Deep Ensembles

Description

Deep ensembles estimate how uncertain a neural network's predictions are by training several copies of the same network independently and comparing what they predict. Each copy starts from different random initial weights, sees the training data in a different random order, and is trained to output a full predictive distribution, such as a probability for each class or a predicted value with its spread, rather than a single answer. The ensemble's prediction is the average of its members. Where members agree the prediction is usually more reliable, though agreement alone is no guarantee, and where they disagree the model is showing uncertainty about which answer is right, known as epistemic or model uncertainty. In the original study, ensembles of about five networks were better calibrated than single networks, meaning their stated confidence matched how often they were right, and expressed higher uncertainty on inputs from classes absent from the training data.

Example Use Cases

Reliability

An autonomous vehicle team runs five independently trained obstacle-detection networks on each camera frame. When the members' outputs disagree beyond a set level, the vehicle slows and requests remote assistance instead of acting on the averaged detection.

A research team estimating body composition from MRI scans trains an ensemble of regression networks that each predict a measurement and its uncertainty, and flags scans where the ensemble is highly uncertain for manual checking before the values enter a population study.

Transparency

A bank's underwriters see the spread of an ensemble's predicted default probabilities alongside each credit score. Scores with a wide spread are labelled low-confidence and sent for manual review, so staff can see where the model is unsure.

Limitations

Training and running M networks multiplies compute, memory and latency by roughly M. A five-member ensemble costs about five times a single model, which can rule it out for very large models or for real-time and edge deployment.
Members trained on the same data share its blind spots. On inputs far from the training data they can agree on a confident wrong answer, so low disagreement is not evidence that an input is familiar.
Applies only where the networks can be trained from scratch, with access to the training data and the compute for several full training runs. It cannot be added to an existing pretrained or third-party model.
Disagreement among a small number of members is a noisy estimate, and members that end up too alike understate uncertainty. Calibration has to be checked on held-out data, not assumed from the ensemble.
Uncertainty depends on the training objective and what each member predicts. In regression, variance across member means alone omits the within-member variability included in the original probabilistic method.

Resources

Research Papers

Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles

Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell•Dec 5, 2016

Deep neural networks (NNs) are powerful black box predictors that have recently achieved impressive performance on a wide spectrum of tasks. Quantifying predictive uncertainty in NNs is a challenging and yet unsolved problem. Bayesian NNs, which learn a distribution over weights, are currently the state-of-the-art for estimating predictive uncertainty; however these require significant modifications to the training procedure and are computationally expensive compared to standard (non-Bayesian) NNs. We propose an alternative to Bayesian NNs that is simple to implement, readily parallelizable, requires very little hyperparameter tuning, and yields high quality predictive uncertainty estimates. Through a series of experiments on classification and regression benchmarks, we demonstrate that our method produces well-calibrated uncertainty estimates which are as good or better than approximate Bayesian NNs. To assess robustness to dataset shift, we evaluate the predictive uncertainty on test examples from known and unknown distributions, and show that our method is able to express higher uncertainty on out-of-distribution examples. We demonstrate the scalability of our method by evaluating predictive uncertainty estimates on ImageNet.

Software Packages

torch-uncertainty

Open-source framework for uncertainty and deep learning models in PyTorch 🌱 - torch-uncertainty/torch-uncertainty

Tutorials

Deep ensembles segmentation tutorial using the MUAD dataset - TorchUncertainty Docs

Documentations

deep_ensembles - TorchUncertainty Docs

Related Techniques

Name	Description	Assurance Goals
Bootstrap Ensembles	Bootstrap ensembles estimate how much a model's predictions depend on the particular sample of data it was trained on. The training set is resampled with replacement many times, each resample (a bootstrap replicate) holding as many cases as the original so that some cases are repeated and others left out; one model is trained on each replicate with the same settings, and every new input is passed through all of the models. The spread of the models' predictions for an input, reported as a standard deviation or as a range between percentiles, estimates the uncertainty in the model's prediction for that input, and inputs where the models disagree are those the training data constrain least. The spread covers uncertainty in the fitted model only, not the random variation in individual outcomes, so an interval meant to contain a case's actual outcome must add an estimate of that noise, for example from the models' errors on held-out data. Putting a confidence interval around a measured performance score is a different procedure, which resamples the evaluation set and retrains nothing; see Bootstrap Confidence Intervals for Evaluation Metrics.	Reliability Transparency Fairness
Jackknife Resampling	Jackknife resampling, also called leave-one-out resampling, shows how much a result computed from a dataset depends on any single record. The result can be a model's coefficients, its predictions or a performance score; it is recomputed n times on a dataset of n records, leaving out a different record each time and refitting the model where the result depends on one. The spread of the recomputed results, after a standard rescaling, estimates the bias and variance of the original result, meaning how far it would move with a slightly different sample. Records whose removal moves the result most are flagged as influential, and may be outliers or data errors worth checking. The method assumes the records are independent and that no single record changes the result abruptly.	Reliability Transparency
Monte Carlo Dropout	Monte Carlo dropout estimates how uncertain a neural network is about each prediction. Dropout randomly switches off some of the network's units and is normally used only in training, but here it stays on at prediction time. Each forward pass (one run of an input through the network) therefore uses a different random pattern of units, and the input is run through typically tens to a hundred times. The average gives the prediction, and the spread across runs is read as epistemic uncertainty, meaning uncertainty from the model's limited knowledge rather than noise in the data. The method approximates Bayesian inference, but the spread is an estimate whose reliability should be checked against held-out errors.	Reliability