Research map / Label-efficient and robust learning / Recognition and long-tail
Structured Prediction Energy Networks
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 |
|---|---|---|---|
| Use SPEN energy-based gradient-descent prediction instead of feed-forward MLP/BR/LR/DMF | multi-label classification · Bibtex, Bookmarks, Delicious | F1 (macro): +3.3 vs MLP on Bibtex, +0.6 on Bookmarks, -0.3 on Delicious | Mixed |
| Use the SPEN with the mirror-descent-based inference and SSVM training on Yeast, compared to exact/LP/LBP/DMF inference | Yeast multi-label · Yeast, 10-fold CV | Hamming error: 20.0 vs EXACT 20.2 and LP 20.5; better than LBP 24.3 and DMF 23 | No clear effect |
| Replace MLP-based features with SPEN global energy network (measurement matrix C1 with nonlinear energy) | synthetic mutual exclusivity task · 1.5k training examples | F1: 91.5 vs 81.6 for 3-layer MLP | Helped |
| Replace SPEN with a 3-layer MLP on the synthetic task with large training set | synthetic mutual exclusivity task · 15k training examples | F1: 96.7 SPEN vs 96.3 MLP | No clear effect |
| Use HardTanh instead of ReLU in the global energy network (ablation) | SPEN structure learning on synthetic data · synthetic mutual exclusivity | interpretability of measurement matrix | Helped |
| Terminate prediction when 90% of examples in batch converge (vs running to full convergence) (ablation) | SPEN test-time prediction · Bibtex | speedup at accuracy: 3-fold speedup at little degradation in accuracy | Helped |
| Add extra entropy term to test-time prediction (soft predictions) for SPEN | SPEN on Delicious · Delicious | F1 (macro) | Helped |
| Initialize SPEN optimization at MLP output (ablation) | SPEN test-time prediction · Bibtex | iterations to converge: curves shift about 5 iterations to the left | Helped |
| Use deep mean field with pretrained MLP unary potentials (ablation) | DMF baseline · Bibtex/Bookmarks/Delicious | F1 (macro) | Helped |
| Use SPEN with ReLU in global energy network instead of HardTanh (ablation) | SPEN structure learning · synthetic mutual exclusivity | interpretability of measurement matrix | Hurt |
Closest papers by meaning
- End-to-End Learning for Structured Prediction Energy Networks (ICML 2017 · similarity 0.92)
- Search-Guided, Lightly-Supervised Training of Structured Prediction Energy Networks (NeurIPS 2019 · similarity 0.91)
- Learning Approximate Inference Networks for Structured Prediction (ICLR 2018 · similarity 0.90)
- Learning with Structural Labels for Learning with Noisy Labels (CVPR 2024 · similarity 0.85)
- Conformal Structured Prediction (ICLR 2025 · similarity 0.85)
- SOInter: A Novel Deep Energy-Based Interpretation Method for Explaining Structured Output Models (ICLR 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.