Bias Is a Subspace, Not a Coordinate: A Geometric Rethinking of Post-hoc Debiasing in Vision-Language Models

AI summary

Overview

Research area: Fairness and bias mitigation in Vision-Language Models (VLMs), specifically post-hoc, training-free debiasing of frozen multimodal embeddings. The work sits at the intersection of computer vision, multimodal representation learning, and algorithmic fairness.

Technical level: Advanced. The paper assumes familiarity with linear probing, null-space projection, and embedding geometry, though its central argument (bias is distributed across directions rather than concentrated in coordinates) can be grasped without the linear algebra.

Scope: A systematic reproduction of the coordinate-wise debiasing method SFID (Jung et al.), a geometric re-analysis of how bias is represented in VLM embeddings, and a proposed subspace-projection alternative called SPD, evaluated on zero-shot classification, text-to-image retrieval, and text-to-image generation. Code is available at github.com/zhendashen896/SPD.

What This Paper Is About

Post-hoc debiasing methods for VLMs typically assume that a sensitive attribute such as gender, race, or age is encoded in a small set of individual embedding coordinates, and that replacing those coordinates with neutral values removes the bias. This paper tests that assumption directly on frozen CLIP-style embeddings and finds it fails on three counts, then proposes instead to treat bias as occupying a linear subspace and to project embeddings onto its orthogonal complement.

The goal is a training-free, interpretable debiasing procedure that achieves more complete bias removal than coordinate editing while preserving the semantic quality that makes VLM representations useful for downstream tasks.

Key Contributions

1. A systematic reproduction of SFID that falsifies its three core assumptions. The authors show that (A1) attributes are not encoded in disjoint dimensions, (A2) the important dimensions for a given attribute shift across datasets, and (A3) replacing the top-m coordinates leaves substantial linearly decodable attribute signal. Concretely, on FairFace CLIP ViT-B/32 embeddings (D = 512), the top-100 dimensions selected for age, gender, and race overlap heavily (A∩G = 31, G∩R = 37, A∩R = 20, and 11 shared by all three).

2. The geometric reframing that bias is subspace-structured rather than coordinate-sparse. The paper argues that discrete coordinate replacement is the wrong primitive because sensitive information is redundantly distributed across many correlated directions, and it recasts debiasing as an operation on learned directions rather than indices.

3. SPD (Subspace Projection Debiasing), a training-free framework with three stages. Bias subspace identification via Iterative Null-space Projection (INLP) using iterative logistic classifiers; projection of embeddings onto the null space of that subspace; and neutral reinjection of a low-confidence mean (estimated with a Random Forest, as in SFID) along the removed directions to keep embeddings on-manifold.

4. Extensive evaluation across three task families and three backbones. Zero-shot multi-class classification on FACET (52 occupation classes), text-to-image retrieval on Flickr30K (1,000 test images), and text-to-image generation with SDXL and CoDi (83 occupations from Bias-in-Bios, 10 images per prompt), on CLIP ResNet-50, CLIP ViT-B/32, and XVLM, against baselines including SFID, DeAR, CLIP-clip, and Prompt-Debias.

Main Findings

• Feature entanglement is severe: The top-m = 100 dimensions selected for age, gender, and race on FairFace CLIP ViT-B/32 overlap substantially, so debiasing one attribute necessarily distorts dimensions critical for the others — directly contradicting SFID's assumption A1.

• Dimensions drift across datasets: Using identical Random Forest procedures, the top-100 gender dimensions from FairFace and FACET overlap by 40 indices, versus a random-selection expectation of approximately 19.5. At m = 50 the overlap is 24 versus an expectation of roughly 4.9. The authors interpret this as evidence of genuine index shift rather than mere re-ranking, violating assumption A2.

• Coordinate replacement barely removes bias: Applying SFID with m = 100 to FairFace embeddings reduces logistic-regression probe accuracy by less than 1% relative to original embeddings across race (0.7086 vs 0.7144), gender (0.9449 vs 0.9466), and age (0.6023 vs 0.6023). These remain far above random baselines (50% for binary gender, 14.3% for 7-class race, 11.1% for 9-class age), even though 100 of 512 dimensions (19.5%) were replaced — violating assumption A3.

• Deeper projection removes more bias but costs semantics: In probe evaluation, SPD at r = 1 leaves target attribute accuracy nearly unchanged; r = 5 substantially reduces target accuracy while non-target attributes change by less than 1% in most cases; r = 10 reduces target accuracy further but also degrades non-target attributes. For example, replacing race with SPD at r = 10 drops race probe accuracy to 0.1913 but also drops gender to 0.6832 and age to 0.4745.

• Zero-shot classification fairness gains: SPD achieves the lowest ΔDP across all three backbones. CLIP ResNet-50: 9.55 vs baseline 11.08 (13.8% improvement) and SFID 9.63 (13.1%). CLIP ViT-B/32: 9.94 vs baseline 11.60 (14.3%) and SFID 10.15 (12.5%). XVLM: 9.85 vs baseline 11.72 (16.0%) and SFID 9.91 (15.4%). Accuracy changes are small in all cases.

• Retrieval skew reductions: SPD reduces Skew@100 on all backbones. CLIP ResNet-50: 0.1177 vs baseline 0.1883 (37.5% improvement) and SFID 0.1414 (24.9%). CLIP ViT-B/32: 0.0699 vs baseline 0.1721 (59.4%) and SFID 0.0744 (56.8%), with a simultaneous gain in R@1 (59.68 vs baseline 58.91). XVLM: 0.1859 vs baseline 0.2355 (21.1%) and SFID 0.2032 (13.7%).

• Generation quality and fairness: On SDXL, SPD yields an overall gender mismatch rate of 0.78 versus SFID's 0.84 and the baseline's 2.35, with composite mismatch 1.67 vs 4.42 baseline. DeAR is reported at 44.64 overall mismatch and 99.81 composite, and Prompt-D at 42.53 and 58.49. On CoDi, SPD reports 2.53 overall mismatch vs baseline 5.54, with Skew dropping from 84.94 to 81.20.

• Aggregate claim: The abstract reports an average improvement of 18.5% across four fairness metrics while maintaining minimal loss in task performance relative to the best debiasing baseline.

• Neutral reinjection helps semantics, not just fairness: The ablation shows reinjection slightly raises accuracy under similar ΔDP relative to projection alone (CLIP ResNet-50: 51.44 vs 50.16 accuracy at 9.55 vs 9.61 ΔDP; XVLM: 54.32 vs 53.72 at 9.85 vs 9.87).

• Threshold choice is stable: On XVLM, τ = 0.7 gives 54.32 accuracy and 9.85 ΔDP; τ = 0.6 gives 53.81 and 9.84; τ = 0.8 gives 54.34 and 9.94; τ = 0.9 gives 54.63 and 10.26. The authors adopt τ = 0.7 throughout.

Methodology in Plain English

The researchers begin by borrowing an existing post-hoc method, SFID, and stress-testing its implicit assumptions. SFID trains a Random Forest (100 trees) to rank embedding dimensions by Gini importance, picks the top-m, and at inference replaces those coordinates with mean values taken from low-confidence samples (those below a confidence threshold τ = 0.7). The authors check three things: whether the top dimensions for different attributes overlap, whether the top dimensions for the same attribute agree across two datasets (FairFace and FACET), and whether replacing the top dimensions actually stops a simple linear classifier from recovering the attribute.

All three checks fail, so they design a replacement. Instead of picking coordinates, they run Iterative Null-space Projection: train a linear classifier to predict the attribute from the embeddings, extract the directions its weights span (via QR decomposition), project the embeddings onto the null space of those directions to erase them, and repeat for T iterations or until the classifier drops to chance. A limited number of directions r is removed in practice to trade bias suppression against utility.

Because attributes and useful semantics are partially entangled, pure projection can damage task-relevant information. To counter this, they add back a single neutral component: the mean embedding of low-confidence samples, projected into the removed subspace. That term is identical for every sample, so it cannot reintroduce attribute-discriminative variability — it only recenters the embeddings. The final representation is the projected embedding plus this constant reinjection.

Evaluation uses the same setup as SFID: FACET for zero-shot classification with ΔDP as the fairness metric, Flickr30K for retrieval with Skew@100, and SDXL and CoDi for generation with mismatch rates and Skew. SPD uses r = 5 for all downstream tasks, with 10 independent runs for retrieval and generation and 1,000 bootstrap samples for classification.

Why This Matters

Impact on research: The paper challenges a widely used and intuitively appealing debiasing primitive — edit the offending coordinates — and supplies diagnostic evidence that it underperforms for structural reasons, not just hyperparameter reasons. That reframing matters beyond VLMs, since coordinate-selection debiasing appears in other modalities and in tabular fairness work. The paper also provides a second, more complete result than the linear-probe tables: SPD still degrades non-target probe accuracy at r = 10, so the trade-off is real rather than eliminable.

Real-world applications:

• Image search and stock-photo retrieval systems, where a neutral query such as "a person in a suit" should not return a demographically skewed result set.
• Text-to-image content generation platforms, where profession prompts should not default to one gender, and where gender-specified prompts should be honored reliably.
• Hiring, advertising, and recommendation pipelines that rely on frozen vision-language embeddings to match people or products to concepts.
• Content moderation and auditing tools that need to measure and reduce demographic skew in multimodal systems without retraining expensive foundation models.

Industry relevance: SPD is training-free and closed-form once the lightweight linear probes are fit, which makes it deployable on frozen production models where fine-tuning is infeasible or costly. The authors report it applies to any frozen encoder or decoder, and the accompanying code release lowers the barrier to adoption. The fact that the paper's strongest results come from reproducing and outperforming a published method gives practitioners a concrete reason to switch rather than merely add another baseline.

Future Directions

• Choosing projection depth automatically. The paper treats r as a controllable fairness–utility knob and shows r = 10 harms non-target attributes, but it does not report a principled method for selecting r per attribute, dataset, or backbone; r = 5 is fixed for all downstream tasks without a stated selection procedure. Deriving an adaptive stopping criterion (for example, from the INLP convergence condition of reaching 1/C) is an open step.

• Nonlinear bias structure. INLP extracts linearly decodable directions. The paper does not test whether nonlinear probes can recover attribute information from the projected embeddings, which would bound how much of the bias structure SPD can reach.

• More attributes and intersecting identities. The diagnostic study on FairFace covers three attributes (age, gender, race), and the downstream generation study covers gender only, with 83 occupations from Bias-in-Bios. The paper discusses the multi-attribute case as motivation but does not report results for intersecting identities.

• Broader backbone and modality coverage. Evaluation covers CLIP ResNet-50, CLIP ViT-B/32, XVLM, SDXL, and CoDi. Whether the learned subspace remains stable across larger open VLMs, video models, or audio-visual encoders is not reported.

Target Audience

This paper is most valuable to fairness and responsible-AI researchers working on multimodal systems, and to ML engineers who need to reduce demographic bias in deployed vision-language models without fine-tuning. It is also a useful case study for practitioners of interpretability and representation geometry, because it demonstrates how a "which dimensions matter" analysis can mislead when information is distributed. Readers will benefit most if they are comfortable with linear probes and projection operators; those unfamiliar with null-space projection can still follow the core argument, but the methodology sections will require the linear algebra.

Authors’ abstract

Vision-Language Models (VLMs) have become indispensable for multimodal reasoning, yet their representations often encode and amplify demographic biases, resulting in biased associations and misaligned predictions in downstream tasks. Such behavior undermines fairness and distorts the intended alignment between vision and language. Recent post-hoc approaches attempt to mitigate bias by replacing the most attribute-correlated embedding coordinates with neutral values. However, our systematic analysis reveals three critical limitations of this coordinate-wise approach: feature entanglement, poor cross-dataset generalization, and incomplete bias removal. We find that bias is not localized to a few coordinates but is instead distributed across a few linear subspaces. To address these limitations, we propose $\textbf{S}$ubspace $\textbf{P}$rojection $\textbf{D}$ebiasing ($\textbf{SPD}$), a geometrically principled framework that identifies and removes the entire subspace of linearly decodable bias while reinserting a neutral mean component to preserve semantic fidelity. Extensive experiments across zero-shot classification, text-to-image retrieval, and image generation validate the effectiveness of SPD: our method achieves more robust debiasing with an average improvement of $18.5\%$ across four fairness metrics, while maintaining minimal loss in task performance compared to the best debiasing baseline.