{"id":1,"date":"2022-12-06T19:19:58","date_gmt":"2022-12-06T19:19:58","guid":{"rendered":"http:\/\/blog.vaniila-ai.catie-na.fr\/?p=1"},"modified":"2022-12-08T13:50:44","modified_gmt":"2022-12-08T13:50:44","slug":"scan-apprentissage-contrastif-non-supervise","status":"publish","type":"post","link":"https:\/\/blog.vaniila-ai.catie-na.fr\/?p=1","title":{"rendered":"SCAN : apprentissage contrastif non supervis\u00e9"},"content":{"rendered":"\n[et_pb_section fb_built=\u00a0\u00bb1&Prime; custom_padding_last_edited=\u00a0\u00bbon|tablet\u00a0\u00bb admin_label=\u00a0\u00bbHeader\u00a0\u00bb _builder_version=\u00a0\u00bb4.18.0&Prime; _module_preset=\u00a0\u00bbdefault\u00a0\u00bb use_background_color_gradient=\u00a0\u00bbon\u00a0\u00bb background_color_gradient_stops=\u00a0\u00bbrgba(0,0,0,0) 0%|#000000 86%\u00a0\u00bb background_color_gradient_overlays_image=\u00a0\u00bbon\u00a0\u00bb background_image=\u00a0\u00bbhttp:\/\/blog.vaniila-ai.catie-na.fr\/wp-content\/uploads\/2022\/12\/web-developer-28.jpg\u00a0\u00bb custom_padding=\u00a0\u00bb5%||||false|false\u00a0\u00bb custom_padding_tablet=\u00a0\u00bb60px||||false|false\u00a0\u00bb custom_padding_phone=\u00a0\u00bb60px||||false|false\u00a0\u00bb collapsed=\u00a0\u00bbon\u00a0\u00bb global_colors_info=\u00a0\u00bb{}\u00a0\u00bb][et_pb_row _builder_version=\u00a0\u00bb4.18.0&Prime; _module_preset=\u00a0\u00bbdefault\u00a0\u00bb global_colors_info=\u00a0\u00bb{}\u00a0\u00bb][et_pb_column type=\u00a0\u00bb4_4&Prime; _builder_version=\u00a0\u00bb4.18.0&Prime; _module_preset=\u00a0\u00bbdefault\u00a0\u00bb global_colors_info=\u00a0\u00bb{}\u00a0\u00bb][et_pb_text _builder_version=\u00a0\u00bb4.19.2&Prime; _module_preset=\u00a0\u00bbe6504a1b-67eb-4b3d-b023-bcab277610b6&Prime; header_4_font=\u00a0\u00bbArchivo|700||on|||||\u00a0\u00bb header_4_text_color=\u00a0\u00bbgcid-f1414204-51c0-48ff-bc68-c545a86d03e7&Prime; header_4_font_size=\u00a0\u00bb14px\u00a0\u00bb header_4_letter_spacing=\u00a0\u00bb1px\u00a0\u00bb header_4_line_height=\u00a0\u00bb1.5em\u00a0\u00bb text_orientation=\u00a0\u00bbcenter\u00a0\u00bb custom_margin=\u00a0\u00bb||0px||false|false\u00a0\u00bb hover_enabled=\u00a0\u00bb0&Prime; global_colors_info=\u00a0\u00bb{%22gcid-f1414204-51c0-48ff-bc68-c545a86d03e7%22:%91%22header_4_text_color%22%93}\u00a0\u00bb text_font=\u00a0\u00bb|||on|||||\u00a0\u00bb sticky_enabled=\u00a0\u00bb0&Prime;]<h1><strong><span style=\"color: #ffcc00;\">Exemples de paires contrastives <\/span><\/strong><\/h1>\n<h1><strong><span style=\"color: #ffcc00;\">g\u00e9n\u00e9r\u00e9es avec ces augmentations<\/span><\/strong><\/h1>[\/et_pb_text][\/et_pb_column][\/et_pb_row][et_pb_row _builder_version=\u00a0\u00bb4.19.2&Prime; _module_preset=\u00a0\u00bbdefault\u00a0\u00bb filter_opacity=\u00a0\u00bb75%\u00a0\u00bb global_colors_info=\u00a0\u00bb{}\u00a0\u00bb][et_pb_column type=\u00a0\u00bb4_4&Prime; _builder_version=\u00a0\u00bb4.19.2&Prime; _module_preset=\u00a0\u00bbdefault\u00a0\u00bb global_colors_info=\u00a0\u00bb{}\u00a0\u00bb][et_pb_text _builder_version=\u00a0\u00bb4.19.2&Prime; _module_preset=\u00a0\u00bbc16985e1-e0d6-4022-964d-e2bfc04fa633&Prime; header_font=\u00a0\u00bbRoboto|700|||||||\u00a0\u00bb header_text_color=\u00a0\u00bbgcid-6171fd24-b893-4d22-843a-f4129850a5c1&Prime; header_font_size=\u00a0\u00bb75px\u00a0\u00bb header_line_height=\u00a0\u00bb1.2em\u00a0\u00bb text_orientation=\u00a0\u00bbcenter\u00a0\u00bb hover_enabled=\u00a0\u00bb0&Prime; header_font_size_tablet=\u00a0\u00bb40px\u00a0\u00bb header_font_size_phone=\u00a0\u00bb24px\u00a0\u00bb header_font_size_last_edited=\u00a0\u00bbon|desktop\u00a0\u00bb global_colors_info=\u00a0\u00bb{%22gcid-6171fd24-b893-4d22-843a-f4129850a5c1%22:%91%22header_text_color%22%93}\u00a0\u00bb sticky_enabled=\u00a0\u00bb0&Prime;]<h2 style=\"text-align: left;\"><span style=\"color: #ffffff;\"><strong>La m\u00e9thode<\/strong><\/span><\/h2>[\/et_pb_text][et_pb_text _builder_version=\u00a0\u00bb4.19.2&Prime; _module_preset=\u00a0\u00bbdefault\u00a0\u00bb text_text_color=\u00a0\u00bb#FFFFFF\u00a0\u00bb global_colors_info=\u00a0\u00bb{}\u00a0\u00bb]<h1 style=\"text-align: justify;\"><span style=\"color: #ffffff;\">\u00a0<\/span><\/h1>\n<p style=\"text-align: justify;\"><span style=\"color: #ffffff;\">L\u2019algorithme SCAN est une m\u00e9thode de clustering non supervis\u00e9, c\u2019est \u00e0 dire qu\u2019elle associe des \u00e9tiquettes \u00e0 des images, sans n\u00e9cessiter d\u2019exemples.<\/span><\/p>\n<p style=\"text-align: justify;\"><span style=\"color: #ffffff;\"> Elle est compos\u00e9e de trois \u00e9tapes :<\/span><\/p>\n<ul style=\"text-align: justify;\">\n<li><span style=\"color: #ffffff;\"><span style=\"font-weight: 600;\" data-token-index=\"0\" class=\"notion-enable-hover\" data-reactroot=\"\">Vectorisation<\/span> : La premi\u00e8re \u00e9tape de traitement sert \u00e0 vectoriser les images, c\u2019est \u00e0 dire transformer chaque image en vecteur de N dimensions, tels que les images qui pr\u00e9sentent des caract\u00e9ristiques communes correspondent \u00e0 des vecteurs proches en N dimensions.<\/span><\/li>\n<li><span style=\"color: #ffffff;\">Cette \u00e9tape est assur\u00e9e dans l\u2019article original par une adaptation de la m\u00e9thode SimCLR. Cependant, n\u2019importe quelle m\u00e9thode de vectorisation non-supervis\u00e9e peut \u00eatre employ\u00e9e \u00e0 la place.<\/span><\/li>\n<li><span style=\"color: #ffffff;\">Le r\u00e9sultat de la vectorisation permet de trouver, pour chaque image, les k plus proches voisins et leurs vecteurs, qui serviront de donn\u00e9es d\u2019entr\u00e9e pour l\u2019\u00e9tape suivante.<\/span><\/li>\n<li><span style=\"color: #ffffff;\"><span style=\"font-weight: 600;\" data-token-index=\"0\" class=\"notion-enable-hover\" data-reactroot=\"\">Clustering<\/span> : A partir des vecteurs de l\u2019\u00e9tape pr\u00e9c\u00e9dente et en se basant sur l\u2019extraction des k plus proches voisins, cette \u00e9tape assigne un cluster \u00e0 chaque image, en optimisant deux crit\u00e8res:<\/span>\n<ul>\n<li><span style=\"color: #ffffff;\">Une image et ses voisins sont assign\u00e9s au m\u00eame cluster<\/span><\/li>\n<li><span style=\"color: #ffffff;\">Un param\u00e8tre d\u2019entropie permet de r\u00e9partir les images entres tous les clusters (\u00e9viter que toutes les images soient assign\u00e9es au m\u00eame cluster).<\/span><\/li>\n<\/ul>\n<\/li>\n<li><span style=\"color: #ffffff;\"><span style=\"font-weight: 600;\" data-token-index=\"0\" class=\"notion-enable-hover\" data-reactroot=\"\">Self-supervised Learning<\/span> : En ne prenant en compte que les images pour lesquelles le mod\u00e8le est le plus sur de son choix, il r\u00e9apprend les classes des autres exemples. A chaque \u00e9poque d\u2019entrainement, de plus en plus d\u2019images seront \u00e9ligibles \u00e0 servir d\u2019exemple pour cette \u00e9tape.<\/span><\/li>\n<\/ul>\n<p style=\"text-align: justify;\"><span style=\"color: #ffffff;\">On peut remarquer que la premi\u00e8re \u00e9tape est cruciale, car elle permet de refl\u00e9ter les caract\u00e9ristiques des images dans un espace multi-dimensionnel, utilis\u00e9 par la suite pour le clustering. Si les caract\u00e9ristiques des images qui doivent servir \u00e0 les rapprocher ne sont pas correctement prises en compte dans cette \u00e9tape de vectorisation, les \u00e9tapes suivantes ne produiront pas le r\u00e9sultat d\u00e9sir\u00e9.<\/span><\/p>\n<p style=\"text-align: justify;\"><span style=\"color: #ffffff;\"><\/span><\/p>\n<p style=\"text-align: justify;\"><\/p>\n<p style=\"text-align: justify;\"><span style=\"font-weight: 600; color: #ffffff;\" data-token-index=\"0\" class=\"notion-enable-hover\" data-reactroot=\"\"><\/span><\/p>[\/et_pb_text][et_pb_text _builder_version=\u00a0\u00bb4.19.2&Prime; _module_preset=\u00a0\u00bbc16985e1-e0d6-4022-964d-e2bfc04fa633&Prime; header_font=\u00a0\u00bbRoboto|700|||||||\u00a0\u00bb header_text_color=\u00a0\u00bbgcid-6171fd24-b893-4d22-843a-f4129850a5c1&Prime; header_font_size=\u00a0\u00bb75px\u00a0\u00bb header_line_height=\u00a0\u00bb1.2em\u00a0\u00bb text_orientation=\u00a0\u00bbcenter\u00a0\u00bb hover_enabled=\u00a0\u00bb0&Prime; header_font_size_tablet=\u00a0\u00bb40px\u00a0\u00bb header_font_size_phone=\u00a0\u00bb24px\u00a0\u00bb header_font_size_last_edited=\u00a0\u00bbon|desktop\u00a0\u00bb global_colors_info=\u00a0\u00bb{%22gcid-6171fd24-b893-4d22-843a-f4129850a5c1%22:%91%22header_text_color%22%93}\u00a0\u00bb sticky_enabled=\u00a0\u00bb0&Prime;]<h2 style=\"text-align: left;\"><span style=\"color: #ffffff;\"><strong>Tests<\/strong><\/span><\/h2>[\/et_pb_text][et_pb_text _builder_version=\u00a0\u00bb4.19.2&Prime; _module_preset=\u00a0\u00bbdefault\u00a0\u00bb text_text_color=\u00a0\u00bb#FFFFFF\u00a0\u00bb hover_enabled=\u00a0\u00bb0&Prime; global_colors_info=\u00a0\u00bb{}\u00a0\u00bb sticky_enabled=\u00a0\u00bb0&Prime;]<h3 style=\"text-align: justify;\"><em><span style=\"color: #ffffff;\">Impl\u00e9mentation<\/span><\/em><\/h3>\n<p style=\"text-align: justify;\"><span style=\"color: #ffffff;\">Nous avons utilis\u00e9 l\u2019impl\u00e9mentation open source des auteurs pour effectuer nos tests : <a href=\"https:\/\/github.com\/wvangansbeke\/Unsupervised-Classification\" style=\"color: #ffffff;\">https:\/\/github.com\/wvangansbeke\/Unsupervised-Classification<\/a><\/span><\/p>\n<h3 style=\"text-align: justify;\"><em><span style=\"color: #ffffff;\">Donn\u00e9es<\/span><\/em><\/h3>\n<p style=\"text-align: justify;\"><span style=\"color: #ffffff;\">Nous avons utilis\u00e9 le jeu de donn\u00e9es lymparza2, compos\u00e9 de 10 vues et 6 classes de d\u00e9fauts.<\/span><\/p>\n<h3 style=\"text-align: justify;\"><em><span style=\"color: #ffffff;\">Mod\u00e8le<\/span><\/em><\/h3>\n<p style=\"text-align: justify;\"><span style=\"color: #ffffff;\">Le mod\u00e8le utilise un Backbone resnet18, avec une sortie en 2048 dimensions<\/span><\/p>\n<p style=\"text-align: justify;\"><span style=\"font-weight: 600; color: #ffffff;\" data-token-index=\"0\" class=\"notion-enable-hover\" data-reactroot=\"\">Augmentations<\/span><\/p>\n<p style=\"text-align: justify;\"><span data-token-index=\"0\" class=\"notion-enable-hover\" data-reactroot=\"\" style=\"color: #ffffff;\">L\u2019apprentissage contrastif sur lequel repose cette m\u00e9thode n\u00e9cessite de g\u00e9n\u00e9rer des paires d\u2019images qui appartiennent au m\u00eame groupe. En pratique, on utilise g\u00e9n\u00e9ralement une m\u00eame image de base, modifi\u00e9e de deux mani\u00e8res diff\u00e9rentes \u00e0 l\u2019aide d\u2019augmentations prises dans un ensemble pr\u00e9-d\u00e9termin\u00e9.<\/span><\/p>\n<p style=\"text-align: justify;\"><span data-token-index=\"0\" class=\"notion-enable-hover\" data-reactroot=\"\" style=\"color: #ffffff;\">Pour nos tests nous avons utilis\u00e9 les augmentations suivantes:<\/span><\/p>\n<ul style=\"text-align: justify;\">\n<li><span data-token-index=\"0\" class=\"notion-enable-hover\" data-reactroot=\"\" style=\"color: #ffffff;\">Miroir horizontal et vertical<\/span><\/li>\n<li><span data-token-index=\"0\" class=\"notion-enable-hover\" data-reactroot=\"\" style=\"color: #ffffff;\">Rotation de 180\u00b0<\/span><\/li>\n<li><span data-token-index=\"0\" class=\"notion-enable-hover\" data-reactroot=\"\" style=\"color: #ffffff;\">ColorJitter<\/span><\/li>\n<li><span data-token-index=\"0\" class=\"notion-enable-hover\" data-reactroot=\"\" style=\"color: #ffffff;\">Bruit gaussien<\/span><\/li>\n<li><span data-token-index=\"0\" class=\"notion-enable-hover\" data-reactroot=\"\" style=\"color: #ffffff;\">Jigsaw (m\u00e9lange de parties de l\u2019image \u00e0 la mani\u00e8re d\u2019un jeu de taquin)<\/span><\/li>\n<\/ul>\n<p style=\"text-align: justify;\"><span data-token-index=\"0\" class=\"notion-enable-hover\" data-reactroot=\"\" style=\"color: #ffffff;\">Exemples de paires contrastives g\u00e9n\u00e9r\u00e9es avec ces augmentations<\/span><\/p>\n<p style=\"text-align: justify;\"><span style=\"font-weight: 600; color: #ffffff;\" data-token-index=\"0\" class=\"notion-enable-hover\" data-reactroot=\"\"><\/span><\/p>[\/et_pb_text][\/et_pb_column][\/et_pb_row][et_pb_row column_structure=\u00a0\u00bb1_3,1_3,1_3&Prime; _builder_version=\u00a0\u00bb4.19.2&Prime; _module_preset=\u00a0\u00bbdefault\u00a0\u00bb global_colors_info=\u00a0\u00bb{}\u00a0\u00bb][et_pb_column type=\u00a0\u00bb1_3&Prime; _builder_version=\u00a0\u00bb4.19.2&Prime; _module_preset=\u00a0\u00bbdefault\u00a0\u00bb global_colors_info=\u00a0\u00bb{}\u00a0\u00bb][et_pb_image src=\u00a0\u00bbhttp:\/\/blog.vaniila-ai.catie-na.fr\/wp-content\/uploads\/2022\/12\/1_1.png\u00a0\u00bb title_text=\u00a0\u00bb1_1&Prime; _builder_version=\u00a0\u00bb4.19.2&Prime; _module_preset=\u00a0\u00bbdefault\u00a0\u00bb global_colors_info=\u00a0\u00bb{}\u00a0\u00bb][\/et_pb_image][\/et_pb_column][et_pb_column type=\u00a0\u00bb1_3&Prime; _builder_version=\u00a0\u00bb4.19.2&Prime; _module_preset=\u00a0\u00bbdefault\u00a0\u00bb global_colors_info=\u00a0\u00bb{}\u00a0\u00bb][et_pb_image src=\u00a0\u00bbhttp:\/\/blog.vaniila-ai.catie-na.fr\/wp-content\/uploads\/2022\/12\/1_2.png\u00a0\u00bb title_text=\u00a0\u00bb1_2&Prime; _builder_version=\u00a0\u00bb4.19.2&Prime; _module_preset=\u00a0\u00bbdefault\u00a0\u00bb global_colors_info=\u00a0\u00bb{}\u00a0\u00bb][\/et_pb_image][\/et_pb_column][et_pb_column type=\u00a0\u00bb1_3&Prime; _builder_version=\u00a0\u00bb4.19.2&Prime; _module_preset=\u00a0\u00bbdefault\u00a0\u00bb global_colors_info=\u00a0\u00bb{}\u00a0\u00bb][et_pb_image src=\u00a0\u00bbhttp:\/\/blog.vaniila-ai.catie-na.fr\/wp-content\/uploads\/2022\/12\/1_3.png\u00a0\u00bb title_text=\u00a0\u00bb1_3&Prime; _builder_version=\u00a0\u00bb4.19.2&Prime; _module_preset=\u00a0\u00bbdefault\u00a0\u00bb global_colors_info=\u00a0\u00bb{}\u00a0\u00bb][\/et_pb_image][\/et_pb_column][\/et_pb_row][et_pb_row _builder_version=\u00a0\u00bb4.19.2&Prime; _module_preset=\u00a0\u00bbdefault\u00a0\u00bb global_colors_info=\u00a0\u00bb{}\u00a0\u00bb][et_pb_column type=\u00a0\u00bb4_4&Prime; _builder_version=\u00a0\u00bb4.19.2&Prime; _module_preset=\u00a0\u00bbdefault\u00a0\u00bb global_colors_info=\u00a0\u00bb{}\u00a0\u00bb][et_pb_text _builder_version=\u00a0\u00bb4.19.2&Prime; _module_preset=\u00a0\u00bbdefault\u00a0\u00bb text_text_color=\u00a0\u00bb#FFFFFF\u00a0\u00bb hover_enabled=\u00a0\u00bb0&Prime; global_colors_info=\u00a0\u00bb{}\u00a0\u00bb sticky_enabled=\u00a0\u00bb0&Prime;]<h3 style=\"text-align: justify;\"><em><span style=\"color: #ffffff;\">Test jeu de donn\u00e9es complet<\/span><\/em><\/h3>\n<p style=\"text-align: justify;\">Nous avons commenc\u00e9 par tester la m\u00e9thode sur le jeu de donn\u00e9es complet<\/p>\n<p style=\"text-align: justify;\">Le r\u00e9sultat de ce test donne bien des clusters identifiables, mais qui correspondent aux diff\u00e9rentes vues du dataset, qui sont plus facilement reconnaissables par le mod\u00e8le que les d\u00e9fauts.<span data-token-index=\"0\" class=\"notion-enable-hover\" data-reactroot=\"\" style=\"color: #ffffff;\"><\/span><\/p>\n<p style=\"text-align: justify;\"><span style=\"font-weight: 600; color: #ffffff;\" data-token-index=\"0\" class=\"notion-enable-hover\" data-reactroot=\"\"><\/span><\/p>[\/et_pb_text][\/et_pb_column][\/et_pb_row][et_pb_row column_structure=\u00a0\u00bb1_3,1_3,1_3&Prime; _builder_version=\u00a0\u00bb4.19.2&Prime; _module_preset=\u00a0\u00bbdefault\u00a0\u00bb global_colors_info=\u00a0\u00bb{}\u00a0\u00bb][et_pb_column type=\u00a0\u00bb1_3&Prime; _builder_version=\u00a0\u00bb4.19.2&Prime; _module_preset=\u00a0\u00bbdefault\u00a0\u00bb global_colors_info=\u00a0\u00bb{}\u00a0\u00bb][et_pb_image src=\u00a0\u00bbhttp:\/\/blog.vaniila-ai.catie-na.fr\/wp-content\/uploads\/2022\/12\/1_4.png\u00a0\u00bb title_text=\u00a0\u00bb1_4&Prime; _builder_version=\u00a0\u00bb4.19.2&Prime; _module_preset=\u00a0\u00bbdefault\u00a0\u00bb background_color=\u00a0\u00bb#FFFFFF\u00a0\u00bb custom_margin=\u00a0\u00bb18px||18px||true|false\u00a0\u00bb global_colors_info=\u00a0\u00bb{}\u00a0\u00bb][\/et_pb_image][\/et_pb_column][et_pb_column type=\u00a0\u00bb1_3&Prime; _builder_version=\u00a0\u00bb4.19.2&Prime; _module_preset=\u00a0\u00bbdefault\u00a0\u00bb global_colors_info=\u00a0\u00bb{}\u00a0\u00bb][et_pb_image src=\u00a0\u00bbhttp:\/\/blog.vaniila-ai.catie-na.fr\/wp-content\/uploads\/2022\/12\/1_5.png\u00a0\u00bb title_text=\u00a0\u00bb1_5&Prime; _builder_version=\u00a0\u00bb4.19.2&Prime; _module_preset=\u00a0\u00bbdefault\u00a0\u00bb background_color=\u00a0\u00bb#FFFFFF\u00a0\u00bb global_colors_info=\u00a0\u00bb{}\u00a0\u00bb][\/et_pb_image][\/et_pb_column][et_pb_column type=\u00a0\u00bb1_3&Prime; _builder_version=\u00a0\u00bb4.19.2&Prime; _module_preset=\u00a0\u00bbdefault\u00a0\u00bb global_colors_info=\u00a0\u00bb{}\u00a0\u00bb][et_pb_image src=\u00a0\u00bbhttp:\/\/blog.vaniila-ai.catie-na.fr\/wp-content\/uploads\/2022\/12\/1_6.png\u00a0\u00bb title_text=\u00a0\u00bb1_6&Prime; _builder_version=\u00a0\u00bb4.19.2&Prime; _module_preset=\u00a0\u00bbdefault\u00a0\u00bb background_color=\u00a0\u00bb#FFFFFF\u00a0\u00bb global_colors_info=\u00a0\u00bb{}\u00a0\u00bb][\/et_pb_image][\/et_pb_column][\/et_pb_row][et_pb_row _builder_version=\u00a0\u00bb4.19.2&Prime; _module_preset=\u00a0\u00bbdefault\u00a0\u00bb global_colors_info=\u00a0\u00bb{}\u00a0\u00bb][et_pb_column type=\u00a0\u00bb4_4&Prime; _builder_version=\u00a0\u00bb4.19.2&Prime; _module_preset=\u00a0\u00bbdefault\u00a0\u00bb global_colors_info=\u00a0\u00bb{}\u00a0\u00bb][et_pb_text _builder_version=\u00a0\u00bb4.19.2&Prime; _module_preset=\u00a0\u00bbdefault\u00a0\u00bb text_text_color=\u00a0\u00bb#FFFFFF\u00a0\u00bb hover_enabled=\u00a0\u00bb0&Prime; global_colors_info=\u00a0\u00bb{}\u00a0\u00bb sticky_enabled=\u00a0\u00bb0&Prime;]<h3 style=\"text-align: justify;\"><em><span style=\"color: #ffffff;\">Test sur une seule vue<\/span><\/em><\/h3>\n<p style=\"text-align: justify;\">Pour \u00e9viter que le mod\u00e8le ne se focalise sur les diff\u00e9rences entre vues, nous avons \u00e9galement test\u00e9 en n\u2019utilisant qu\u2019une seule vue. Le but est alors de diff\u00e9rencier les comprim\u00e9s normaux de ceux pr\u00e9sentant des d\u00e9faut, pour cette vue.<\/p>\n<p style=\"text-align: justify;\">Dans ce cas, l\u2019\u00e9tape de vectorisation ne converge pas et ne donne pas de s\u00e9paration entre les deux classes.<br \/><span data-token-index=\"0\" class=\"notion-enable-hover\" data-reactroot=\"\" style=\"color: #ffffff;\"><\/span><\/p>\n<p style=\"text-align: justify;\">L\u2019\u00e9tape de clustering proprement dite collapse, c\u2019est \u00e0 dire qu\u2019elle finit par assigner toutes les images au m\u00eame cluster.<\/p>\n<p style=\"text-align: justify;\"><span style=\"font-weight: 600; color: #ffffff;\" data-token-index=\"0\" class=\"notion-enable-hover\" data-reactroot=\"\"><\/span><\/p>[\/et_pb_text][\/et_pb_column][\/et_pb_row][et_pb_row column_structure=\u00a0\u00bb1_3,1_3,1_3&Prime; _builder_version=\u00a0\u00bb4.19.2&Prime; _module_preset=\u00a0\u00bbdefault\u00a0\u00bb global_colors_info=\u00a0\u00bb{}\u00a0\u00bb][et_pb_column type=\u00a0\u00bb1_3&Prime; _builder_version=\u00a0\u00bb4.19.2&Prime; _module_preset=\u00a0\u00bbdefault\u00a0\u00bb global_colors_info=\u00a0\u00bb{}\u00a0\u00bb][et_pb_image src=\u00a0\u00bbhttp:\/\/blog.vaniila-ai.catie-na.fr\/wp-content\/uploads\/2022\/12\/1_7.png\u00a0\u00bb title_text=\u00a0\u00bb1_7&Prime; _builder_version=\u00a0\u00bb4.19.2&Prime; _module_preset=\u00a0\u00bbdefault\u00a0\u00bb background_color=\u00a0\u00bb#FFFFFF\u00a0\u00bb custom_margin=\u00a0\u00bb29px||29px||true|false\u00a0\u00bb global_colors_info=\u00a0\u00bb{}\u00a0\u00bb][\/et_pb_image][\/et_pb_column][et_pb_column type=\u00a0\u00bb1_3&Prime; _builder_version=\u00a0\u00bb4.19.2&Prime; _module_preset=\u00a0\u00bbdefault\u00a0\u00bb global_colors_info=\u00a0\u00bb{}\u00a0\u00bb][et_pb_image src=\u00a0\u00bbhttp:\/\/blog.vaniila-ai.catie-na.fr\/wp-content\/uploads\/2022\/12\/1_8.png\u00a0\u00bb title_text=\u00a0\u00bb1_8&Prime; _builder_version=\u00a0\u00bb4.19.2&Prime; _module_preset=\u00a0\u00bbdefault\u00a0\u00bb background_color=\u00a0\u00bb#FFFFFF\u00a0\u00bb global_colors_info=\u00a0\u00bb{}\u00a0\u00bb][\/et_pb_image][\/et_pb_column][et_pb_column type=\u00a0\u00bb1_3&Prime; _builder_version=\u00a0\u00bb4.19.2&Prime; _module_preset=\u00a0\u00bbdefault\u00a0\u00bb global_colors_info=\u00a0\u00bb{}\u00a0\u00bb][et_pb_image src=\u00a0\u00bbhttp:\/\/blog.vaniila-ai.catie-na.fr\/wp-content\/uploads\/2022\/12\/1_9.png\u00a0\u00bb title_text=\u00a0\u00bb1_9&Prime; _builder_version=\u00a0\u00bb4.19.2&Prime; _module_preset=\u00a0\u00bbdefault\u00a0\u00bb background_color=\u00a0\u00bb#FFFFFF\u00a0\u00bb global_colors_info=\u00a0\u00bb{}\u00a0\u00bb][\/et_pb_image][\/et_pb_column][\/et_pb_row][et_pb_row _builder_version=\u00a0\u00bb4.19.2&Prime; _module_preset=\u00a0\u00bbdefault\u00a0\u00bb global_colors_info=\u00a0\u00bb{}\u00a0\u00bb][et_pb_column type=\u00a0\u00bb4_4&Prime; _builder_version=\u00a0\u00bb4.19.2&Prime; _module_preset=\u00a0\u00bbdefault\u00a0\u00bb global_colors_info=\u00a0\u00bb{}\u00a0\u00bb][et_pb_text _builder_version=\u00a0\u00bb4.19.2&Prime; _module_preset=\u00a0\u00bbc16985e1-e0d6-4022-964d-e2bfc04fa633&Prime; header_font=\u00a0\u00bbRoboto|700|||||||\u00a0\u00bb header_text_color=\u00a0\u00bbgcid-6171fd24-b893-4d22-843a-f4129850a5c1&Prime; header_font_size=\u00a0\u00bb75px\u00a0\u00bb header_line_height=\u00a0\u00bb1.2em\u00a0\u00bb text_orientation=\u00a0\u00bbcenter\u00a0\u00bb hover_enabled=\u00a0\u00bb0&Prime; header_font_size_tablet=\u00a0\u00bb40px\u00a0\u00bb header_font_size_phone=\u00a0\u00bb24px\u00a0\u00bb header_font_size_last_edited=\u00a0\u00bbon|desktop\u00a0\u00bb global_colors_info=\u00a0\u00bb{%22gcid-6171fd24-b893-4d22-843a-f4129850a5c1%22:%91%22header_text_color%22%93}\u00a0\u00bb sticky_enabled=\u00a0\u00bb0&Prime;]<h2 style=\"text-align: left;\"><span style=\"color: #ffffff;\"><strong>Conclusions<\/strong><\/span><\/h2>[\/et_pb_text][et_pb_text _builder_version=\u00a0\u00bb4.19.2&Prime; _module_preset=\u00a0\u00bbdefault\u00a0\u00bb text_text_color=\u00a0\u00bb#FFFFFF\u00a0\u00bb global_colors_info=\u00a0\u00bb{}\u00a0\u00bb]<p>La m\u00e9thode SCAN ne semble pas adapt\u00e9e pour traiter ce jeu de donn\u00e9es. Les d\u00e9tails des d\u00e9fauts \u00e0 identifier sont trop difficiles \u00e0 identifier sans supervision. Une piste de progr\u00e8s consisterait \u00e0 remplacer la tache pr\u00e9texte (actuellement via SimCLR) par une autre t\u00e2che, pour obtenir une meilleure vectorisation.<\/p>\n<p style=\"text-align: justify;\"><span style=\"font-weight: 600; color: #ffffff;\" data-token-index=\"0\" class=\"notion-enable-hover\" data-reactroot=\"\"><\/span><\/p>[\/et_pb_text][\/et_pb_column][\/et_pb_row][et_pb_row _builder_version=\u00a0\u00bb4.18.0&Prime; _module_preset=\u00a0\u00bbdefault\u00a0\u00bb max_width=\u00a0\u00bb700px\u00a0\u00bb global_colors_info=\u00a0\u00bb{}\u00a0\u00bb][et_pb_column type=\u00a0\u00bb4_4&Prime; _builder_version=\u00a0\u00bb4.18.0&Prime; _module_preset=\u00a0\u00bbdefault\u00a0\u00bb global_colors_info=\u00a0\u00bb{}\u00a0\u00bb][et_pb_divider color=\u00a0\u00bbgcid-6171fd24-b893-4d22-843a-f4129850a5c1&Prime; divider_weight=\u00a0\u00bb4px\u00a0\u00bb _builder_version=\u00a0\u00bb4.18.0&Prime; _module_preset=\u00a0\u00bbdefault\u00a0\u00bb max_width=\u00a0\u00bb300px\u00a0\u00bb module_alignment=\u00a0\u00bbcenter\u00a0\u00bb global_colors_info=\u00a0\u00bb{%22gcid-6171fd24-b893-4d22-843a-f4129850a5c1%22:%91%22color%22%93}\u00a0\u00bb][\/et_pb_divider][\/et_pb_column][\/et_pb_row][et_pb_row _builder_version=\u00a0\u00bb4.19.2&Prime; _module_preset=\u00a0\u00bbdefault\u00a0\u00bb global_colors_info=\u00a0\u00bb{}\u00a0\u00bb][et_pb_column type=\u00a0\u00bb4_4&Prime; _builder_version=\u00a0\u00bb4.19.2&Prime; _module_preset=\u00a0\u00bbdefault\u00a0\u00bb global_colors_info=\u00a0\u00bb{}\u00a0\u00bb][et_pb_text _builder_version=\u00a0\u00bb4.19.2&Prime; _module_preset=\u00a0\u00bbc16985e1-e0d6-4022-964d-e2bfc04fa633&Prime; header_font=\u00a0\u00bbRoboto|700|||||||\u00a0\u00bb header_text_color=\u00a0\u00bbgcid-6171fd24-b893-4d22-843a-f4129850a5c1&Prime; header_font_size=\u00a0\u00bb75px\u00a0\u00bb header_line_height=\u00a0\u00bb1.2em\u00a0\u00bb text_orientation=\u00a0\u00bbcenter\u00a0\u00bb hover_enabled=\u00a0\u00bb0&Prime; header_font_size_tablet=\u00a0\u00bb40px\u00a0\u00bb header_font_size_phone=\u00a0\u00bb24px\u00a0\u00bb header_font_size_last_edited=\u00a0\u00bbon|desktop\u00a0\u00bb global_colors_info=\u00a0\u00bb{%22gcid-6171fd24-b893-4d22-843a-f4129850a5c1%22:%91%22header_text_color%22%93}\u00a0\u00bb sticky_enabled=\u00a0\u00bb0&Prime;]<h2 style=\"text-align: left;\"><span style=\"color: #ffffff;\"><strong>Bibliographie<\/strong><\/span><\/h2>[\/et_pb_text][et_pb_text _builder_version=\u00a0\u00bb4.19.2&Prime; _module_preset=\u00a0\u00bbdefault\u00a0\u00bb text_text_color=\u00a0\u00bb#FFFFFF\u00a0\u00bb global_colors_info=\u00a0\u00bb{}\u00a0\u00bb]<p>VAN GANSBEKE, Wouter, VANDENHENDE, Simon, GEORGOULIS, Stamatios, <span style=\"font-style: italic;\" data-token-index=\"1\" class=\"notion-enable-hover\" data-reactroot=\"\">et al.<\/span> Scan: Learning to classify images without labels. In : <span style=\"font-style: italic;\" data-token-index=\"3\" class=\"notion-enable-hover\" data-reactroot=\"\">European Conference on Computer Vision<\/span>. Springer, Cham, 2020. p. 268-285.<\/p>\n<p style=\"text-align: justify;\"><span style=\"font-weight: 600; color: #ffffff;\" data-token-index=\"0\" class=\"notion-enable-hover\" data-reactroot=\"\"><\/span><\/p>[\/et_pb_text][\/et_pb_column][\/et_pb_row][\/et_pb_section]\n","protected":false},"excerpt":{"rendered":"<p>zdgaigfuefuafgliuQGVCUF<br \/>\nFOPFJOZIFHOZMFH<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_et_pb_use_builder":"on","_et_pb_old_content":"<!-- wp:paragraph -->\n<p>Welcome to WordPress. This is your first post. Edit or delete it, then start writing!<\/p>\n<!-- \/wp:paragraph -->","_et_gb_content_width":"","footnotes":""},"categories":[6],"tags":[],"class_list":["post-1","post","type-post","status-publish","format-standard","hentry","category-vision"],"_links":{"self":[{"href":"https:\/\/blog.vaniila-ai.catie-na.fr\/index.php?rest_route=\/wp\/v2\/posts\/1","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/blog.vaniila-ai.catie-na.fr\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/blog.vaniila-ai.catie-na.fr\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/blog.vaniila-ai.catie-na.fr\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/blog.vaniila-ai.catie-na.fr\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=1"}],"version-history":[{"count":18,"href":"https:\/\/blog.vaniila-ai.catie-na.fr\/index.php?rest_route=\/wp\/v2\/posts\/1\/revisions"}],"predecessor-version":[{"id":65,"href":"https:\/\/blog.vaniila-ai.catie-na.fr\/index.php?rest_route=\/wp\/v2\/posts\/1\/revisions\/65"}],"wp:attachment":[{"href":"https:\/\/blog.vaniila-ai.catie-na.fr\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=1"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/blog.vaniila-ai.catie-na.fr\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=1"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/blog.vaniila-ai.catie-na.fr\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=1"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}