Web-crawled datasets have enabled remarkable generalization capabilities in recent image-text models such as CLIP (Contrastive Language-Image pre-training) or Flamingo, but little is known about the dataset creation processes. In this work, we introduce a testbed of six publicly available data sources---YFCC, LAION, Conceptual Captions, WIT, RedCaps, Shutterstock---to investigate how pre-training distributions induce robustness in CLIP. We find that the performance of the pre-training data varies substantially across distribution shifts, with no single data source dominating. Moreover, we systematically study the interactions between these data sources and find that mixing multiple sources does not necessarily yield better models, but rather dilutes the robustness of the best individual data source. We complement our empirical findings with theoretical insights from a simple setting, where combining the training data also results in diluted robustness. In addition, our theoretical model provides a candidate explanation for the success of the CLIP-based data filtering technique recently employed in the LAION dataset. Overall our results demonstrate that simply gathering a large amount of data from the web is not the most effective way to build a pre-training dataset for robust generalization, necessitating further study into dataset design.

Vision-language pretrained models have seen remarkable success, but their application to safety-critical settings is limited by their lack of interpretability. To improve the interpretability of vision-language models such as CLIP, we propose a multi-modal information bottleneck (M2IB) approach that learns latent representations that compress irrelevant information while preserving relevant visual and textual features. We demonstrate how M2IB can be applied to attribution analysis of vision-language pretrained models, increasing attribution accuracy and improving the interpretability of such models when applied to safety-critical domains such as healthcare. Crucially, unlike commonly used unimodal attribution methods, M2IB does not require ground truth labels, making it possible to audit representations of vision-language pretrained models when multiple modalities but no ground-truth data is available. Using CLIP as an example, we demonstrate the effectiveness of M2IB attribution and show that it outperforms gradient-based, perturbation-based, and attention-based attribution methods both qualitatively and quantitatively.

Pre-trained vision-language models (e.g., CLIP) have shown promising zero-shot generalization in many downstream tasks with properly designed text prompts. Instead of relying on hand-engineered prompts, recent works learn prompts using the training data from downstream tasks. While effective, training on domain-specific data reduces a model's generalization capability to unseen new domains. In this work, we propose test-time prompt tuning (TPT), a method that can learn adaptive prompts on the fly with a single test sample. TPT optimizes the prompt by minimizing the entropy with confidence selection so that the model has consistent predictions across different augmented views of each test sample. In evaluating generalization to natural distribution shifts, TPT improves the zero-shot top-1 accuracy of CLIP by 3.6\% on average, surpassing previous prompt tuning approaches that require additional task-specific training data.

During the CL of VLMs, we need not only to prevent the catastrophic forgetting on incrementally learned knowledge but also to preserve the zero-shot ability of VLMs.

Advanced image editing techniques, particularly inpainting, are essential for seam-lessly removing unwanted elements while preserving visual integrity.

PInNED dataset has the potential to advance the field of visual navigation and Embodied AI.

Pre-trained vision-language models (e.g., CLIP) have shown powerful zero-shot transfer capabilities. But they still struggle with domain shifts and typically require labeled data to adapt to downstream tasks, which could be costly. In this work, we aim to leverage unlabeled data that naturally spans multiple domains to enhance the transferability of vision-language models. Under this unsupervised multi-domain setting, we have identified inherent model bias within CLIP, notably in its visual and text encoders. Specifically, we observe that CLIP's visual encoder tends to prioritize encoding domain over discriminative category information, meanwhile its text encoder exhibits a preference for domain-relevant classes.