Research map / Label-efficient and robust learning / Noisy labels
Robust-by-Design Distributional Learning from Contaminated Samples
Nevena Gligić, Arya Farahi
Read the paperCheck your idea against 137,217 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 a robust distributional learning framework that recovers the target distribution from contaminated mixed samples without discarding most of the data, as stated in the abstract.
Problem
Distributional learning from observed samples that are contaminated mixtures of the target population (e.g., from retrieval, scraping, automated filtering), where standard objectives either converge to the wrong target or discard most data.
Approach
Designs distributional objectives that are robust by construction to contamination, using the full contaminated sample rather than only a small trusted subset; details beyond the abstract are unavailable.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add audit-based residual correction to the MMD objective instead of using naive MMD on the contaminated pool | MMD-based linear regression · Contaminated linear regression | l2 error: 0.259 -> 0.110 | Improved |
| use full pool with corrected objective instead of audit-only training (ablation) | CC-MMD regression · Contaminated linear regression | l2 error: 0.169 -> 0.110 | Improved |
| use bounded-kernel MMD instead of squared error or robust losses (Huber, Tukey) | Regression under contamination · Contaminated linear regression | l2 error: 0.733 -> 0.259 (Tukey vs naive MMD) | Improved |
| use Huber or Tukey robust losses instead of MSE | Regression under contamination · Concrete strength prediction (epsilon=0.6) | test MSE: 0.76 -> 0.77 (Huber), 0.95 (Tukey) | No clear effect |
| use CC-MMD instead of next-best baseline at high contamination | MLP regressor · Concrete strength prediction (epsilon=0.5) | test MSE: 34% reduction at epsilon=0.5; 25% at epsilon=0.6 | Improved |
| use CC-MMD at no contamination | MLP regressor · Concrete strength prediction (epsilon=0) | test MSE | No clear effect |
| use CC-MMD instead of naive MMD | IAF generator on MNIST with contamination · MNIST generation, epsilon=0.8 | test MMD^2: 0.046 -> 0.00371 | Improved |
| use CC-MMD instead of ELBO or adversarial training | VAE and GMMN on MNIST with K-MNIST contamination · MNIST generation, epsilon=0.60 | FID | Mixed |
| use CC-MMD instead of audit-only training at very high contamination (ablation) | IAF generator · MNIST generation, contamination sweep | test MMD^2 | Mixed |
| increase audit size (ablation) | CC-MMD · MNIST generation (audit-size ablation) | test MMD^2 / FID | Improved |
| increase proxy AUROC (ablation) | CC-MMD · MNIST generation (proxy-quality ablation) | test MMD^2 / FID | Improved |
Experimental setup
Baselines: standard distributional objectives trusted-subset training
Closest papers by meaning
- Robust Conformal Outlier Detection under Contaminated Reference Data (ICML 2025 · similarity 0.86)
- The Power of Iterative Filtering for Supervised Learning with (Heavy) Contamination (NeurIPS 2025 · similarity 0.85)
- Corruptions of Supervised Learning Problems: Typology and Mitigations (NeurIPS 2026 · similarity 0.85)
- Learning from Noisy Labels via Conditional Distributionally Robust Optimization (NeurIPS 2024 · similarity 0.84)
- Beyond Clean Training Data: A Versatile and Model-Agnostic Framework for Out-of-Distribution Detection with Contaminated Training Data (CVPR 2025 · similarity 0.84)
- Revealing Distribution Discrepancy by Sampling Transfer in Unlabeled Data (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.
