Research map / Label-efficient and robust learning / Recognition and long-tail
Do deep networks transfer invariances across classes?
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 GIT augmentation to LDAM+DRS | LDAM+DRS · K49-BG-LT | balanced test accuracy: 54.08 ± 1.21 -> 58.86 ± 1.11 | Helped |
| Add GIT augmentation to LDAM+DRS | LDAM+DRS · K49-DIL-LT | balanced test accuracy: 50.44 ± 1.24 -> 56.76 ± 1.11 | Helped |
| Add GIT augmentation to LDAM+DRS | LDAM+DRS · GTSRB-LT | balanced test accuracy: 77.25 ± 1.29 -> 81.39 ± 0.98 | Helped |
| Add GIT augmentation to LDAM+DRS | LDAM+DRS · CIFAR-10-LT | balanced test accuracy: 76.73 ± 0.74 -> 78.76 ± 0.19 | Helped |
| Add GIT augmentation to LDAM+DRS | LDAM+DRS · CIFAR-100-LT | balanced test accuracy: 43.21 ± 0.31 -> 44.35 ± 0.21 | Helped |
| Add GIT augmentation to LDAM+DRS | LDAM+DRS · TinyImageNet-LT | balanced test accuracy: 20.54 ± 0.21 -> 21.99 ± 0.23 | Helped |
| Add GIT augmentation to LDAM+DRS | LDAM+DRS · iNaturalist-2018 validation | Top-1 accuracy: 57.545 -> 53.051 | Hurt |
| Add GIT augmentation to LDAM+DRS | LDAM+DRS · iNaturalist-2018 validation | Top-5 accuracy: 77.487 -> 73.946 | Hurt |
| Disable the class-size cutoff and augment every class (ablation) | CE+DRS+GIT · GTSRB-LT | balanced test accuracy: 75.19 ± 0.50 -> 66.22 ± 1.12 | Hurt |
| Disable the class-size cutoff and augment every class (ablation) | CE+DRS+GIT · CIFAR-10-LT | balanced test accuracy: 77.25 ± 0.18 -> 76.29 ± 0.20 | Hurt |
| Disable the class-size cutoff and augment every class (ablation) | CE+DRS+GIT · CIFAR-100-LT | balanced test accuracy: 42.73 ± 0.27 -> 39.77 ± 0.16 | Hurt |
| Replace GIT augmentation with RandAugment (ablation) | CE+DRS on CIFAR-LT · CIFAR-10-LT and CIFAR-100-LT | balanced test accuracy: CIFAR-10-LT: 77.25 ± 0.18 vs 78.78 ± 0.99; CIFAR-100-LT: 42.73 ± 0.27 vs 45.53 ± 0.20 | Mixed |
Closest papers by meaning
- Learning Invariances in Neural Networks from Training Data (NeurIPS 2020 · similarity 0.89)
- Understanding Invariance via Feedforward Inversion of Discriminatively Trained Classifiers (ICML 2021 · similarity 0.89)
- In What Ways Are Deep Neural Networks Invariant and How Should We Measure This? (NeurIPS 2022 · similarity 0.87)
- Grounding inductive biases in natural images: invariance stems from variations in data (NeurIPS 2021 · similarity 0.87)
- Theoretical Understanding of Learning from Adversarial Perturbations (ICLR 2024 · similarity 0.85)
- Exploiting Inter-sample and Inter-feature Relations in Dataset Distillation (CVPR 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.