Neural Rule Inducer: Zero-Shot Logical Rule Induction
Scientists have unveiled the Neural Rule Inducer (NRI), a pretrained model designed for zero-shot logical rule induction. In contrast to conventional Inductive Logic Programming (ILP) techniques that necessitate retraining for every new task, NRI leverages domain-agnostic statistical characteristics to depict literals, facilitating generalization across different identities and quantities. This model incorporates a statistical encoder alongside a parallel slot-based decoder, which maintains the permutation invariance of logical disjunction. Additionally, the use of Product T-norm relaxation renders rule execution differentiable, enabling comprehensive training focused on prediction accuracy. The performance of NRI is assessed through rule recovery and robustness evaluations.
Key facts
- Neural Rule Inducer (NRI) is a pretrained model for zero-shot rule induction.
- NRI uses domain-agnostic statistical properties like class-conditional rates, entropy, and co-occurrence.
- The model consists of a statistical encoder and a parallel slot-based decoder.
- Parallel decoding preserves permutation invariance of logical disjunction.
- Product T-norm relaxation makes rule execution differentiable.
- NRI is evaluated on rule recovery and robustness.
- Existing ILP methods are transductive and require retraining for each new task.
- NRI generalizes across variable identities and counts without retraining.
Entities
—