Seokhyeon Jeong

Sunghwan Joo, Seokhyeon Jeong, Juyeon Heo et al. · cambridge

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.

Juhyeon Park, Seokhyeon Jeong, Taesup Moon

An image 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 shown to be efficient in mitigating the spurious correlation of classifiers. However, the acquisition of image-based group-balanced datasets is costly, which hinders the general applicability of the LLR method. In this work, we propose to perform LLR based on text datasets built with large language models to debias a general image classifier. To that end, we demonstrate that text can generally be a proxy for its corresponding image beyond the image-text joint embedding space, which is achieved with a linear projector that ensures orthogonality between its weight and the modality gap of the joint embedding space. In addition, we propose a systematic validation procedure that checks whether the generated words are compatible with the embedding space of CLIP and the image classifier, which is shown to be effective for improving debiasing performance. We dub these procedures as TLDR (Text-based Last layer retraining for Debiasing image classifieRs) and show our method achieves the performance that is competitive with the LLR methods that require group-balanced image dataset for retraining. Furthermore, TLDR outperforms other baselines that involve training the last layer without any group annotated dataset. Codes: https://github.com/beotborry/TLDR

Jaeseok Byun, Seokhyeon Jeong, Wonjae Kim et al.

Composed Image Retrieval (CIR) aims to retrieve a target image based on a reference image and conditioning text, enabling controllable image searches. The mainstream Zero-Shot (ZS) CIR methods bypass the need for expensive training CIR triplets by projecting image embeddings into the text token embedding space, forming a composed query for retrieval. However, we highlight an inherent limitation in these projection-based CIR: a task discrepancy of text encoders between the original pre-training task of the encoders (text $\leftrightarrow$ image) and the target CIR task (image + text $\leftrightarrow$ image), which potentially negatively impacts CIR performance. To reduce such a discrepancy, a naive solution would be to train both image and text encoders with CIR triplets in a supervised manner. Instead, we introduce Reducing Task Discrepancy of Text Encoders (RTD), an efficient text-only post-hoc framework that complements projection-based CIR methods. We devise a novel target-anchored text contrastive learning designed to enhance the capability of the text encoder for CIR. We also propose two key enhancements: (1) a hard negative-based refined batch sampling strategy and (2) a refined concatenation scheme to further mitigate training-inference discrepancy. Integrating RTD into state-of-the-art projection-based methods achieves performance comparable to, or even surpassing, resource-intensive state-of-the-art synthetic CIR triplet-based approaches only with 23 minutes of additional training on 4 A100 GPUs (up to $100\times$ faster in training). Our code will be available upon acceptance.