Neural Bayesian Anomaly Mitigation: A Robust Loss that Doubles as an Unsupervised Contamination Classifier
Researchers have developed Neural Bayesian Anomaly Mitigation (NBAM), a novel loss function designed to improve the robustness of supervised machine learning models against data contamination. NBAM not only makes models tolerant to corrupted data, similar to existing robust losses like Huber or Student's t-test, but also functions as an unsupervised classifier to identify which specific observations are corrupted. The method utilizes a Bayesian latent-switch mixture model to achieve this, outperforming baseline robust losses on the CIFAR-10 dataset with significant contamination rates. AI
IMPACT This research introduces a method to improve data quality in machine learning, potentially leading to more reliable models trained on real-world, often noisy, datasets.