Weakly Supervised Concept Learning for Object-centric Visual Reasoning

Researchers present a weakly supervised learning approach that combines neural networks with symbolic AI for object-centric reasoning tasks, requiring only 1% of typical labels while outperforming foundation models in domain generalization. The method bridges perception and logical reasoning by using slot-based architectures and VAEs to ground symbolic outputs for frameworks like Inductive Logic Programming.

This research addresses a fundamental challenge in neurosymbolic AI: efficiently bridging the gap between raw sensory perception and symbolic reasoning without requiring extensive labeled datasets. The approach demonstrates that perception stages in two-stage neurosymbolic systems can operate effectively with minimal human supervision by leveraging self-supervised learning mechanisms that compete for interpretability alongside task performance.

The broader context reflects growing interest in hybrid AI systems that combine neural networks' pattern recognition capabilities with symbolic AI's reasoning transparency and sample efficiency. Traditional end-to-end differentiable approaches struggle with interpretability and require substantial labeled data, while decoupled systems have faced bottlenecks from expensive perception labeling. This work directly tackles that bottleneck through innovative weak supervision.

For the AI development community, these findings carry substantial implications. The ability to achieve strong performance with 1% supervision while maintaining robustness under domain shift suggests practical pathways for deploying AI systems in data-scarce or frequently shifting environments. The method's compatibility with multiple reasoning frameworks—ILP, Decision Trees, Bayesian Networks—indicates broad applicability across different problem domains requiring explainable AI solutions.

Developers and researchers should monitor whether these results generalize to more complex real-world datasets beyond the synthetic benchmarks tested. The performance advantage over foundation models specifically in domain generalization settings suggests opportunities for building more efficient, interpretable systems for specialized applications where transfer learning alone proves insufficient. Future work will likely explore scaling these approaches to larger datasets and more complex reasoning tasks.

• →Weakly supervised neurosymbolic approach achieves strong performance with only 1% of typical label requirements.
• →Method combines slot-based architectures with VAE self-supervision for interpretable symbol grounding in reasoning tasks.
• →Outperforms foundation model baselines in domain generalization despite using significantly less supervision.
• →Two-stage decoupling of perception and reasoning avoids optimization issues while maintaining symbolic interpretability.
• →Approach demonstrates robustness to substantial domain shift while supporting multiple reasoning frameworks.