Calibrating deep classifiers with dynamic confidence propagation and adaptive normalization.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Deep neural networks (DNNs) show progress in confidence calibration.
• Conventional methods (e.g., temperature scaling, histogram binning) struggle in dynamic, open-world settings due to static parameters and data assumptions.

Purpose of the Study:

• To address limitations of existing confidence calibration methods in dynamic environments.
• To propose a novel framework, Dynamic Confidence Propagation and Alternating Normalization (DCP-AN), for improved DNN calibration.

Main Methods:

• Introduced a bidirectional alternating propagation mechanism for sample-class confidence synergy.
• Implemented entropy-driven horizontal and KL-divergence-weighted vertical normalization.
• Developed an adaptive temperature field with dynamic coefficients for differential calibration.
• Ensured theoretical convergence within 15 iterations.

Main Results:

• Achieved a 10.3% boost in tail-class accuracy on ImageNet-LT.
• Reduced expected calibration error by 56% on ImageNet-LT.
• Decreased domain discrepancy by 24% and improved target domain accuracy by 5.5% in cross-domain adaptation.
• Demonstrated low GPU latency (1.05 ms) and memory overhead (<0.5 MB).

Conclusions:

• DCP-AN significantly enhances DNN confidence calibration in dynamic and open-world scenarios.
• The framework offers substantial improvements in accuracy and calibration error reduction.
• DCP-AN is computationally efficient and suitable for real-time applications.