Research map / Label-efficient and robust learning / Recognition and long-tail
Generative Calibration of Inaccurate Annotation for 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 GCIA, a variational generative model that calibrates inaccurate label distribution annotations for label distribution learning using Gaussian/Dirichlet mixtures and a confusion matrix.
Problem
Label distribution learning suffers when training label distributions are inaccurate due to annotator randomness and subjectivity, degrading model accuracy and generalization.
Approach
Model features as a Gaussian mixture and ground-truth label distributions as a Dirichlet mixture tied by a shared categorical variable (smoothness assumption), and model annotator errors as a confusion matrix applied to the true distribution; infer the true distribution via variational inference (ELBO) and use it to calibrate noisy labels.
Key results
- On the sj dataset at P=4, GCIA-recovered labels reduce KL divergence from 0.5583 (noisy) to 0.1247 versus 0.1444 for KMeans calibration.
- GCIA improves recovered label distributions over noisy data on most datasets, but struggles to improve the Rho ranking metric and performs averagely on Twitter-LDL.
- Ablation removing the categorical variable z or confusion matrix Pi degrades prediction performance relative to the full GCIA model.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply GCIA calibration to noisy labels | Label distributions · SJAFFE (P=4) | Cheb: 0.1649 -> 0.1248 | Helped |
| Apply GCIA calibration to noisy labels | Label distributions · Movie (P=4) | Cheb: 0.2230 -> 0.1676 | Helped |
| Apply GCIA calibration to noisy labels | Label distributions · Twitter-LDL (P=4) | Cheb: 0.1272 -> 0.1426 | Hurt |
| Apply GCIA calibration to noisy labels | Label distributions · Yeast-heat (P=4) | Cheb: 0.1592 -> 0.1437 | Helped |
| Apply GCIA calibration to noisy labels | Label distributions · Yeast-spo (P=4) | Cheb: 0.1643 -> 0.1485 | Helped |
| Apply GCIA calibration to noisy labels | Label distributions · Yeast-cold (P=4) | Cheb: 0.2707 -> 0.2440 | Helped |
| Add confusion matrix Π (ablation) | GCIA · SJAFFE, Movie, Yeast-cold, Yeast-spo, and Yeast-heat (P=4) | prediction performance | Helped |
| Add categorical variable z (ablation) | GCIA · SJAFFE, Movie, Yeast-cold, Yeast-spo, and Yeast-heat (P=4) | prediction performance | Helped |
| Increase the annotation-noise level P | GCIA calibration · SJAFFE and Movie | recovery performance | Mixed |
Surprising findings
- Improvements in Spearman rank (Rho) are inconsistent, and the method fails to improve ranking on several datasets when label differences are small.
Limitations
- Performance on Twitter-LDL is average because highly similar images were removed during construction
- Ranking metric Rho is not reliably improved when label values differ little
- Evaluation relies on synthetic noise generated by Gaussian resampling rather than real annotator errors
Experimental setup
Datasets: SJAFFE Twitter-LDL Movie Yeast-heat Yeast-cold Yeast-spo Yeast-spo5 Nature-Scene
Benchmarks: SJAFFE Twitter-LDL Movie Yeast-heat Yeast-cold Yeast-spo Yeast-spo5
Metrics: Chebyshev distance Clark distance Canberra metric KL divergence cosine similarity intersection similarity Spearman rank correlation
Baselines: KMeans calibration 3WD-LDL SABFGS Duo-LDL LDL-LRR
Builds on: LEVI label distribution learning variational autoencoder
Closest papers by meaning
- Approximately Correct Label Distribution Learning (ICML 2025 · similarity 0.90)
- Adaptive-Grained Label Distribution Learning (AAAI 2025 · similarity 0.89)
- Predicting Label Distribution from Ternary Labels (NeurIPS 2024 · similarity 0.89)
- Divide and Conquer: Learning Label Distribution with Subtasks (ICML 2025 · similarity 0.88)
- LIMEFLDL: A Local Interpretable Model-Agnostic Explanations Approach for Label Distribution Learning (ICML 2025 · similarity 0.87)
- Generative Modeling Reinvents Supervised Learning: Label Repurposing with Predictive Consistency Learning (ICML 2025 · similarity 0.86)
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.