Research map / Architectures and efficiency / CNN design

DD-CAM: Minimal Sufficient Explanations for Vision Models Using Delta Debugging

NeurIPS 2026 · Architectures and efficiency › CNN design · approach: Biologically inspired CNN explanation · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes DD-CAM, a gradient-free CAM method that uses delta debugging to find a 1-minimal sufficient subset of feature maps or ViT patch tokens and builds saliency maps from that subset.

Problem

Existing CAM methods aggregate contributions from all internal units, producing diffuse saliency maps that do not show which units are necessary for a prediction.

Approach

Adapt delta debugging from software debugging to find a 1-minimal set of representational units whose joint activation preserves the top-1 prediction, with a one-pass search for models whose linear heads make units non-interacting and a standard recursive search for models with interacting units.

Key results

• On CNNs with linear heads, DD-CAM reports the highest ADCC (0.8087) and lowest Average Drop (0.1393) among the compared CAM methods.
• On ChestX-ray14, DD-CAM reaches IoU 0.263, precision 0.307, and recall 0.692, about 45% and 22% above the strongest baseline in IoU and precision.
• On Vision Transformers, DD-CAM gets the lowest Average Drop (0.0832) but runs about 1.9 s per image versus 0.04 s for Grad-CAM.

Surprising findings

• Score-CAM obtains a slightly higher ADCC than DD-CAM on ViTs (0.8103 vs 0.7985), though with larger regions and more computation.

Limitations

• Saliency maps are upsampled from the final feature map, which can cause spatial imprecision.
• The method is white-box and requires access to internal activations and weights.
• 1-minimality does not guarantee the minimum-cardinality set.

Experimental setup

Datasets: ImageNet ILSVRC2012 validation subset NIH ChestX-ray14

Benchmarks: ImageNet ILSVRC2012 validation (2,000 images) NIH ChestX-ray14 (1,000 annotated radiographs)

Metrics: Average Drop Coherency Complexity ADCC Increase in Confidence ADD IoU Precision Recall Percentage Highlighted

Baselines: Grad-CAM Grad-CAM++ XGrad-CAM Layer-CAM Score-CAM Ablation-CAM Recipro-CAM

Builds on: Grad-CAM Ablation-CAM ViT-ReciproCAM delta debugging

Closest papers by meaning

• Finer-CAM: Spotting the Difference Reveals Finer Details for Visual Explanation (CVPR 2025 · similarity 0.91)
• CAPE: CAM as a Probabilistic Ensemble for Enhanced DNN Interpretation (CVPR 2024 · similarity 0.89)
• DiffCAM: Data-Driven Saliency Maps by Capturing Feature Differences (CVPR 2025 · similarity 0.88)
• Empowering CAM-Based Methods with Capability to Generate Fine-Grained and High-Faithfulness Explanations (AAAI 2024 · similarity 0.88)
• Emphasizing Discriminative Features for Dataset Distillation in Complex Scenarios (CVPR 2025 · similarity 0.87)
• Show and Tell: Visually Explainable Deep Neural Nets via Spatially-Aware Concept Bottleneck Models (CVPR 2025 · similarity 0.86)