Research map / Architectures and efficiency / RNNs and normalization
On Controllable Sparse Alternatives to Softmax
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 softmax attention with sparsegen-lin | seq2seq translation and summarization · FR-EN | BLEU: 36.38 -> 37.27 | Helped |
| Replace softmax attention with sparsegen-lin | seq2seq translation · EN-FR | BLEU: 36.00 -> 35.78 | Hurt |
| Replace softmax attention with sparsegen-lin | seq2seq summarization · Gigaword | ROUGE-1: 34.80 -> 35.90 | Helped |
| Replace softmax attention with sparsegen-lin | seq2seq summarization · Gigaword | ROUGE-2: 16.64 -> 17.57 | Helped |
| Replace softmax attention with sparsegen-lin | seq2seq summarization · DUC 2003 | ROUGE-2: 9.22 -> 9.00 | Hurt |
| Replace softmax attention with sparsegen-lin | seq2seq summarization · DUC 2004 | ROUGE-1: 30.68 -> 31.85 | Helped |
| Increase the sparsegen-lin sparsity control λ | attention weights · machine translation and summarization | attention sparsity | Helped |
| Use hinge loss instead of Huber loss with sparsemax | multilabel classification · synthetic datasets | F-score | Helped |
| Use sparsehourglass hinge loss instead of sparsemax Huber loss | multilabel classification · synthetic datasets | F-score | Helped |
| Use hinge-loss variants instead of softmax+log | multilabel classification · synthetic datasets | Jensen-Shannon divergence | Helped |
| Use sparsehourglass hinge loss instead of sparsemax hinge loss | multilabel classification · synthetic datasets | F-score | Helped |
| Use sparsemax hinge loss instead of sparsemax Huber loss | multilabel classification · Scene | F-score: 0.70 -> 0.68 | Hurt |
Closest papers by meaning
- MultiMax: Sparse and Multi-Modal Attention Learning (ICML 2024 · similarity 0.86)
- Taming the Sigmoid Bottleneck: Provably Argmaxable Sparse Multi-Label Classification (AAAI 2024 · similarity 0.85)
- Global Minimizers of $\ell^p$-Regularized Objectives Yield the Sparsest ReLU Neural Networks (NeurIPS 2025 · similarity 0.85)
- $\epsilon$-Softmax: Approximating One-Hot Vectors for Mitigating Label Noise (NeurIPS 2024 · similarity 0.85)
- Navigating Extremes: Dynamic Sparsity in Large Output Spaces (NeurIPS 2024 · similarity 0.84)
- MaxQ: Multi-Axis Query for N:M Sparsity Network (CVPR 2024 · 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.