Research map / Label-efficient and robust learning / Recognition and long-tail
Interpreting Deep Visual Representations via Network Dissection
Bolei Zhou, David Bau, Aude Oliva, Antonio B. Torralba
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| apply random orthogonal rotation to conv5 representation (ablation) | AlexNet Places205 conv5 · unique detectors | number of unique detectors: 80% fewer | Hurt |
| rotate pool5 representation then retrain fc layers (ablation) | AlexNet Places365 · Places365 val | top-1 accuracy: 51.9 -> 51.7 | No clear effect |
| add batch normalization (ablation) | Places205-AlexNet · conv5 interpretability | unique detectors | Hurt |
| add batch normalization (ablation) | Places205-AlexNet · Places205 val | top-1 accuracy: 50.0 -> 50.5 | No clear effect |
| add dropout in FC layers (ablation) | Places205-AlexNet · conv5 interpretability | object detectors | Improved |
| add dropout in FC layers (ablation) | Places205-AlexNet · Places205 val | top-1 accuracy: 49.2 -> 50.0 | Improved |
| change random initialization (ablation) | Places205-AlexNet · conv5 interpretability | unique detectors | No clear effect |
| triple conv5 units (256 to 768) (ablation) | AlexNet-GAP · Places365 conv5 | unique detectors: 256 -> 768 | Improved |
| raise conv5 units to 1024 and 2048 (ablation) | AlexNet-GAP-Wide · Places365 conv5 | unique detectors: 1024 and 2048 | No clear effect |
| use deeper architecture (ResNet over AlexNet) | supervised CNNs · Places and ImageNet last conv layer | unique detectors | Improved |
| train on Places instead of ImageNet | supervised CNNs · last conv layer | unique detectors | Improved |
| use self-supervised instead of supervised training | AlexNet · conv5 | object detectors | Hurt |
Closest papers by meaning
- Network Dissection: Quantifying Interpretability of Deep Visual Representations (CVPR 2017 · similarity 0.97)
- PICNN: A Pathway towards Interpretable Convolutional Neural Networks (AAAI 2024 · similarity 0.89)
- Representational Similarity via Interpretable Visual Concepts (ICLR 2025 · similarity 0.89)
- Measuring Per-Unit Interpretability at Scale Without Humans (NeurIPS 2024 · similarity 0.88)
- Visual Concept Connectome (VCC): Open World Concept Discovery and their Interlayer Connections in Deep Models (CVPR 2024 · similarity 0.87)
- A Convolutional Neural Network Interpretable Framework for Human Ventral Visual Pathway Representation (AAAI 2024 · similarity 0.87)
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.
