Research map / Architectures and efficiency / RNNs and normalization
Efficient softmax approximation for GPUs
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 |
|---|---|---|---|
| Replace full softmax with adaptive softmax | Text8 language model · Text8 | training time: 83 min -> 30 min | Helped |
| Replace full softmax with adaptive softmax | Text8 language model · Text8 | perplexity: 144 -> 147 | Hurt |
| Replace full softmax with adaptive softmax | Language model · Europarl Bulgarian | training time: 58 -> 18 | Helped |
| Replace full softmax with adaptive softmax | Language model · Europarl Bulgarian | perplexity: 37 -> 37 | No clear effect |
| Replace full softmax with adaptive softmax | Language model · Europarl Danish | perplexity: 37 -> 35 | Helped |
| Replace full softmax with adaptive softmax | Language model · Europarl German | training time: 802 -> 110 | Helped |
| Replace full softmax with adaptive softmax | Language model · Europarl Czech | perplexity: 62 -> 62 | No clear effect |
| Replace full softmax with adaptive softmax | One Billion Word LSTM language model · One Billion Word test | perplexity: 43.7 -> 43.9 | No clear effect |
| Use a 2-layer LSTM | Adaptive softmax language model · One Billion Word test | perplexity: 43.9 -> 39.8 | Helped |
| Use D-softmax [*] instead of adaptive softmax | Text8 language model · Text8 | training time: 30 min -> 54 min | Hurt |
Closest papers by meaning
- GANQ: GPU-Adaptive Non-Uniform Quantization for Large Language Models (ICML 2025 · similarity 0.85)
- Pruning Large Language Models with Semi-Structural Adaptive Sparse Training (AAAI 2025 · similarity 0.85)
- SqueezeLLM: Dense-and-Sparse Quantization (ICML 2024 · similarity 0.85)
- Search for Efficient Large Language Models (NeurIPS 2024 · similarity 0.84)
- Beyond Linear Approximations: A Novel Pruning Approach for Attention Matrix (ICLR 2025 · similarity 0.84)
- AdaSplash: Adaptive Sparse Flash Attention (ICML 2025 · similarity 0.84)
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.