Recent advances in image classification have achieved impressive performance by leveraging large-scale datasets. However, it has been repeatedly pointed out that these models often learn spurious correlations inherent in the training data. In particul...
Recent advances in image classification have achieved impressive performance by leveraging large-scale datasets. However, it has been repeatedly pointed out that these models often learn spurious correlations inherent in the training data. In particular, when the prediction is made based on incidental visual attributes such as background, lighting, or gender rather than core visual features, model performance may degrade severely, especially for minority groups or in out-of-distribution settings.
To address this issue, recent studies have proposed Last-Layer Retraining (LLR)-based debiasing techniques. LLR retrains only the final linear classification layer using a small, group-balanced dataset, allowing effective debiasing without re-training the entire model. A representative example, DFR, demonstrated strong performance by using a small image dataset balanced across combinations of class labels and spurious attributes (i.e., “groups”).
However, these approaches face practical limitations. In order to construct a group-balanced dataset, one must annotate not only class labels but also spurious attributes for each image—a task known as group annotation. This process is costly, labor-intensive, and often infeasible in practice due to the rarity of minority group samples or the difficulty in clearly defining spurious features.
To overcome these limitations, we propose TLDR (Text-based Last-layer Retraining for Debiasing Image Classifiers), a novel approach that mitigates model bias without requiring additional image collection or group annotation. Instead of relying on images, TLDR uses large language models (LLMs) to generate a diverse set of textual expressions corresponding to class labels and spurious attributes. These texts are embedded into CLIP’s multimodal space and then projected into the feature space of a given image classifier through a closed-form linear projector.
Importantly, our method introduces a theoretical analysis of the modality gap between image and text embeddings and designs a linear projector that considers this gap to ensure effective alignment between modalities. We also implement a Validation of Embedding Alignment (VEA) procedure that filters generated texts based on their semantic alignment within both the CLIP and classifier embedding spaces, thereby improving the quality and robustness of the retraining data.
We validate our method on multiple benchmark datasets with inherent biases—Waterbirds, CelebA, and SpuCoAnimals. TLDR achieves performance comparable to or better than existing LLR-based methods, without any group annotation. Moreover, TLDR can be applied post hoc to already trained models, making it highly practical and scalable.
This study presents a novel framework that effectively debiases image classifiers using only text, without relying on costly annotated images. TLDR offers a promising direction for building fairer and more generalizable AI systems, especially in real-world domains such as medical imaging, satellite vision, and face recognition, where group annotations are difficult to obtain.