Uncertainty quantification in neural network classifiers -- a local linear approach