BiasFlow: Geometric Monitoring and Backbone Regularization for Spurious Feature Reliance

Forum topic · 小凯 · 2026-10-07

Summary

BiasFlow is a hook-based toolkit for measuring how frozen deep learning backbones behave when new classification heads are trained on biased data. It introduces geometric diagnostics including class-attribute centroid alignment (IBMI), within-class centroid separation (W-IBMI), and feature projection sensitivity, paired with BiasFlow Regularization (BFR), a supervised, composable class-conditioned centroid-alignment penalty. On reported small-scale benchmarks, adding BFR improved or preserved average worst-group accuracy (WGA), with gains up to +26.0 percentage points on UrbanCars. An independent stress test froze CelebA-Std backbones and trained new heads on biased data: BFR combined with GroupDRO raised WGA from 40.7% to 64.1% while male-probe accuracy dropped from 92.5% to 72.2%, indicating reduced attribute encoding, though attribute information remains partially recoverable and cross-task results are mixed. A controlled synthetic-watermark ImageNet experiment additionally improved watermark-bias accuracy by +23.0 percentage points under matched training. The paper (arXiv:2610.06846) is by Haojin Deng, Zhiping Lin, and Yimin Yang.

Paper Overview

Field: Machine Learning Authors: Haojin Deng, Zhiping Lin, Yimin Yang Published: 2026-10-05 arXiv: 2610.06846

Summary

Worst-group accuracy (WGA) evaluates a trained predictor but does not characterize how its frozen backbone behaves when a new head is learned on top of it. The authors introduce BiasFlow, a hook-based toolkit for monitoring:

IBMI — class–attribute centroid alignment
W-IBMI — within-class centroid separation
Feature projection sensitivity

These diagnostics are paired with BiasFlow Regularization (BFR), a supervised, composable class-conditioned centroid-alignment penalty.

Key Findings

On reported small-scale benchmarks, adding BFR improved or preserved average WGA, with gains up to +26.0 percentage points on UrbanCars.
In an independent stress test, a frozen CelebA-Std backbone with a new head trained on biased data: BFR + GroupDRO raised WGA from 40.7% to 64.1%, while male-probe accuracy dropped from 92.5% to 72.2%.
Attribute information remains partially recoverable, and cross-task results are mixed.
A controlled synthetic-watermark ImageNet experiment further improved watermark-bias accuracy by +23.0 percentage points under matched training conditions.

Original Abstract

BiasFlow introduces hook-based geometric monitoring (IBMI, W-IBMI) and a composable centroid-alignment regularizer (BFR) to characterize and reduce spurious feature reliance in frozen backbones when new heads are learned on biased data.

--- *Collected automatically on 2026-10-07.*

BiasFlow: Geometric Monitoring and Backbone Regularization for Spurious Feature Reliance

Summary

Paper Overview

Summary

Key Findings

Original Abstract

Tags