Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Right for the Right Reasons: Avoiding Reasoning Shortcuts via Prototypical Neurosymbolic AI

Luca Andolfi, Eleonora Giunchiglia

TL;DR

Prototypical Neurosymbolic AI tackles reasoning shortcuts in neurosymbolic learning by grounding neural predicates to class prototypes and updating embeddings with both background knowledge and proximity to labelled exemplars. The approach yields a disentangled, annotation-efficient NeSy model that reduces deterministic shortcuts and improves concept–semantic alignment, as demonstrated on rsbench tasks MNIST-EvenOdd, Kand-Logic, and BDD-OIA. Theoretical analysis quantifies shortcut possibilities and empirical results show substantial improvements under very limited supervision. This yields a practical, scalable strategy for safe and reliable neurosymbolic reasoning, especially in low-data regimes.

Abstract

Neurosymbolic AI is growing in popularity thanks to its ability to combine neural perception and symbolic reasoning in end-to-end trainable models. However, recent findings reveal these are prone to shortcut reasoning, i.e., to learning unindented concepts--or neural predicates--which exploit spurious correlations to satisfy the symbolic constraints. In this paper, we address reasoning shortcuts at their root cause and we introduce prototypical neurosymbolic architectures. These models are able to satisfy the symbolic constraints (be right) because they have learnt the correct basic concepts (for the right reasons) and not because of spurious correlations, even in extremely low data regimes. Leveraging the theory of prototypical learning, we demonstrate that we can effectively avoid reasoning shortcuts by training the models to satisfy the background knowledge while taking into account the similarity of the input with respect to the handful of labelled datapoints. We extensively validate our approach on the recently proposed rsbench benchmark suite in a variety of settings and tasks with very scarce supervision: we show significant improvements in learning the right concepts both in synthetic tasks (MNIST-EvenOdd and Kand-Logic) and real-world, high-stake ones (BDD-OIA). Our findings pave the way to prototype grounding as an effective, annotation-efficient strategy for safe and reliable neurosymbolic learning.

Right for the Right Reasons: Avoiding Reasoning Shortcuts via Prototypical Neurosymbolic AI

TL;DR

Prototypical Neurosymbolic AI tackles reasoning shortcuts in neurosymbolic learning by grounding neural predicates to class prototypes and updating embeddings with both background knowledge and proximity to labelled exemplars. The approach yields a disentangled, annotation-efficient NeSy model that reduces deterministic shortcuts and improves concept–semantic alignment, as demonstrated on rsbench tasks MNIST-EvenOdd, Kand-Logic, and BDD-OIA. Theoretical analysis quantifies shortcut possibilities and empirical results show substantial improvements under very limited supervision. This yields a practical, scalable strategy for safe and reliable neurosymbolic reasoning, especially in low-data regimes.

Abstract

Neurosymbolic AI is growing in popularity thanks to its ability to combine neural perception and symbolic reasoning in end-to-end trainable models. However, recent findings reveal these are prone to shortcut reasoning, i.e., to learning unindented concepts--or neural predicates--which exploit spurious correlations to satisfy the symbolic constraints. In this paper, we address reasoning shortcuts at their root cause and we introduce prototypical neurosymbolic architectures. These models are able to satisfy the symbolic constraints (be right) because they have learnt the correct basic concepts (for the right reasons) and not because of spurious correlations, even in extremely low data regimes. Leveraging the theory of prototypical learning, we demonstrate that we can effectively avoid reasoning shortcuts by training the models to satisfy the background knowledge while taking into account the similarity of the input with respect to the handful of labelled datapoints. We extensively validate our approach on the recently proposed rsbench benchmark suite in a variety of settings and tasks with very scarce supervision: we show significant improvements in learning the right concepts both in synthetic tasks (MNIST-EvenOdd and Kand-Logic) and real-world, high-stake ones (BDD-OIA). Our findings pave the way to prototype grounding as an effective, annotation-efficient strategy for safe and reliable neurosymbolic learning.

Paper Structure

This paper contains 33 sections, 3 theorems, 32 equations, 16 figures, 12 tables.

Table of Contents

  1. Introduction
  2. Problem Statement
  3. Notation.
  4. Setting.
  5. Prototypical Neurosymbolic AI
  6. Reasoning over Distances
  7. Counting Deterministic Reasoning Shortcuts
  8. Experimental Analysis
  9. Related Work
  10. Conclusion
  11. Proof Theorem \ref{['th:embedding_update']}
  12. Statement.
  13. Proof.
  14. Proof Corollary \ref{['th:embedding_update_sl']}
  15. Proof.
  16. ...and 18 more sections

Key Result

Theorem 4.1

Let $\mathbf{y}$ be the output of a prototypical NeSy predictor reasoning over the knowledge $\textsc{K}$ and the set of concepts $\mathcal{C} = [h_1] \times \ldots \times [h_k]$. If the predictor is trained using $\mathcal{L}^{\text{\rm{NeSy}}}(\mathbf{y}, \textsc{K})$, then for every $i \in [k]$ t

Figures (16)

  • Figure 1: Ground truth generation process (in black).
  • Figure 2: Prototype configuration for Example \ref{['ex:animals']}.
  • Figure 3: MNIST-EvenOdd concept confusion matrices.
  • Figure 4: F1(C) on MNIST-EvenOdd across unlabelled data ratios.
  • Figure 5: F1(C) and Cls(C) on BDD-OIA across different percentages of unlabelled datapoints. The green dotted line represents the baseline non-supervised DPL.
  • ...and 11 more figures

Theorems & Definitions (6)

  • Example 1
  • Example 2
  • Theorem 4.1
  • Corollary 4.2
  • Example 3
  • Proposition 5.1