PEER: Towards reliable and efficient inference via Patience-Based Early Exiting with Rejection

Area of Science:

• Artificial Intelligence
• Biomedical Informatics
• Machine Learning

Background:

• Biomedical AI models require a balance between inference efficiency and prediction reliability.
• Patience-based early exiting (PABEE) accelerates inference but struggles with uncertain predictions.

Purpose of the Study:

• To introduce PEER (Patience-based Early Exiting with Rejection), a unified framework combining PABEE with a rejection mechanism.
• To enable efficient and reliable AI inference in biomedical applications without retraining.

Main Methods:

• PEER integrates a rejection mechanism into PABEE using a patience counter to track prediction consistency.
• Models either make a prediction or reject uncertain inputs, avoiding unreliable final-layer predictions.
• Evaluated on 11 biomedical datasets (text and images) using Transformer backbones, measuring accuracy, macro-F1, and speed-up ratio.

Main Results:

• PEER consistently enhances reliability while maintaining early exiting efficiency gains.
• Achieved 90.73% accuracy on MIMIC-III by rejecting 2.79% of uncertain samples, outperforming the baseline.
• Demonstrated an 80% speed-up ratio in high-efficiency settings with comparable performance.
• Successfully abstained from uncertain cases, improving prediction trustworthiness across diverse datasets and modalities.
• Showcased generalization across various architectures, scales, and modalities (language and vision).

Conclusions:

• PEER provides a simple, architecture-agnostic framework for fast and trustworthy AI inference.
• Its generalizability across language and vision models shows significant potential for clinical decision support systems.