Research map / Learning theory and algorithms / Combinatorial and search optimization
Pareto Manifold Learning: Tackling multiple tasks via ensembles of single-task models
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 weight-space interpolation between single-task predictors with preference-weighted task losses | PaMaL · Cityscapes test | mIoU: 70.35 vs 71.13 | Mixed |
| Use weight-space interpolation between single-task predictors with preference-weighted task losses | PaMaL · Cityscapes test | Abs Err: 0.0141 vs 0.0141 | No clear effect |
| Use weight-space interpolation between single-task predictors with preference-weighted task losses | PaMaL · Cityscapes test | Rel Err: 38.644 vs 54.520 | Hurt |
| Add multi-forward training and set λ=0 on symmetric MultiMNIST (ablation) | PaMaL · MultiMNIST validation | HyperVolume: 0.9129 -> 0.9184 | Helped |
| Increase multi-forward window from W=4 to W=5 at λ=0 (ablation) | PaMaL · MultiMNIST validation | HyperVolume: 0.9184 -> 0.9156 | Hurt |
| Increase regularization strength from λ=0 to λ=10 at W=2 (ablation) | PaMaL · MultiMNIST validation | HyperVolume: 0.9129 -> 0.8723 | Hurt |
| Add monotonicity regularization at W=2, λ=5 (ablation) | PaMaL · Census validation | HyperVolume: 0.6526 -> 0.6581 | Helped |
| Increase multi-forward window from W=2 to W=5 at λ=0 (ablation) | PaMaL · Census validation | HyperVolume: 0.6526 -> 0.6494 | Hurt |
| Use strong regularization λ=10 with only W=2 samples (ablation) | PaMaL · Census validation | HyperVolume: 0.6581 -> 0.6553 | Hurt |
| Add multi-forward training without regularization (ablation) | PaMaL · MultiMNIST | HyperVolume | Helped |
| Add multi-forward training without regularization (ablation) | PaMaL · Census | HyperVolume | Hurt |
| Use gradient balancing for unequal task loss scales (ablation) | PaMaL · Illustrative two-task problem | Pareto-front coverage | Helped |
Closest papers by meaning
- Efficient Pareto Manifold Learning with Low-Rank Structure (ICML 2024 · similarity 0.93)
- Pareto Low-Rank Adapters: Efficient Multi-Task Learning with Preferences (ICLR 2025 · similarity 0.88)
- Towards Principled Task Grouping for Multi-Task Learning (NeurIPS 2026 · similarity 0.88)
- Delve into the Applicability of Advanced Optimizers for Multi-Task Learning (NeurIPS 2026 · similarity 0.86)
- DMTG: One-Shot Differentiable Multi-Task Grouping (ICML 2024 · similarity 0.85)
- Bayesian Uncertainty for Gradient Aggregation in Multi-Task Learning (ICML 2024 · similarity 0.85)
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.