Augmentation by Counterfactual Explanation - Fixing an Overconfident Classifier.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Overconfident AI models pose risks in critical applications like healthcare and autonomous driving.
• Models must accurately reflect uncertainty for ambiguous in-distribution and out-of-distribution (OOD) samples.
• Existing methods struggle with nuanced uncertainty quantification for novel or boundary cases.

Purpose of the Study:

• To address overconfidence in AI classifiers by leveraging counterfactual explanations.
• To enhance model uncertainty estimation for improved safety in critical AI systems.
• To maintain predictive performance while refining uncertainty characteristics.

Main Methods:

• Proposed fine-tuning a pre-trained classifier with augmentations from a counterfactual explainer (ACE).
• Evaluated the approach on detecting far-out-of-distribution (far-OOD), near-out-of-distribution (near-OOD), and ambiguous samples.
• Focused on improving uncertainty measures without sacrificing classification accuracy.

Main Results:

• The revised model demonstrated significantly improved uncertainty quantification.
• Effectively identified ambiguous, far-OOD, and near-OOD samples with greater accuracy.
• Achieved competitive performance compared to state-of-the-art uncertainty estimation techniques.

Conclusions:

• Counterfactual explanations offer a viable method for correcting overconfident AI classifiers.
• The ACE-based fine-tuning approach enhances model reliability for real-world deployment.
• This technique is promising for developing safer and more trustworthy AI systems.