Existing Approaches
- Traditional Methods: Approaches, like resampling and reweighing are characterized by their interpretability and the advantage of being training-free, making them accessible and easy to implement in certain contexts. However, they suffer from significant limitations such as a lack of scalability and inconsistency in results. Moreover, these methods often require manual bias annotations and demonstrate poor generalization across different datasets or domains, which severely restricts their applicability in real-world, large-scale systems.
- Supervised Methods: Supervised approaches offer more targeted and reliable debiasing strategies. These methods allow for fine-grained performance control and typically produce consistent outcomes due to their reliance on labeled data. However, this dependency introduces the drawback of requiring bias annotations, which can be labor-intensive, subjective, and difficult to obtain at scale. Additionally, despite their strengths, supervised methods often face challenges in generalizing across unseen data or domains, limiting their effectiveness in dynamic or evolving environments.
- Unsupervised Methods: Unsupervised approaches present a compelling alternative by removing the dependency on bias labels, thus enabling more scalable and generalizable solutions. These methods are particularly well-suited for diverse and complex datasets, offering improved control over bias mitigation without the manual overhead associated with annotation. They excel in adapting to new data distributions, which is crucial for deployment in real-world applications. However, a notable trade-off is the increased model complexity, which can make these systems harder to interpret and maintain.
Fill-Up: Balancing Long-Tailed Data with Generative Models
This paper addresses the challenge of non-uniform data distributions in image classification by leveraging generative models to augment underrepresented classes, thereby balancing the dataset and mitigating bias.
The process begins by analyzing the class distribution within the dataset to identify minority classes, which is essential since the approach relies on explicit bias attributes or class labels. To generate additional samples for these underrepresented classes, the authors employ textual inversion-a technique that enables pre-trained text-to-image diffusion models to learn new visual concepts from a small number of images by associating them with unique text tokens. This allows the model to synthesize diverse, realistic images for minority classes, effectively “filling in” gaps in the data distribution. The augmented dataset, comprising both real and synthetic images, is then used to train an image classifier, exposing it to a more balanced representation of all classes and thereby eliminating bias.
Debiasing Classifiers by Amplifying Bias with Latent Diffusion and Large Language Models
The DiffuBias framework presents a novel approach to mitigating bias in classifiers by generating and leveraging bias-conflict samples through a combination of pre-trained latent diffusion models and large language models (LLMs).
The methodology has four key steps. First, it trains a classifier on the biased dataset using Generalized Cross Entropy (GCE) loss to amplify bias patterns and identify bias-conflict samples-those with the highest loss values. Next, these bias-conflict samples are captioned using an LLM (Llava), and irrelevant captions are filtered out to ensure they contain class-related keywords. Third, new bias-conflict samples are generated with Stable Diffusion, guided by the filtered captions. Finally, the original dataset is augmented with these synthetic samples, and a new classifier is trained on this balanced set using standard cross-entropy loss to improve fairness and reduce bias.
However, this approach has several limitations: its effectiveness depends on the quality of the bias-conflict sample identification and the semantic accuracy of the generated captions; the reliance on generative models introduces risks of creating unrealistic or low-fidelity synthetic samples, which may not fully capture the complexity of real-world bias
References
Ko, D., Lee, D., Park, N., Shim, W., & Kim, J. (2024). Debiasing classifiers by amplifying bias with latent diffusion and large language models. arXiv. https://doi.org/10.48550/arXiv.2411.16079
Shin, J., Kang, M., & Park, J. (2023). Fill-Up: Balancing Long-Tailed Data with Generative Models. arXiv. https://doi.org/10.48550/arXiv.2306.07200