FastEnsemble: Benchmarking and Accelerating Ensemble-based Uncertainty Estimation for Image-to-Image Translation

Estimating prediction uncertainty and confidence of deep learning models is crucial for mission-critical machine learning applications, such as biomedical imaging for diagnostics or therapy, and self-driving cars. However, making robust uncertainty estimation is complicated given the variety of learning objectives, data modalities, types of data corruption. Previous studies often addressed such a challenge by restricting datasets to standard ones like CIFAR or ImageNet. While convenient, it is doubtful whether the same conclusion holds for real-life datasets, in which more complicated image generation tasks are involved. This paper presents a different perspective to evaluate how confidence and uncertainty estimators behave under distribution shifts, focusing on the biomedical imaging domain. Specifically, we test a series of pair-wise cell imaging datasets using a new metric to compare existing models. In addition, we introduce FastEnsemble, a fast ensemble method which only requires less than $8\%$ of the full-ensemble training time to generate a new ensemble member. Our experiments show that the proposed fast ensemble method is able to substantially improve the speed vs. quality trade-off.