ConvAD: Robust AI Explanations for Vision

• ConvAD is a dual-framework system offering both post-hoc activation-deactivation for in-distribution, causally valid CNN explanations and concept-aware visual anomaly detection with human-interpretable outputs.
• The Activation-Deactivation module integrates seamlessly with CNNs, turning off internal activations for masked feature regions to ensure robust, state-consistent explanations.
• In its anomaly detection mode, CONVAD leverages concept bottleneck models alongside pixelwise and semantic analyses to achieve high performance and clear, actionable insights.

ConvAD refers to two distinct yet influential frameworks in explainable artificial intelligence for computer vision. In post-hoc explanation of classifiers, ConvAD is a drop-in module implementing the Activation-Deactivation (AD) paradigm, which yields robust and causally sound explanations by “deactivating” internal activations corresponding to masked input features without generating out-of-distribution mutants (Chanchal et al., 1 Oct 2025). In the domain of Visual Anomaly Detection (VAD), CONVAD denotes the Concept-Aware Visual Anomaly Detection system, which unifies concept bottleneck models (CBMs) with VAD to provide both pixelwise and semantic, human-interpretable anomaly explanations (Stropeni et al., 25 Nov 2025). This article articulates the algorithms, mathematical formulations, practical workflows, and comparative empirical analyses for both frameworks.

1. Activation-Deactivation ConvAD: Framework and Motivation

The ConvAD module implements the Activation-Deactivation (AD) paradigm for explaining convolutional neural networks (CNNs) (Chanchal et al., 1 Oct 2025). In contrast to input-occlusion methods—where input regions are masked and the classifier is queried with perturbed, out-of-distribution inputs—AD performs “model-occlusion.” Here, a binary mask $M$ is propagated through the CNN in lockstep with feature maps, such that activations corresponding to masked regions are internally “switched off” via elementwise gating at specific checkpoints.

This approach addresses two persistent issues in traditional black-box explanation methods:

• Out-of-distribution mutants: Static masking (e.g., zero, min, mean pixel values) yields $x'$ that lie outside the training distribution, leading to unreliable explanations.
• Domain-specific hyperparameters: The choice of occlusion value often requires domain knowledge and profoundly influences results.

By ensuring that the original input $x$ always remains in-distribution and activations for masked regions are precisely and systematically deactivated within the network, ConvAD yields explanations that are both robust and causal.

2. Algorithmic Workflow and Mathematical Formulation

ConvAD operates as a drop-in module, requiring no retraining or fine-tuning of the original CNN $N$. At each checkpoint—after every convolution, and before/after spatial dimension changes—a lock-stepped binary gating is performed:

1. Input: $(x, M^{(0)})$, where $x \in \mathbb{R}^{H\times W\times C_{in}}$ is the input image and $M^{(0)} \in \{0,1\}^{H\times W}$ the initial mask.
2. Position Attribution: For each layer $i$, the fraction $P_i(z^{(i)}, M^{(0)})_{a,b}$ of the receptive field for feature map position $(a,b)$ that remains unmasked is computed using a convolution of $x'$0 with a mean kernel corresponding to the receptive field size.
3. Gate Construction: For each location,

$x'$1

where $x'$2 is a threshold (default $x'$3).

1. Gating: Internal activations are gated via Hadamard product:

$x'$4

The flow is formalized as:

$x$4

This ensures that, for $x'$5 (all-ones mask), $x'$6, so ConvAD is a strict restriction of the original model, guaranteeing decision function invariance.

3. Theoretical Properties and Causal Guarantees

The AD paradigm, as operationalized in ConvAD, admits several formal properties (Chanchal et al., 1 Oct 2025):

• Decision-function invariance: Adding ConvAD to $x'$7 does not alter its output for unmasked images, i.e., $x'$8 if $x'$9.
• Causal explanation validity: Any AD explanation meets EXIC1–EXIC3 causal explanation conditions over all possible occlusions, and AD-explanations are prime-implicant explanations. This grants grounded interpretability: the highlighted features are minimal sets sufficient for the model’s actual prediction.

4. Empirical Performance and Comparative Evaluation

Extensive evaluations were performed using state-of-the-art CNNs (ResNet-50, RegNetY-12GF, EfficientNet-V2-S) on canonical image classification datasets (ImageNet-1k, ImageNet-v2, CalTech-256, PASCAL-VOC). ConvAD explanations were compared with occlusion-based masking (min, max, mean, zero) on metrics including $x$0-robustness, explanation size, and confidence drop-off.

Representative results:

Model	p-Robustness (AD)	p-Robustness (Best Mask)	Relative Gain (%)
ResNet-50 (y=0.9)	0.60	0.35	71
EfficientNet-V2 (y=0.9)	0.65	0.40	62.5

ConvAD leads to consistent improvements (30–40% mean, up to 62.5% best-case) in robustness, with only modest increases in explanation size (2–20%, largest at high confidence thresholds). The method reliably hits requested confidence thresholds while preserving higher robustness than any occlusion value baseline (Chanchal et al., 1 Oct 2025).

5. Practical Usage, Limitations, and Extensibility

ConvAD is architecturally agnostic, requiring only lightweight checkpoint logic and thus minimal computational overhead at inference. Key practical considerations include:

• Threshold T selection: Default $x$1 disables only fully masked regions; $x$2 may be necessary to handle mixed receptive fields ("leakage") due to padding or partial overlap, controlling for inadvertent feature influence.
• Explanation size: Larger explanations can increase annotation effort but may enhance interpretability.
• Extensibility: The AD paradigm generalizes to transformer-based and MLP architectures by introducing analogous gating mechanisms on internal activations.
• No retraining: Operates entirely as a post-hoc insertion.

ConvAD is susceptible to leakage if sliding receptive fields cover both masked and unmasked regions; careful thresholding mitigates this.

6. CONVAD for Explainable Visual Anomaly Detection

In the VAD setting, CONVAD signifies a different system—Concept-Aware Visual Anomaly Detection—that unifies Concept Bottleneck Models (CBMs) with feature-based anomaly localization (Stropeni et al., 25 Nov 2025). The system delivers:

• Visual explanations: Student–teacher feature matching highlights anomalous pixels.
• Semantic explanations: Concept vectors $x$3 (e.g., "scratch," "hole") explain anomalies in human-interpretable terms.

Key technical features include:

• Automated concept dataset construction via VLM (Gemma3) and CLIP encoding.
• Synthetic anomaly generation using generative editing (Nano Banana) to maintain unsupervised VAD constraints.
• Dual-branch architecture: a concept-based "bottleneck" classifier and a feature-matching visual branch.
• Intervention capacity: domain experts can override uncertain concepts at test-time to boost accuracy.

Evaluation on benchmarks (MVTec AD, VisA, Real-IAD) yields I-AUC up to 0.97 (fully supervised CBM branch), state-of-the-art pixel-level anomaly localization (P-AUC 0.97), and substantial gains when combining limited real and synthetic anomalies.

7. Limitations, Insights, and Prospects

For ConvAD (AD paradigm), explanation scale and potential leakage in composite receptive fields may require application-specific threshold calibration (Chanchal et al., 1 Oct 2025). For CONVAD, automated concept extraction can misrepresent subtle defects, and distributional shifts from synthetic anomalies may limit generalization for certain categories, suggesting supplementation by real examples or more refined generative pipelines (Stropeni et al., 25 Nov 2025). Both frameworks advance the field toward robust, causally valid, and semantically tractable explanations in computer vision.

ConvAD and CONVAD offer theoretically sound, empirically validated, and practically deployable solutions to enduring challenges in explainable AI and anomaly detection, supporting deeper model transparency with minimal architectural disruption.