Neural Bayesian Anomaly Mitigation: A Robust Loss that Doubles as an Unsupervised Contamination Classifier

A new loss function called Neural Bayesian Anomaly Mitigation (NBAM) can simultaneously train supervised models to be robust to corrupted data and identify which individual samples are contaminated, according to research published on arXiv [1]. The method, introduced by Samuel Alan Kossoff Leeney, is derived from a Bayesian latent-switch mixture model [1]. The marginal likelihood of this model defines a robust supervised loss, while the associated posterior provides an unsupervised contamination classifier that returns a calibrated per-sample contamination posterior [2]. Unlike engineered robust losses such as Huber or Student-t, which make models tolerant of contamination but cannot flag corrupted observations, NBAM learns a structured contamination model [3]. A key mechanism is a learned input-dependent prior that captures the spatial locality of contamination, meaning samples near known corruptions are more likely to be flagged [4]. An Occam penalty emerges automatically from the marginal-likelihood maximisation and regularises against over-flagging without requiring a tunable coefficient [3]. NBAM requires no contamination-rate tuning; everything follows from the Bayesian mixture model [4]. On CIFAR-10 with asymmetric label contamination, NBAM recovered the structure of the corruption process without supervision. The contamination posterior separated clean from corrupted samples, and the learned anomaly head identified the direction of every label-flip pair [2]. The method outperformed four robust-loss baselines at contamination rates of 0.2 to 0.6 [1]. The problem of estimating contamination in unlabeled data has been studied from a Bayesian perspective before. Prior work introduced a method for estimating the posterior distribution of the contamination factor using a Dirichlet Process Gaussian Mixture Model on anomaly detector scores, demonstrating well-calibrated uncertainty estimates across 22 datasets [5]. Leeney has also previously explored Bayesian contamination flagging at the likelihood level in the context of radio-frequency interference mitigation for astrophysical observations, where a Bernoulli prior was used to simultaneously flag and manage corrupted data points [6]. The NBAM submission file is 226 KB [1].