Equivalent error bars for neural network classifiers trained by Bayesian inference

The topic of this paper is the problem of outlier detection for neural networks trained by Bayesian inference. I will show that marginalization is not a good method to get moderated probabilities for classes in outlying regions. The reason why marginalization fails to indicate outliers is analysed and an alternative measure, that is a more reliable indicator for outliers, is proposed. A simple artiicial classiication problem is used to visualize the diierences. Finally both methods are used to classify a real world problem, where outlier detection is mandatory.