New Neurosymbolic Framework Boosts Swin Transformers with Focal Set Reasoning for Hierarchical Classification

Deep neural networks often produce overconfident predictions and violate logical constraints, especially in hierarchical image classification where predictions across fine and coarse levels must be coherent. To address this, researchers introduce a novel neurosymbolic approach that extends Swin Transformers with focal set reasoning—data-driven sets within the learned embedding space that capture epistemic uncertainty over multiple plausible fine-grained classes. These focal sets feed into a belief-theoretic layer that uses fuzzy membership functions and t-norm conjunctions to enforce consistency between fine and coarse predictions. A learnable loss adaptively trades off symbolic structure with data-driven evidence, optimizing calibration, mass regularization, and logical consistency.

The framework maintains accuracy on par with standard transformer baselines while delivering more calibrated and interpretable predictions. Experimental results show reduced overconfidence and high logical consistency across hierarchical outputs. This is the first unified model combining focal set reasoning with fuzzy logic for hierarchical vision tasks, offering a practical step toward deep learning that is both accurate and epistemically aware. The full paper (36 pages) is available on arXiv (2506.16383).