Research map / Learning theory and algorithms / Matrix and tensor methods
SVD-Softmax: Fast Softmax Approximation on Large Vocabulary Neural Networks
Kyuhong Shim, Minjae Lee, Iksoo Choi, Yoonho Boo, Wonyong Sung
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 |
|---|---|---|---|
| widen preview window W from 16 to 32 (D=256) (ablation) | SVD-softmax · WikiText-2 LM D=256 | NLL (SVD): 4.518 -> 4.441 | Improved |
| widen preview window W from 64 to 128 (D=1024) (ablation) | SVD-softmax · WikiText-2 LM D=1024 | NLL (SVD): 3.789 -> 3.761 | Improved |
| increase full-view vectors N from 32768 to 65536 (800K vocab) (ablation) | SVD-softmax · OBW 800K LM | NLL (SVD): 4.718 -> 4.690 | Improved |
| increase full-view vectors N from 1024 to 2048 (8K vocab) (ablation) | SVD-softmax · OBW 8K LM | NLL (SVD): 2.698 -> 2.687 | Improved |
| reduce W and N to 50 and 1000 (ablation) | SVD-softmax · De-En NMT | perplexity: 10.57 -> 11.04 | Hurt |
| reduce W and N to 50 and 1000 (ablation) | SVD-softmax · De-En NMT | BLEU: 21.98 -> 22.00 | No clear effect |
| replace full-softmax with SVD-softmax (W=256, N=16K) | softmax layer · GPU timing V=262k D=2k | time (ms): 14.12 -> 4.43 | Improved |
| replace full-softmax with SVD-softmax (W=256, N=16K) | softmax layer · CPU timing V=262k D=2k | time (ms): 1541.43 -> 303.57 | Improved |
Closest papers by meaning
- Navigating with Graph Representations for Fast and Scalable Decoding of Neural Language Models (NeurIPS 2018 · similarity 0.90)
- Efficient softmax approximation for GPUs (ICML 2017 · similarity 0.89)
- Learning to Screen for Fast Softmax Inference on Large Vocabulary Neural Networks (ICLR 2019 · similarity 0.88)
- Von Mises-Fisher Loss for Training Sequence to Sequence Models with Continuous Outputs (ICLR 2019 · similarity 0.88)
- Basis Sharing: Cross-Layer Parameter Sharing for Large Language Model Compression (ICLR 2025 · similarity 0.85)
- WSVD: Weighted Low-Rank Approximation for Fast and Efficient Execution of Low-Precision Vision-Language Models (ICLR 2026 · 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.
