Research map / Label-efficient and robust learning / Recognition and long-tail
Concentration Distribution Learning from Label Distributions
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.
Introduces background concentration as an absolute description-degree term appended to label distributions, and proposes CDL-LD, an evidential Dirichlet-based neural model that learns concentration distributions directly from existing LDL datasets.
Problem
Label distributions in label distribution learning (LDL) are relative and sum to one, so they omit the absolute intensity of labels and the share of unknown/background content, causing ambiguous instances to share identical targets.
Approach
Adds a background concentration term mu (the description degree of the complement of all labels) to each label distribution to form a concentration distribution, and learns it from ordinary LDL data by placing a Dirichlet prior over apparent label distributions whose parameters come from a neural confidence network trained with an adjusted MSE (evidential) loss.
Key results
- CDL-LD ranks 1st in 45 of 48 metric-dataset comparisons (93.75%) and achieves the lowest average rank on all four metrics.
- On SJA_c, CDL-LD reaches Chebyshev 0.1153 and KL 0.0722 versus best baseline values of 0.2504 and 0.2021 (LCLR and SA-IIS).
- Friedman statistics (22.26, 40.21, 43.94, 21.87) exceed the 2.913 critical value, rejecting equal performance across methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CDL-LD instead of the strongest baseline | SJA_c · SJA_c | Chebyshev: 0.2504 -> 0.1153 | Helped |
| Use CDL-LD instead of the strongest baseline | SJA_c · SJA_c | Clark: 0.7821 -> 0.5336 | Helped |
| Use CDL-LD instead of the strongest baseline | SJA_c · SJA_c | KL: 0.2021 -> 0.0722 | Helped |
| Use CDL-LD instead of the strongest baseline | SJA_c · SJA_c | Cosine: 0.8824 -> 0.9740 | Helped |
| Use CDL-LD instead of the strongest baseline | Scene · Scene | Chebyshev: 0.4765 -> 0.3948 | Helped |
| Use CDL-LD instead of the strongest baseline | Scene · Scene | Clark: 2.4907 -> 1.9714 | Helped |
| Use CDL-LD instead of the strongest baseline | Scene · Scene | KL: 1.1354 -> 0.4918 | Helped |
| Use CDL-LD instead of the strongest baseline | Scene · Scene | Cosine: 0.5787 -> 0.6063 | Helped |
| Use CDL-LD rather than LDL baselines | CDL-LD · 12 LDL datasets | average rank | Helped |
| Use CDL-LD rather than LDL baselines | CDL-LD · Alp, Cdc, Col, Dia, Elu, Hea, Spo, Spo5 | Chebyshev | Helped |
Surprising findings
- Baselines were given a noisy ground-truth background concentration plus random noise, so the comparison is constructed rather than based on native concentration predictions.
- On the SJAFFE-based SJA dataset CDL-LD ranks 7th on Chebyshev, so it is not uniformly best.
Limitations
- Baseline background concentrations are simulated by appending the ground-truth last-class degree with random noise, so comparisons on concentration prediction are not against native CD methods.
- Concentration distribution ground truth exists only for the newly built SJA_c dataset; the other 11 datasets treat the last label as the background, an assumption that is not validated.
- The generalization bound is derived under strong assumptions (uniform spread of background concentration over classes and n tending to infinity).
Experimental setup
Datasets: Alpha Cdc Cold Diau Elu Heat Spo Spo5 SJAFFE SBU_3DFE Scene SJA_c
Benchmarks: Alpha Cdc Cold Diau Elu Heat Spo Spo5 SJAFFE SBU_3DFE Scene SJA_c
Metrics: Chebyshev Clark KL divergence Cosine similarity
Baselines: LDLLC SA-IIS LCLR LDLFs LDLLDM DLDL
Builds on: label distribution learning evidential deep learning SA-IIS DLDL
Closest papers by meaning
- Approximately Correct Label Distribution Learning (ICML 2025 · similarity 0.90)
- Divide and Conquer: Learning Label Distribution with Subtasks (ICML 2025 · similarity 0.86)
- Generative Calibration of Inaccurate Annotation for Label Distribution Learning (AAAI 2024 · similarity 0.86)
- Adaptive-Grained Label Distribution Learning (AAAI 2025 · similarity 0.85)
- LIMEFLDL: A Local Interpretable Model-Agnostic Explanations Approach for Label Distribution Learning (ICML 2025 · similarity 0.85)
- Predicting Label Distribution from Ternary Labels (NeurIPS 2024 · 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.