When we train a machine learning model to be "fair," we are usually forcing it to ignore a specific demographic shortcut present in a specific dataset. For example, we might penalize a facial recognition model so it stops using skin tone as a proxy for predicting a target label.
However, when we move that model—or its training methodology—to a new domain, task, or data modality, the fairness guarantees almost always fall apart. A fairness intervention that works beautifully on continuous image pixels fails completely when applied to discrete tabular data or sequential text. Even within the same modality, a debiasing technique designed for dermatological photos cannot easily transfer to ultrasound scans or MRIs.
The core challenge of modern algorithmic fairness research is generalization. To build truly equitable AI, we must understand why fairness mechanisms are so tightly bound to their original domains, and design new frameworks that can transfer across modalities and tasks.
The Geometry of Modalities: Why Vision Fixes Fail on Tabular Data
Many of the most successful fairness interventions in computer vision rely on manipulating continuous latent spaces. For instance, Hyungjun Joo et al. utilize invertible disentanglement modules to separate sensitive demographic attributes from core task features in a model's latent space (Constructing Fair Latent Space for Intersection of Fairness and Explainability). Similarly, Amirarsalan Rajabi et al. use a reconstruction-based U-Net architecture to subtly scrub bias directly from image datasets before a classifier ever sees them (Through a fair looking-glass: mitigating bias in image datasets).
These methods work because image data is continuous and highly redundant. A neural network can smoothly perturb a latent vector or alter a few pixels to hide a protected attribute (like gender or age) without destroying the semantic features needed to predict the target label.
Image Modality (Continuous, Redundant)
[ Pixels / Latent Vectors ] ---> Smooth Perturbation ---> Bias Scrubbed (Features Intact)
Tabular Modality (Discrete, Sparse)
[ Zip Code / Age / Income ] ---> Discrete Shift ---> Semantic Meaning Broken
But as the authors of these studies note, this geometric intuition fails when applied to non-visual domains like finance or healthcare tabular data. Tabular data consists of discrete, sparse, and highly correlated features (e.g., zip codes, income brackets, or laboratory values). You cannot "smoothly perturb" a discrete category without fundamentally changing its meaning or breaking the logical correlations the model relies on for accuracy.
This modality gap extends down to the optimization algorithms themselves. Mojtaba Kolahdouzi et al. demonstrated that adaptive optimizers like Adam and RMSProp inherently converge to fairer models than standard Stochastic Gradient Descent (SGD) when trained on imbalanced computer vision datasets (Some Optimizers are More Equal: Understanding the Role of Optimizers in Group Fairness). However, because the loss surfaces of large language models (LLMs) and tabular networks behave with entirely different optimization dynamics, these natural fairness advantages do not automatically cross modality boundaries.
The Intra-Modality Barrier: Domain Shifts in Medical Imaging
Even when we restrict our focus to a single modality like medical imaging, fairness mechanisms remain stubborn non-generalizers. The physical properties of how data is captured create unique demographic "shortcuts" for different clinical tasks.
Consider two different approaches to medical vision fairness:
- Ultrasound: Zikang Xu et al. developed APPLE, a framework that applies adversarial perturbations to latent embeddings to eliminate sex and age bias in ultrasound diagnostics (Fair ultrasound diagnosis via adversarial protected attribute aware perturbations on latent embeddings).
- Dermatology: Another team led by Zikang Xu et al. introduced FairAdaBN, which uses adaptive batch normalization layers to specialize a model's internal scaling parameters for different skin tones (FairAdaBN: Mitigating unfairness with adaptive batch normalization and its application to dermatological disease classification).
Why can't we just use the same tool for both? Because the underlying manifestation of bias is completely different.
In dermatology, bias is primarily a surface-level color and lighting issue. As William Thong et al. point out, traditional fairness metrics fail here because they treat skin color as a simple light-to-dark scale, ignoring the critical dimension of hue (red-to-yellow variations) which heavily skews diagnostic performance (Beyond Skin Tone: A Multidimensional Measure of Apparent Skin Color). Adaptive batch normalization works well here because it can normalize these surface-level color distributions across subgroups.
In contrast, an ultrasound image contains no skin tone or hue; its biases stem from acoustic impedance variations caused by differing tissue densities, body mass indexes, or anatomical structures across age and sex groups. An intervention designed to normalize surface color distributions is useless against deep structural variations.
When a model experiences a technological domain shift—such as moving from retinal fundus cameras to CT scanners—the relationship between demographic attributes and image features changes entirely. Yu Tian et al. addressed this by developing Fair Identity Attention (FIA) to explicitly decouple identity features from domain-specific imaging artifacts (FairDomain: Achieving Fairness in Cross-Domain Medical Image Segmentation and Classification). They proved that unless a mechanism explicitly accounts for how domain shifts alter demographic signatures, fairness metrics will degrade rapidly upon deployment.
Decoupling Concepts and Parameters for Generalization
To overcome these boundaries, recent work focuses on a powerful intuition: decoupling the "what" from the "who." If a model can cleanly isolate task-specific concepts from demographic backgrounds, its fairness properties become significantly more transferable.
One way to achieve this is through generative data augmentation. Janet Wang et al. leveraged diffusion models to extract abstract "lesion concepts" from data dominated by majority skin types, and then translated those exact pathological concepts onto synthetic images of underrepresented minority skin tones (From Majority to Minority: A Diffusion-based Augmentation for Underrepresented Groups in Skin Lesion Analysis). By mathematically separating the disease concept from the demographic background, they allowed downstream classifiers to learn invariant, fair representations.
[ Majority Skin + Lesion ] ---> Extract "Lesion Concept" (The What)
|
v
[ Minority Skin (The Who) ] ---> Diffusion Translation ---> [ Fair Synthetic Training Data ]
On the architectural side, researchers are isolating fairness at the parameter level. Zengqun Zhao et al. introduced AIM-Fair, which uses selective fine-tuning on synthetic data to update only the specific model parameters responsible for demographic fairness, leaving the core parameters that handle real-world visual features untouched (AIM-Fair: Advancing Algorithmic Fairness via Selectively Fine-Tuning Biased Models with Contextual Synthetic Data).
Taking this structural approach to its logical conclusion, Yi Sheng developed BiaslessNAS, an automated machine learning framework that co-optimizes data sampling, training algorithms, and neural network architectures simultaneously (Data-Algorithm-Architecture Co-Optimization for Fair Neural Networks on Skin Lesion Dataset). This work proves that fairness is not just a data or algorithmic problem—the very wiring of a neural network determines its propensity to latch onto demographic shortcuts.
Ultimately, the field is moving away from ad-hoc, single-dataset fairness patches. By treating fairness as a problem of domain invariance and structural decoupling, researchers are laying the groundwork for models that remain equitable, no matter the task or modality they encounter.
Key Foundational Work
Fairness for Image Generation with Uncertain Sensitive Attributes (2021) — contributed 1 concept elements to this direction.
Adversarial Removal of Demographic Attributes Revisited (2019) — contributed 1 concept elements to this direction.
On the Fairness of Swarm Learning in Skin Lesion Classification (2021) — contributed 1 concept elements to this direction.