New CC-MMD Method Learns from Contaminated Data

Researchers have developed a new method called contamination-corrected maximum mean discrepancy (CC-MMD) to address learning from datasets that contain unknown contaminants. This approach combines proxy scores with audited corrections to accurately estimate the maximum mean discrepancy (MMD) from a latent clean sample, even when the data is heavily corrupted. CC-MMD does not require calibration assumptions on proxy scores or structural assumptions on the contaminant, and it has been shown to improve parameter recovery and prediction in contaminated regression tasks. The method enables generative models trained on mixed data to recover the clean target distribution, demonstrating that contaminated data need not be discarded. AI

IMPACT Enables more robust training of machine learning models on real-world, often contaminated, datasets.

RANK_REASON Academic paper detailing a new statistical method for machine learning. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv stat.ML →

• arXiv
• CC-MMD

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →