Frequency-Aware Gradient Rectification for Robust Model Calibration
A new training framework called Frequency-aware Gradient Rectification (FGR) addresses the challenge of maintaining reliable confidence estimates in deep neural networks when deployed in real-world scenarios where distribution shifts occur. Existing calibration methods often require access to target domains, limiting their practicality. FGR is target-agnostic, using low-pass filtering on a subset of training images to reduce spurious high-frequency cues and promote domain-invariant feature learning. To prevent degradation of in-distribution calibration due to information loss, FGR treats ID calibration as a hard constraint and resolves conflicting parameter updates through geometric projection. The approach is detailed in arXiv:2508.19830v2.
Key facts
- FGR is a target-agnostic training framework for robust calibration.
- It uses low-pass filtering to diminish spurious high-frequency cues.
- FGR encourages learning of domain-invariant features.
- ID calibration is treated as a hard constraint.
- Conflicting parameter updates are rectified via geometric projection.
- The method is described in arXiv:2508.19830v2.
- Existing methods often rely on access to target domains.
- Distribution shifts cause unreliable confidence estimates in deep neural networks.
Entities
—