Research map / Label-efficient and robust learning / Recognition and long-tail
Addressing the Loss-Metric Mismatch with Adaptive Loss Alignment
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 |
|---|---|---|---|
| adapt the loss online using a validation classification-error reward | ALA on CIFAR-10 · CIFAR-10 | classification error (%): 6.95 -> 6.79 | Helped |
| adapt the loss directly against AUCPR | ALA on CIFAR-10 · CIFAR-10 | AUCPR (%): 94.2 -> 94.9±0.14 | Helped |
| adapt the distance-mixture loss with a Recall@k reward | Margin metric-learning framework · Stanford Online Products | Recall@1 (%): 76.3 -> 78.9 | Helped |
| transfer the CIFAR-10-trained ALA policy to ImageNet | RMSProp classifier · ImageNet | Top-1 accuracy (%): 73.9 -> 74.3 | Helped |
| replace learned loss parameters with a random matrix (ablation) | ALA classification on CIFAR-10 · CIFAR-10 | classification error (%): 6.79±0.07 -> 8.23±0.41 | Hurt |
| use the prediction confusion matrix instead of learned loss parameters (ablation) | ALA classification on CIFAR-10 · CIFAR-10 | classification error (%): 6.79±0.07 -> 7.42±0.04 | Hurt |
| train with one network rather than multiple child networks (ablation) | ALA classification on CIFAR-10 · CIFAR-10 | classification error (%): 6.79±0.07 -> 6.85±0.09 | Hurt |
| continue a second run while reusing or fine-tuning the policy (ablation) | ALA on CIFAR-10 · CIFAR-10 | testing classification error (%): 6.79±0.07 -> 6.72±0.04; 6.79±0.07 -> 6.72±0.02 | Helped |
| replace distance-mixture parameterization with focal weighting (ablation) | Margin + ALA · Stanford Online Products | Recall@1 (%): 78.9 -> 77.9 | Hurt |
| reduce the controller from two layers to one (ablation) | ALA on CIFAR-10 · CIFAR-10 | classification error increase (ΔError): +0.06 | Hurt |
| reduce the controller from two layers to one (ablation) | Margin + ALA · Stanford Online Products | Recall@1 change (ΔRecall): -0.5 | Hurt |
| increase the controller from two layers to three (ablation) | ALA · CIFAR-10 and Stanford Online Products | classification error increase (ΔError) and Recall@1 change (ΔRecall): -0.03 ΔError; -0.1 ΔRecall | No clear effect |
Closest papers by meaning
- Leaving the Nest: Going beyond Local Loss Functions for Predict-Then-Optimize (AAAI 2024 · similarity 0.85)
- Decision-Focused Learning with Directional Gradients (NeurIPS 2024 · similarity 0.85)
- Adaptive Preference Scaling for Reinforcement Learning with Human Feedback (NeurIPS 2024 · similarity 0.84)
- MetaOptimize: A Framework for Optimizing Step Sizes and Other Meta-parameters (ICML 2025 · similarity 0.84)
- A Survey of Learning Criteria Going beyond the Usual Risk (Abstract Reprint) (AAAI 2024 · similarity 0.84)
- Gradient Descent as Loss Landscape Navigation: a Normative Framework for Deriving Learning Rules (NeurIPS 2025 · similarity 0.83)
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.