{"about":{"site":"https://codewithpapers.app","non_affiliation":"Code with Papers and Syntology are not affiliated with, endorsed by, or sponsored by Papers with Code, Meta, or the pwc-archive mirror.","licence":"CC BY-SA 4.0","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","attribution":"https://codewithpapers.app/attribution","modified":"archive material modified by Syntology; see the attribution page"},"url":"/paper/clip-surgery-for-better-explainability-with","title":"A Closer Look at the Explainability of Contrastive Language-Image Pre-training","arxiv_id":"2304.05653","date":"2023-04-12","proceeding":null,"authors":["Yi Li","Hualiang Wang","Yiqun Duan","Jiheng Zhang","Xiaomeng Li"],"abstract":"Contrastive language-image pre-training (CLIP) is a powerful vision-language model that has shown great benefits for various tasks. However, we have identified some issues with its explainability, which undermine its credibility and limit the capacity for related tasks. Specifically, we find that CLIP tends to focus on background regions rather than foregrounds, with noisy activations at irrelevant positions on the visualization results. These phenomena conflict with conventional explainability methods based on the class attention map (CAM), where the raw model can highlight the local foreground regions using global supervision without alignment. To address these problems, we take a closer look at its architecture and features. Based on thorough analyses, we find the raw self-attentions link to inconsistent semantic regions, resulting in the opposite visualization. Besides, the noisy activations are owing to redundant features among categories. Building on these insights, we propose the CLIP Surgery for reliable CAM, a method that allows surgery-like modifications to the inference architecture and features, without further fine-tuning as classical CAM methods. This approach significantly improves the explainability of CLIP, surpassing existing methods by large margins. Besides, it enables multimodal visualization and extends the capacity of raw CLIP on open-vocabulary tasks without extra alignment. The code is available at https://github.com/xmed-lab/CLIP_Surgery.","url_abs":"https://arxiv.org/abs/2304.05653v2","url_pdf":"https://arxiv.org/pdf/2304.05653v2.pdf","source":{"archive":"pwc-archive (Hugging Face), CC BY-SA 4.0","snapshot":"2025-07-28","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","row_kind":"abstracts"},"code_links":[{"paper_slug":"clip-surgery-for-better-explainability-with","repo_url":"https://github.com/xmed-lab/clip_surgery","is_official":1,"mentioned_in_paper":1,"mentioned_in_github":1,"framework":"pytorch","reach":{"status":"ok"}},{"paper_slug":"clip-surgery-for-better-explainability-with","repo_url":"https://github.com/xmed-lab/clipn","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"pytorch","reach":{"status":"ok","spdx":"MIT"}}],"tasks":[{"task_slug":"interactive-segmentation","task_name":"Interactive Segmentation"},{"task_slug":"language-modelling","task_name":"Language Modelling"},{"task_slug":"open-vocabulary-semantic-segmentation","task_name":"Open Vocabulary Semantic Segmentation"},{"task_slug":"open-vocabulary-semantic-segmentation-1","task_name":"Open-Vocabulary Semantic Segmentation"},{"task_slug":"segmentation","task_name":"Segmentation"},{"task_slug":"semantic-segmentation","task_name":"Semantic Segmentation"},{"task_slug":"zero-shot-segmentation","task_name":"Zero Shot Segmentation"},{"task_slug":"zero-shot-learning","task_name":"Zero-Shot Learning"}],"methods":[{"method_slug":"clip","method_name":"CLIP"}],"datasets_introduced":[],"methods_introduced":[],"results":[{"leaderboard":"/sota/open-vocabulary-semantic-segmentation-on-coco","task":"Open Vocabulary Semantic Segmentation","dataset":"COCO-Stuff-171","model":"CLIP Surgery (original CLIP without any fine-tuning)","rank_in_archive_order":6,"of":7,"metrics":{"mIoU":"21.9"},"uses_additional_data":false},{"leaderboard":"/sota/open-vocabulary-semantic-segmentation-on","task":"Open Vocabulary Semantic Segmentation","dataset":"Cityscapes","model":"CLIP Surgery (CLIP without any fine-tuning)","rank_in_archive_order":4,"of":5,"metrics":{"mIoU":"31.4"},"uses_additional_data":false},{"leaderboard":"/sota/open-vocabulary-semantic-segmentation-on-1","task":"Open Vocabulary Semantic Segmentation","dataset":"PASCAL Context-59","model":"CLIP Surgery (original CLIP without any fine-tuning)","rank_in_archive_order":24,"of":24,"metrics":{"mIoU":"29.3"},"uses_additional_data":false},{"leaderboard":"/sota/zero-shot-segmentation-on-ade20k-training","task":"Zero Shot Segmentation","dataset":"ADE20K training-free zero-shot segmentation","model":"CLIPSurgery","rank_in_archive_order":4,"of":5,"metrics":{"mIoU":"12.9"},"uses_additional_data":false}],"syntology":{"syntology_url":"https://syntology.ai/paper/2304.05653","atlas_url":"https://app.syntology.ai/?focus=2304.05653","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}