BiasFlow: Geometric Monitoring and Backbone Regularization for Spurious Feature Reliance
BiasFlow regularizes backbone centroids, lifting worst-group accuracy by up to 26 points on UrbanCars.
BiasFlow is a hook-based toolkit that monitors class-attribute centroid alignment (IBMI), within-class separation (W-IBMI), and feature-projection sensitivity. BiasFlow Regularization adds a supervised class-conditional centroid-alignment penalty. Across small benchmarks it improves or preserves mean worst-group accuracy, with gains up to 26.0 points on UrbanCars. Freezing CelebA-Std backbones and retraining heads, BFR plus GroupDRO lifts worst-group accuracy from 40.7% to 64.1% while Male probe accuracy drops from 92.5% to 72.2%.
- IBMI tracks centroid alignment but is not causal feature reliance.
- BFR plus GroupDRO raises CelebA worst-group accuracy from 40.7% to 64.1%.
- Male probe accuracy falls from 92.5% to 72.2%; attributes stay recoverable.
- Synthetic-watermark ImageNet shift accuracy improves by 23.0 points.
From the original article
1 min read at arxiv.orgWorst-group accuracy (WGA) evaluates a trained predictor but does not characterize how its frozen backbone behaves when a new head is learned. We introduce BiasFlow, a hook-based toolkit for monitoring class-attribute centroid alignment (IBMI), within-class centroid separation (W-IBMI), and feature-projection sensitivity.
By Haojin Deng · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.