Uncertainty quantification for convolutional neural networks based on test time augmentation with multiobjective evolutionary algorithms

doi:10.1016/j.asoc.2026.115799

Uncertainty quantification for convolutional neural networks based on test time augmentation with multiobjective evolutionary algorithms

2026-06-20

0

PRE

AI

DOI：10.1016/j.asoc.2026.115799

Original

Original request for help

Save

Abstract

En 中文

• A novel method based on test time augmentation with multiobjective evolutionary algorithms is proposed to evaluate the uncertainty for convolutional neural networks in regression tasks. Specifically, in the proposed method, the objectives of minimizing the difference between the average of the predictive distribution and the ground truth and maximizing the correlation between the variance of the predictive distribution and the predicted error are taken as the two primary objectives. A multiobjective evolutionary algorithm is employed to optimize the two objectives such that the appropriate transformed images can be automatically determined when generating the associated predictive distribution. • A thorough analysis is conducted to assess the effectiveness of the proposed method. The experimental results demonstrate that the proposed method can provide a set of trade-off solutions for regression and uncertainty estimation tasks, indicating that the proposed method possesses the ability to determine the appropriate transformed images to use when generating the predictive distribution. In addition, the proposed method also yields some interesting findings (see the discussion section), which could encourage the soft computing community to explore novel approaches to solve a much wider range of issues in the future.

Uncertainty quantification for convolutional neural networks based on test time augmentation with multiobjective evolutionary algorithms

Abstract

Journal

Organization