Research map / Architectures and efficiency / RNNs and normalization
B-Cos Networks: Alignment Is All We Need for Interpretability
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 |
|---|---|---|---|
| Replace standard linear transforms with B-cos transforms | VGG-11 · ImageNet | Top-1 accuracy: 69.0 -> 69.6 | Helped |
| Replace standard linear transforms with B-cos transforms | ResNet-34 · ImageNet | Top-1 accuracy: 73.3 -> 71.7 | Hurt |
| Replace standard linear transforms with B-cos transforms | DenseNet-121 · ImageNet | Top-1 accuracy: 74.4 -> 73.3 | Hurt |
| Replace standard linear transforms with B-cos transforms | InceptionNet · ImageNet | Top-1 accuracy: 77.3 -> 75.4 | Hurt |
| Add MaxOut units (ablation) | B-cos network · CIFAR-10 test | accuracy (%): 91.5 -> 93.2 | Helped |
| Increase B from 1.25 to 2.50 (ablation) | B-cos MaxOut network · CIFAR-10 test | accuracy (%): 93.8 -> 92.4 | Hurt |
| Increase B (ablation) | B-cos network · CIFAR-10 test | localisation | Helped |
| Train B-cos DenseNet-121 longer using a cosine learning-rate schedule and warm-up (ablation) | B-cos DenseNet-121 · ImageNet | Top-1 accuracy (%): 73.3 -> 74.4 | Helped |
| Remove MaxOut units (ablation) | B-cos DenseNet-121 · ImageNet | Top-1 accuracy (%): 74.4 -> 72.6 | Hurt |
| Increase growth factor from 20 to 32 with 2 MaxOut units (ablation) | B-cos DenseNet-121 · ImageNet | Top-1 accuracy (%): 72.8 -> 74.3 | Helped |
| Use growth factor 22 with 2 MaxOut units instead of the conventional DenseNet-121 (ablation) | DenseNet-121 · ImageNet | Top-1 accuracy (%): 74.4 -> 73.2 | Hurt |
| Use model-inherent B-cos explanations instead of post-hoc attribution methods | B-cos architectures · ImageNet grid pointing game | localisation | Helped |
Closest papers by meaning
- B-cosification: Transforming Deep Neural Networks to be Inherently Interpretable (NeurIPS 2024 · similarity 0.93)
- PICNN: A Pathway towards Interpretable Convolutional Neural Networks (AAAI 2024 · similarity 0.84)
- Neurons to Words: A Novel Method for Automated Neural Network Interpretability and Alignment (AAAI 2025 · similarity 0.84)
- Network Inversion of Convolutional Neural Nets (Student Abstract) (AAAI 2025 · similarity 0.83)
- Measuring Per-Unit Interpretability at Scale Without Humans (NeurIPS 2024 · similarity 0.83)
- Tracing Representation Progression: Analyzing and Enhancing Layer-Wise Similarity (ICLR 2025 · 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.