Research map / Label-efficient and robust learning / Recognition and long-tail
Explaining Deep Convolutional Neural Networks via Latent Visual-Semantic Filter Attention
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 |
|---|---|---|---|
| Use LaViSE instead of NetDissect | CNN filter explanations · Places365, Broden, ResNet-18 layer4 | human-evaluated top-1 precision: 0.70 -> 0.74 | Helped |
| Use LaViSE instead of NetDissect | CNN filter explanations · MS COCO, ResNet-18 layer4 | human-evaluated top-1 precision: 0.42 -> 0.70 | Helped |
| Use LaViSE instead of NetDissect | CNN filter explanations · MS COCO, ResNet-50 layer4 | human-evaluated top-1 precision: 0.38 -> 0.68 | Helped |
| Use LaViSE instead of NetDissect | CNN filter explanations · ImageNet, VG, ResNet-50 layer4 | human-evaluated top-1 precision: 0.24 -> 0.46 | Helped |
| Use LaViSE instead of NetDissect | CNN filter explanations · ImageNet, VG, ResNet-50 layer3 | human-evaluated top-1 precision: 0.20 -> 0.26 | Helped |
| Add filter attention instead of using original images without masking (ablation) | LaViSE · MS COCO, ResNet-18 layer4 | Recall (Top-20): 0.675 -> 0.776 | Helped |
| Add filter attention instead of image masking (ablation) | LaViSE · ImageNet, VG, ResNet-50 layer4 | Recall (Top-20): 0.213 -> 0.373 | Helped |
| Add filter attention rather than activation masking (ablation) | LaViSE · MS COCO, ResNet-18 layer4 | Recall (Top-20): 0.721 -> 0.776 | Helped |
| Add filter attention rather than activation masking (ablation) | LaViSE · ImageNet, VG, ResNet-18 layer4 | Recall (Top-20): 0.382 -> 0.429 | Helped |
| Add filter attention rather than activation masking (ablation) | LaViSE · ImageNet, VG, ResNet-50 layer4 | Recall (Top-20): 0.333 -> 0.373 | Helped |
| Add filter attention rather than activation masking (ablation) | LaViSE · ImageNet, VG, ResNet-50 layer3 | Recall (Top-20): 0.210 -> 0.207 | Hurt |
| Use a higher convolutional layer rather than layer1 (ablation) | LaViSE on ResNet-50 · ImageNet, VG | Recall (Top-20): 0.092 -> 0.373 | Helped |
Closest papers by meaning
- Restyling Unsupervised Concept Based Interpretable Networks with Generative Models (ICLR 2025 · similarity 0.87)
- PICNN: A Pathway towards Interpretable Convolutional Neural Networks (AAAI 2024 · similarity 0.87)
- Language Model as Visual Explainer (NeurIPS 2024 · similarity 0.87)
- LatentLens: Revealing Highly Interpretable Visual Tokens in LLMs (ICML 2026 · similarity 0.87)
- Semantic-Enriched Latent Visual Reasoning (ICML 2026 · similarity 0.86)
- Show and Tell: Visually Explainable Deep Neural Nets via Spatially-Aware Concept Bottleneck Models (CVPR 2025 · similarity 0.86)
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.