{"created":"2025-01-19T01:13:04.960802+00:00","metadata":{"_oai":{"id":"oai:ipsj.ixsq.nii.ac.jp:00212061","sets":["581:10433:10440"]},"path":["10440"],"owner":"44499","recid":"212061","title":["深層ニューラルネットワークによるクラスと幾何変換の同時分類確率を利用した分布外検知"],"pubdate":{"attribute_name":"公開日","attribute_value":"2021-07-15"},"_buckets":{"deposit":"af56801b-42e7-4221-b3ea-3a6cb3ef163b"},"_deposit":{"id":"212061","pid":{"type":"depid","value":"212061","revision_id":0},"owners":[44499],"status":"published","created_by":44499},"item_title":"深層ニューラルネットワークによるクラスと幾何変換の同時分類確率を利用した分布外検知","author_link":["540058","540060","540056","540055","540059","540057"],"item_titles":{"attribute_name":"タイトル","attribute_value_mlt":[{"subitem_title":"深層ニューラルネットワークによるクラスと幾何変換の同時分類確率を利用した分布外検知"},{"subitem_title":"Out-of-distribution Detection Using Joint Probability between Class and Geometric Transformation","subitem_title_language":"en"}]},"item_keyword":{"attribute_name":"キーワード","attribute_value_mlt":[{"subitem_subject":"[一般論文] 分布外検知，深層ニューラルネットワーク，自己教師あり学習，多クラス分類","subitem_subject_scheme":"Other"}]},"item_type_id":"2","publish_date":"2021-07-15","item_2_text_3":{"attribute_name":"著者所属","attribute_value_mlt":[{"subitem_text_value":"東京大学工学系研究科技術経営戦略学専攻松尾研究室"},{"subitem_text_value":"東京大学工学系研究科技術経営戦略学専攻松尾研究室"},{"subitem_text_value":"東京大学工学系研究科技術経営戦略学専攻松尾研究室"}]},"item_2_text_4":{"attribute_name":"著者所属(英)","attribute_value_mlt":[{"subitem_text_value":"Department of Technology Management for Innovation, Graduate School of Engineering, The University of Tokyo","subitem_text_language":"en"},{"subitem_text_value":"Department of Technology Management for Innovation, Graduate School of Engineering, The University of Tokyo","subitem_text_language":"en"},{"subitem_text_value":"Department of Technology Management for Innovation, Graduate School of Engineering, The University of Tokyo","subitem_text_language":"en"}]},"item_language":{"attribute_name":"言語","attribute_value_mlt":[{"subitem_language":"jpn"}]},"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/212061/files/IPSJ-JNL6207001.pdf","label":"IPSJ-JNL6207001.pdf"},"date":[{"dateType":"Available","dateValue":"2023-07-15"}],"format":"application/pdf","billing":["billing_file"],"filename":"IPSJ-JNL6207001.pdf","filesize":[{"value":"591.4 kB"}],"mimetype":"application/pdf","priceinfo":[{"tax":["include_tax"],"price":"660","billingrole":"5"},{"tax":["include_tax"],"price":"330","billingrole":"6"},{"tax":["include_tax"],"price":"0","billingrole":"8"},{"tax":["include_tax"],"price":"0","billingrole":"44"}],"accessrole":"open_date","version_id":"27b2388e-5d99-4c5b-bf81-0d21c0d5b8ce","displaytype":"detail","licensetype":"license_note","license_note":"Copyright (c) 2021 by the Information Processing Society of Japan"}]},"item_2_creator_5":{"attribute_name":"著者名","attribute_type":"creator","attribute_value_mlt":[{"creatorNames":[{"creatorName":"岡本, 弘野"}],"nameIdentifiers":[{}]},{"creatorNames":[{"creatorName":"鈴木, 雅大"}],"nameIdentifiers":[{}]},{"creatorNames":[{"creatorName":"松尾, 豊"}],"nameIdentifiers":[{}]}]},"item_2_creator_6":{"attribute_name":"著者名(英)","attribute_type":"creator","attribute_value_mlt":[{"creatorNames":[{"creatorName":"Hirono, Okamoto","creatorNameLang":"en"}],"nameIdentifiers":[{}]},{"creatorNames":[{"creatorName":"Masahiro, Suzuki","creatorNameLang":"en"}],"nameIdentifiers":[{}]},{"creatorNames":[{"creatorName":"Yutaka, Matsuo","creatorNameLang":"en"}],"nameIdentifiers":[{}]}]},"item_2_source_id_9":{"attribute_name":"書誌レコードID","attribute_value_mlt":[{"subitem_source_identifier":"AN00116647","subitem_source_identifier_type":"NCID"}]},"item_resource_type":{"attribute_name":"資源タイプ","attribute_value_mlt":[{"resourceuri":"http://purl.org/coar/resource_type/c_6501","resourcetype":"journal article"}]},"item_2_source_id_11":{"attribute_name":"ISSN","attribute_value_mlt":[{"subitem_source_identifier":"1882-7764","subitem_source_identifier_type":"ISSN"}]},"item_2_description_7":{"attribute_name":"論文抄録","attribute_value_mlt":[{"subitem_description":"分布外検知はあるデータが入力されたときに，そのデータが特定の分布からのデータ（分布内データ）かそれ以外の分布からのデータ（分布外データ）かに分類するタスクである．近年の研究により，深層ニューラルネットワークを使った分類器は，分布外データを入力としたとき，分布内データを入力とする場合と比較して，(1)クラス分類の出力が一様分布に近づき，(2)幾何変換を予測することが難しいという2つの仮説のもとで，分布外検知が行えることが明らかになっている．しかし，(1)において，クラス分類するのが難しいデータセットを用いる場合，分布内データにも分類器の出力が一様分布になるようなものが存在する．また，(2)において，near-distribution outliersと呼ばれる分布外データを使った場合，幾何変換が予測できてしまう場合がある．このとき，(1)または(2)の仮説をもとにした先行研究は，分布外データと分布内データの見分けがつかなくなってしまう問題が発生する．筆者らは，それぞれの先行研究において分布外データの検知精度が低くなる条件が異なるので，両者の手法を組み合わせることで，両者の欠点を補えると仮定した．この仮定に基づき，筆者らは(1)と(2)の仮説を同時に活用する指標を用いることを提案する．具体的には，筆者らは入力が与えられたときのクラスと幾何変換の同時確率を求め，これをもとにした新しいスコアを提案する．実験では，様々なネットワーク構造とデータセットを使って実験を行い，(1)と(2)のそれぞれの仮説に基づく先行研究よりも，両方の仮説に基づく本手法のほうが，AUROCを指標として平均6.7%，分布外データの検知精度が高くなることを示す．","subitem_description_type":"Other"}]},"item_2_description_8":{"attribute_name":"論文抄録(英)","attribute_value_mlt":[{"subitem_description":"Out-of-distribution detection is a task to categorize the input data into data from a specific distribution (in-distribution data) or data from another distribution (out-of-distribution data). Recent studies have shown that a classifier using deep neural networks can detect out-of-distribution data under the following two hypotheses: (1) the output of the classifier approaches a uniform distribution, and (2) it is difficult to predict geometric transformations, when the inputs are out-of-distribution data. However, in (1), when using datasets that are difficult to classify, there are some in-distribution data that result in a uniform distribution of the classifier's output. In (2), the use of near-distribution outliers, which are out-of-distribution data, can lead to the correct prediction of geometric transformations. In this case, previous studies based on hypotheses (1) or (2) cause problems in distinguishing between out-of-distribution and in-distribution data. We hypothesized that the combination of the two methods could compensate for the low detection accuracy of out-of-distribution data in each previous study. Therefore, we propose to use a metric that utilizes both hypotheses (1) and (2) simultaneously. Specifically, we calculate the joint probability of class and geometric transformations given the input and propose a new anomaly score based on it. We conduct experiments using various network structures and datasets and show that the method based on both hypotheses is more accurate in detecting out-of-distribution data than the method based on either (1) or (2) by 6.7% on average.","subitem_description_type":"Other"}]},"item_2_biblio_info_10":{"attribute_name":"書誌情報","attribute_value_mlt":[{"bibliographicPageEnd":"1392","bibliographic_titles":[{"bibliographic_title":"情報処理学会論文誌"}],"bibliographicPageStart":"1382","bibliographicIssueDates":{"bibliographicIssueDate":"2021-07-15","bibliographicIssueDateType":"Issued"},"bibliographicIssueNumber":"7","bibliographicVolumeNumber":"62"}]},"relation_version_is_last":true,"item_2_identifier_registration":{"attribute_name":"ID登録","attribute_value_mlt":[{"subitem_identifier_reg_text":"10.20729/00211955","subitem_identifier_reg_type":"JaLC"}]},"weko_creator_id":"44499"},"links":{},"id":212061,"updated":"2025-01-19T17:34:21.535826+00:00"}