaneously, producing systems that are better performing, more data-efficient and more explanable. In that sense, this axis is most directly related to the knowledge representations and reasoning axis, since the combination of symbolic representations with machine learning methods (e.g., through knowledge graph embeddings, graph neural networks, or neuro-symbolic architectures) is among the most common forms of hybridation. It is also closely linked to the explicability and interpretability axis, since the symbolic components of a hybrid system naturally provide reasoning traces and structured justifications that can serve as the basis for explanations, and since human-AI hybrid systems are partly motivated by the need for transparent, accountable decision-making. Finally, the NLP (Natural Language Processing) and LLMs axis is increasingly relevant as LLMs are being integrated with symbolic components in retrieval-augmented and tool-augmented architectures.