Neurosymbolic Object-Centric Learning with Distant Supervision
This work addresses the challenge of object-centric learning without object-level supervision for AI systems requiring structured reasoning, representing a novel integration of perceptual and symbolic components.
The paper tackles the problem of learning object-centric representations from raw unstructured perceptual data using only distant supervision, and shows that their neurosymbolic model DeepObjectLog outperforms neural and neurosymbolic baselines across various generalization settings.
Relational learning enables models to generalize across structured domains by reasoning over objects and their interactions. While recent advances in neurosymbolic reasoning and object-centric learning bring us closer to this goal, existing systems rely either on object-level supervision or on a predefined decomposition of the input into objects. In this work, we propose a neurosymbolic formulation for learning object-centric representations directly from raw unstructured perceptual data and using only distant supervision. We instantiate this approach in DeepObjectLog, a neurosymbolic model that integrates a perceptual module, which extracts relevant object representations, with a symbolic reasoning layer based on probabilistic logic programming. By enabling sound probabilistic logical inference, the symbolic component introduces a novel learning signal that further guides the discovery of meaningful objects in the input. We evaluate our model across a diverse range of generalization settings, including unseen object compositions, unseen tasks, and unseen number of objects. Experimental results show that our method outperforms neural and neurosymbolic baselines across the tested settings.