Research map / Optimization / Optimizers
Revisiting Fairness in Multitask Learning: A Performance-Driven Approach for Variance Reduction
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes PIVRG, a gradient aggregation method for multi-task learning that uses cross-task performance variance as a fairness signal to dynamically reweight task gradients, and shows it also improves existing MTL methods when plugged in.
Problem
Task imbalance in multi-task learning: loss-based and gradient-based methods still leave some tasks under-optimized and show large variance in per-task performance drop relative to single-task learning.
Approach
Uses per-task performance drop relative to STL as a performance-level fairness indicator; derives dynamic softmax-style weights (with temperature and bounds) that reweight an inverse-utility gradient aggregation objective, with a closed-form-like least-squares solve for the update direction and a Pareto-stationarity convergence argument.
Key results
- On NYUv2 (3-task), PIVRG reaches delta m of -6.50% and mean rank 2.33, the only method beating STL on all 3 tasks and 9 metrics.
- On CelebA (40-task), PIVRG achieves delta m of -0.96% with mean rank 3.15, the first negative delta m reported there; on QM9 delta m is 33.6% vs 57.9% for FairGrad.
- On MT10 multi-task RL, PIVRG reaches a 0.96 ± 0.02 success rate, above the 0.91 of Nash-MTL and the 0.90 STL baseline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use PIVRG instead of FairGrad | MTAN · NYUv2 | ∆m(%): -4.66 -> -6.50 | Helped |
| use PIVRG instead of FairGrad | MTAN · CityScapes | ∆m(%): 5.18 -> -0.54 | Helped |
| use PIVRG instead of IMTL-G | 9-layer CNN · CelebA | ∆m(%): 0.84 -> -0.96 | Helped |
| use PIVRG instead of FairGrad | MTL · QM9 | ∆m(%): 57.9 -> 33.6 | Helped |
| use PIVRG instead of Nash-MTL | SAC · MT10 | success rate: 0.91 ± 0.03 -> 0.96 ± 0.02 | Helped |
| add performance-informed weighting to LS (ablation) | LS · NYUv2 | ∆m(%): 5.59 -> -0.06 | Helped |
| add performance-informed weighting to RLW (ablation) | RLW · NYUv2 | ∆m(%): 7.78 -> 4.52 | Helped |
| add performance-informed weighting to DWA (ablation) | DWA · NYUv2 | ∆m(%): 3.57 -> 0.78 | Helped |
| add performance-informed weighting to UW (ablation) | UW · NYUv2 | ∆m(%): 4.05 -> -1.71 | Helped |
| add performance-informed weighting to MGDA (ablation) | MGDA · NYUv2 | ∆m(%): 1.38 -> -3.45 | Helped |
| add performance-informed weighting to Nash-MTL (ablation) | Nash-MTL · NYUv2 | ∆m(%): -4.04 -> -5.77 | Helped |
Surprising findings
- Plugging the performance-informed weights into LS, DWA, UW, MGDA and Nash-MTL improves their delta m and reduces the variance of per-task performance drop on NYUv2.
Limitations
- Authors note there are many ways to incorporate performance-level information and only one fairness indicator (performance variance) is explored.
- Assumes task gradients are linearly independent until Pareto stationarity, and convergence relies on that assumption and L-smoothness.
- Weights for benchmarks without validation sets are computed from training-set performance, which may be less reliable.
Experimental setup
Datasets: NYUv2 CityScapes QM9 CelebA MT10 (Meta-World)
Benchmarks: NYUv2 CityScapes CelebA QM9 MT10
Metrics: delta m (%) mean rank variance of per-task performance drop mIoU pixel accuracy success rate
Baselines: LS SI RLW DWA UW MGDA PCGrad GradDrop CAGrad IMTL-G MoCo Nash-MTL
Builds on: Nash-MTL FairGrad MGDA FAMO IMTL-G
Closest papers by meaning
- Fair Resource Allocation in Multi-Task Learning (ICML 2024 · similarity 0.91)
- FairMT: Fairness for Heterogeneous Multi-Task Learning (NeurIPS 2026 · similarity 0.91)
- NTKMTL: Mitigating Task Imbalance in Multi-Task Learning from Neural Tangent Kernel Perspective (NeurIPS 2025 · similarity 0.88)
- Towards Consistent Multi-Task Learning: Unlocking the Potential of Task-Specific Parameters (CVPR 2025 · similarity 0.88)
- Selective Task Group Updates for Multi-Task Optimization (ICLR 2025 · similarity 0.87)
- Bayesian Uncertainty for Gradient Aggregation in Multi-Task Learning (ICML 2024 · similarity 0.87)
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.