Research map / Architectures and efficiency / Attention and transformers
B-Cos Alignment for Inherently Interpretable CNNs and Vision Transformers
Moritz Böhle, Navdeeppal Singh, Mario Fritz, Bernt Schiele
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 |
|---|---|---|---|
| add MaxOut units to plain B-cos network (ablation) | B-cos network (B=2) · CIFAR-10 | Accuracy (%): 91.5 -> 93.2 | Improved |
| increase B from 1.25 to 2.5 (ablation) | MaxOut B-cos network · CIFAR-10 | Accuracy (%): 93.8 -> 92.4 | Hurt |
| increase B (alignment pressure) (ablation) | B-cos network · CIFAR-10 | Grid pointing localisation: higher B improves localisation | Improved |
| train B-cos ResNet-152 for 600 epochs with augmentation instead of 90 epochs | B-cos ResNet-152 · ImageNet val | Top-1 accuracy (%): 76.6 -> 80.2 | Improved |
| convert ResNet-152 to B-cos (90 epochs) | ResNet-152 · ImageNet val (90 epochs) | Top-1 accuracy (%): -1.7 | Hurt |
| convert ResNext-50-32x4d to B-cos | ResNext-50-32x4d · ImageNet val (90 epochs) | Top-1 accuracy (%): -2.0 | Hurt |
| convert ResNet-50 to B-cos | ResNet-50 · ImageNet val (90 epochs) | Top-1 accuracy (%): -0.2 | No clear effect |
| convert ConvNext-Tiny to B-cos (600-epoch protocol) | ConvNext-Tiny · ImageNet val (600 epochs, ConvNext-style recipe) | Top-1 accuracy (%): -5.0 | Hurt |
| remove bias terms from baseline CNNs (as in B-cos) (ablation) | ResNet-101 · ImageNet val (90 epochs) | Top-1 accuracy (%): -1.1 -> +6.2 | Improved |
| convert small ViT-Ti to B-cos | ViT-Ti · ImageNet val (90 epochs, Simple-ViT protocol) | Top-1 accuracy (%): -10.3 | Hurt |
| convert large ViT_C-B to B-cos | ViT_C-B · ImageNet val (90 epochs, Simple-ViT protocol) | Top-1 accuracy (%): +0.3 | No clear effect |
| add convolutional stem of four layers to B-cos ViT-Ti (ablation) | B-cos ViT-Ti · ImageNet val (90 epochs, Simple-ViT protocol) | Top-1 accuracy (%): 60.0 -> 67.3 | Improved |
Closest papers by meaning
- B-Cos Networks: Alignment Is All We Need for Interpretability (CVPR 2022 · similarity 0.97)
- B-cosification: Transforming Deep Neural Networks to be Inherently Interpretable (NeurIPS 2024 · similarity 0.94)
- PICNN: A Pathway towards Interpretable Convolutional Neural Networks (AAAI 2024 · similarity 0.85)
- No Head Left Behind – Multi-Head Alignment Distillation for Transformers (AAAI 2024 · similarity 0.84)
- ScaleKD: Strong Vision Transformers Could Be Excellent Teachers (NeurIPS 2024 · similarity 0.83)
- Token Transformation Matters: Towards Faithful Post-hoc Explanation for Vision Transformer (CVPR 2024 · similarity 0.83)
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.
