TL;DR
Deep ensembles are a popular technique for improving neural network performance and uncertainty estimation. However, this paper argues that the benefits of deep ensembles, particularly in terms of robustness and uncertainty quantification on out-of-distribution data, can be replicated by larger, single neural networks. This suggests that deep ensembles might not be as uniquely necessary as previously thought, and their advantages may stem more from increased model capacity than from the ensemble effect itself.
What Problem Does This Paper Address?
In machine learning, especially in safety-critical applications, achieving high accuracy and reliable uncertainty estimates is paramount. Deep ensembles, which combine predictions from multiple neural networks, have emerged as a powerful tool for improving both. It's widely believed that the diversity of predictions within an ensemble is key to these benefits, particularly for handling out-of-distribution (OOD) data shifts and providing well-calibrated uncertainty estimates. However, the question remains: are these benefits truly unique to ensembles, or can they be achieved through other means, like simply using a larger, more capable single model? This paper aims to rigorously investigate this question, challenging the conventional wisdom that deep ensembles are inherently superior for these specific tasks.
Methods
The authors employ a systematic approach to dissect the contributions of ensemble diversity to predictive performance, uncertainty quantification (UQ), and robustness. They decompose ensemble uncertainty into "ensemble diversity" and "average single model uncertainty." By controlling for the latter, they analyze how much of the UQ benefit can be attributed to diversity itself. They also investigate the relationship between ensemble diversity and the expected improvement from increasing model capacity. For robustness, they examine whether deep ensembles exhibit "effective robustness," meaning they outperform the expected OOD performance predicted by their InD performance, a relationship well-established for single models. Statistical tests, including kernel ridge regression and Monte Carlo permutation tests, are used to validate their findings across various metrics and datasets.
Datasets
The paper utilizes several standard and OOD datasets for training and evaluation:
- In-distribution (InD) datasets:
- CIFAR10
- ImageNet
- Out-of-distribution (OOD) datasets:
- CIFAR10.1 (reproduction of CIFAR10)
- CINIC10 (derived from ImageNet)
- CIFAR10C (synthetically corrupted CIFAR10)
- ImageNetV2 (reproduction of ImageNet)
- ImageNetC (synthetically corrupted ImageNet)
Key Insights and Ideas
- Ensemble diversity doesn't directly drive OOD uncertainty: The paper's most striking finding is that, after accounting for the uncertainty of individual models, ensemble diversity (e.g., variance in predictions) does not significantly differ between in-distribution and out-of-distribution data. This challenges the hypothesis that diversity is the primary driver of improved OOD uncertainty estimates.
- Ensemble diversity correlates with increased model capacity: Ensemble diversity is found to be strongly correlated with the expected performance improvements gained by simply increasing the capacity of a single model. This suggests that the benefits attributed to diversity might be a proxy for having a more powerful model overall.
- Deep ensembles are not "effectively robust": The study demonstrates that deep ensembles do not exhibit "effective robustness" beyond what is expected from a single model with comparable InD performance. Their OOD performance, across various metrics (accuracy, NLL, Brier score, calibration), follows the same deterministic relationship to InD performance as single models.
- Bayesian model averaging interpretation is limited: The authors show that deep ensembles do not exhibit the expected behavior of Bayesian model averaging, which should ideally show higher posterior variance for OOD data. This suggests that deep ensembles are an approximation, not a direct implementation, of BMA.
Key Results
- The conditional distributions of ensemble diversity on InD and OOD data are found to be "extremely similar" and "not distinguishable" after controlling for average single model uncertainty.
- The percentage increase in expected conditional diversity from OOD to InD data is often very small, on the order of ~1%, and statistically insignificant in many cases.
- The paper shows that performance improvements from ensembling and increasing single model capacity are highly correlated (Pearson's R of 0.81 on InD data, 0.76 on OOD data for one experiment).
- Across various metrics (0-1 error, NLL, Brier score, rESCE), deep ensembles do not show better OOD performance than single models when conditioned on InD performance.
Limitations and Future Work
- Focus on neural network ensembles: The authors acknowledge that their findings are specific to neural network ensembles and may not generalize to other ensemble methods like random forests or gradient boosted decision trees. Future work could explore these other methods.
- Overparameterized regime: The study focuses on overparameterized neural networks. Investigating the effect of overparameterization directly by using ensembles of weak learners is suggested for future research.
- Interpretation of uncertainty: The paper cautions against directly interpreting ensemble diversity as purely epistemic uncertainty, as the observed patterns do not align with theoretical expectations for OOD data.
- Practical considerations: While deep ensembles may not offer unique benefits, they remain a convenient and reliable "black box" method for improving performance. The choice between ensembles and larger single models might ultimately depend on practical factors like computational budget and ease of implementation. The authors state they have "yet to find evidence for any reason to prefer the use of deep ensembles over an appropriately chosen single model" beyond convenience.
Summary updated