TLDR: Researchers Anchit Jain and Stephen Bates propose a novel frequentist method, based on bootstrap, to estimate epistemic uncertainty in machine learning, which is typically measured by Bayesian mutual information. They prove this frequentist measure is asymptotically equivalent to mutual information, offering a simpler computational strategy. Their work also provides a theoretical explanation for the practical success of deep ensembles by showing they primarily capture uncertainty due to training stochasticity, a dominant component of epistemic uncertainty.
Understanding and quantifying uncertainty in machine learning predictions is crucial, especially as these systems are deployed in complex real-world scenarios. Researchers Anchit Jain and Stephen Bates from MIT have introduced a novel approach to measure epistemic uncertainty, a type of uncertainty that arises from limited training data and can be reduced by gathering more information. Their work, titled “Frequentist Validity of Epistemic Uncertainty Estimators,” offers a new perspective on a long-standing challenge in the field.
Machine learning uncertainty is typically categorized into two main types: aleatoric and epistemic. Aleatoric uncertainty stems from the inherent unpredictability in the data itself, meaning even with perfect models, some level of uncertainty would remain. Epistemic uncertainty, on the other hand, is due to the model’s lack of knowledge, often because it hasn’t seen enough training data in a particular region. Distinguishing between these is vital because it dictates the appropriate action: high aleatoric uncertainty suggests the task is inherently difficult and might require richer data, while high epistemic uncertainty signals that collecting more training data is likely to improve the model’s performance.
A popular and principled way to measure epistemic uncertainty is through mutual information (MI) between the response variable and model parameters. However, MI is a fundamentally Bayesian concept, requiring access to the posterior distribution of model parameters, which is notoriously difficult to compute in practice. This intractability has led to significant efforts in developing approximate Bayesian computational techniques, but these often necessitate changes to model architectures or training procedures, potentially compromising predictive accuracy.
A Frequentist Approach to a Bayesian Problem
Jain and Bates propose an innovative solution: a frequentist measure of epistemic uncertainty based on the bootstrap method. The bootstrap is a statistical technique that involves resampling data to estimate the sampling distribution of an estimator. The core theoretical breakthrough in their paper is a novel asymptotic expansion demonstrating that their proposed frequentist measure and the Bayesian mutual information are asymptotically equivalent. This means that as the amount of data grows, these two seemingly different measures converge to the same value. This equivalence provides a frequentist interpretation for mutual information and opens up new, computationally simpler strategies for approximating it.
The bootstrap estimator is remarkably straightforward to implement. It involves creating multiple “bootstrapped” datasets by resampling from the original training data, training models on each, and then using these models to compute the mutual information. This approach requires minimal engineering effort and imposes no restrictions on the underlying model architecture, making it a versatile complement to existing MI estimators.
Connecting with Deep Ensembles
The research also sheds light on the practical success of deep ensembles, a widely-used heuristic for uncertainty quantification in deep neural networks. Deep ensembles involve training multiple models with the same architecture and data but with different random seeds (e.g., for initialization or data shuffling). The paper shows that their bootstrap-based estimator can be naturally decomposed into components: one arising from data sampling variability and another from the stochasticity in the training procedure (like random seeds). Their experiments reveal that the component due to training stochasticity is often the dominant one, and deep ensembles effectively capture this portion of epistemic uncertainty. This provides a new frequentist motivation and theoretical backing for why deep ensembles perform so well in practice.
Also Read:
- AI Models That Understand Their Own Ignorance: A New Era of Trustworthy Machine Learning
- Information Mechanics: A Model-Free Approach to Machine Learning
Experimental Validation
The authors conducted several experiments to validate their approach. They showed that the variance in predictions from models trained on bootstrapped datasets correlates strongly with the variance from models trained on truly independent redraws of data. In active learning tasks, where algorithms select data points to label to improve model accuracy, their bootstrap measure performed comparably to established methods like Monte Carlo Dropout and deep ensembles, all significantly outperforming random data acquisition. Furthermore, their decomposition analysis confirmed that the uncertainty arising from training stochasticity is indeed the primary contributor to overall epistemic uncertainty, and this is what deep ensembles effectively capture.
The paper also explores the potential of data attribution methods, such as influence functions, to approximate epistemic uncertainty. While these methods showed promise as a proof-of-concept, they tended to underestimate the uncertainty, likely because they don’t fully account for the randomness introduced by the training process itself.
This research offers a significant step forward in understanding and quantifying epistemic uncertainty. By bridging the gap between frequentist and Bayesian perspectives, it provides both theoretical grounding and practical tools for developing more reliable and uncertainty-aware machine learning systems. For more details, you can read the full paper here.
