Neurosymbolic Epistemic Deep Learning for Hierarchical Image Classification
A recent study introduces a comprehensive framework that combines neurosymbolic and epistemic modeling, enhancing Swin Transformers through focal set reasoning and differentiable fuzzy logic for hierarchical image classification. This approach tackles the shortcomings of deep neural networks, which frequently generate overly confident predictions and breach logical or structural constraints, particularly in hierarchical classifications where both fine and coarse predictions need to align. By integrating data-driven focal sets into the learned embedding space, the framework effectively captures epistemic uncertainty across various plausible fine-grained categories. Additionally, a belief-theoretic layer that utilizes fuzzy membership functions and t-norm conjunctions promotes consistency among predictions across different levels of the hierarchy. The research can be found on arXiv with the identifier 2605.16383.
Key facts
- Proposes first unified neurosymbolic and epistemic modeling framework for hierarchical image classification.
- Augments Swin Transformers with focal set reasoning and differentiable fuzzy logic.
- Addresses overconfident predictions and violation of logical constraints in deep neural networks.
- Induces data-driven focal sets within embedding space to capture epistemic uncertainty.
- Uses belief-theoretic layer with fuzzy membership functions and t-norm conjunctions.
- Ensures consistency between fine-grained and coarse-grained predictions.
- Published on arXiv with identifier 2605.16383.
- Focuses on hierarchical classification tasks.
Entities
Institutions
- arXiv