{"metadata":{"_oai":{"id":"oai:ipsj.ixsq.nii.ac.jp:00241065","sets":["1164:3206:11566:11793"]},"path":["11793"],"owner":"44499","recid":"241065","title":["Sigmoid AttentionによるAttention修正機構を導入したDINOの性能評価"],"pubdate":{"attribute_name":"公開日","attribute_value":"2024-11-22"},"_buckets":{"deposit":"97e66753-43ce-42e7-a9cc-881c66fd2992"},"_deposit":{"id":"241065","pid":{"type":"depid","value":"241065","revision_id":0},"owners":[44499],"status":"published","created_by":44499},"item_title":"Sigmoid AttentionによるAttention修正機構を導入したDINOの性能評価","author_link":["663017","663019","663021","663018","663020","663023","663016","663022"],"item_titles":{"attribute_name":"タイトル","attribute_value_mlt":[{"subitem_title":"Sigmoid AttentionによるAttention修正機構を導入したDINOの性能評価"},{"subitem_title":"Performance Evaluation of DINO with Attention Modification Mechanism Introduced by Sigmoid Attention","subitem_title_language":"en"}]},"item_type_id":"4","publish_date":"2024-11-22","item_4_text_3":{"attribute_name":"著者所属","attribute_value_mlt":[{"subitem_text_value":"住友電気工業株式会社"},{"subitem_text_value":"株式会社ユニコ"},{"subitem_text_value":"住電通信エンジニアリング株式会社"},{"subitem_text_value":"住友電気工業株式会社"}]},"item_4_text_4":{"attribute_name":"著者所属(英)","attribute_value_mlt":[{"subitem_text_value":"Sumitomo Electric Industries, Ltd.","subitem_text_language":"en"},{"subitem_text_value":"UNICO Co.,Ltd.","subitem_text_language":"en"},{"subitem_text_value":"Sumiden Communication Engineering Co.,Ltd.","subitem_text_language":"en"},{"subitem_text_value":"Sumitomo Electric Industries, Ltd.","subitem_text_language":"en"}]},"item_language":{"attribute_name":"言語","attribute_value_mlt":[{"subitem_language":"jpn"}]},"item_publisher":{"attribute_name":"出版者","attribute_value_mlt":[{"subitem_publisher":"情報処理学会","subitem_publisher_language":"ja"}]},"publish_status":"0","weko_shared_id":-1,"item_file_price":{"attribute_name":"Billing file","attribute_type":"file","attribute_value_mlt":[{"url":{"url":"https://ipsj.ixsq.nii.ac.jp/record/241065/files/IPSJ-CG24196006.pdf","label":"IPSJ-CG24196006.pdf"},"format":"application/pdf","billing":["billing_file"],"filename":"IPSJ-CG24196006.pdf","filesize":[{"value":"2.0 MB"}],"mimetype":"application/pdf","priceinfo":[{"tax":["include_tax"],"price":"0","billingrole":"28"},{"tax":["include_tax"],"price":"0","billingrole":"44"}],"accessrole":"open_login","version_id":"72fad787-f407-481d-8334-dceb78a0a655","displaytype":"detail","licensetype":"license_note","license_note":"Copyright (c) 2024 by the Institute of Electronics, Information and Communication Engineers This SIG report is only available to those in membership of the SIG."}]},"item_4_creator_5":{"attribute_name":"著者名","attribute_type":"creator","attribute_value_mlt":[{"creatorNames":[{"creatorName":"松本, 悠希"}],"nameIdentifiers":[{}]},{"creatorNames":[{"creatorName":"尾澤, 知憲"}],"nameIdentifiers":[{}]},{"creatorNames":[{"creatorName":"奥野, 拓也"}],"nameIdentifiers":[{}]},{"creatorNames":[{"creatorName":"三浦, 勝司"}],"nameIdentifiers":[{}]}]},"item_4_creator_6":{"attribute_name":"著者名(英)","attribute_type":"creator","attribute_value_mlt":[{"creatorNames":[{"creatorName":"Yuki, Matsumoto","creatorNameLang":"en"}],"nameIdentifiers":[{}]},{"creatorNames":[{"creatorName":"Tomokazu, Ozawa","creatorNameLang":"en"}],"nameIdentifiers":[{}]},{"creatorNames":[{"creatorName":"Takuya, Okuno","creatorNameLang":"en"}],"nameIdentifiers":[{}]},{"creatorNames":[{"creatorName":"Katsushi, Miura","creatorNameLang":"en"}],"nameIdentifiers":[{}]}]},"item_4_source_id_9":{"attribute_name":"書誌レコードID","attribute_value_mlt":[{"subitem_source_identifier":"AN10100541","subitem_source_identifier_type":"NCID"}]},"item_4_textarea_12":{"attribute_name":"Notice","attribute_value_mlt":[{"subitem_textarea_value":"SIG Technical Reports are nonrefereed and hence may later appear in any journals, conferences, symposia, etc."}]},"item_resource_type":{"attribute_name":"資源タイプ","attribute_value_mlt":[{"resourceuri":"http://purl.org/coar/resource_type/c_18gh","resourcetype":"technical report"}]},"item_4_source_id_11":{"attribute_name":"ISSN","attribute_value_mlt":[{"subitem_source_identifier":"2188-8949","subitem_source_identifier_type":"ISSN"}]},"item_4_description_7":{"attribute_name":"論文抄録","attribute_value_mlt":[{"subitem_description":"本論文ではラベル無し自己教師あり学習の一つである DINO（ViT ベース）に Sigmoid Attention による Attention の修正機構を導入することを提案する．Sigmoid Attention は Self-Attention におけるクエリとキーを内積したテンソルにシグモイド関数を適用したものである．その後，逆シグモイド変換を行い，通常のソフトマックス関数を適用する．このような演算過程により Vision Transformer のネットワークを阻害することなく，人による柔軟な Attention の修正が可能となる．本提案は人が Sigmoid Attention を修正し，これを教師データとしてファインチューニング（Human In the Loop）することで，人の知見をネットワークに反映させ，精度向上や Attention の視覚的説明性の向上に貢献するものである．提案手法の性能を評価するために ImageNet から 3 クラスを抽出し検討を行った．","subitem_description_type":"Other"}]},"item_4_description_8":{"attribute_name":"論文抄録(英)","attribute_value_mlt":[{"subitem_description":"In this paper, we propose to introduce Attention Modification Mechanism using Sigmoid Attention to DINO (ViT based), which is a type of unlabeled self-supervised learning. Sigmoid Attention is applied Sigmoid function to tensor of dot product Query and Key in Self-Attention. After, it is applied inverse Sigmoid function and Softmax function in that order. By such calculation process, Vision Transformer network isn't impeded, and flexible modification of Attention by humans is possible. This proposal allows humans to modify Sigmoid Attention and use it as training data for fine-tuning, thereby reflecting human knowledge in the network and contributing to improved accuracy and visual explanation of Attention. In order to evaluate the proposal method performance, we extracted three classes from ImageNet and examined them.","subitem_description_type":"Other"}]},"item_4_biblio_info_10":{"attribute_name":"書誌情報","attribute_value_mlt":[{"bibliographicPageEnd":"6","bibliographic_titles":[{"bibliographic_title":"研究報告コンピュータグラフィックスとビジュアル情報学（CG）"}],"bibliographicPageStart":"1","bibliographicIssueDates":{"bibliographicIssueDate":"2024-11-22","bibliographicIssueDateType":"Issued"},"bibliographicIssueNumber":"6","bibliographicVolumeNumber":"2024-CG-196"}]},"relation_version_is_last":true,"weko_creator_id":"44499"},"id":241065,"updated":"2025-01-19T07:45:18.130148+00:00","links":{},"created":"2025-01-19T01:45:37.032795+00:00"}