Introduction
In modern computer vision, transfer learning is the standard paradigm. Developers typically utilize models pre-trained on massive datasets like ImageNet—such as VGG16, ResNet50, or InceptionV3—and fine-tune them for specific downstream tasks. However, this practice carries the risk of inheriting societal biases embedded in the original training data. While previous research often examines sensitive variables like race, age, or gender in isolation, this study adopts an intersectional approach. By investigating how overlapping identities (e.g., "Black, 20–29, Male") are represented internally, the authors uncover how deep learning architectures encode and potentially propagate complex biases.
Methodology
The study employs three primary techniques to interrogate the internal representations of pre-trained models:
- Linear Classifier Probes: Simple linear layers are attached to intermediate model layers and trained to identify 126 intersectional groups. High accuracy in these probes indicates that the model's internal features are heavily influenced by demographic attributes.
- Neuron Activation Profiles (NAPs): By averaging activations for specific intersectional groups and subtracting the global average, the researchers isolate the visual features the model associates with specific demographics.
- Topographic Activation Maps: Using UMAP dimensionality reduction, NAPs are visualized as 2D heatmaps, providing a qualitative view of how the model distinguishes between different ages, races, and genders.
The evaluation utilizes the FairFace dataset, comprising 86,744 images balanced across seven racial groups, nine age groups, and two genders.
Key Insights: The Architecture of Bias
Age as the Dominant Signal
Across all tested models, age emerged as the most distinguishable feature. The researchers suggest that CNNs are naturally adept at detecting the highly structured visual patterns associated with aging, such as skin texture (wrinkles) and changes in facial geometry.
The "Middle Layer" Peak
Bias is not uniformly distributed throughout a neural network. Probing reveals that demographic representations are weakest in the initial and final layers, reaching a peak in the middle layers. This suggests that as the model constructs increasingly complex features, it inadvertently builds representations that are highly specific to demographic groups.
Ethnicity and Gender Nuances
The models frequently exhibited "ethnicity confusion," particularly among groups like Latino/Hispanic, White, Indian, and Middle Eastern. This indicates a blurry internal "map" of race that relies on shared visual cues. Gender differences were most pronounced in middle-aged groups (10–59), likely due to the prevalence of gendered social cues like hairstyles and makeup in those age ranges.
Results and Error Analysis
While linear probes achieved near 100% training accuracy in middle layers for models like ResNet50, validation accuracy remained significantly lower. This suggests that while the models encode bias, it is often tied to specific training samples rather than a generalized linear rule.
Specific misclassifications often involved age-related errors, such as mistaking individuals in their 60s for those over 70. Additionally, certain ethnicities were more prone to being mislabeled when the subject was older.
Limitations and Future Directions
The study identifies several areas for further exploration:
- Domain Specificity: The current findings are limited to facial recognition; it is unclear how these biases manifest in general object detection.
- Probe Complexity: Linear probes may underestimate bias; non-linear probes could reveal even deeper demographic encoding.
- Architecture Scope: Future work will investigate whether Vision Transformers (ViTs) inherit similar intersectional biases compared to the CNNs studied here.
By highlighting these internal biases, the research underscores the importance of auditing pre-trained models before they are deployed in sensitive real-world applications.