Research map / Adversarial robustness and security / Attacks on text and graphs
Informative Dropout for Robust Representation Learning: A Shape-bias Perspective
Read the paperCheck 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 |
|---|---|---|---|
| Add InfoDrop to the baseline | PACS multi-source domain generalization · PACS target domains | accuracy: Baseline: 95.98, 77.87, 74.86, 70.17; + InfoDrop: 96.11, 80.27, 76.54, 76.38 | Helped |
| Add InfoDrop to ProtoNet | ProtoNet · CUB 5-way few-shot classification, no augmentation | accuracy: 67.13 -> 70.94 (5-shot); 51.62 -> 52.40 (1-shot) | Helped |
| Add InfoDrop to MatchingNet | MatchingNet · CUB 5-way few-shot classification, no augmentation | accuracy: 71.18 -> 72.32 (5-shot); 57.81 -> 57.88 (1-shot) | Helped |
| Add InfoDrop to RelationNet | RelationNet · CUB 5-way few-shot classification, no augmentation | accuracy: 69.85 -> 73.72 (5-shot); 56.71 -> 59.21 (1-shot) | Helped |
| Add InfoDrop to a vanilla CNN | Caltech-256 corruption classification · Random image corruptions | accuracy: Gaussian noise 66.38 -> 69.58; defocus blur 65.97 -> 62.52 | Mixed |
| Add InfoDrop to adversarial training | CIFAR-10 · ℓ∞ = 8/255 | accuracy: 42.05 -> 43.19 | Helped |
| Set InfoDrop temperature to 1.0 (ablation) | ProtoNet · mini-ImageNet→CUB 5-way 1-shot | accuracy: 35.24 -> 37.89 | Helped |
| Set temperature to 0.1 rather than 1.0 (ablation) | InfoDrop with ProtoNet · mini-ImageNet→CUB 5-way 1-shot | accuracy: 37.89 -> 36.33 | Hurt |
| Use infinite temperature, making the method regular Dropout (ablation) | InfoDrop with ProtoNet · mini-ImageNet→CUB 5-way 1-shot | accuracy: 37.89 -> 35.54 | Hurt |
| Apply InfoDrop to the first residual block (ablation) | ResNet18 · PACS sketch-to-art domain generalization | accuracy: 27.34 -> 31.05 | Helped |
| Apply InfoDrop to two or three residual blocks instead of one (ablation) | ResNet18 · PACS sketch-to-art domain generalization | accuracy: 31.05 -> 28.85 (2 blocks); 31.05 -> 29.28 (3 blocks) | Hurt |
| Increase shape bias by tuning temperature (ablation) | InfoDrop models · PACS photo-source domain generalization | accuracy | Mixed |
Closest papers by meaning
- Does enhanced shape bias improve neural network robustness to common corruptions? (ICLR 2021 · similarity 0.90)
- The Origins and Prevalence of Texture Bias in Convolutional Neural Networks (NeurIPS 2020 · similarity 0.89)
- ImageNet-trained CNNs are not biased towards texture: Revisiting feature reliance through controlled suppression (NeurIPS 2025 · similarity 0.89)
- Shape or Texture: Understanding Discriminative Features in CNNs (ICLR 2021 · similarity 0.88)
- DRoP: Distributionally Robust Data Pruning (ICLR 2025 · similarity 0.87)
- Visual Anagrams Reveal Hidden Differences in Holistic Shape Processing Across Vision Models (NeurIPS 2025 · 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.