Background

Uncertainty Estimation is the task of effectively measuring the uncertainty a model exhibits.

Problem Definition

Performing Uncertainty Estimation requires creating multiple models or multiple inference processes, making it excessively costly. Therefore, it is difficult to perform Uncertainty Estimation on large pretrained models.

Solution

Evidential Deep Learning is a method that allows a model to efficiently estimate uncertainty. However, it requires a separate pretraining process, making it difficult to apply directly in our situation. In this research, we propose a Variational AutoEncoder that enables a pretrained model to perform Evidential Deep Learning with minimal cost.

Achievements

• Made it possible to perform with the same uncertainty estimation performance as existing methods but at a much lower cost.
• Currently under review for CVPR.

My Role

• As the first author, I was responsible for the planning, experimentation, and writing of the research paper.
• Implemented various baselines for Uncertainty Estimation.
• Implemented my methodology in pretrained image classification models and LLMs.