BiasFlow: Geometric Monitoring and Backbone Regularization for Spurious Feature Reliance
BiasFlow：以幾何監控與骨幹正規化解決模型對虛假特徵的依賴
While Worst-Group Accuracy (WGA) evaluates trained predictors, it fails to capture how a frozen backbone behaves under new heads. The authors present BiasFlow, a hook-based toolkit using geometric diagnostics (IBMI, W-IBMI) to monitor class-attribute alignment. Paired with BiasFlow Regularization (BFR), a class-conditional alignment penalty, it mitigates spurious feature reliance. BFR improves UrbanCars WGA by up to +26.0 pp and boosts frozen CelebA-Std backbone WGA from 40.7% to 64.1% when evaluated with fresh classification heads.
Key points
Geometric Monitoring
Uses hook-based diagnostics (IBMI, W-IBMI) to analyze class-attribute centroid alignment and feature-projection sensitivity.
BiasFlow Regularization
Introduces a supervised, composable class-conditional centroid-alignment penalty to optimize backbone representations.
Substantial Bias Mitigation
Improves frozen CelebA-Std backbone WGA from 40.7% to 64.1% while successfully reducing biased probe accuracy.
Watermark Shift Resilience
Boosts synthetic-watermark ImageNet watermark-shift accuracy by +23.0 pp under matched training protocols.
How it works
| 標準骨幹網路 (Standard) | BiasFlow 正規化骨幹 (BFR + GroupDRO) | |
|---|---|---|
| Worst-Group Accuracy (WGA) | 40.7% | 64.1% |
| Male Probe Accuracy (Lower is better) | 92.5% | 72.2% |
| UrbanCars WGA Improvement | 基準值 (Baseline) | 最高提升 +26.0 pp |
Why it matters
Deep learning models frequently fail when relying on spurious correlations (such as background or watermarks). BiasFlow addresses this vulnerability from a representation geometry perspective, providing tools to both diagnose and regularize models. By enabling backbones to resist biased head retraining, it offers a practical pathway toward building fairer and more robust computer vision systems.
Who it affects
- AI Researcher
- AI Developer
How to use it
- 1Evaluating pre-trained vision models for reliance on spurious features such as watermarks and backgrounds.
- 2Training robust image classifiers on datasets prone to group bias or correlated attribute imbalances.
Limitations & caveats
- Attribute information remains recoverable in representations, and cross-task results are mixed.
- The W-IBMI diagnostic metric is scale-dependent and cannot independently establish attribute removal.
Related
Less Decoder is More Encoder: Extracting Robust 3D Geometric Representations via Novel View Synthesis
減少解碼器反而增強編碼器：從新視角合成中提煉強大三維幾何表徵
This paper reveals how expressive decoders dilute geometric learning in Novel View Synthesis, and proposes SNAP—a self-supervised framework that restricts decoders to force encoders to learn robust 3D representations.
4DCodeBench: Benchmarking AI Agents on 4D Inverse Graphics and Dynamic Scene Code Generation
4DCodeBench：評估 AI Agent 動態場景 4D 反向圖形學與程式碼生成能力的全新基準
4DCodeBench is a new benchmark designed to evaluate AI agents' ability to reconstruct 4D dynamic scenes from videos by generating executable graphics and physics code.
What Should World Models Forget? Stratified Retention for Continual Adaptation
世界模型該遺忘什麼？以「分層保留」實現持續適應環境的能力
This paper argues that world models must not avoid all forgetting, proposing 'stratified retention' to distinguish permanent physical laws from dynamic, environment-specific facts that require timely revision.