Research map / Architectures and efficiency / CNN design
Optimising for Interpretability: Convolutional Dynamic Alignment Networks
Moritz Böhle, 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 |
|---|---|---|---|
| increase temperature T from 10 to 1000 (ablation) | S-CoDA-SQ · CIFAR-10 | accuracy: 93.6% -> 93.2% | Hurt |
| increase temperature (alignment pressure) (ablation) | S-CoDA-SQ · CIFAR-10 | interpretability (localisation) | Improved |
| use eDAU weight-bounded (WB) non-linearity instead of SQ/L2 with shared B (ablation) | S-CoDA · CIFAR-10 | accuracy: 93.2% -> 94.0% | Improved |
| add learnt patch embedding E(x) as input (ablation) | S-eCoDA-WB · CIFAR-10 | accuracy: 94.0% -> 94.1% | Improved |
| use WB rescaling instead of SQ | L-CoDA · Imagenet | GPU memory: 30.0GB -> 9.7GB | Improved |
| use WB rescaling instead of SQ | L-CoDA · Imagenet | training time: 14.1 minutes -> 8.7 minutes | Improved |
| add data augmentation | XL-CoDA-SQ · TinyImagenet | accuracy: 54.4% -> 58.4% | Improved |
| replace last 4 ResNet-56 segments by CoDA layers (R1-C4) (ablation) | ResNet-56 · CIFAR-10 | accuracy: 93.4% -> 93.8% | No clear effect |
| replace last 5 ResNet-50 blocks by CoDA layers (R0-C5) (ablation) | ResNet-50 · Imagenet | accuracy: 76.1% -> 71.4% | Hurt |
| replace more ResNet segments by CoDA layers (ablation) | ResNet hybrids · CIFAR-10 and Imagenet | interpretable depth (localisation) | Improved |
| use CoDA layers rather than piece-wise linear ResNet blocks in the head (ablation) | hybrid networks · CIFAR-10 | localisation | Improved |
Closest papers by meaning
- Convolutional Dynamic Alignment Networks for Interpretable Classifications (CVPR 2021 · similarity 0.98)
- CLKN: Cascaded Lucas-Kanade Networks for Image Alignment (CVPR 2017 · similarity 0.86)
- CoDA: From Text-to-Image Diffusion Models to Training-Free Dataset Distillation (ICLR 2026 · similarity 0.85)
- Residual Alignment: Uncovering the Mechanisms of Residual Networks (NeurIPS 2023 · similarity 0.85)
- B-Cos Networks: Alignment Is All We Need for Interpretability (CVPR 2022 · similarity 0.84)
- iFAN: Image-Instance Full Alignment Networks for Adaptive Object Detection (AAAI 2020 · similarity 0.84)
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.
