Research map / Label-efficient and robust learning / Recognition and long-tail
Two Fists, One Heart: Multi-Objective Optimization Based Strategy Fusion for Long-tailed Learning
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.
Frames long-tailed learning as multi-objective optimization over head/medium/tail performance and fuses multiple existing long-tailed strategies via MOOSF (adaptive weighting, gradient orthogonal projection, and per-class output selection).
Problem
Long-tailed image classification suffers a trade-off between head and tail class performance, and existing single strategies show biased per-frequency accuracy.
Approach
Treat long-tailed learning as a multi-objective problem over head, medium, and tail performance, and fuse heterogeneous LTL strategies in a multi-task framework using three modules: HICA (cosine-similarity-based softmax weighting), GHOP (orthogonal gradient projection), and EOSS (per-class selection of the best strategy's output).
Key results
- On CIFAR-100-LT at IR=100, MOOSF(CE+BS) reaches 52.1% overall accuracy, +11.7 over the average of its two constituent strategies.
- On ImageNet-LT, MOOSF(CE+BS) achieves 57.1% overall accuracy (+10.1 relative gain), above BCL at 55.6%.
- On iNaturalist 2018, MOOSF(CE+BS) reaches 72.5% vs. 71.8% for BCL and 69.6% for KPS.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| fusing two strategies via MOOSF instead of single strategy | CE, BS, KPS, BCL · CIFAR-100-LT IR=100 | Accuracy (All): +6.6 to +11.7 over average of the pair | Helped |
| MOOSF multi-objective fusion vs linear-weighted MTL fusion (ablation) | CE+BS · CIFAR-100-LT IR=100 | Accuracy (All): 52.1 vs 48.0 | Helped |
| fusing more strategies (3 to 5) instead of two (ablation) | MOOSF on CIFAR-100-LT · CIFAR-100-LT | Accuracy (All): general decline relative to two-strategy fusion | Hurt |
| fusion of simple strategies (CE, BS) with MOOSF | ImageNet-LT and iNaturalist 2018 · ImageNet-LT / iNaturalist 2018 | Accuracy (All): +10.1 on ImageNet-LT; +8.3 on iNaturalist 2018 | Helped |
| MOOSF output conflict resolution (EOSS) and gradient projection in fusion (ablation) | strategy fusion · CIFAR-100-LT IR=100 | qualitative: multi-objective optimization benefit | Helped |
| fusion of CE with other strategies (2-strategy pairs) | MOOSF pairs in Table 2 · CIFAR-100-LT IR=50 with KPS pair | Accuracy (All): MOOSF(CE+KPS) 48.3 at IR=50 (δ1 4.5, δ2 -2.2) vs 65.3 at IR=10 | Mixed |
Surprising findings
- Fusing more than two strategies (3-5) generally underperforms fusing two strategies on CIFAR-100-LT, despite still improving over individual strategies.
- Simple strategies such as CE and BS, once fused, surpass advanced baselines like BCL and SHIKE.
Limitations
- Multi-strategy fusion adds parameters and computation relative to single strategies.
- Performance gains from fusion are capped by strategy conflicts and Pareto frontier complexity, and fusing many strategies degrades results.
- Theoretical results rely on assumptions (Lipschitz loss, compact parameter space) and the convergence proof is loose.
Experimental setup
Datasets: CIFAR-100-LT ImageNet-LT iNaturalist 2018
Benchmarks: CIFAR-100-LT ImageNet-LT iNaturalist 2018
Metrics: top-1 accuracy head accuracy medium accuracy tail accuracy
Baselines: CE CE-DRW LDAM-DRW BS RIDE (3 experts) BCL KPS SHIKE MTL fusion
Builds on: Balanced Softmax BCL KPS SHIKE LDAM-DRW
Closest papers by meaning
- Long-Tailed Learning as Multi-Objective Optimization (AAAI 2024 · similarity 0.87)
- Breaking Long-Tailed Learning Bottlenecks: A Controllable Paradigm with Hypernetwork-Generated Diverse Experts (NeurIPS 2024 · similarity 0.87)
- PSMGD: Periodic Stochastic Multi-Gradient Descent for Fast Multi-Objective Optimization (AAAI 2025 · similarity 0.85)
- SMOG: Scalable Meta-Learning for Multi-Objective Bayesian Optimization (NeurIPS 2026 · similarity 0.84)
- Improving Pareto Set Learning for Expensive Multi-objective Optimization via Stein Variational Hypernetworks (AAAI 2025 · similarity 0.84)
- Preference-Guided Diffusion for Multi-Objective Offline 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.