Research map / Label-efficient and robust learning / Recognition and long-tail

Divide and Conquer: Learning Label Distribution with Subtasks

ICML 2025 · Label-efficient and robust learning › Recognition and long-tail · approach: Partial and multi-label 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 S-LDL, which builds label-distribution subtasks by partitioning the label space via learned masks and normalizing slices, then uses them as auxiliary supervision in shallow and deep LDL frameworks.

Problem

Label distribution learning (LDL) either relies on auxiliary tasks requiring domain-specific prior knowledge, or lacks extra supervision, limiting performance and generalizability; scarce and ambiguous LDL datasets make improvement hard.

Approach

Generate subtask label distributions without expert knowledge by learning binary masks over the label space (penalizing local ignorance and mask similarity), slicing and renormalizing (NSUM) the primary label distributions, and training a plug-and-play subtask-auxiliary loss alongside the primary LDL loss; NSUM is shown to be the only normalization enabling reconstruction of the primary distributi

Key results

• On JAFFE, S-LDSVR improves Spearman from 0.5175 to 0.5237 and Cosine from 0.9549 to 0.9580 over LDSVR under the shallow regime.
• On Natural Scene, S-LDSVR raises Spearman from 0.4997 to 0.5236 and reduces Chebyshev distance from 0.4899 to 0.4692.
• On sBU 3DFE, S-HR and S-LDLM improve over their base methods on 0/1 loss, e.g., S-LDLM 0.4809 vs LDLM 0.5258 (reported as significant in several cases).

Measured effects reported in the paper

Surprising findings

• Having too many auxiliary subtasks (large T) degrades performance, likely due to overfitting.
• S-LDL4C is unstable on the Flickr dataset, where LDL4C itself fails, attributed to dataset sparsity and entropy operations.

Limitations

• Shallow-regime S-LDL is intuition-driven and relies on naive concatenation, which the deep regime addresses.
• Not applicable to large label spaces with continuous labels yielding unimodal distributions (e.g., age estimation) without binning.
• Gains on some methods (e.g., SCL) are smaller, attributed to reliance on shallow methods in prediction.

Experimental setup

Datasets: JAFFE Natural Scene Yeast diau emotion6 sBU 3DFE Flickr

Benchmarks: JAFFE Natural Scene Yeast diau sBU 3DFE Flickr

Metrics: Chebyshev distance Clark distance Canberra KL divergence cosine similarity intersection Spearman Kendall 0/1 loss error probability

Baselines: LDSVR AA-kNN LDLFs DF-BFGS KLD SCL QFD2 LRR CJS HR LDLM

Builds on: LDSVR HOMER RAkEL KLD LDL4C

Closest papers by meaning

• Adaptive-Grained Label Distribution Learning (AAAI 2025 · similarity 0.92)
• Approximately Correct Label Distribution Learning (ICML 2025 · similarity 0.91)
• Generative Calibration of Inaccurate Annotation for Label Distribution Learning (AAAI 2024 · similarity 0.88)
• LIMEFLDL: A Local Interpretable Model-Agnostic Explanations Approach for Label Distribution Learning (ICML 2025 · similarity 0.87)
• GLDL: Graph Label Distribution Learning (AAAI 2024 · similarity 0.87)
• Concentration Distribution Learning from Label Distributions (ICML 2025 · similarity 0.86)