Research map / Learning theory and algorithms / Combinatorial and search optimization
Differentiable Dynamic Programming for Structured Prediction and Attention
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 |
|---|---|---|---|
| Train the alignment classifier end-to-end using the relaxed DTW loss rather than using a pretrained classifier | DTW audio-to-score alignment · Bach 10 test | Mean absolute deviation: 3.69 ± 2.85 -> 1.07 ± 0.61 | Helped |
| Use ℓ₂² instead of negentropy regularization (ablation) | Structured attention · WMT14 FR-EN newstest2014 | BLEU: 27.96 -> 27.19 | Hurt |
| Replace standard softmax attention with negentropy structured attention | Neural machine translation · WMT14 FR-EN newstest2014 | BLEU: 27.96 -> 27.96 | No clear effect |
| Use entropy regularization rather than ℓ₂² regularization (ablation) | NER structured prediction · CoNLL03 | F1: English: 90.86 vs 90.80; Spanish: 85.51 vs 86.68; German: 76.01 vs 77.35; Dutch: 86.58 vs 87.56 (surrogate) | Mixed |
| Use the relaxed loss rather than the surrogate convex loss (ablation) | Negentropy NER · CoNLL03 | F1: English 90.80 -> 90.47; Spanish 86.68 -> 86.20; German 77.35 -> 77.56; Dutch 87.56 -> 87.37 | Mixed |
| Use the relaxed loss rather than the surrogate convex loss (ablation) | ℓ₂²-regularized NER · CoNLL03 | F1: English 90.86 -> 89.49; Spanish 85.51 -> 84.07; German 76.01 -> 76.91; Dutch 86.58 -> 85.90 | Mixed |
| Use ℓ₂² regularization instead of negentropy (ablation) | Structured attention · WMT14 FR-EN | Attention sparsity | Helped |
| Use ℓ₂² regularization instead of negentropy (ablation) | DP-based structured prediction · CoNLL03 NER | Prediction sparsity | Helped |
Closest papers by meaning
- Scaling Structured Inference with Randomization (ICML 2022 · similarity 0.89)
- d$^2$p: Structured Soft Attention Is All You Need (ICML 2026 · similarity 0.88)
- LP-SparseMAP: Differentiable Relaxed Optimization for Sparse Structured Prediction (ICML 2020 · similarity 0.86)
- Structured Linear CDEs: Maximally Expressive and Parallel-in-Time Sequence Models (NeurIPS 2025 · similarity 0.85)
- BPQP: A Differentiable Convex Optimization Framework for Efficient End-to-End Learning (NeurIPS 2024 · similarity 0.85)
- ASGO: Adaptive Structured Gradient Optimization (NeurIPS 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.