BiasFlow: Geometric Monitoring and Backbone Regularization for Spurious Feature Reliance
BiasFlow tests whether a frozen vision backbone still lets a biased new head recover spurious attributes.
The paper introduces diagnostics for centroid alignment and feature-projection sensitivity, while warning that its IBMI score is not causal evidence of reliance. Its regularizer, BFR, is reported to improve or preserve mean worst-group accuracy on small benchmarks, with a +26.0 point gain on UrbanCars. In a CelebA-Std frozen-backbone stress test, BFR+GroupDRO raises WGA from 40.7% to 64.1% and lowers Male probe accuracy from 92.5% to 72.2%, though attribute information remains recoverable. The authors also report a +23.0 point gain on a synthetic-watermark ImageNet shift, within matched training protocols. ArXiv · AI/CL/LG's note
The paper introduces diagnostics for centroid alignment and feature-projection sensitivity, while warning that its IBMI score is not causal evidence of reliance. Its regularizer, BFR, is reported to improve or preserve mean worst-group accuracy on small benchmarks, with a +26.0 point gain on UrbanCars. In a CelebA-Std frozen-backbone stress test, BFR+GroupDRO raises WGA from 40.7% to 64.1% and lowers Male probe accuracy from 92.5% to 72.2%, though attribute information remains recoverable. The authors also report a +23.0 point gain on a synthetic-watermark ImageNet shift, within matched training protocols. ArXiv · AI/CL/LG's note
score 4