Research map / Label-efficient and robust learning / Recognition and long-tail
PRIME: Deep Imbalanced Regression with Proxies
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 PRIME, a proxy-based representation learning framework for deep imbalanced regression that arranges learnable proxies uniformly over targets while preserving their ordinal structure, and aligns sample features to proxies, optionally combined with classification-style imbalance techniques.
Problem
Deep imbalanced regression: models trained on skewed continuous targets collapse minority-target features into majority regions, yielding poor minority-target accuracy and poorly ordered representations.
Approach
Learnable proxies placed at quantile-spaced target values are trained with a SNE-style ordinal loss plus a spreading term, then each sample feature is aligned to proxies via target-similarity-weighted cross-entropy, which turns regression into a classification-like objective so class-imbalance losses (PRW, CB, LDAM) can be applied.
Key results
- On AgeDB-DIR, PRIME reaches MAE 7.09 overall and GM 4.39, versus 7.10 MAE for the best prior RankSim-style baseline (and 7.34 for ConR), with PRIME+PRW reaching the best Median/Few results.
- On IMDB-WIKI-DIR with CLIP ViT-B/32 backbone, PRIME achieves MAE 6.42 and GM 3.49, outperforming IM-Context GPT2-localized (MAE 6.05 overall is better on All, but PRIME is better on Many, Median and Few is worse); PRIME improves Few-category GM over IM-Context GPT2 (3.35 vs 4.90 on AgeDB-DIR).
- On STS-B-DIR, PRIME variants reach Pearson 0.750-0.753 overall, above RankSim (0.745).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PRIME instead of RankSim | ResNet50 regression model · AgeDB-DIR All | MAE: 7.10±0.05 -> 7.09±0.08 | Helped |
| Use PRIME instead of RankSim | ResNet50 regression model · AgeDB-DIR All | GM: 4.53±0.07 -> 4.39±0.08 | Helped |
| Use PRIME instead of RankSim | ResNet50 regression model · IMDB-WIKI-DIR All | MAE: 7.43±0.04 -> 7.36±0.05 | Helped |
| Use PRIME instead of RankSim | ResNet50 regression model · IMDB-WIKI-DIR All | GM: 4.14±0.03 -> 3.98±0.04 | Helped |
| Use PRIME instead of Balanced MSE | ResNet50 encoder-decoder · NYUD2-DIR All | RMSE: 1.307±0.021 -> 1.292±0.020 | Helped |
| Add proxy-wise re-weighting | PRIME · NYUD2-DIR All | δ1: 0.687±0.004 -> 0.689±0.003 | Helped |
| Add proxy-wise re-weighting | PRIME · STS-B-DIR All | MSE: 0.970±0.004 -> 0.967±0.004 | Helped |
| Add proxy loss (ablation) | M1 on AgeDB-DIR · AgeDB-DIR All | MAE: 7.34±0.08 -> 7.24±0.09 | Helped |
| Assign proxy targets uniformly rather than randomly (ablation) | M2 on AgeDB-DIR · AgeDB-DIR All | MAE: 7.24±0.09 -> 7.23±0.12 | Helped |
| Use distance-proportional alignment instead of one-hot alignment (ablation) | M3 on AgeDB-DIR · AgeDB-DIR All | MAE: 7.23±0.12 -> 7.09±0.08 | Helped |
| Add proxy loss (ablation) | M1 on AgeDB-DIR · AgeDB-DIR All | GM: 4.66±0.07 -> 4.61±0.12 | Helped |
| Assign proxy targets uniformly rather than randomly (ablation) | M2 on AgeDB-DIR · AgeDB-DIR All | GM: 4.61±0.12 -> 4.55±0.12 | Helped |
Surprising findings
- Non-learnable centroid-based proxies match or slightly beat learnable proxies in the Median category on AgeDB-DIR, though they degrade substantially in the Few category.
- On IMDB-WIKI-DIR with CLIP backbone, the IM-Context GPT2-localized variant attains a lower overall MAE (6.05 on AgeDB-DIR) than PRIME's Many/Median results would suggest, indicating mixed comparisons across backbones.
Limitations
- Proxy count and loss hyperparameters (lambda_p, lambda_a, tau_f, tau_t, alpha) are set empirically per dataset and require tuning.
- Alignment remains affected by sample imbalance for minority targets, so class-imbalance techniques are needed to reach the best minority results.
- Theoretical bound assumes optimally positioned proxies (L_proxy = 0), with the non-optimal extension relying on an approximation argument.
Experimental setup
Datasets: AgeDB-DIR IMDB-WIKI-DIR NYUD2-DIR STS-B-DIR
Benchmarks: AgeDB-DIR IMDB-WIKI-DIR NYUD2-DIR STS-B-DIR
Metrics: MAE GM RMSE delta1 threshold accuracy MSE Pearson correlation Spearman correlation
Baselines: SQInv LDS FDS Balanced MSE RankSim VIR ConR HCA IM-Context
Builds on: ProxyNCA Balanced MSE RankSim ConR LDAM
Closest papers by meaning
- Improve Representation for Imbalanced Regression through Geometric Constraints (CVPR 2025 · similarity 0.87)
- ConR: Contrastive Regularizer for Deep Imbalanced Regression (ICLR 2024 · similarity 0.85)
- Learning under Localized Minority Imbalance (NeurIPS 2026 · similarity 0.85)
- Balancing the Scales: A Theoretical and Algorithmic Framework for Learning from Imbalanced Data (ICML 2025 · similarity 0.85)
- Leveraging Group Classification with Descending Soft Labeling for Deep Imbalanced Regression (AAAI 2025 · similarity 0.85)
- Balanced Data, Imbalanced Spectra: Unveiling Class Disparities with Spectral Imbalance (ICML 2024 · similarity 0.85)
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.