\textit{FocaLogic}: Logic-Based Interpretation of Visual Model Decisions

TL;DR

FocaLogic tackles the interpretability of visual models by identifying minimal visual regions (visual focuses) that decisively influence predictions and encoding them as compact logical expressions. The framework is model-agnostic and training-free, using SAM-based segmentation, iterative region pruning, and a final states to logic translation $T(V)$, complemented by quantitative metrics including $\mathcal{P}$, $\mathcal{R}$, and $\mathcal{D}$. It demonstrates that focusing patterns become more concentrated with training, generalization improves focus precision, and biases or adversarial perturbations induce anomalous or distracted focuses, all while offering scalable, automated evaluation. The approach provides a principled, quantitative lens for diagnosing model behavior, biases, and robustness, with broad applicability to white-box and black-box settings and clear pathways for systematic analysis across architectures. ${T(V)}$ captures the decision logic in a compact form, enabling interpretable, structured explanations and facilitating practical deployment in high-stakes contexts.

Abstract

Interpretability of modern visual models is crucial, particularly in high-stakes applications. However, existing interpretability methods typically suffer from either reliance on white-box model access or insufficient quantitative rigor. To address these limitations, we introduce FocaLogic, a novel model-agnostic framework designed to interpret and quantify visual model decision-making through logic-based representations. FocaLogic identifies minimal interpretable subsets of visual regions-termed visual focuses-that decisively influence model predictions. It translates these visual focuses into precise and compact logical expressions, enabling transparent and structured interpretations. Additionally, we propose a suite of quantitative metrics, including focus precision, recall, and divergence, to objectively evaluate model behavior across diverse scenarios. Empirical analyses demonstrate FocaLogic's capability to uncover critical insights such as training-induced concentration, increasing focus accuracy through generalization, and anomalous focuses under biases and adversarial attacks. Overall, FocaLogic provides a systematic, scalable, and quantitative solution for interpreting visual models.

Figures (13)

• Figure 1: The overall workflow of FocaLogic. Given a target model and an input image, FocaLogic first localizes key visual regions that constitute the visual focus of the model on the image. It then derives a logical expression from the visual focus to interpret the model's decision-making behavior. Finally, it introduces a comprehensive set of metrics to evaluate the model's performance on the image from a focus-based perspective. $I_i$ denotes the $i^\text{th}$ visual region of image $I$.
• Figure 2: An example of visual focus refinement.
• Figure 3: Common types of visual model behavior from the focus-based perspective, with the ground-truth focus in blue and the model's visual focus in red.
• Figure 4: Horizontal comparisons on the localized visual focuses of ResNet18 and ViT-L16 in common scenarios. For the baselines, the localized regions are marked red. For FocaLogic, regions shared among all final states are marked red, and unique regions are marked white.
• Figure 5: Horizontal comparisons on the localized visual focuses of ResNet18 in different scenarios.