This repository contains code used to run experiments in Can You Trust Your Model's Uncertainty? Evaluating Predictive Uncertainty Under Dataset Shift, which benchmarks a handful of methods for training models with robust uncertainty on varying datasets with off-distribution holdout test-sets.

The study compared the following methods and datasets:

1. Vanilla: Maximum-likelihood DNN.

2. Temperature Scaling: Vanilla model calibrated on in-distribution as described in Guo et al. 2017

3. Ensemble of vanilla models as described in Lakshminarayanan et al. 2017

4. Dropout: MC Dropout as described by Gal & Ghahramani

5. SVI: Mean field BNN optimized by stochastic variational inference.

6. LL-Dropout and LL-SVI: Simplified variants of Dropout and SVI where the method is only applied to the last layer.

1. MNIST and rotated MNIST.
2. CIFAR-10 with corruptions by Hendrycks & Dietterich 2019
3. ImageNet 2012 with corruptions by Hendrycks & Dietterich 2019
4. Criteo's ad-click prediction dataset with synthetic random feature corruptions.
5. 20 Newsgroups text with out-of-distribution data from LM1B.

Abstract: Modern machine learning methods including deep learning have achieved great success in predictive accuracy for supervised learning tasks, but may still fall short in giving useful estimates of their predictive uncertainty. Quantifying uncertainty is especially critical in real-world settings, which often involve input distributions that are shifted from the training distribution due to a variety of factors including sample bias and non-stationarity. In such settings, well calibrated uncertainty estimates convey information about when a model's output should (or should not) be trusted. Many probabilistic deep learning methods, including Bayesian-and non-Bayesian methods, have been proposed in the literature for quantifying predictive uncertainty, but to our knowledge there has not previously been a rigorous large-scale empirical comparison of these methods under dataset shift. We present a large-scale benchmark of existing state-of-the-art methods on classification problems and investigate the effect of dataset shift on accuracy and calibration. We find that traditional post-hoc calibration does indeed fall short, as do several other previous methods. However, some methods that marginalize over models give surprisingly strong results across a broad spectrum of tasks.

To facilitate follow-up work, we also make available all of our trained models along with their predictions on each dataset at the GCS bucket: uq-benchmark-2019.

The predictions files are structured as follows:

mnist_model_predictions.hdf5 and cifar_model_predictions.hdf5 are HDF5 file with dataset group hierarchy:

METHOD > DATASET > {labels, probs}

where METHOD corresponds to the modeling method, and DATASET corresponds to an MNIST data split (or off-distribution variant) with posible corruptions.

Labels have shape [5, N] where N is the number of samples in the dataset, and 5 is the number of independently trained models for verifying reproducibility. Probabilities datasets have shape [5, N, 10] where 10 is the number of MNIST / CIFAR classes.

criteo_model_predictions.hdf5 is organized identically to MNIST and CIFAR except the shape of the probabilities array is [5, N] since Criteo is not a multiclass problem.

imagenet_predictions.hdf5 is organized identically to the other datasets, except the probabilities array has shape [N, 1000] and the labels are a vector of length N. For the ensemble, dropout, and SVI methods, probabilties represent the means over ensemble members/samples. Probabilities less than 1e-6 have been rounded to 0 to increase gzip compressibilty.