(Implicit) Ensembles of Ensembles: Epistemic Uncertainty Collapse in Large Models

TL;DR

This paper investigates the phenomenon where larger deep learning models exhibit a collapse in epistemic uncertainty, challenging assumptions about model size and uncertainty quantification, and proposes implicit ensembling to address this issue.

Contribution

The paper introduces the concept of implicit ensembling as an explanation for uncertainty collapse and develops techniques to extract diverse sub-models from large models to recover epistemic uncertainty.

Findings

Uncertainty collapse occurs in larger models across various architectures.

Implicit ensembling can explain and mitigate the uncertainty collapse.

Extracted sub-models from large models can recover epistemic uncertainty.

Abstract

Epistemic uncertainty is crucial for safety-critical applications and data acquisition tasks. Yet, we find an important phenomenon in deep learning models: an epistemic uncertainty collapse as model complexity increases, challenging the assumption that larger models invariably offer better uncertainty quantification. We introduce implicit ensembling as a possible explanation for this phenomenon. To investigate this hypothesis, we provide theoretical analysis and experiments that demonstrate uncertainty collapse in explicit ensembles of ensembles and show experimental evidence of similar collapse in wider models across various architectures, from simple MLPs to state-of-the-art vision models including ResNets and Vision Transformers. We further develop implicit ensemble extraction techniques to decompose larger models into diverse sub-models, showing we can thus recover epistemic uncertainty. We explore the implications of these findings for uncertainty estimation.