Dominance-Relative Bayesian Self-Correction A Pattern-Level Error Correction Mechanism for Neural Classifiers

We present a Bayesian self-correction layer that operates on the output distributionof a neural classifier. Rather than applying a global correction, the mechanism treatsthe dominant pattern as a hypothesis and the remaining probability mass as the opposingevidence. A Bayes-style update re-weights the dominant pattern relative to this opposition,then redistributes the residual probability to non-dominant patterns according to learnedpriors. An adaptive bias term — piecewise linear for small errors and logarithmicallydampened for large ones — is accumulated over training to reduce the mean absolute errorbetween the raw and corrected distributions. Experiments on a 16-class Geometric MLPshow a post-training MAE of 0.0079, average confidence of 0.999, and a reduction oflinear-mode MAE from 0.0088 to 0.0051, while log-mode MAE converges stably.