Towards causally linking architectural parametrizations to algorithmic bias in neural networks

Hao Liang, Josue Ortega Caro, Vikram Maheshri, Ankit B. Patel, Guha Balakrishnan

Introduction

Algorithmic bias is frequently attributed to "bad data"—unbalanced or non-representative training sets. However, this research demonstrates that the mathematical structure of a neural network itself can be a primary driver of bias. By employing a causal framework, the authors show that changing a single architectural hyperparameter—the convolutional kernel size—alters how a model perceives image frequencies, leading to significant performance disparities across race and gender when the model encounters out-of-distribution (OOD) noise.

Convolutional Neural Networks (CNNs) possess an implicit bias regarding spatial frequencies. Smaller kernels are typically more sensitive to high-frequency details, such as sharp edges and fine textures. This study investigates whether these architectural choices causally lead to algorithmic bias, particularly when a model relies on high-frequency features that vary across demographic groups.

Methodology

The researchers proposed a three-step experimental causal framework to isolate the effects of architecture on bias.

Overview of the causal framework involving architectural interventions, OOD perturbations, and causal analysis.
  1. Architectural Intervention: Multiple versions of standard CNNs (ResNet-34, VGG-16, DenseNet-121) were trained. These models were identical except for the First Layer Kernel Size (FLKS), which was varied from $3 \times 3$ to $11 \times 11$.
  2. OOD Perturbation: To expose implicit biases, the models were tested using Adversarial Attacks (CW and FGSM) and Fourier Energy Injection (adding noise to specific frequency bands). These perturbations force the model to reveal which image features it prioritizes for classification.
  3. Causal Analysis: Multivariable linear regression was used to model the relationship between kernel size and performance across subgroups. Regression coefficients ($\beta$) quantified the causal effect of kernel size on generalization bias for specific races and genders.

Datasets

The study utilized two major face analysis benchmarks: FairFace, which is balanced across seven race groups and two gender groups, and UTKFace, containing over 20,000 images. Inverse sampling was used on UTKFace to mitigate existing sampling biases, ensuring that any observed bias stemmed from the architecture rather than data imbalance.

A central finding is that different demographic groups possess distinct "frequency signatures" in the context of CNN feature extraction. Models consistently focused less on high-frequency details for Black and Male subgroups compared to others.

Modifying the kernel size—even solely in the first layer—significantly shifts the frequency content of learned features. As the kernel size increases, the model becomes more robust to high-frequency noise but potentially more reliant on coarser features that may not generalize equally across groups. This aligns with the Fourier Uncertainty Principle, where smaller kernels result in features spanning a greater range in the frequency domain, making them "high-frequency hunters" that are more susceptible to sharp, pixel-level noise.

Results and Causal Significance

The experiments revealed several critical trends regarding how architecture dictates robustness and fairness:

  • Spectral Concentration: As FLKS increased from 3 to 11, the perturbation spectra became more concentrated at low frequencies.
  • Disparate Robustness: In adversarial tests, the "perturbation distance" (the amount of noise required to fool the model) was significantly lower for the Black subgroup. This indicates that models are inherently "brittler" for certain demographics based on their architectural tuning.
  • Causal Impact: Regression analysis showed that the impact of kernel size on performance was most significant for Black and Indian subgroups. For instance, the causal effect ($\beta$ values) for Black and Indian females was often ~100% higher than for East Asian females.
Accuracy vs. frequency of injected noise for different First Layer Kernel Sizes (FLKS) across Black and Indian demographics.

These trends remained consistent across ResNet, VGG, and DenseNet architectures, suggesting that this is a fundamental property of CNN design rather than an artifact of a specific model.

Conclusion and Future Work

This research proves that bias can emerge purely from the interaction between a model's functional form and natural variations in human features, even with perfectly balanced data. However, the study's scope was limited to face analysis and a specific hyperparameter (kernel size).

Future work should explore whether these frequency-bias links exist in other domains like medical imaging or autonomous driving. Additionally, investigating other architectural components—such as activation functions, depth, or attention mechanisms—could provide a more holistic understanding of architectural bias. Finally, testing against real-world OOD artifacts like motion blur or sensor noise will be essential for translating these findings into more robust and fair AI systems.

Summary updated

BibTeX
@inproceedings{liang2024towards,
  author = {Hao Liang and Josue Ortega Caro and Vikram Maheshri and Ankit B. Patel and Guha Balakrishnan},
  title = {{Towards causally linking architectural parametrizations to algorithmic bias in neural networks}},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision},
  year = {2024},
  doi = {10.1109/wacv57701.2024.00464},
}
Figure