Research map / Architectures and efficiency / CNN design
B-cosification: Transforming Deep Neural Networks to be Inherently Interpretable
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes B-cosification, converting pretrained CNNs/ViTs and CLIP into inherently interpretable B-cos models via fine-tuning with B=2 and bias removal, achieving B-cos-level interpretability at a fraction of training cost.
Problem
Inherently interpretable B-cos networks must be trained from scratch, which is costly for large and foundation models; the paper asks whether pretrained DNNs can be fine-tuned to similar interpretability.
Approach
Show that standard DNNs can be rewritten as functionally equivalent B-cos models (6-channel color input via mean-normalization trick, ReLU kept as MaxOut special case, no weight normalization), then fine-tune after raising B to 2 and removing biases, which yields faithful, colored, localized explanations.
Key results
- B-cosified ResNet-18 reaches 71.5% ImageNet top-1 vs 68.7% for B-cos from scratch (+2.8pp) with 3.1x speedup.
- Averaged across architectures, B-cosified models outperform B-cos from scratch by 2.5pp with an average 5.2x training speedup.
- B-cosified models achieve localisation 32.7-71.0pp above conventional DNNs and on par with B-cos models; B-cosified CLIP RN-50 zero-shot performance is competitive, outperforming Text2Concept on most datasets.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| B-cosify pretrained ResNet-18 instead of using standard ResNet-18 | ResNet-18 · ImageNet validation | top-1 accuracy (%): 69.8 -> 71.5 | Helped |
| B-cosify pretrained ResNet-18 instead of training B-cos ResNet-18 from scratch | ResNet-18 · ImageNet validation | top-1 accuracy (%): 68.7 -> 71.5 | Helped |
| B-cosify pretrained DenseNet-121 instead of training B-cos DenseNet-121 from scratch | DenseNet-121 · ImageNet validation | top-1 accuracy (%): 73.6 -> 76.3 | Helped |
| B-cosify pretrained ViT-S instead of training B-cos ViT-S from scratch | ViT-S · ImageNet validation | top-1 accuracy (%): 69.2 -> 75.2 | Helped |
| set B to 2 rather than 1 during B-cosification (ablation) | ResNet-18 · ImageNet validation | GridPG localisation: 33.9{ $\pm$0.2} -> 87.6{ $\pm$0.2} | Helped |
| set B to 2 rather than 1 during B-cosification (ablation) | ResNet-18 · ImageNet validation | accuracy: 70.6{ $\pm$0.1} -> 71.5{ $\pm$0.1} | Helped |
| remove biases rather than retain them (ablation) | B-cosified ResNet-18 · ImageNet validation | Localisation: 47.2{ $\pm$0.5} -> 87.6{ $\pm$0.2} | Helped |
| remove biases rather than retain them (ablation) | B-cosified ResNet-18 · ImageNet validation | Accuracy: 71.2{ $\pm$0.1} -> 71.5{ $\pm$0.1} | No clear effect |
| replace ReLU with Identity (ablation) | B-cosified ResNet-18 · ImageNet validation | top-1 accuracy (%): 71.5 -> 63.4 | Hurt |
| replace BatchNorm Uncentered with BatchNorm Centered (ablation) | B-cosified ResNet-18 · ImageNet validation | Localisation: 86.8 -> 28.8 | Hurt |
| change the Global Average Pool order (ablation) | B-cosified ResNet-18 · ImageNet validation | Localisation: 86.8 -> 63.8 | Hurt |
| remove the Logit Bias layer (ablation) | B-cosified ResNet-18 · ImageNet validation | top-1 accuracy (%): 71.5 -> 70.8 | Hurt |
Surprising findings
- Setting B=2 and removing biases immediately before fine-tuning matched learnable-B and bias-decay strategies, suggesting the simple recipe suffices.
- Larger B-cosified ViT_c models (B, L) did not yield speedups over B-cos training despite comparable accuracy.
Limitations
- Larger ViT_c models did not achieve training speedups over from-scratch B-cos.
- Standard CLIP zero-shot performance is generally higher than B-cosified CLIP on many datasets.
- Weight normalization invariance argument holds only if every B-cos layer is followed by normalization, which is not guaranteed; unnormalized weights used.
Experimental setup
Datasets: ImageNet CC3M VOC2007 CLIP Benchmark (38 datasets)
Benchmarks: ImageNet GridPG CLIP Benchmark VOC2007 EPG
Metrics: top-1 accuracy GridPG localisation zero-shot accuracy linear probe accuracy EPG score speedup
Baselines: B-cos models trained from scratch Standard torchvision models Guided Backprop GradCAM Integrated Gradients Text2Concept Standard CLIP
Builds on: B-cos networks SigLIP loss Text2Concept GridPG
Closest papers by meaning
- PICNN: A Pathway towards Interpretable Convolutional Neural Networks (AAAI 2024 · similarity 0.87)
- Training the Untrainable: Introducing Inductive Bias via Representational Alignment (NeurIPS 2025 · similarity 0.85)
- Unveiling the Unseen: Identifiable Clusters in Trained Depthwise Convolutional Kernels (ICLR 2024 · similarity 0.85)
- Measuring Per-Unit Interpretability at Scale Without Humans (NeurIPS 2024 · similarity 0.85)
- Dataset Distillation for Pre-Trained Self-Supervised Vision Models (NeurIPS 2025 · similarity 0.85)
- Scaling White-Box Transformers for Vision (NeurIPS 2024 · 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.