[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"detail-sidebar-cat-0-en-105":3,"doc-seo-661240-105":59,"doc-detail-661240-en":130},{"code":4,"msg":5,"data":6},0,"success",[7,13,18,23,28,33,38,43,48,51,55],{"id":8,"doc_module":4,"doc_module_name":9,"category_name":10,"show_sort_weight":11,"slug":12},1,"Document","Story & Novel",90,"story-novel",{"id":14,"doc_module":4,"doc_module_name":9,"category_name":15,"show_sort_weight":16,"slug":17},2,"Literature",80,"literature",{"id":19,"doc_module":4,"doc_module_name":9,"category_name":20,"show_sort_weight":21,"slug":22},4,"Exam",70,"exam",{"id":24,"doc_module":4,"doc_module_name":9,"category_name":25,"show_sort_weight":26,"slug":27},5,"Comic",60,"comic",{"id":29,"doc_module":4,"doc_module_name":9,"category_name":30,"show_sort_weight":31,"slug":32},6,"Technology",50,"technology",{"id":34,"doc_module":4,"doc_module_name":9,"category_name":35,"show_sort_weight":36,"slug":37},7,"Healthcare",40,"healthcare",{"id":39,"doc_module":4,"doc_module_name":9,"category_name":40,"show_sort_weight":41,"slug":42},8,"Research & Report",30,"research-report",{"id":44,"doc_module":4,"doc_module_name":9,"category_name":45,"show_sort_weight":46,"slug":47},9,"Religion & Spirituality",20,"religion-spirituality",{"id":46,"doc_module":4,"doc_module_name":9,"category_name":49,"show_sort_weight":46,"slug":50},"World Cup","world-cup",{"id":52,"doc_module":4,"doc_module_name":9,"category_name":53,"show_sort_weight":52,"slug":54},10,"Lifestyle","lifestyle",{"id":56,"doc_module":4,"doc_module_name":9,"category_name":57,"show_sort_weight":24,"slug":58},19,"General","general",{"code":4,"msg":60,"data":61},"ok",{"site_id":62,"language":63,"slug":64,"title":65,"keywords":66,"description":67,"schema_data":68,"social_meta":123,"head_meta":125,"extra_data":127,"updated_unix":129},105,"en","vision-language-dataset-distillation","Vision-Language Dataset Distillation","","Dataset distillation reduces large-scale datasets to smaller synthetic subsets while preserving enough information to train a new model from scratch quickly. Existing approaches largely target image classification benchmarks with discrete labels, but modern large vision-language datasets lack a natural class structure. This work introduces the first vision-language dataset distillation method based on trajectory matching, jointly distilling image-text pairs in a contrastive setting. Low-Rank Adaptation (LoRA) further improves matching efficiency for complex vision-language models, yielding strong gains on Flickr30K and COCO retrieval tasks.",{"@graph":69,"@context":122},[70,84,105],{"@type":71,"itemListElement":72},"BreadcrumbList",[73,77,79,82],{"item":74,"name":75,"@type":76,"position":8},"https://docshare.wps.com","Home","ListItem",{"item":78,"name":9,"@type":76,"position":14},"https://docshare.wps.com/document/",{"item":80,"name":40,"@type":76,"position":81},"https://docshare.wps.com/document/research-report/",3,{"item":83,"name":65,"@type":76,"position":19},"https://docshare.wps.com/document/vision-language-dataset-distillation/661240/",{"url":83,"name":65,"@type":85,"image":86,"author":91,"headline":65,"publisher":94,"fileFormat":97,"inLanguage":63,"description":67,"dateModified":98,"datePublished":99,"encodingFormat":97,"isAccessibleForFree":100,"interactionStatistic":101},"DigitalDocument",{"url":87,"@type":88,"width":89,"height":90},"https://docshare.wps.com/thumbnails/vision-language-dataset-distillation/661240.png","ImageObject",300,407,{"name":92,"@type":93},"นรินทร์","Person",{"url":74,"name":95,"@type":96},"DocShare","Organization","application/pdf","2026-10-29","2026-10-28",true,{"@type":102,"interactionType":103,"userInteractionCount":19},"InteractionCounter",{"@type":104},"ViewAction",{"@type":106,"mainEntity":107},"FAQPage",[108,114,118],{"name":109,"@type":110,"acceptedAnswer":111},"What problem does dataset distillation address in large-scale learning?","Question",{"text":112,"@type":113},"Dataset distillation creates concise summaries that preserve critical information from a large dataset, reducing reliance on irrelevant data.","Answer",{"name":115,"@type":110,"acceptedAnswer":116},"What is the main challenge for vision-language dataset distillation compared with prior work?",{"text":117,"@type":113},"Vision-language datasets do not provide a discrete set of classes, so distillation cannot rely on class-specific boundaries.",{"name":119,"@type":110,"acceptedAnswer":120},"How does the proposed method distill vision-language data without discrete classes?",{"text":121,"@type":113},"It jointly distills image-text pairs using a contrastive formulation and matches the by-products of training, leveraging LoRA-based matching for efficiency.","https://schema.org",{"og:url":83,"og:type":124,"og:title":65,"og:site_name":95,"og:description":67},"article",{"robots":126,"canonical":83},"index,follow",{"doc_id":128,"site_id":62},661240,1793161224,{"code":4,"msg":5,"data":131},{"doc_id":128,"user_id":132,"nickname":92,"user_avatar":133,"doc_module":4,"category_id":39,"category_name":40,"doc_title":65,"doc_description":67,"doc_content":134,"file_id":135,"file_url":136,"file_type":137,"file_size":138,"view_count":19,"is_deleted":4,"is_public":8,"is_downloadable":8,"audit_status":8,"page_count":139,"language":140,"language_code":63,"site_id":62,"html_lang":63,"table_of_contents":141,"faqs":142,"seo_title":143,"seo_description":67,"update_tm":129,"read_time":144},2336475104957,"https://ap-avatar.wpscdn.com/avatar/22000c4c6bd8a5076e1?x-image-process=image/resize,m_fixed,w_180,h_180&k=1787554080175789136","arXiv :2308 .07545v4 [ cs .CV] 20 Aug 2024  \nVision-Language Dataset Distillation  \nXindi Wu 1 Byron Zhang 1 Zhiwei Deng2 Olga Russakovsky 1  \n1 Princeton University 2 Google DeepMind  \n{xindiw, zishuoz, [olgarus}@princeton. edu](olgarus}@princeton. edu), [zhiweideng@google. com](zhiweideng@google. com)  \nReviewed on OpenReview: [https: // openreview. net/ forum? id= 6L6cD65pot](https: // openreview. net/ forum? id= 6L6cD65pot)  \nWebsite: [https: // princetonvisualai. github. io/ multimodal_ dataset_ distillation/](https: // princetonvisualai. github. io/ multimodal_ dataset_ distillation/)  \nAbstract  \nDataset distillation methods reduce large-scale datasets to smaller sets of synthetic data, preserving sufficient information to quickly train a new model from scratch. However, prior work on dataset distillation has focused exclusively on image classification datasets, whereas modern large-scale datasets are primarily vision-language datasets. In this work, we design the first vision-language dataset distillation method, building on the idea of trajectory matching. A key challenge is that vision-language datasets do not have a set of discrete classes.  \nTo overcome this, our proposed method jointly distills image-text pairs in a contrastive formulation. Further, we leverage Low-Rank Adaptation (LoRA) matching to enable more efficient and effective trajectory matching in complex modern vision-language models. Since there are no existing baselines, we compare our distillation approach with three adapted vision-language coreset selection methods. We demonstrate significant improvements on the challenging Flickr30K and COCO retrieval benchmarks: for example, on Flickr30K, the best coreset selection method selecting 1000 image-text pairs for training achieves only 5.6% image-to-text retrieval accuracy (i.e., recall@1); in contrast, our dataset distillation almost doubles that to 9 .9% with just 100 training pairs, an order of magnitude fewer.  \n1 Introduction  \nData = Information + Irrelevant Data (Wright & Ma, 2022)  \nDataset distillation aims to create concise summaries of data that preserve most of the critical information of the entire dataset. It holds paramount importance in the era of big data as it addresses the challenge posed by “Data = Information + Irrelevant Data” (Wright & Ma, 2022), where we often need to learn the useful information in an ocean of non-critical data. The recent growth of dataset distillation methods, e.g., (Wang et al. , 2018 ; Cazenavette et al. , 2022 ; Nguyen et al. , 2020) has primarily focused on image classification datasets, capturing class-specific information to build discriminative boundaries. Considering the recent progress in multimodal machine learning, where we witness the explosion of vision-language datasets in which the majority of image pixels may belong to irrelevant contextual elements and may further lack corresponding textual descriptions, a significant necessity arises to efficiently distill this vast amount of data. A well-distilled multimodal dataset simplifies complex vision-language interactions and emphasizes the most salient connections, making it more effective for models to learn cross-modal representations.  \nWhy is it hard? The first key challenge and the main difference from prior dataset distillation methods (Wanget al. , 2018 ; Cazenavette et al. , 2022) is that vision-language datasets do not contain a discrete set of classes to ground the distillation process. Instead, these datasets contain complex cross-modal connections and redundancies, requiring a co-distillation approach to capture their interdependencies effectively. Second, the complexity of cross-modal representations and vision-language models (VLMs) leads to computational challenges. Prior dataset distillation methods operate on low-resolution images (typically 28x28 or 32x32, as in MNIST (LeCun et al. , 1998) or CIFAR (Krizhevsky et al. , 2009)) and nevertheless suffer from significant computational c","cbCaiaBN2YWnVxTS","https://ap.wps.com/l/cbCaiaBN2YWnVxTS","pdf",11235594,31,"English","# Abstract\n# 1 Introduction\n## Why is it hard?\n## Figure 1: Dataset Distillation Comparison\n## Our work\n## Contributions","[{\"question\":\"What problem does dataset distillation address in large-scale learning?\",\"answer\":\"Dataset distillation creates concise summaries that preserve critical information from a large dataset, reducing reliance on irrelevant data.\"},{\"question\":\"What is the main challenge for vision-language dataset distillation compared with prior work?\",\"answer\":\"Vision-language datasets do not provide a discrete set of classes, so distillation cannot rely on class-specific boundaries.\"},{\"question\":\"How does the proposed method distill vision-language data without discrete classes?\",\"answer\":\"It jointly distills image-text pairs using a contrastive formulation and matches the by-products of training, leveraging LoRA-based matching for efficiency.\"}]","Vision-Language Dataset Distillation | PDF",78]