{"metadata":{"_oai":{"id":"oai:ipsj.ixsq.nii.ac.jp:00227251","sets":["581:11107:11117"]},"path":["11117"],"owner":"44499","recid":"227251","title":["行動認識における深層学習モデル訓練時の最適なsoftmax温度パラメータ"],"pubdate":{"attribute_name":"公開日","attribute_value":"2023-08-15"},"_buckets":{"deposit":"834cbf4b-47bf-4efe-84f0-94f1c82792b7"},"_deposit":{"id":"227251","pid":{"type":"depid","value":"227251","revision_id":0},"owners":[44499],"status":"published","created_by":44499},"item_title":"行動認識における深層学習モデル訓練時の最適なsoftmax温度パラメータ","author_link":["605166","605165"],"item_titles":{"attribute_name":"タイトル","attribute_value_mlt":[{"subitem_title":"行動認識における深層学習モデル訓練時の最適なsoftmax温度パラメータ"},{"subitem_title":"Optimal Temperature Parameter of Softmax while Training Deep Learning Model in Activity Recognition","subitem_title_language":"en"}]},"item_keyword":{"attribute_name":"キーワード","attribute_value_mlt":[{"subitem_subject":"[一般論文（推薦論文，特選論文）] Softmax関数，温度パラメータ，ニューラルネットワーク，行動認識","subitem_subject_scheme":"Other"}]},"item_type_id":"2","publish_date":"2023-08-15","item_2_text_3":{"attribute_name":"著者所属","attribute_value_mlt":[{"subitem_text_value":"福井大学大学院工学研究科"}]},"item_2_text_4":{"attribute_name":"著者所属(英)","attribute_value_mlt":[{"subitem_text_value":"Graduate School of Engineering, University of Fukui","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/227251/files/IPSJ-JNL6408004.pdf","label":"IPSJ-JNL6408004.pdf"},"date":[{"dateType":"Available","dateValue":"2025-08-15"}],"format":"application/pdf","billing":["billing_file"],"filename":"IPSJ-JNL6408004.pdf","filesize":[{"value":"1.4 MB"}],"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":"c267d90b-a5bf-44d8-9ddc-f3dcedbc9514","displaytype":"detail","licensetype":"license_note","license_note":"Copyright (c) 2023 by the Information Processing Society of Japan"}]},"item_2_creator_5":{"attribute_name":"著者名","attribute_type":"creator","attribute_value_mlt":[{"creatorNames":[{"creatorName":"長谷川, 達人"}],"nameIdentifiers":[{}]}]},"item_2_creator_6":{"attribute_name":"著者名(英)","attribute_type":"creator","attribute_value_mlt":[{"creatorNames":[{"creatorName":"Tatsuhito, Hasegawa","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_publisher_15":{"attribute_name":"公開者","attribute_value_mlt":[{"subitem_publisher":"情報処理学会","subitem_publisher_language":"ja"}]},"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":"深層学習はハイパーパラメータが膨大であり，適切に使いこなすには熟練の技能が必要となる．本研究では，膨大なハイパーパラメータの中でも未解明な点が多いsoftmax関数の温度パラメータTと特徴マップの次元数Mに焦点を当てる．特に行動認識ではモデルサイズを調整することは少なくなく，TとMの関係の解明は重要である．深層学習モデルを出力の分散の観点から理論的に考察した結果，出力層のパラメータはMの制約を受けて最適化されており，最適なTの設定はこの制約を緩和できる可能性があると考えた．本研究では，様々な行動認識データセットやモデル構造において，TとMの関係を実験的に検証する．実験の結果，T=1の従来の設定ではモデルの最良のパフォーマンスを発揮しきれていないこと，Mの増加にともない最適なTも増加すること，最適なTにおいてはsoftmax関数の入力の分布が安定していること等を明らかにした．実験結果をもとに，出力層にLayer Normalizationを挿入することでMの影響を緩和する手法を新たに提案し，追加実験を経て提案手法の有効性を示した．","subitem_description_type":"Other"}]},"item_2_description_8":{"attribute_name":"論文抄録(英)","attribute_value_mlt":[{"subitem_description":"Deep learning has many hyperparameters and requires skilled operators to use it properly. In this study, we focus on the temperature parameter T of the softmax function and M: the number of dimensions of the feature map, which are still largely unexplained among many hyperparameters. In particular, it is important to clarify the relationship between T and M because, in activity recognition, model scale M is commonly adjusted. Theoretical consideration of deep learning models in terms of output variance suggests that output layer parameters are optimized under the constraint of M, and that optimal T may deregulate this constraint. In this study, we experimentally verify the relationship between T and M on various activity recognition datasets and model architectures. Experimental results show that the model does not perform at its best in the conventional setting of T=1, that the optimal T increases as M increases, and that the distribution of inputs to the softmax function is stable at the optimal T. Based on the experimental results, we proposed a new method to mitigate the effect of M by inserting a layer normalization after the output layer and demonstrated the effectiveness of the proposed method through additional experiments.","subitem_description_type":"Other"}]},"item_2_biblio_info_10":{"attribute_name":"書誌情報","attribute_value_mlt":[{"bibliographicPageEnd":"1192","bibliographic_titles":[{"bibliographic_title":"情報処理学会論文誌"}],"bibliographicPageStart":"1182","bibliographicIssueDates":{"bibliographicIssueDate":"2023-08-15","bibliographicIssueDateType":"Issued"},"bibliographicIssueNumber":"8","bibliographicVolumeNumber":"64"}]},"relation_version_is_last":true,"item_2_identifier_registration":{"attribute_name":"ID登録","attribute_value_mlt":[{"subitem_identifier_reg_text":"10.20729/00227142","subitem_identifier_reg_type":"JaLC"}]},"weko_creator_id":"44499"},"id":227251,"updated":"2025-01-19T12:12:34.106213+00:00","links":{},"created":"2025-01-19T01:26:31.308585+00:00"}