Research map / Architectures and efficiency / RNNs and normalization
Towards Binary-Valued Gates for Robust LSTM 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 |
|---|---|---|---|
| Use Gumbel-Softmax gates for the input and forget gates instead of standard soft gates | AWD-LSTM · Penn Treebank test | perplexity: 57.3 -> 56.1 | Helped |
| Use Gumbel-Softmax gates for the input and forget gates instead of standard soft gates | Encoder-decoder LSTM · WMT'14 English→German | BLEU: 21.89 -> 22.43 | Helped |
| Use Gumbel-Softmax gates for the input and forget gates instead of standard soft gates | Encoder-decoder LSTM · IWSLT’14 German→English | BLEU: 31.00 -> 31.95 | Helped |
| Use sharpened sigmoid gates instead of standard soft gates | AWD-LSTM · Penn Treebank test | perplexity: 57.3 -> 57.5 | Hurt |
| Use sharpened sigmoid gates instead of standard soft gates | Encoder-decoder LSTM · IWSLT’14 German→English | BLEU: 31.00 -> 29.73 | Hurt |
| Add a continuous cache pointer | G²-LSTM · Penn Treebank test | perplexity: 56.1 -> 52.1 | Helped |
| Add a continuous cache pointer | AWD-LSTM · Penn Treebank test | perplexity: 57.3 -> 52.8 | Helped |
| Also push output gates toward binary values (ablation) | G²-LSTM · Machine translation | BLEU | Hurt |
| Round input- and forget-gate parameters (ablation) | G²-LSTM · Penn Treebank | perplexity: 52.1 -> 52.2 | Hurt |
| Round and clip input- and forget-gate parameters (ablation) | G²-LSTM · Penn Treebank | perplexity: 52.1 -> 52.8 | Hurt |
| Apply SVD rank 64 to input- and forget-gate matrices (ablation) | G²-LSTM · Penn Treebank | perplexity: 52.1 -> 56.0 | Hurt |
| Round input- and forget-gate parameters (ablation) | G²-LSTM · IWSLT’14 German→English | BLEU: 31.95 -> 31.44 | Hurt |
Closest papers by meaning
- xLSTM: Extended Long Short-Term Memory (NeurIPS 2024 · similarity 0.89)
- Wider and Deeper, Cheaper and Faster: Tensorized LSTMs for Sequence Learning (NeurIPS 2017 · similarity 0.88)
- Structured Sparsification of Gated Recurrent Neural Networks (AAAI 2020 · similarity 0.88)
- Recurrent neural networks: vanishing and exploding gradients are not the end of the story (NeurIPS 2024 · similarity 0.86)
- Accelerating Linear Recurrent Neural Networks for the Edge with Unstructured Sparsity (ICML 2025 · similarity 0.85)
- Soft Prompt Recovers Compressed LLMs, Transferably (ICML 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.