Confidence Calibration of Hand Gesture Recognition Networks Based on Ensemble Learning

초록

Confidence calibration refers to the degree of matching between predicted confidence and actual correctness probability. The ensemble of neural networks, also known as the ensemble of models, is a common method for estimating predictive uncertainty. However, this method requires a lot of modifications to the training procedure, and the computational cost is also expensive compared with other existing methods. We propose an approximate ensemble method that is simple to implement and results in well-calibrated confidence estimates without introducing various sets of models. More specifically, the historical model weights in the training phase are selected to simulate a new set of model parameters for uncertainty estimation. We show that our proposed method can produce high-quality predictive uncertainty estimates comparable to standard ensemble methods through a series of experiments on the Jester benchmark dataset. © 2023, The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd.

키워드