Proximity-Informed Calibration for Deep Neural Networks

Proximity bias is a widespread problem: Across 504 pretrained ImageNet models, neural networks systematically show overconfidence on data from sparse regions of the training distribution, meaning the model's confidence levels are less reliable precisely where data is rarest—a critical issue for safety-critical applications.

Popular calibration methods don't fully solve the problem: Standard calibration techniques like temperature scaling fail to address proximity bias, suggesting that current best practices for improving model reliability have a significant blind spot that could affect real-world deployment.

Transformer models are more vulnerable than CNNs: The research found that Transformer-based architectures are more susceptible to proximity bias than traditional CNN models, which has implications for organizations choosing between different model types for high-stakes applications.

A new solution and measurement approach are needed: The paper proposes ProCal, a plug-and-play algorithm designed specifically to correct for proximity bias, along with a new metric (PIECE) to measure calibration quality across different data regions—providing practical tools for practitioners to identify and address this problem.

The issue reflects deeper overfitting patterns: Models overfit more heavily on low-proximity (rare) samples than common ones, suggesting that proximity bias stems from fundamental learning dynamics rather than just confidence estimation, which may require rethinking how models are trained for reliability.