Research map / Architectures and efficiency / RNNs and normalization
Empowering CAM-Based Methods with Capability to Generate Fine-Grained and High-Faithfulness Explanations
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 FG-CAM, which extends CAM-based explanation methods to shallow layers and the input layer by propagating last-layer explanation components backward through adjacent layers with z+/z-beta rules, plus a denoising variant using SVD.
Problem
CAM-based explanations are coarse-grained at deep layers because they use global per-channel weights, while LRP-style explanations are fine-grained but have low faithfulness.
Approach
Compute CAM explanation components at the last convolutional layer, then layer by layer redistribute their importance to higher-resolution previous layers using z+/z-beta relevance rules, filtering non-contributing pixels, so any CAM variant yields fine-grained, high-faithfulness maps; denoising applies SVD truncation to the components before redistribution.
Key results
- In Layer13, FG-Grad-CAM reaches 19.65 A.D. and 32.35 A.I. versus Grad-CAM's 83.80 A.D. and 3.25 A.I. on 2000 ILSVRC2012 val images.
- FG-Grad-CAM at the input layer achieves Over-all 0.52089 in deletion/insertion, versus 0.45369 for LRP, 0.37572 for CLRP and 0.29340 for SGLRP.
- Localization Proportion on 500 single-box images: FG-Score-CAM with denoising reaches 0.66983 versus 0.48385 for LRP.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply FG-CAM using Grad-CAM weights instead of Grad-CAM alone | VGG-16 with batch normalization · ILSVRC2012 val, Layer13 | A.D.: 83.80 -> 19.65 | Helped |
| Apply FG-CAM using Grad-CAM weights instead of Grad-CAM alone | VGG-16 with batch normalization · ILSVRC2012 val, Layer13 | A.I.: 3.25 -> 32.35 | Helped |
| Apply FG-CAM using Score-CAM weights instead of Score-CAM alone | VGG-16 with batch normalization · ILSVRC2012 val, Layer13 | A.D.: 57.06 -> 23.50 | Helped |
| Apply FG-CAM using Score-CAM weights instead of Score-CAM alone | VGG-16 with batch normalization · ILSVRC2012 val, Layer13 | A.I.: 12.00 -> 25.85 | Helped |
| Add denoising to FG-Grad-CAM (ablation) | FG-Grad-CAM · ILSVRC2012 val, Layer13 | A.D.: 19.65 -> 17.90 | Helped |
| Add denoising to FG-Grad-CAM (ablation) | FG-Grad-CAM · ILSVRC2012 val, input layer | Over-all: 0.52089 -> 0.50115 | No clear effect |
| Add denoising to FG-Score-CAM (ablation) | FG-Score-CAM · ILSVRC2012 val, Layer13 | A.D.: 23.50 -> 19.24 | Helped |
| Add denoising to FG-Score-CAM (ablation) | FG-Score-CAM · ILSVRC2012 val, input layer | Over-all: 0.49628 -> 0.49423 | No clear effect |
| Apply FG-CAM using Grad-CAM weights instead of LRP | VGG-16 with batch normalization · ILSVRC2012 val, input layer | Over-all: 0.45369 -> 0.52089 | Helped |
Surprising findings
- Grad-CAM and Score-CAM become unusable at shallow layers, while Layer-CAM's explanations become noisier as the target layer gets shallower.
- CLRP and SGLRP over-remove target-class pixels, yielding lower faithfulness than plain LRP despite higher localization Proportion.
Limitations
- Evaluation is limited to a single model (VGG-16 with batch normalization) and ILSVRC2012 validation images.
- Denoising uses a fixed top-10% singular value truncation, with no sensitivity analysis reported.
- Localization metric measures only the fraction of attention inside bounding boxes and cannot judge edge fit.
Experimental setup
Datasets: ILSVRC2012 val
Benchmarks: ILSVRC2012 val
Metrics: Average Drop Average Increase deletion/insertion Over-all (AUC difference) Proportion (bounding-box localization)
Baselines: Grad-CAM Score-CAM Layer-CAM LRP CLRP SGLRP
Builds on: Grad-CAM Score-CAM Layer-CAM LRP CLRP SGLRP
Closest papers by meaning
- Finer-CAM: Spotting the Difference Reveals Finer Details for Visual Explanation (CVPR 2025 · similarity 0.88)
- DD-CAM: Minimal Sufficient Explanations for Vision Models Using Delta Debugging (NeurIPS 2026 · similarity 0.88)
- CAPE: CAM as a Probabilistic Ensemble for Enhanced DNN Interpretation (CVPR 2024 · similarity 0.85)
- DiffCAM: Data-Driven Saliency Maps by Capturing Feature Differences (CVPR 2025 · similarity 0.84)
- Attention Guided CAM: Visual Explanations of Vision Transformer Guided by Self-Attention (AAAI 2024 · similarity 0.84)
- Q-SENN: Quantized Self-Explaining Neural Networks (AAAI 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.