VisionLogic: Discovering and Grounding Decision-Relevant Visual Concepts

VisionLogic introduces a novel neuralsymbolic framework that produces faithful, hierarchical explanations as global logical rules over causally validated concepts, addressing the limitation of prior methods that rely solely on correlational signals. The gist: VISIONLOGIC is the first framework to deliver both causally validated concepts and interpretable logic-rule explanations.

How it works

VISIONLOGIC operates in two main stages to transform neuron activations into interpretable logical rules. First, it learn[s] activation thresholds that abstract neuron activations into predicates, which are then used to induce class-level logical rules from these predicates. These predicates provide an intermediate symbolic representation of the model’s internal reasoning.

Second, the framework ground[s] these predicates into high-level visual concepts via ablation-based causal tests. This involves:

1. Starting with an initial bounding box likely to influence the predicate.

2. Perturbing the region with random noise or similar masking strategies and checking whether this flips the predicate’s truth value. A transition from activation to deactivation provides causal evidence that the region is critical for the predicate.

3. Proposing an efficient algorithm to iteratively refine the bounding box for more precise localization.

4. Using segmentation methods such as Mask R-CNN or SAM to validate the intersection of the segmentation and refined box, thereby strengthening causal validity.

Deriving Predicates from Neuron Activations

The process begins by analyzing final-layer activations, where the logit is computed as a function of channel contributions: Fc(x) = WcZ(x) + bc. The framework then converts real-valued activations into binary predicates, denoted as pj(x) ∈ [0, 1], which serve as logical atoms. Instead of fixing ad hoc thresholds, VISIONLOGIC learns per-channel thresholds Tj and sharpness sj > 0 via a differentiable gate: p˜j(x) = σsj (zj (x − Tj)) or the hard predicate pj(x) = I(zj (x ≥ Tj). The learning objective minimizes a loss function that includes distillation from the frozen teacher network, stability terms to keep thresholds near an initial seed T(0), and a compact predicate set penalty using a group lasso over rank variants.

Learning Logical Rules and Inference Score

Once the predicate vocabulary P is learned, VISIONLOGIC induces symbolic rules for class-level decision making. For each class c, it evaluates all predicates on correctly classified training examples to define a Conjunctive Normal Form (DNF) capturing the class’s training patterns: ∀x, v∈Vc i:vi=1 pi(x) ∧ i:vi=0 ¬pi(x) ! =⇒ Label(x) = c. To summarize these patterns, it builds a class profile by counting predicate appearances across the class-c clauses in Eq. 6 and sorting predicates by frequency to obtain a ranking Rc(pi). The inference score for a test input x is computed as: S(x, c) = 1/P(x) Σ pj∈P (x) Rc(pj), and the predicted class is the one that minimizes this score.

Grounding Predicates to Vision Concepts

The final stage links abstract predicates to interpretable visual features through causal validation. This is achieved by testing whether a specific region is necessary for a predicate by replacing it with random noise to obtain x', and recomputing the predicate pj(x'). A flip from activation to deactivation signifies that the region is causally important for pj(x). The algorithm iteratively refines bounding boxes using this test, and further refinement can incorporate segmentation masks (e.g., SAM) to confirm causality by intersecting the mask with the refined box and repeating the intervention. This process results in consistent, causally supported visual concepts across images of the same class.

Empirical Validation

The framework is validated through extensive human studies and model performance checks. In human evaluations, VISIONLOGIC's concept explanations significantly enhance participants’ understanding of model behavior over prior concept-based methods, achieving high utility scores, particularly in identifying bias and novel strategies. Experimentally, across different architectures like CNNs and ViTs, VISIONLogic largely retains the original model’s predictive performance, maintaining over 90% top-5 test accuracy on covered images while providing explanations that are both causally grounded and human-understandable. Furthermore, analysis of adversarial attacks shows that misclassification is often explained by Type A or Type B causes related to the deactivation or activation of specific predicates. In evaluating logical rules, VISIONLOGIC provides the first explicit, global interpretable rules for large vision models such as CNNs and ViTs at a scale where prior rule-extraction approaches were limited.

As a fastidious and diligent researcher, I have analyzed the VISIONLOGIC framework. It addresses the critical limitation of existing concept-based methods by introducing causal validation through ablation-based tests, leading to robust, interpretable logical rules over concepts.

Based on this scientific paper, here are the specific improvements that can be made to AI systems:

1. The core improvement is shifting from purely correlational explanations (like TCAV or ACE) to a framework of causally validated concepts. This moves explanations from statistical associations to genuine causal links between visual features and model decisions.

2. The system generates global logical rules defined over these concepts, rather than just local attribution maps (like Grad-CAM).

Specific Capabilities of the Improved AI System:

1. An AI system can generate a set of compact, high-level logical rules that approximate the base model's class-level decision-making process (e.g., If Predicate A and Predicate B are true, then the prediction is Class C).

2. The system provides human-understandable explanations for complex decisions by grounding these rules in specific, causally validated visual concepts (e.g., instead of showing a heat map, it shows The presence of 'squirrel tail' and 'bird beak' causes this image to be classified as 'Bird').

3. The system can perform robust model probing by identifying the root causes (Type A vs. Type B) of misclassification under different adversarial attacks (like Pixelation or PGD), allowing researchers to understand the underlying logic that makes a model vulnerable or robust.

4. The system offers enhanced trust and accountability for high-stakes applications (e.g., medical diagnostics, autonomous driving) by providing explanations that are not just plausible but causally grounded and validated through iterative refinement using techniques like Mask R-CNN or SAM for precise localization.

5. The system can be generalized across different vision architectures (CNNs, ViTs, Swin Transformers) while maintaining its symbolic logic structure, ensuring the interpretability framework is not tied to a single network design.