Effective method to evaluate the reliability of AI without compromising accuracy

A study conducted by Yasushi Esaki et al. was selected for the presentation at The 27th International Conference on Artificial Intelligence and Statistics (AISTATS) 2024.

In recent years, artificial intelligence (AI) has been applied to various fields. However, since the predictions made by AI are not always correct, it is very important to evaluate how reliable the predictions of AI are before practical operation, such as image recognition. Some methods have been proposed to accurately evaluate the reliability of predictions. However, these methods tend to result in lower prediction accuracy compared to usual methods that focus on increasing prediction accuracy.
Therefore, in this study, we propose a method to accurately evaluate the reliability of high-performance AI obtained by the usual methods without degrading the prediction accuracy. The proposed method not only evaluates the reliability, but also can estimate whether the low reliability is due to insufficient knowledge in AI or the difficulty of the problem. Our method is expected to contribute to higher reliability and performance of AI in the future.

Title: Accuracy-preserving Calibration via Statistical Modeling on Probability Simplex
Authors: Esaki, Y., Nakamura, A., Kawano, K., Tokuhisa, R., Kutsuna, T.
Appears in: International Conference on Artificial Intelligence and Statistics
Presented: May 3, 2024