Research map / Architectures and efficiency / Attention and transformers
R-Drop: Regularized Dropout for Neural Networks
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 R-Drop KL consistency loss | Transformer big · WMT14 En-De | BLEU: 29.12 -> 30.91 | Helped |
| add R-Drop | Transformer · IWSLT 8 tasks | BLEU: 32.44 -> 34.68 | Helped |
| add R-Drop in fine-tuning | BERT-base · GLUE | avg score: 82.85 -> 84.06 | Helped |
| add R-Drop in fine-tuning | RoBERTa-large · GLUE | avg score: 88.93 -> 89.73 | Helped |
| add R-Drop in fine-tuning | BART · CNN/DailyMail | ROUGE-1: 44.16 -> 44.51 | Helped |
| add R-Drop | Transformer LM · WikiText-103 | test perplexity: 26.62 -> 24.94 | Helped |
| add R-Drop in fine-tuning | ViT-B/16 · ImageNet | accuracy: 83.97 -> 84.38 | Helped |
| add KL term (vs doubled forward without KL) (ablation) | Transformer · IWSLT14 De-En | BLEU: 34.93 -> 37.25 | Helped |
| regularize 3 sub-models instead of 2 (ablation) | R-Drop · IWSLT14 De-En | BLEU: 37.25 -> 37.30 | No clear effect |
| apply R-Drop every k steps (k>1) instead of every step (ablation) | R-Drop · IWSLT14 De-En | BLEU | Hurt |
| increase KL weight alpha from 1 to 5 (ablation) | R-Drop · IWSLT14 De-En | BLEU: 36.05 -> 37.25 | Helped |
| increase KL weight alpha from 5 to 10 (ablation) | R-Drop · IWSLT14 De-En | BLEU: 37.25 -> 36.95 | Hurt |
Closest papers by meaning
- Dropout with Expectation-linear Regularization (ICLR 2017 · similarity 0.92)
- Dropout Reduces Underfitting (ICML 2023 · similarity 0.90)
- The Implicit and Explicit Regularization Effects of Dropout (ICML 2020 · similarity 0.89)
- Adversarial Dropout for Recurrent Neural Networks (AAAI 2019 · similarity 0.89)
- Fraternal Dropout (ICLR 2018 · similarity 0.89)
- Stochastic Modified Equations and Dynamics of Dropout Algorithm (ICLR 2024 · similarity 0.88)
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.