Auditing Pointwise Reliability Subsequent to Training

To use machine learning in high stakes applications (e.g. medicine), we need tools for building confidence in the system and evaluating whether it is reliable. Methods to improve model reliability are often applied at train time (e.g. using Bayesian inference to obtain uncertainty estimates). An alternative is to audit a fixed model subsequent to training. In this paper, we describe resampling uncertainty estimation (RUE), an algorithm to audit the pointwise reliability of predictions. Intuitively, RUE estimates the amount that a single prediction would change if the model had been fit on different training data drawn from the same distribution by using the gradient and Hessian of the model's loss on training data. Experimentally, we show that RUE more effectively detects inaccurate predictions than existing tools for auditing reliability subsequent to training. We also show that RUE can create predictive distributions that are competitive with state-of-the-art methods like Monte Carlo dropout, probabilistic backpropagation, and deep ensembles, but does not depend on specific algorithms at train-time like these methods do.