Research map / Architectures and efficiency / RNNs and normalization
Learning to Screen for Fast Softmax Inference on Large Vocabulary Neural Networks
Patrick Chen, Si Si, Sanjiv Kumar, Yang Li, Cho-Jui Hsieh
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 learned context-cluster screening and exact softmax over the selected candidate subset instead of full softmax | L2S · PTB-Small | speedup: 10.6x | Improved |
| Use learned context-cluster screening and exact softmax over the selected candidate subset instead of full softmax | L2S · PTB-Large | speedup: 45.3x | Improved |
| Use learned context-cluster screening and exact softmax over the selected candidate subset instead of full softmax | L2S · NMT: DE-EN | speedup: 20.4x | Improved |
| Use L2S instead of FGD in beam-search decoding | NMT: DE-EN, beam 5 · NMT: DE-EN, beam 5 | speedup: 2.9x -> 13.4x | Improved |
| Use L2S instead of FGD in beam-search decoding | NMT: EN-VE, beam 5 · NMT: EN-VE, beam 5 | speedup: 4.6x -> 20x | Improved |
| Use L2S for perplexity computation with low-rank logits outside the candidate set | PTB-Small language model · PTB-Small perplexity | speedup: 1x -> 5.69x | Improved |
| Use L2S for perplexity computation with low-rank logits outside the candidate set | PTB-Large language model · PTB-Large perplexity | PPL: 78.32 -> 80.09 | Hurt |
| Increase the number of clusters from 50 to 100 while adjusting the time budget (ablation) | L2S · cluster-count experiment | P@1: 0.997 -> 0.998 | No clear effect |
| Increase the number of clusters from 100 to 250 while adjusting the time budget (ablation) | L2S · cluster-count experiment | P@1: 0.998 -> 0.994 | No clear effect |
| Replace end-to-end learned screening with spherical k-means clustering (ablation) | L2S · PTB-Small | speedup: 10.6x -> 4x | Hurt |
| Replace end-to-end learned screening with spherical k-means clustering (ablation) | L2S · PTB-Large | speedup: 45.3x -> 6.9x | Hurt |
| Replace end-to-end learned screening with spherical k-means clustering (ablation) | L2S · NMT: DE-EN | speedup: 20.4x -> 13.8x | Hurt |
Closest papers by meaning
- Navigating with Graph Representations for Fast and Scalable Decoding of Neural Language Models (NeurIPS 2018 · similarity 0.88)
- The Closeness of In-Context Learning and Weight Shifting for Softmax Regression (NeurIPS 2024 · similarity 0.85)
- Adaptive Layer Sparsity for Large Language Models via Activation Correlation Assessment (NeurIPS 2024 · similarity 0.84)
- Discovering Sparsity Allocation for Layer-wise Pruning of Large Language Models (NeurIPS 2024 · similarity 0.84)
- A Paradigm Shift in Machine Translation: Boosting Translation Performance of Large Language Models (ICLR 2024 · similarity 0.84)
- Beyond Linear Approximations: A Novel Pruning Approach for Attention Matrix (ICLR 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.
