BiasFlow: Geometric Monitoring and Backbone Regularization for Spurious Feature Reliance
arXiv cs.AI · 6 Oct 2026, 03:59 AEST
BiasFlow introduces a toolkit for monitoring class-attribute centroid alignment and feature-projection sensitivity. The method shows improvements in worst-group accuracy across various benchmarks, particularly in biased data scenarios.
Key facts
- Worst-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.
- Across the reported small-scale benchmarks, adding BFR improves or preserves mean WGA, with gains up to +26.0 pp on UrbanCars.
- The principal independent stress test freezes CelebA-Std backbones and trains fresh heads on biased data: BFR+GroupDRO improves WGA from 40.7% to 64.1%, while Male probe accuracy decreases from 92.5% to 72.2%.
- A controlled synthetic-watermark ImageNet experiment additionally improves watermark-shift accuracy by +23.0 pp under matched training.
Read aloud
BiasFlow introduces a toolkit for monitoring class-attribute centroid alignment and feature-projection sensitivity. The method shows improvements in worst-group accuracy across various benchmarks, particularly in biased data scenarios. Worst-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. Across the reported small-scale benchmarks, adding BFR improves or preserves mean WGA, with gains up to +twenty six point zero pp on UrbanCars. The principal independent stress test freezes CelebA-Std backbones and trains fresh heads on biased data: BFR+GroupDRO improves WGA from forty point seven percent to sixty four point one percent, while Male probe accuracy decreases from ninety two point five percent to seventy two point two percent. A controlled synthetic-watermark ImageNet experiment additionally improves watermark-shift accuracy by +twenty three point zero pp under matched training.
Sources in this cluster
- BiasFlow: Geometric Monitoring and Backbone Regularization for Spurious Feature Reliance
arXiv cs.AI · 6 Oct 2026, 03:59 AEST