Revisiting Confidence Estimation: Towards Reliable Failure Prediction.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Deep Neural Networks

Background:

• Reliable confidence estimation is crucial for risk-sensitive AI applications.
• Deep neural networks often exhibit overconfidence in incorrect predictions and for out-of-distribution (OOD) samples.
• Existing confidence calibration and OOD detection methods have limitations.

Purpose of the Study:

• To identify and address the detrimental effect of common confidence estimation methods on misclassification error detection.
• To improve the confidence separation between correctly classified and misclassified samples.
• To enhance the reliability of deep learning models in real-world scenarios.

Main Methods:

• Investigated the phenomenon where confidence estimation methods harm failure prediction.
• Analyzed how popular calibration and OOD detection techniques affect confidence separation.
• Proposed a novel approach of finding flat minima to enlarge the confidence gap.

Main Results:

• Demonstrated that most confidence estimation methods negatively impact misclassification detection.
• Showcased that popular methods reduce confidence separation between correct and incorrect predictions.
• Achieved state-of-the-art failure prediction performance using the flat minima approach across various classification settings.

Conclusions:

• The proposed method of finding flat minima significantly improves failure prediction.
• This work provides a strong baseline for reliable confidence estimation in deep learning.
• The study bridges the understanding between model calibration, OOD detection, and failure prediction.