Research map / Label-efficient and robust learning / Recognition and long-tail
Consistent Explanations by Contrastive Learning
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add contrastive Grad-CAM consistency loss | ResNet50 · ImageNet val | CH (%): 54.77 -> 71.75 | Helped |
| Add contrastive Grad-CAM consistency loss | ResNet50 · ImageNet val | Top-1 accuracy (%): 76.13 -> 74.60 | Hurt |
| Add contrastive Grad-CAM consistency loss | ResNet50 · ImageNet val | Insertion AUC: 0.4860 -> 0.5216 | Helped |
| Add contrastive Grad-CAM consistency loss | ResNet50 · UnRel (28 ImageNet-overlap categories) | CH (%): 51.66 -> 74.20 | Helped |
| Add contrastive Grad-CAM consistency loss | ResNet50 · CUB-200, FGVC-Aircraft, Cars-196, VGG Flowers-102 | accuracy: 80.09 $\pm$ 0.89 -> 81.49 $\pm$ 0.09; 83.65 $\pm$ 0.15 -> 85.72 $\pm$ 0.20; 89.71 $\pm$ 0.14 -> 90.28 $\pm$ 0.08; 96.09 | Helped |
| Add contrastive Grad-CAM consistency loss | ResNet50 · CUB-200 5-shot and Cars-196 5-shot | accuracy: 51.7 $\pm$ 0.3 -> 55.2 $\pm$ 0.3; 34.3 $\pm$ 0.4 -> 36.9 $\pm$ 0.4 | Helped |
| Remove negative heatmaps from the contrastive loss (ablation) | CGC, ResNet18 · ImageNet val | CH (%): 65.83 -> 44.08 | Hurt |
| Remove negative heatmaps from the contrastive loss (ablation) | CGC, ResNet50 · ImageNet val | CH (%): 71.75 -> 39.84 | Hurt |
| Remove negative heatmaps from the contrastive loss (ablation) | CGC, ResNet50 · ImageNet val | Top-1 accuracy (%): 74.60 -> 74.80 | No clear effect |
| Set lambda to 0, removing the CGC loss (ablation) | ResNet18 · ImageNet | CH: 65.83 -> 54.47 | Hurt |
| Increase lambda from 1.0 to 2.0 (ablation) | ResNet18 · ImageNet | CH: 65.83 -> 52.60 | Hurt |
| Use additional unlabeled images for the CGC loss | ResNet50 initialized from SwAV · ImageNet val (1% labeled training) | Top-1 accuracy (%): 54.00 -> 55.18 | Helped |
Closest papers by meaning
- How to Probe: Simple Yet Effective Techniques for Improving Post-hoc Explanations (ICLR 2025 · similarity 0.87)
- Keep the Faith: Faithful Explanations in Convolutional Neural Networks for Case-Based Reasoning (AAAI 2024 · similarity 0.86)
- SLICE: Stabilized LIME for Consistent Explanations for Image Classification (CVPR 2024 · similarity 0.85)
- Post-hoc Part-Prototype Networks (ICML 2024 · similarity 0.85)
- Empowering CAM-Based Methods with Capability to Generate Fine-Grained and High-Faithfulness Explanations (AAAI 2024 · similarity 0.85)
- Representational Difference Explanations (NeurIPS 2025 · similarity 0.85)
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.