Research map / Label-efficient and robust learning / Recognition and long-tail
Adaptive-Grained Label Distribution 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.
Proposes AGLDL, which adaptively coarsens noisy label distributions into ordinal coarse-grained labels, learns them with an off-the-shelf classifier, and refines predictions back into label distributions, improving label distribution learning under noisy supervision.
Problem
Label distribution learning (LDL) trained directly on collected label distributions that are unreliable due to annotator subjectivity and label enhancement assumptions, leading to unpredictable generalization.
Approach
Decompose LDL into coarsening label distributions into ordinal coarse-grained labels (via a fuzzy-dependency-based metric FDGI), learning those labels with any classifier, and refining the coarse prediction into a fine-grained distribution using an unreliable LDL predictor's output within the coarse interval.
Key results
- AGLDL achieves rank 1 on Emotion6, Painting and M2B across most measures, e.g. Emotion6 KL 0.427 vs 0.476 for the next best method.
- AGLDL performs weaker on JAFFE and Movie (high fuzzy dependency datasets), where it is comparable to the top method.
- Under increasing label noise (Figure 4), AGLDL performs best in most cases and is relatively insensitive to noise.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use AGLDL instead of LDL-LDM | Label distribution learning · JAFFE | KL: 0.044 ± 0.003 -> 0.049 ± 0.003 | Hurt |
| Use AGLDL instead of LDL-LDM | Label distribution learning · Movie | KL: 0.098 ± 0.002 -> 0.101 ± 0.002 | Hurt |
| Use AGLDL instead of LDL-LRR | Label distribution learning · Emotion6 | KL: 0.476 ± 0.067 -> 0.427 ± 0.012 | Helped |
| Use AGLDL instead of LDL-LRR | Label distribution learning · Painting | KL: 0.557 ± 0.016 -> 0.553 ± 0.010 | No clear effect |
| Use AGLDL instead of LDL-LRR | Label distribution learning · M2B | KL: 0.562 ± 0.055 -> 0.561 ± 0.019 | No clear effect |
| Use AGLDL instead of LDL-LRR | Label distribution learning · FBP5500 | KL: 0.107 ± 0.004 -> 0.134 ± 0.010 | Hurt |
| Replace adaptive label coarsening with uniform discretization (ablation) | AGLDL · JAFFE, Movie, M2B | KL | Hurt |
| Replace adaptive label coarsening with quantile discretization (ablation) | AGLDL · JAFFE, Movie, M2B | KL | Hurt |
| Replace adaptive label coarsening with k-means discretization (ablation) | AGLDL · JAFFE, Movie, M2B | KL | Mixed |
Surprising findings
- Performance of the alternative discretizations first improves and then degrades as the number of coarse labels increases, indicating a granularity trade-off.
Limitations
- Performs weakly on datasets with high fuzzy dependency such as JAFFE and Movie.
- Refinement uses Monte Carlo sampling (L=20) and the exact expectation is O(KM), adding cost.
- Theoretical results rely on uniform-noise assumptions for the true LDD within an interval.
Experimental setup
Datasets: JAFFE Movie Emotion6 Painting M2B FBP5500
Benchmarks: JAFFE Movie Emotion6 Painting M2B FBP5500
Metrics: Cheb Clark KL Cosine Intersec
Baselines: SABFGS AAkNN BD-LDL LDL-LRR LDL-LDM
Builds on: fuzzy rough feature selection LogisticAT ordinal regression label enhancement
Closest papers by meaning
- Divide and Conquer: Learning Label Distribution with Subtasks (ICML 2025 · similarity 0.92)
- Label Distribution Learning by Exploiting Label Correlations (AAAI 2018 · similarity 0.92)
- Approximately Correct Label Distribution Learning (ICML 2025 · similarity 0.90)
- Generative Calibration of Inaccurate Annotation for Label Distribution Learning (AAAI 2024 · similarity 0.89)
- Predicting Label Distribution from Ternary Labels (NeurIPS 2024 · similarity 0.89)
- LIMEFLDL: A Local Interpretable Model-Agnostic Explanations Approach for Label Distribution Learning (ICML 2025 · similarity 0.88)
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.