A classifier may depend on incidental features stemming from a strong correlation between the feature and the classification target in the training dataset. Recently, Last Layer Retraining (LLR) with group-balanced datasets is known to be efficient in mitigating the spurious correlation of classifiers. However, the acquisition of group-balanced datasets is costly, which hinders the applicability of the LLR method. In this work, we propose to perform LLR based on text datasets built with large language models for a general image classifier. We demonstrate that text can be a proxy for its corresponding image beyond the image-text joint embedding space, such as CLIP. Based on this, we use generated texts to train the final layer in the embedding space of the arbitrary image classifier. In addition, we propose a method of filtering the generated words to get rid of noisy, imprecise words, which reduces the effort of inspecting each word. We dub these procedures as TLDR (\textbf{T}ext-based \textbf{L}ast layer retraining for \textbf{D}ebiasing image classifie\textbf{R}s) and show our method achieves the performance that is comparable to those of the LLR methods that also utilize group-balanced image dataset for retraining. Furthermore, TLDR outperforms other baselines that involve training the last linear layer without a group annotated dataset.
Neural network interpretation methods, particularly feature attribution methods, are known to be fragile with respect to adversarial input perturbations. To address this, several methods for enhancing the local smoothness of the gradient while training have been proposed for attaining \textit{robust} feature attributions. However, the lack of considering the normalization of the attributions, which is essential in their visualizations, has been an obstacle to understanding and improving the robustness of feature attribution methods. In this paper, we provide new insights by taking such normalization into account. First, we show that for every non-negative homogeneous neural network, a naive $\ell_2$-robust criterion for gradients is \textit{not} normalization invariant, which means that two functions with the same normalized gradient can have different values. Second, we formulate a normalization invariant cosine distance-based criterion and derive its upper bound, which gives insight for why simply minimizing the Hessian norm at the input, as has been done in previous work, is not sufficient for attaining robust feature attribution. Finally, we propose to combine both $\ell_2$ and cosine distance-based criteria as regularization terms to leverage the advantages of both in aligning the local gradient. As a result, we experimentally show that models trained with our method produce much more robust interpretations on CIFAR-10 and ImageNet-100 without significantly hurting the accuracy, compared to the recent baselines. To the best of our knowledge, this is the first work to verify the robustness of interpretation on a larger-scale dataset beyond CIFAR-10, thanks to the computational efficiency of our method.