Research map / Architectures and efficiency / RNNs and normalization
Structured in Space, Randomized in Time: Leveraging Dropout in RNNs for Efficient Training
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 |
|---|---|---|---|
| Structure non-recurrent dropout within batches while varying masks across time steps (ablation) | Zaremba Medium LSTM · PTB test | perplexity: 82.70 -> 83.01 | Hurt |
| Add structured recurrent-hidden dropout to structured non-recurrent dropout (ablation) | Zaremba Medium LSTM · PTB test | perplexity: 83.01 -> 80.84 | Helped |
| Add structured recurrent-hidden dropout to structured non-recurrent dropout (ablation) | Zaremba Large LSTM · PTB test | perplexity: 82.39 -> 76.05 | Helped |
| Use structured non-recurrent and recurrent-hidden dropout | Zaremba Large LSTM · PTB test | perplexity: 78.40 -> 76.05 | Helped |
| Use structured non-recurrent and recurrent-hidden dropout | Zaremba Medium LSTM · PTB test | overall speedup: 1 -> 1.45 | Helped |
| Use structured non-recurrent and recurrent-hidden dropout | Zaremba Large LSTM · PTB test | overall speedup: 1 -> 1.64 | Helped |
| Structure only the non-recurrent dropout (ablation) | OpenNMT LSTM · IWSLT'14 De-En | BLEU: 28.13 -> 27.54 | Hurt |
| Add structured recurrent-hidden dropout to structured non-recurrent dropout (ablation) | OpenNMT LSTM · IWSLT'14 De-En | BLEU: 27.54 -> 28.46 | Helped |
| Add structured recurrent-hidden dropout to structured non-recurrent dropout (ablation) | OpenNMT LSTM · IWSLT'15 En-Vi | BLEU: 26.08 -> 26.20 | Helped |
| Move input dropout to the concatenated CNN and embedding output and structure it within batches (ablation) | BiLSTM-CNN-CRF · CoNLL-2003 | F1-score: 89.26 -> 89.32 | Helped |
| Add structured recurrent dropout in both BiLSTM directions (ablation) | BiLSTM-CNN-CRF · CoNLL-2003 | F1-score: 89.32 -> 89.49 | Helped |
Closest papers by meaning
- Adversarial Dropout for Recurrent Neural Networks (AAAI 2019 · similarity 0.92)
- Fraternal Dropout (ICLR 2018 · similarity 0.92)
- Recurrent neural networks: vanishing and exploding gradients are not the end of the story (NeurIPS 2024 · similarity 0.88)
- Accelerating Linear Recurrent Neural Networks for the Edge with Unstructured Sparsity (ICML 2025 · similarity 0.87)
- Dynamic Sparse No Training: Training-Free Fine-tuning for Sparse LLMs (ICLR 2024 · similarity 0.87)
- Exploiting Symmetric Temporally Sparse BPTT for Efficient RNN Training (AAAI 2024 · similarity 0.87)
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.