{"id":6420,"date":"2022-12-25T20:45:33","date_gmt":"2022-12-25T12:45:33","guid":{"rendered":"https:\/\/i007.vip:15443\/wordpress\/?p=6420"},"modified":"2022-12-25T20:45:33","modified_gmt":"2022-12-25T12:45:33","slug":"%e4%b8%80%e6%96%87%e7%9c%8b%e6%87%8225%e4%b8%aa%e7%a5%9e%e7%bb%8f%e7%bd%91%e7%bb%9c%e6%a8%a1%e5%9e%8b","status":"publish","type":"post","link":"https:\/\/i007.cc\/wordpress\/archives\/6420","title":{"rendered":"\u4e00\u6587\u770b\u61c225\u4e2a\u795e\u7ecf\u7f51\u7edc\u6a21\u578b"},"content":{"rendered":"<div class=\"article-header-box\">\n<div class=\"article-header\">\n<div class=\"article-title-box\">\n<p id=\"articleContentId\" class=\"title-article\"><a href=\"https:\/\/blog.csdn.net\/qq_35082030\/article\/details\/73368962\">\u539f\u6587\u5730\u5740<\/a><\/p>\n<\/div>\n<div class=\"article-info-box\">\n<div class=\"article-bar-top\"><img decoding=\"async\" class=\"article-type-img\" src=\"https:\/\/csdnimg.cn\/release\/blogv2\/dist\/pc\/img\/reprint.png\" alt=\"\" \/><\/p>\n<div class=\"bar-content\"><span class=\"c-gray\">\u7f6e\u9876<\/span><a class=\"follow-nickName vip-name\" title=\"\u5218\u70ab320\" href=\"https:\/\/fjiang.blog.csdn.net\/\" target=\"_blank\" rel=\"noopener\">\u5218\u70ab320<\/a><img decoding=\"async\" class=\"article-time-img article-heard-img\" src=\"https:\/\/csdnimg.cn\/release\/blogv2\/dist\/pc\/img\/newUpTime2.png\" alt=\"\" \/><span class=\"time\">\u5df2\u4e8e\u00a02022-02-13 11:07:33\u00a0\u4fee\u6539<\/span><img decoding=\"async\" class=\"article-read-img article-heard-img\" src=\"https:\/\/csdnimg.cn\/release\/blogv2\/dist\/pc\/img\/articleReadEyes2.png\" alt=\"\" \/><span class=\"read-count\">211299<\/span><a id=\"blog_detail_zk_collection\" class=\"un-collection\" data-report-click=\"{\"mod\":\"popu_823\",\"spm\":\"1001.2101.3001.4232\",\"ab\":\"new\"}\"><\/a><img decoding=\"async\" class=\"article-collect-img article-heard-img un-collect-status isdefault\" src=\"https:\/\/csdnimg.cn\/release\/blogv2\/dist\/pc\/img\/tobarCollect2.png\" alt=\"\" \/>\u00a0<span class=\"name\">\u6536\u85cf<\/span>\u00a0<span class=\"get-collection\">1879<\/span><\/div>\n<\/div>\n<div class=\"blog-tags-box\">\n<div class=\"tags-box artic-tag-box\"><span class=\"label\">\u5206\u7c7b\u4e13\u680f\uff1a<\/span>\u00a0<a class=\"tag-link\" href=\"https:\/\/blog.csdn.net\/qq_35082030\/category_6977140.html\" target=\"_blank\" rel=\"noopener\">\u6df1\u5ea6\u5b66\u4e60<\/a>\u00a0<span class=\"label\">\u6587\u7ae0\u6807\u7b7e\uff1a<\/span>\u00a0<a class=\"tag-link\" href=\"https:\/\/so.csdn.net\/so\/search\/s.do?q=%E7%A5%9E%E7%BB%8F%E7%BD%91%E7%BB%9C&#038;t=blog&#038;o=vip&#038;s=&#038;l=&#038;f=&#038;viparticle=\" target=\"_blank\" rel=\"noopener\">\u795e\u7ecf\u7f51\u7edc<\/a>\u00a0<a class=\"tag-link\" href=\"https:\/\/so.csdn.net\/so\/search\/s.do?q=%E6%B7%B1%E5%BA%A6%E5%AD%A6%E4%B9%A0&#038;t=blog&#038;o=vip&#038;s=&#038;l=&#038;f=&#038;viparticle=\" target=\"_blank\" rel=\"noopener\">\u6df1\u5ea6\u5b66\u4e60<\/a>\u00a0<a class=\"tag-link\" href=\"https:\/\/so.csdn.net\/so\/search\/s.do?q=LSTM&#038;t=blog&#038;o=vip&#038;s=&#038;l=&#038;f=&#038;viparticle=\" target=\"_blank\" rel=\"noopener\">LSTM<\/a>\u00a0<a class=\"tag-link\" href=\"https:\/\/so.csdn.net\/so\/search\/s.do?q=%E5%AF%B9%E6%8A%97%E7%A5%9E%E7%BB%8F%E7%BD%91%E7%BB%9C&#038;t=blog&#038;o=vip&#038;s=&#038;l=&#038;f=&#038;viparticle=\" target=\"_blank\" rel=\"noopener\">\u5bf9\u6297\u795e\u7ecf\u7f51\u7edc<\/a>\u00a0<a class=\"tag-link\" href=\"https:\/\/so.csdn.net\/so\/search\/s.do?q=%E6%B7%B1%E5%BA%A6%E4%BF%A1%E5%BF%B5%E7%BD%91%E7%BB%9C&#038;t=blog&#038;o=vip&#038;s=&#038;l=&#038;f=&#038;viparticle=\" target=\"_blank\" rel=\"noopener\">\u6df1\u5ea6\u4fe1\u5ff5\u7f51\u7edc<\/a><\/div>\n<\/div>\n<div class=\"operating\"><a class=\"href-article-edit slide-toggle\">\u7248\u6743<\/a><\/div>\n<\/div>\n<\/div>\n<\/div>\n<div id=\"blogHuaweiyunAdvert\">\n<div class=\"column-group active\" data-id=\"142302\" data-url=\"https:\/\/aistudio.csdn.net?utm_source=blog_detail\" data-report-click=\"{\"spm\":\"1001.2101.3001.8542\"}\">\n<div class=\"column-group-item\">\n<div class=\"item-l\"><a class=\"item-target\" title=\"\u767e\u5ea6\u98de\u6868AI Studio\u793e\u533a\"><img decoding=\"async\" class=\"item-target\" src=\"https:\/\/devpress.csdnimg.cn\/3456490327f745e78d59b3232e07662a.png\" alt=\"\" \/><span class=\"title item-target\"><span class=\"tit\">\u767e\u5ea6\u98de\u6868AI Studio\u793e\u533a<\/span><span class=\"dec\">\u6587\u7ae0\u5df2\u88ab\u767e\u5ea6\u98de\u6868 AI Studio\u793e\u533a\u6536\u5f55 ?iPad\u3001\u673a\u68b0\u952e\u76d8\u3001\u65e0\u7ebf\u9f20\u6807\uff0c\u9650\u65f6\u514d\u8d39\u9886\uff01<\/span><\/span><\/a><\/div>\n<div class=\"item-r\"><a class=\"item-target article-column-bt join-huawei-community\">\u52a0\u5165\u793e\u533a<\/a><\/div>\n<\/div>\n<\/div>\n<\/div>\n<div id=\"blogColumnPayAdvert\">\n<div class=\"column-group\">\n<div class=\"column-group-item column-group0 column-group-item-one\">\n<div class=\"item-l\"><a class=\"item-target\" title=\"\u6df1\u5ea6\u5b66\u4e60\" href=\"https:\/\/blog.csdn.net\/qq_35082030\/category_6977140.html\" target=\"_blank\" rel=\"noopener\"><img decoding=\"async\" class=\"item-target\" src=\"https:\/\/img-blog.csdnimg.cn\/20201014180756922.png?x-oss-process=image\/resize,m_fixed,h_64,w_64\" alt=\"\" \/><span class=\"title item-target\"><span class=\"tit\">\u6df1\u5ea6\u5b66\u4e60<\/span><span class=\"dec\">\u4e13\u680f\u6536\u5f55\u8be5\u5185\u5bb9<\/span><\/span><\/a><\/div>\n<div class=\"item-m\">3 \u7bc7\u6587\u7ae021 \u8ba2\u9605<\/div>\n<div class=\"item-r\"><a class=\"item-target article-column-bt articleColumnFreeBt\" data-id=\"6977140\">\u8ba2\u9605\u4e13\u680f<\/a><\/div>\n<\/div>\n<\/div>\n<\/div>\n<article class=\"baidu_pl\">\n<div id=\"article_content\" class=\"article_content clearfix\">\n<div id=\"content_views\" class=\"markdown_views prism-atom-one-light\">\n<h1><a name=\"t0\"><\/a><a id=\"1__0\"><\/a>1. \u5f15\u8a00<\/h1>\n<p>\u5728<a class=\"hl hl-1\" href=\"https:\/\/so.csdn.net\/so\/search?q=%E6%B7%B1%E5%BA%A6%E5%AD%A6%E4%B9%A0&#038;spm=1001.2101.3001.7020\" target=\"_blank\" rel=\"noopener\" data-tit=\"\u6df1\u5ea6\u5b66\u4e60\" data-pretit=\"\u6df1\u5ea6\u5b66\u4e60\">\u6df1\u5ea6\u5b66\u4e60<\/a>\u5341\u5206\u706b\u70ed\u7684\u4eca\u5929\uff0c\u4e0d\u65f6\u4f1a\u6d8c\u73b0\u51fa\u5404\u79cd\u65b0\u578b\u7684\u4eba\u5de5\u795e\u7ecf\u7f51\u7edc\uff0c\u60f3\u8981\u5b9e\u65f6\u4e86\u89e3\u8fd9\u4e9b\u65b0\u578b\u795e\u7ecf\u7f51\u7edc\u7684\u67b6\u6784\u8fd8\u771f\u662f\u4e0d\u5bb9\u6613\u3002\u5149\u662f\u77e5\u9053\u5404\u5f0f\u5404\u6837\u7684\u795e\u7ecf\u7f51\u7edc\u6a21\u578b\u7f29\u5199\uff08\u5982\uff1aDCIGN\u3001BiLSTM\u3001DCGAN\u2026\u2026\u8fd8\u6709\u54ea\u4e9b\uff1f)\uff0c\u5c31\u5df2\u7ecf\u8ba9\u4eba\u62db\u67b6\u4e0d\u4f4f\u4e86\u3002<\/p>\n<p>\u56e0\u6b64\uff0c\u8fd9\u91cc\u6574\u7406\u51fa\u4e00\u4efd\u6e05\u5355\u6765\u68b3\u7406\u6240\u6709\u8fd9\u4e9b\u67b6\u6784\u3002\u5176\u4e2d\u5927\u90e8\u5206\u662f<a class=\"hl hl-1\" href=\"https:\/\/so.csdn.net\/so\/search?q=%E4%BA%BA%E5%B7%A5%E7%A5%9E%E7%BB%8F%E7%BD%91%E7%BB%9C&#038;spm=1001.2101.3001.7020\" target=\"_blank\" rel=\"noopener\" data-tit=\"\u4eba\u5de5\u795e\u7ecf\u7f51\u7edc\" data-pretit=\"\u4eba\u5de5\u795e\u7ecf\u7f51\u7edc\">\u4eba\u5de5\u795e\u7ecf\u7f51\u7edc<\/a>\uff0c\u4e5f\u6709\u4e00\u4e9b\u5b8c\u5168\u4e0d\u540c\u7684\u602a\u7269\u3002\u5c3d\u7ba1\u6240\u6709\u8fd9\u4e9b\u67b6\u6784\u90fd\u5404\u4e0d\u76f8\u540c\u3001\u529f\u80fd\u72ec\u7279\uff0c\u5f53\u6211\u5728\u753b\u5b83\u4eec\u7684\u8282\u70b9\u56fe\u65f6\u2026\u2026\u5176\u4e2d\u6f5c\u5728\u7684\u5173\u7cfb\u5f00\u59cb\u9010\u6e10\u6e05\u6670\u8d77\u6765\u3002<\/p>\n<p>\u628a\u8fd9\u4e9b\u67b6\u6784\u505a\u6210\u8282\u70b9\u56fe\uff0c\u4f1a\u5b58\u5728\u4e00\u4e2a\u95ee\u9898\uff1a\u5b83\u65e0\u6cd5\u5c55\u793a<a class=\"hl hl-1\" href=\"https:\/\/so.csdn.net\/so\/search?q=%E7%A5%9E%E7%BB%8F%E7%BD%91%E7%BB%9C&#038;spm=1001.2101.3001.7020\" target=\"_blank\" rel=\"noopener\" data-tit=\"\u795e\u7ecf\u7f51\u7edc\" data-pretit=\"\u795e\u7ecf\u7f51\u7edc\">\u795e\u7ecf\u7f51\u7edc<\/a>\u67b6\u6784\u5185\u90e8\u7684\u5de5\u4f5c\u539f\u7406\u3002\u4e3e\u4f8b\u6765\u8bf4\uff0c\u53d8\u5206\u81ea\u7f16\u7801\u673a\uff08VAE\uff1avariational autoencoders \uff09\u770b\u8d77\u6765\u8ddf\u81ea\u7f16\u7801\u673a\uff08AE\uff1aautoencoders\uff09\u5dee\u4e0d\u591a\uff0c\u4f46\u5b83\u4eec\u7684\u8bad\u7ec3\u8fc7\u7a0b\u5374\u5927\u4e0d\u76f8\u540c\u3002\u8bad\u7ec3\u540e\u7684\u6a21\u578b\u5728\u4f7f\u7528\u573a\u666f\u4e0a\u5dee\u522b\u66f4\u5927\uff1aVAE\u662f\u751f\u6210\u5668\uff0c\u901a\u8fc7\u63d2\u5165\u566a\u97f3\u6570\u636e\u6765\u83b7\u53d6\u65b0\u6837\u672c\uff1b\u800cAE\u4ec5\u4ec5\u662f\u628a\u4ed6\u4eec\u6240\u6536\u5230\u7684\u4efb\u4f55\u4fe1\u606f\u4f5c\u4e3a\u8f93\u5165\uff0c\u6620\u5c04\u5230\u201c\u8bb0\u5fc6\u4e2d\u201d\u6700\u76f8\u4f3c\u7684\u8bad\u7ec3\u6837\u672c\u4e0a\u3002<\/p>\n<p>\u5728\u4ecb\u7ecd\u4e0d\u540c\u6a21\u578b\u7684\u795e\u7ecf\u5143\u548c\u795e\u7ecf\u7ec6\u80de\u5c42\u4e4b\u95f4\u7684\u8fde\u63a5\u65b9\u5f0f\u524d\uff0c\u6211\u4eec\u4e00\u6b65\u4e00\u6b65\u6765\uff0c\u5148\u6765\u4e86\u89e3\u4e0d\u540c\u7684\u795e\u7ecf\u5143\u8282\u70b9\u5185\u90e8\u662f\u5982\u4f55\u5de5\u4f5c\u7684\u3002<\/p>\n<h2><a name=\"t1\"><\/a><a id=\"11__9\"><\/a>1.1 \u795e\u7ecf\u5143<\/h2>\n<p>\u5bf9\u4e0d\u540c\u7c7b\u578b\u7684\u795e\u7ecf\u5143\u6807\u8bb0\u4e0d\u540c\u7684\u989c\u8272\uff0c\u53ef\u4ee5\u66f4\u597d\u5730\u5728\u5404\u79cd\u7f51\u7edc\u67b6\u6784\u4e4b\u95f4\u8fdb\u884c\u533a\u5206\u3002\u4f46\u662f\uff0c\u8fd9\u4e9b\u795e\u7ecf\u5143\u7684\u5de5\u4f5c\u65b9\u5f0f\u5374\u662f\u5927\u540c\u5c0f\u5f02\u3002\u5728\u4e0b\u56fe\u7684\u57fa\u672c\u795e\u7ecf\u5143\u7ed3\u6784\u540e\u9762\uff0c\u4f60\u4f1a\u770b\u5230\u8be6\u7ec6\u7684\u8bb2\u89e3\uff1a<\/p>\n<p>\u57fa\u672c\u7684\u4eba\u5de5\u795e\u7ecf\u7f51\u7edc\u795e\u7ecf\u5143\uff08basic neural network cell\uff09\u76f8\u5f53\u7b80\u5355\uff0c\u8fd9\u79cd\u7b80\u5355\u7684\u7c7b\u578b\u53ef\u4ee5\u5728\u5e38\u89c4\u7684\u524d\u9988\u4eba\u5de5\u795e\u7ecf\u7f51\u7edc\u67b6\u6784\u91cc\u9762\u627e\u5230\u3002\u8fd9\u79cd\u795e\u7ecf\u5143\u4e0e\u5176\u5b83\u795e\u7ecf\u5143\u4e4b\u95f4\u7684\u8fde\u63a5\u5177\u6709\u6743\u91cd\uff0c\u4e5f\u5c31\u662f\u8bf4\uff0c\u5b83\u53ef\u4ee5\u548c\u524d\u4e00\u5c42\u795e\u7ecf\u7f51\u7edc\u5c42\u4e2d\u7684\u6240\u6709\u795e\u7ecf\u5143\u6709\u8fde\u63a5\u3002<\/p>\n<p>\u6bcf\u4e00\u4e2a\u8fde\u63a5\u90fd\u6709\u5404\u81ea\u7684\u6743\u91cd\uff0c\u901a\u5e38\u60c5\u51b5\u4e0b\u662f\u4e00\u4e9b\u968f\u673a\u503c\uff08\u5173\u4e8e\u5982\u4f55\u5bf9\u4eba\u5de5\u795e\u7ecf\u7f51\u7edc\u7684\u6743\u91cd\u8fdb\u884c\u521d\u59cb\u5316\u662f\u4e00\u4e2a\u975e\u5e38\u91cd\u8981\u7684\u8bdd\u9898\uff0c\u8fd9\u5c06\u4f1a\u76f4\u63a5\u5f71\u54cd\u5230\u4e4b\u540e\u7684\u8bad\u7ec3\u8fc7\u7a0b\uff0c\u4ee5\u53ca\u6700\u7ec8\u6574\u4e2a\u6a21\u578b\u7684\u6027\u80fd\uff09\u3002\u8fd9\u4e2a\u6743\u91cd\u53ef\u4ee5\u662f\u8d1f\u503c\uff0c\u6b63\u503c\uff0c\u975e\u5e38\u5c0f\uff0c\u6216\u8005\u975e\u5e38\u5927\uff0c\u4e5f\u53ef\u4ee5\u662f\u96f6\u3002\u548c\u8fd9\u4e2a\u795e\u7ecf\u5143\u8fde\u63a5\u7684\u6240\u6709\u795e\u7ecf\u5143\u7684\u503c\u90fd\u4f1a\u4e58\u4ee5\u5404\u81ea\u5bf9\u5e94\u7684\u6743\u91cd\u3002\u7136\u540e\uff0c\u628a\u8fd9\u4e9b\u503c\u90fd\u6c42\u548c\u3002<\/p>\n<p>\u5728\u8fd9\u4e2a\u57fa\u7840\u4e0a\uff0c\u4f1a\u989d\u5916\u52a0\u4e0a\u4e00\u4e2abias\uff0c\u5b83\u53ef\u4ee5\u7528\u6765\u907f\u514d\u8f93\u51fa\u4e3a\u96f6\u7684\u60c5\u51b5\uff0c\u5e76\u4e14\u80fd\u591f\u52a0\u901f\u67d0\u4e9b\u64cd\u4f5c\uff0c\u8fd9\u8ba9\u89e3\u51b3\u67d0\u4e2a\u95ee\u9898\u6240\u9700\u8981\u7684\u795e\u7ecf\u5143\u6570\u91cf\u4e5f\u6709\u6240\u51cf\u5c11\u3002\u8fd9\u4e2abias\u4e5f\u662f\u4e00\u4e2a\u6570\u5b57\uff0c\u6709\u4e9b\u65f6\u5019\u662f\u4e00\u4e2a\u5e38\u91cf\uff08\u7ecf\u5e38\u662f-1\u6216\u80051\uff09\uff0c\u6709\u4e9b\u65f6\u5019\u4f1a\u6709\u6240\u53d8\u5316\u3002\u8fd9\u4e2a\u603b\u548c\u6700\u7ec8\u88ab\u8f93\u5165\u5230\u4e00\u4e2a\u6fc0\u6d3b\u51fd\u6570\uff0c\u8fd9\u4e2a\u6fc0\u6d3b\u51fd\u6570\u7684\u8f93\u51fa\u6700\u7ec8\u5c31\u6210\u4e3a\u8fd9\u4e2a\u795e\u7ecf\u5143\u7684\u8f93\u51fa\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/img-blog.csdnimg.cn\/img_convert\/6018056f8c8ebf90c18f79cea91122f8.png#pic_center\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<h2><a name=\"t2\"><\/a><a id=\"12Convolutional_cells_21\"><\/a>1.2\u5377\u79ef\u795e\u7ecf\u5143\uff08Convolutional cells\uff09<\/h2>\n<p>\u548c\u524d\u9988\u795e\u7ecf\u5143\u975e\u5e38\u76f8\u4f3c\uff0c\u9664\u4e86\u5b83\u4eec\u53ea\u8ddf\u524d\u4e00\u795e\u7ecf\u7ec6\u80de\u5c42\u7684\u90e8\u5206\u795e\u7ecf\u5143\u6709\u8fde\u63a5\u3002\u56e0\u4e3a\u5b83\u4eec\u4e0d\u662f\u548c\u67d0\u4e9b\u795e\u7ecf\u5143\u968f\u673a\u8fde\u63a5\u7684\uff0c\u800c\u662f\u4e0e\u7279\u5b9a\u8303\u56f4\u5185\u7684\u795e\u7ecf\u5143\u76f8\u8fde\u63a5\uff0c\u901a\u5e38\u7528\u6765\u4fdd\u5b58\u7a7a\u95f4\u4fe1\u606f\u3002\u8fd9\u8ba9\u5b83\u4eec\u5bf9\u4e8e\u90a3\u4e9b\u62e5\u6709\u5927\u91cf\u5c40\u90e8\u4fe1\u606f\uff0c\u6bd4\u5982\u56fe\u50cf\u6570\u636e\u3001\u8bed\u97f3\u6570\u636e\uff08\u4f46\u591a\u6570\u60c5\u51b5\u4e0b\u662f\u56fe\u50cf\u6570\u636e\uff09\uff0c\u4f1a\u975e\u5e38\u5b9e\u7528\u3002<\/p>\n<h2><a name=\"t3\"><\/a><a id=\"13__24\"><\/a>1.3 \u89e3\u5377\u79ef\u795e\u7ecf\u5143<\/h2>\n<p>\u6070\u597d\u76f8\u53cd\uff1a\u5b83\u4eec\u662f\u901a\u8fc7\u8ddf\u4e0b\u4e00\u795e\u7ecf\u7ec6\u80de\u5c42\u7684\u8fde\u63a5\u6765\u89e3\u7801\u7a7a\u95f4\u4fe1\u606f\u3002\u8fd9\u4e24\u79cd\u795e\u7ecf\u5143\u90fd\u6709\u5f88\u591a\u526f\u672c\uff0c\u5b83\u4eec\u90fd\u662f\u72ec\u7acb\u8bad\u7ec3\u7684\uff1b\u6bcf\u4e2a\u526f\u672c\u90fd\u6709\u81ea\u5df1\u7684\u6743\u91cd\uff0c\u4f46\u8fde\u63a5\u65b9\u5f0f\u5374\u5b8c\u5168\u76f8\u540c\u3002\u53ef\u4ee5\u8ba4\u4e3a\uff0c\u8fd9\u4e9b\u526f\u672c\u662f\u88ab\u653e\u5728\u4e86\u5177\u5907\u76f8\u540c\u7ed3\u6784\u7684\u4e0d\u540c\u7684\u795e\u7ecf\u7f51\u7edc\u4e2d\u3002\u8fd9\u4e24\u79cd\u795e\u7ecf\u5143\u672c\u8d28\u4e0a\u90fd\u662f\u4e00\u822c\u610f\u4e49\u4e0a\u7684\u795e\u7ecf\u5143\uff0c\u4f46\u662f\uff0c\u5b83\u4eec\u7684\u4f7f\u7528\u65b9\u5f0f\u5374\u4e0d\u540c\u3002<\/p>\n<h2><a name=\"t4\"><\/a><a id=\"14_Pooling_and_interpolating_cells_27\"><\/a>1.4 \u6c60\u5316\u795e\u7ecf\u5143\u548c\u63d2\u503c\u795e\u7ecf\u5143\uff08Pooling and interpolating cells\uff09<\/h2>\n<p>\u7ecf\u5e38\u548c\u5377\u79ef\u795e\u7ecf\u5143\u7ed3\u5408\u8d77\u6765\u4f7f\u7528\u3002\u5b83\u4eec\u4e0d\u662f\u771f\u6b63\u610f\u4e49\u4e0a\u7684\u795e\u7ecf\u5143\uff0c\u53ea\u80fd\u8fdb\u884c\u4e00\u4e9b\u7b80\u5355\u7684\u64cd\u4f5c\u3002<\/p>\n<p>\u6c60\u5316\u795e\u7ecf\u5143\u63a5\u53d7\u5230\u6765\u81ea\u5176\u5b83\u795e\u7ecf\u5143\u7684\u8f93\u51fa\u8fc7\u540e\uff0c\u51b3\u5b9a\u54ea\u4e9b\u503c\u53ef\u4ee5\u901a\u8fc7\uff0c\u54ea\u4e9b\u503c\u4e0d\u80fd\u901a\u8fc7\u3002\u5728\u56fe\u50cf\u9886\u57df\uff0c\u53ef\u4ee5\u7406\u89e3\u6210\u662f\u628a\u4e00\u4e2a\u56fe\u50cf\u7f29\u5c0f\u4e86\uff08\u5728\u67e5\u770b\u56fe\u7247\u7684\u65f6\u5019\uff0c\u4e00\u822c\u8f6f\u4ef6\u90fd\u6709\u4e00\u4e2a\u653e\u5927\u3001\u7f29\u5c0f\u7684\u529f\u80fd\uff1b\u8fd9\u91cc\u7684\u56fe\u50cf\u7f29\u5c0f\uff0c\u5c31\u76f8\u5f53\u4e8e\u8f6f\u4ef6\u4e0a\u7684\u7f29\u5c0f\u56fe\u50cf\uff1b\u4e5f\u5c31\u662f\u8bf4\u6211\u4eec\u80fd\u770b\u5230\u56fe\u50cf\u7684\u5185\u5bb9\u66f4\u52a0\u5c11\u4e86\uff1b\u5728\u8fd9\u4e2a\u6c60\u5316\u7684\u8fc7\u7a0b\u5f53\u4e2d\uff0c\u56fe\u50cf\u7684\u5927\u5c0f\u4e5f\u4f1a\u76f8\u5e94\u5730\u51cf\u5c11\uff09\u3002\u8fd9\u6837\uff0c\u4f60\u5c31\u518d\u4e5f\u4e0d\u80fd\u770b\u5230\u6240\u6709\u7684\u50cf\u7d20\u4e86\uff0c\u6c60\u5316\u51fd\u6570\u4f1a\u77e5\u9053\u4ec0\u4e48\u50cf\u7d20\u8be5\u4fdd\u7559\uff0c\u4ec0\u4e48\u50cf\u7d20\u8be5\u820d\u5f03\u3002<\/p>\n<p>\u63d2\u503c\u795e\u7ecf\u5143\u6070\u597d\u662f\u76f8\u53cd\u7684\u64cd\u4f5c\uff1a\u5b83\u4eec\u83b7\u53d6\u4e00\u4e9b\u4fe1\u606f\uff0c\u7136\u540e\u6620\u5c04\u51fa\u66f4\u591a\u7684\u4fe1\u606f\u3002\u989d\u5916\u7684\u4fe1\u606f\u90fd\u662f\u6309\u7167\u67d0\u79cd\u65b9\u5f0f\u5236\u9020\u51fa\u6765\u7684\uff0c\u8fd9\u5c31\u597d\u50cf\u5728\u4e00\u5f20\u5c0f\u5206\u8fa8\u7387\u7684\u56fe\u7247\u4e0a\u9762\u8fdb\u884c\u653e\u5927\u3002\u63d2\u503c\u795e\u7ecf\u5143\u4e0d\u4ec5\u4ec5\u662f\u6c60\u5316\u795e\u7ecf\u5143\u7684\u53cd\u5411\u64cd\u4f5c\uff0c\u800c\u4e14\uff0c\u5b83\u4eec\u4e5f\u662f\u5f88\u5e38\u89c1\uff0c\u56e0\u4e3a\u5b83\u4eec\u8fd0\u884c\u975e\u5e38\u5feb\uff0c\u540c\u65f6\uff0c\u5b9e\u73b0\u8d77\u6765\u4e5f\u5f88\u7b80\u5355\u3002\u6c60\u5316\u795e\u7ecf\u5143\u548c\u63d2\u503c\u795e\u7ecf\u5143\u4e4b\u95f4\u7684\u5173\u7cfb\uff0c\u5c31\u50cf\u5377\u79ef\u795e\u7ecf\u5143\u548c\u89e3\u5377\u79ef\u795e\u7ecf\u5143\u4e4b\u95f4\u7684\u5173\u7cfb\u3002<\/p>\n<h2><a name=\"t5\"><\/a><a id=\"15Mean_and_standard_deviation_cells_34\"><\/a>1.5\u5747\u503c\u795e\u7ecf\u5143\u548c\u6807\u51c6\u65b9\u5dee\u795e\u7ecf\u5143\uff08Mean and standard deviation cells\uff09\uff08\u4f5c\u4e3a\u6982\u7387\u795e\u7ecf\u5143\u5b83\u4eec\u603b\u662f\u6210\u5bf9\u5730\u51fa\u73b0\uff09<\/h2>\n<p>\u662f\u4e00\u7c7b\u7528\u6765\u63cf\u8ff0\u6570\u636e\u6982\u7387\u5206\u5e03\u7684\u795e\u7ecf\u5143\u3002\u5747\u503c\u5c31\u662f\u6240\u6709\u503c\u7684\u5e73\u5747\u503c\uff0c\u800c\u6807\u51c6\u65b9\u5dee\u63cf\u8ff0\u7684\u662f\u8fd9\u4e9b\u6570\u636e\u504f\u79bb\uff08\u4e24\u4e2a\u65b9\u5411\uff09\u5747\u503c\u6709\u591a\u8fdc\u3002\u6bd4\u5982\uff1a\u4e00\u4e2a\u7528\u4e8e\u56fe\u50cf\u5904\u7406\u7684\u6982\u7387\u795e\u7ecf\u5143\u53ef\u4ee5\u5305\u542b\u4e00\u4e9b\u4fe1\u606f\uff0c\u6bd4\u5982\uff1a\u5728\u67d0\u4e2a\u7279\u5b9a\u7684\u50cf\u7d20\u91cc\u9762\u6709\u591a\u5c11\u7ea2\u8272\u3002\u4e3e\u4e2a\u4f8b\u6765\u8bf4\uff0c\u5747\u503c\u53ef\u80fd\u662f0.5\uff0c\u540c\u65f6\u6807\u51c6\u65b9\u5dee\u662f0.2\u3002\u5f53\u8981\u4ece\u8fd9\u4e9b\u6982\u7387\u795e\u7ecf\u5143\u53d6\u6837\u7684\u65f6\u5019\uff0c\u4f60\u53ef\u4ee5\u628a\u8fd9\u4e9b\u503c\u8f93\u5165\u5230\u4e00\u4e2a\u9ad8\u65af\u968f\u673a\u6570\u751f\u6210\u5668\uff0c\u8fd9\u6837\u5c31\u4f1a\u751f\u6210\u4e00\u4e9b\u5206\u5e03\u57280.4\u548c0.6\u4e4b\u95f4\u7684\u503c\uff1b\u503c\u79bb0.5\u8d8a\u8fdc\uff0c\u5bf9\u5e94\u751f\u6210\u7684\u6982\u7387\u4e5f\u5c31\u8d8a\u5c0f\u3002\u5b83\u4eec\u4e00\u822c\u548c\u524d\u4e00\u795e\u7ecf\u5143\u5c42\u6216\u8005\u4e0b\u4e00\u795e\u7ecf\u5143\u5c42\u662f\u5168\u8fde\u63a5\uff0c\u800c\u4e14\uff0c\u5b83\u4eec\u6ca1\u6709\u504f\u5dee\uff08bias\uff09\u3002<\/p>\n<h2><a name=\"t6\"><\/a><a id=\"16_Recurrent_cells__37\"><\/a>1.6 \u5faa\u73af\u795e\u7ecf\u5143\uff08Recurrent cells \uff09<\/h2>\n<p>\u4e0d\u4ec5\u4ec5\u5728\u795e\u7ecf\u7ec6\u80de\u5c42\u4e4b\u95f4\u6709\u8fde\u63a5\uff0c\u800c\u4e14\u5728\u65f6\u95f4\u8f74\u4e0a\u4e5f\u6709\u76f8\u5e94\u7684\u8fde\u63a5\u3002\u6bcf\u4e00\u4e2a\u795e\u7ecf\u5143\u5185\u90e8\u90fd\u4f1a\u4fdd\u5b58\u5b83\u5148\u524d\u7684\u503c\u3002\u5b83\u4eec\u8ddf\u4e00\u822c\u7684\u795e\u7ecf\u5143\u4e00\u6837\u66f4\u65b0\uff0c\u4f46\u662f\uff0c\u5177\u6709\u989d\u5916\u7684\u6743\u91cd\uff1a\u4e0e\u5f53\u524d\u795e\u7ecf\u5143\u4e4b\u524d\u503c\u4e4b\u95f4\u7684\u6743\u91cd\uff0c\u8fd8\u6709\u5927\u591a\u6570\u60c5\u51b5\u4e0b\uff0c\u4e0e\u540c\u4e00\u795e\u7ecf\u7ec6\u80de\u5c42\u5404\u4e2a\u795e\u7ecf\u5143\u4e4b\u95f4\u7684\u6743\u91cd\u3002\u5f53\u524d\u503c\u548c\u5b58\u50a8\u7684\u5148\u524d\u503c\u4e4b\u95f4\u6743\u91cd\u7684\u5de5\u4f5c\u673a\u5236\uff0c\u4e0e\u975e\u6c38\u4e45\u6027\u5b58\u50a8\u5668\uff08\u6bd4\u5982RAM\uff09\u7684\u5de5\u4f5c\u673a\u5236\u5f88\u76f8\u4f3c\uff0c\u7ee7\u627f\u4e86\u4e24\u4e2a\u6027\u8d28\uff1a<\/p>\n<ul>\n<li>\u7b2c\u4e00\uff0c\u7ef4\u6301\u4e00\u4e2a\u7279\u5b9a\u7684\u72b6\u6001\uff1b<\/li>\n<li>\u7b2c\u4e8c\uff1a\u5982\u679c\u4e0d\u5bf9\u5176\u6301\u7eed\u8fdb\u884c\u66f4\u65b0\uff08\u8f93\u5165\uff09\uff0c\u8fd9\u4e2a\u72b6\u6001\u5c31\u4f1a\u6d88\u5931\u3002<\/li>\n<\/ul>\n<p>\u7531\u4e8e\u5148\u524d\u7684\u503c\u662f\u901a\u8fc7\u6fc0\u6d3b\u51fd\u6570\u5f97\u5230\u7684\uff0c\u800c\u5728\u6bcf\u4e00\u6b21\u7684\u66f4\u65b0\u65f6\uff0c\u90fd\u4f1a\u628a\u8fd9\u4e2a\u503c\u548c\u5176\u5b83\u6743\u91cd\u4e00\u8d77\u8f93\u5165\u5230\u6fc0\u6d3b\u51fd\u6570\uff0c\u56e0\u6b64\uff0c\u4fe1\u606f\u4f1a\u4e0d\u65ad\u5730\u6d41\u5931\u3002\u5b9e\u9645\u4e0a\uff0c\u4fe1\u606f\u7684\u4fdd\u5b58\u7387\u975e\u5e38\u7684\u4f4e\uff0c\u4ee5\u81f3\u4e8e\u4ec5\u4ec5\u56db\u6b21\u6216\u8005\u4e94\u6b21\u8fed\u4ee3\u66f4\u65b0\u8fc7\u540e\uff0c\u51e0\u4e4e\u4e4b\u524d\u6240\u6709\u7684\u4fe1\u606f\u90fd\u4f1a\u6d41\u5931\u6389\u3002<\/p>\n<h2><a name=\"t7\"><\/a><a id=\"17_Long_short_term_memory_cells_45\"><\/a>1.7 \u957f\u77ed\u671f\u8bb0\u5fc6\u795e\u7ecf\u5143\uff08Long short term memory cells\uff09<\/h2>\n<p>\u7528\u4e8e\u514b\u670d\u5faa\u73af\u795e\u7ecf\u5143\u4e2d\u4fe1\u606f\u5feb\u901f\u6d41\u5931\u7684\u95ee\u9898\u3002<\/p>\n<p><a class=\"hl hl-1\" href=\"https:\/\/so.csdn.net\/so\/search?q=LSTM&#038;spm=1001.2101.3001.7020\" target=\"_blank\" rel=\"noopener\" data-tit=\"LSTM\" data-pretit=\"lstm\">LSTM<\/a>\u662f\u4e00\u4e2a\u903b\u8f91\u56de\u8def\uff0c\u5176\u8bbe\u8ba1\u53d7\u5230\u4e86\u8ba1\u7b97\u673a\u5185\u5b58\u5355\u5143\u8bbe\u8ba1\u7684\u542f\u53d1\u3002\u4e0e\u53ea\u5b58\u50a8\u4e24\u4e2a\u72b6\u6001\u7684\u5faa\u73af\u795e\u7ecf\u5143\u76f8\u6bd4\uff0cLSTM\u53ef\u4ee5\u5b58\u50a8\u56db\u4e2a\u72b6\u6001\uff1a\u8f93\u51fa\u503c\u7684\u5f53\u524d\u548c\u5148\u524d\u503c\uff0c\u8bb0\u5fc6\u795e\u7ecf\u5143\u72b6\u6001\u7684\u5f53\u524d\u503c\u548c\u5148\u524d\u503c\u3002\u5b83\u4eec\u90fd\u6709\u4e09\u4e2a\u95e8\uff1a\u8f93\u5165\u95e8\uff0c\u8f93\u51fa\u95e8\uff0c\u9057\u5fd8\u95e8\uff0c\u540c\u65f6\uff0c\u5b83\u4eec\u4e5f\u8fd8\u6709\u5e38\u89c4\u7684\u8f93\u5165\u3002<\/p>\n<p>\u8fd9\u4e9b\u95e8\u5b83\u4eec\u90fd\u6709\u5404\u81ea\u7684\u6743\u91cd\uff0c\u4e5f\u5c31\u662f\u8bf4\uff0c\u4e0e\u8fd9\u79cd\u7c7b\u578b\u7684\u795e\u7ecf\u5143\u7ec6\u80de\u8fde\u63a5\u9700\u8981\u8bbe\u7f6e\u56db\u4e2a\u6743\u91cd\uff08\u800c\u4e0d\u662f\u4e00\u4e2a\uff09\u3002\u8fd9\u4e9b\u95e8\u7684\u5de5\u4f5c\u673a\u5236\u4e0e\u6d41\u95e8\uff08flow gates\uff09\u5f88\u76f8\u4f3c\uff0c\u800c\u4e0d\u662f\u6805\u680f\u95e8\uff08fence gates\uff09\uff1a\u5b83\u4eec\u53ef\u4ee5\u8ba9\u6240\u6709\u7684\u4fe1\u606f\u90fd\u901a\u8fc7\uff0c\u6216\u8005\u53ea\u662f\u901a\u8fc7\u90e8\u5206\uff0c\u4e5f\u53ef\u4ee5\u4ec0\u4e48\u90fd\u4e0d\u8ba9\u901a\u8fc7\uff0c\u6216\u8005\u901a\u8fc7\u67d0\u4e2a\u533a\u95f4\u7684\u4fe1\u606f\u3002<\/p>\n<p>\u8fd9\u79cd\u8fd0\u884c\u673a\u5236\u7684\u5b9e\u73b0\u662f\u901a\u8fc7\u628a\u8f93\u5165\u4fe1\u606f\u548c\u4e00\u4e2a\u57280\u52301\u4e4b\u95f4\u7684\u7cfb\u6570\u76f8\u4e58\uff0c\u8fd9\u4e2a\u7cfb\u6570\u5b58\u50a8\u5728\u5f53\u524d\u95e8\u4e2d\u3002\u8fd9\u6837\uff0c\u8f93\u5165\u95e8\u51b3\u5b9a\u8f93\u5165\u7684\u4fe1\u606f\u6709\u591a\u5c11\u53ef\u4ee5\u88ab\u53e0\u52a0\u5230\u5f53\u524d\u95e8\u503c\u3002\u8f93\u51fa\u95e8\u51b3\u5b9a\u6709\u591a\u5c11\u8f93\u51fa\u4fe1\u606f\u662f\u53ef\u4ee5\u4f20\u9012\u5230\u540e\u9762\u7684\u795e\u7ecf\u7f51\u7edc\u4e2d\u3002\u9057\u5fd8\u95e8\u5e76\u4e0d\u662f\u548c\u8f93\u51fa\u795e\u7ecf\u5143\u7684\u5148\u524d\u503c\u76f8\u8fde\u63a5\uff0c\u800c\u662f\uff0c\u548c\u524d\u4e00\u8bb0\u5fc6\u795e\u7ecf\u5143\u76f8\u8fde\u63a5\u3002\u5b83\u51b3\u5b9a\u4e86\u4fdd\u7559\u591a\u5c11\u8bb0\u5fc6\u795e\u7ecf\u5143\u6700\u65b0\u7684\u72b6\u6001\u4fe1\u606f\u3002\u56e0\u4e3a\u6ca1\u6709\u548c\u8f93\u51fa\u76f8\u8fde\u63a5\uff0c\u4ee5\u53ca\u6ca1\u6709\u6fc0\u6d3b\u51fd\u6570\u5728\u8fd9\u4e2a\u5faa\u73af\u4e2d\uff0c\u56e0\u6b64\u53ea\u4f1a\u6709\u66f4\u5c11\u7684\u4fe1\u606f\u6d41\u5931\u3002<\/p>\n<h2><a name=\"t8\"><\/a><a id=\"18_Gated_recurrent_units_cells_54\"><\/a>1.8 \u95e8\u63a7\u5faa\u73af\u795e\u7ecf\u5143\uff08Gated recurrent units (cells)\uff09<\/h2>\n<p>\u662fLSTM\u7684\u53d8\u4f53\u3002\u5b83\u4eec\u540c\u6837\u4f7f\u7528\u95e8\u6765\u6291\u5236\u4fe1\u606f\u7684\u6d41\u5931\uff0c\u4f46\u662f\u53ea\u7528\u4e24\u4e2a\u95e8\uff1a\u66f4\u65b0\u95e8\u548c\u91cd\u7f6e\u95e8\u3002\u8fd9\u4f7f\u5f97\u6784\u5efa\u5b83\u4eec\u4ed8\u51fa\u7684\u4ee3\u4ef7\u6ca1\u6709\u90a3\u4e48\u9ad8\uff0c\u800c\u4e14\u8fd0\u884c\u901f\u5ea6\u66f4\u52a0\u5feb\u4e86\uff0c\u56e0\u4e3a\u5b83\u4eec\u5728\u6240\u6709\u7684\u5730\u65b9\u4f7f\u7528\u4e86\u66f4\u5c11\u7684\u8fde\u63a5\u3002<\/p>\n<p>\u4ece\u672c\u8d28\u4e0a\u6765\u8bf4LSTM\u548cGRU\u6709\u4e24\u4e2a\u4e0d\u540c\u7684\u5730\u65b9\uff1a<\/p>\n<ul>\n<li>\u7b2c\u4e00\uff1aGRU\u795e\u7ecf\u5143\u6ca1\u6709\u88ab\u8f93\u51fa\u95e8\u4fdd\u62a4\u7684\u9690\u795e\u7ecf\u5143\uff1b<\/li>\n<li>\u7b2c\u4e8c\uff1aGRU\u628a\u8f93\u51fa\u95e8\u548c\u9057\u5fd8\u95e8\u6574\u5408\u5728\u4e86\u4e00\u8d77\uff0c\u5f62\u6210\u4e86\u66f4\u65b0\u95e8\u3002\u6838\u5fc3\u7684\u601d\u60f3\u5c31\u662f\u5982\u679c\u4f60\u60f3\u8981\u4e00\u4e9b\u65b0\u7684\u4fe1\u606f\uff0c\u90a3\u4e48\u4f60\u5c31\u53ef\u4ee5\u9057\u5fd8\u6389\u4e00\u4e9b\u9648\u65e7\u7684\u4fe1\u606f\uff08\u53cd\u8fc7\u6765\u4e5f\u53ef\u4ee5\uff09\u3002<\/li>\n<\/ul>\n<h2><a name=\"t9\"><\/a><a id=\"19_Layers_61\"><\/a>1.9 \u795e\u7ecf\u7ec6\u80de\u5c42(Layers)<\/h2>\n<p>\u5f62\u6210\u4e00\u4e2a\u795e\u7ecf\u7f51\u7edc\uff0c\u6700\u7b80\u5355\u7684\u8fde\u63a5\u795e\u7ecf\u5143\u65b9\u5f0f\u662f\u2014\u2014\u628a\u6240\u6709\u7684\u795e\u7ecf\u5143\u4e0e\u5176\u5b83\u6240\u6709\u7684\u795e\u7ecf\u5143\u76f8\u8fde\u63a5\u3002\u8fd9\u5c31\u597d\u50cfHopfield\u795e\u7ecf\u7f51\u7edc\u548c\u73bb\u5c14\u5179\u66fc\u673a\uff08Boltzmann machines\uff09\u7684\u8fde\u63a5\u65b9\u5f0f\u3002\u5f53\u7136\uff0c\u8fd9\u4e5f\u5c31\u610f\u5473\u7740\u8fde\u63a5\u6570\u91cf\u4f1a\u968f\u7740\u795e\u7ecf\u5143\u4e2a\u6570\u7684\u589e\u52a0\u5448\u6307\u6570\u7ea7\u5730\u589e\u52a0\uff0c\u4f46\u662f\uff0c\u5bf9\u5e94\u7684\u51fd\u6570\u8868\u8fbe\u529b\u4e5f\u4f1a\u8d8a\u6765\u8d8a\u5f3a\u3002\u8fd9\u5c31\u662f\u6240\u8c13\u7684\u5168\u8fde\u63a5\uff08completely (or fully) connected\uff09\u3002<\/p>\n<p>\u7ecf\u5386\u4e86\u4e00\u6bb5\u65f6\u95f4\u7684\u53d1\u5c55\uff0c\u53d1\u73b0\u628a\u795e\u7ecf\u7f51\u7edc\u5206\u89e3\u6210\u4e0d\u540c\u7684\u795e\u7ecf\u7ec6\u80de\u5c42\u4f1a\u975e\u5e38\u6709\u6548\u3002\u795e\u7ecf\u7ec6\u80de\u5c42\u7684\u5b9a\u4e49\u662f\u4e00\u7fa4\u5f7c\u6b64\u4e4b\u95f4\u4e92\u4e0d\u8fde\u63a5\u7684\u795e\u7ecf\u5143\uff0c\u5b83\u4eec\u4ec5\u8ddf\u5176\u5b83\u795e\u7ecf\u7ec6\u80de\u5c42\u6709\u8fde\u63a5\u3002\u8fd9\u4e00\u6982\u5ff5\u5728\u53d7\u9650\u73bb\u5c14\u5179\u66fc\u673a\uff08Restricted Boltzmann Machines\uff09\u4e2d\u6709\u6240\u4f53\u73b0\u3002\u73b0\u5728\uff0c\u4f7f\u7528\u795e\u7ecf\u7f51\u7edc\u5c31\u610f\u5473\u7740\u4f7f\u7528\u795e\u7ecf\u7ec6\u80de\u5c42\uff0c\u5e76\u4e14\u662f\u4efb\u610f\u6570\u91cf\u7684\u795e\u7ecf\u7ec6\u80de\u5c42\u3002\u5176\u4e2d\u4e00\u4e2a\u6bd4\u8f83\u4ee4\u4eba\u56f0\u60d1\u7684\u6982\u5ff5\u662f\u5168\u8fde\u63a5\uff08fully connected or completely connected\uff09\uff0c\u4e5f\u5c31\u662f\u67d0\u4e00\u5c42\u7684\u6bcf\u4e2a\u795e\u7ecf\u5143\u8ddf\u53e6\u4e00\u5c42\u7684\u6240\u6709\u795e\u7ecf\u5143\u90fd\u6709\u8fde\u63a5\uff0c\u4f46\u771f\u6b63\u7684\u5168\u8fde\u63a5\u795e\u7ecf\u7f51\u7edc\u76f8\u5f53\u7f55\u89c1\u3002<\/p>\n<h2><a name=\"t10\"><\/a><a id=\"110_Convolutionally_connected_layers_67\"><\/a>1.10 \u5377\u79ef\u8fde\u63a5\u5c42\uff08Convolutionally connected layers\uff09<\/h2>\n<p>\u76f8\u5bf9\u4e8e\u5168\u8fde\u63a5\u5c42\u8981\u6709\u66f4\u591a\u7684\u9650\u5236\uff1a\u5728\u5377\u79ef\u8fde\u63a5\u5c42\u4e2d\u7684\u6bcf\u4e00\u4e2a\u795e\u7ecf\u5143\u53ea\u4e0e\u76f8\u90bb\u7684\u795e\u7ecf\u5143\u5c42\u8fde\u63a5\u3002\u56fe\u50cf\u548c\u58f0\u97f3\u8574\u542b\u4e86\u5927\u91cf\u7684\u4fe1\u606f\uff0c\u5982\u679c\u4e00\u5bf9\u4e00\u5730\u8f93\u5165\u5230\u795e\u7ecf\u7f51\u7edc\uff08\u6bd4\u5982\uff0c\u4e00\u4e2a\u795e\u7ecf\u5143\u5bf9\u5e94\u4e00\u4e2a\u50cf\u7d20\uff09\u3002\u5377\u79ef\u8fde\u63a5\u7684\u5f62\u6210\uff0c\u53d7\u76ca\u4e8e\u4fdd\u7559\u7a7a\u95f4\u4fe1\u606f\u66f4\u4e3a\u91cd\u8981\u7684\u89c2\u5bdf\u3002\u5b9e\u8df5\u8bc1\u660e\u8fd9\u662f\u4e00\u4e2a\u975e\u5e38\u597d\u7684\u731c\u6d4b\uff0c\u56e0\u4e3a\u73b0\u5728\u5927\u591a\u6570\u57fa\u4e8e\u4eba\u5de5\u795e\u7ecf\u7f51\u7edc\u7684\u56fe\u50cf\u548c\u8bed\u97f3\u5e94\u7528\u90fd\u4f7f\u7528\u4e86\u8fd9\u79cd\u8fde\u63a5\u65b9\u5f0f\u3002\u7136\u800c\uff0c\u8fd9\u79cd\u8fde\u63a5\u65b9\u5f0f\u6240\u9700\u7684\u4ee3\u4ef7\u8fdc\u8fdc\u4f4e\u4e8e\u5168\u8fde\u63a5\u5c42\u7684\u5f62\u5f0f\u3002\u4ece\u672c\u8d28\u4e0a\u6765\u8bb2\uff0c\u5377\u79ef\u8fde\u63a5\u65b9\u5f0f\u8d77\u5230\u91cd\u8981\u6027\u8fc7\u6ee4\u7684\u4f5c\u7528\uff0c\u51b3\u5b9a\u54ea\u4e9b\u7d27\u7d27\u8054\u7cfb\u5728\u4e00\u8d77\u7684\u4fe1\u606f\u5305\u662f\u91cd\u8981\u7684\uff1b\u5377\u79ef\u8fde\u63a5\u5bf9\u4e8e\u6570\u636e\u964d\u7ef4\u975e\u5e38\u6709\u7528\u3002<\/p>\n<p>\u5f53\u7136\u4e86\uff0c\u8fd8\u6709\u53e6\u5916\u4e00\u79cd\u9009\u62e9\uff0c\u5c31\u662f\u968f\u673a\u8fde\u63a5\u795e\u7ecf\u5143\uff08randomly connected neurons\uff09\u3002\u8fd9\u79cd\u5f62\u5f0f\u7684\u8fde\u63a5\u4e3b\u8981\u6709\u4e24\u79cd\u53d8\u4f53\uff1a<\/p>\n<ul>\n<li>\u7b2c\u4e00\uff0c\u5141\u8bb8\u90e8\u5206\u795e\u7ecf\u5143\u8fdb\u884c\u5168\u8fde\u63a5\u3002<\/li>\n<li>\u7b2c\u4e8c\uff0c\u795e\u7ecf\u5143\u5c42\u4e4b\u95f4\u53ea\u6709\u90e8\u5206\u8fde\u63a5\u3002<br \/>\n\u968f\u673a\u8fde\u63a5\u65b9\u5f0f\u6709\u52a9\u4e8e\u7ebf\u6027\u5730\u964d\u4f4e\u4eba\u5de5\u795e\u7ecf\u7f51\u7edc\u7684\u6027\u80fd\uff1b\u5f53\u5168\u8fde\u63a5\u5c42\u9047\u5230\u6027\u80fd\u95ee\u9898\u7684\u65f6\u5019\uff0c\u5728\u5927\u89c4\u6a21\u4eba\u5de5\u795e\u7ecf\u7f51\u7edc\u4e2d\uff0c\u4f7f\u7528\u968f\u673a\u8fde\u63a5\u65b9\u5f0f\u975e\u5e38\u6709\u76ca\u3002\u62e5\u6709\u66f4\u591a\u795e\u7ecf\u5143\u4e14\u66f4\u52a0\u7a00\u758f\u7684\u795e\u7ecf\u5143\u5c42\u5728\u67d0\u4e9b\u60c5\u51b5\u4e0b\u8fd0\u884c\u6548\u679c\u66f4\u597d\uff0c\u7279\u522b\u662f\u5f88\u591a\u7684\u4fe1\u606f\u9700\u8981\u88ab\u5b58\u50a8\u8d77\u6765\uff0c\u4f46\u662f\uff0c\u9700\u8981\u4ea4\u6362\u7684\u4fe1\u606f\u5e76\u4e0d\u591a\uff08\u8fd9\u4e0e\u5377\u79ef\u8fde\u63a5\u5c42\u7684\u8fd0\u884c\u673a\u5236\u5f88\u76f8\u4f3c\uff0c\u4f46\u662f\uff0c\u5b83\u4eec\u662f\u968f\u673a\u7684\uff09\u3002\u975e\u5e38\u7a00\u758f\u7684\u8fde\u63a5\u7f51\u7edc\uff081%\u62162%\uff09\u4e5f\u6709\u88ab\u4f7f\u7528\uff0c\u6bd4\u5982ELMs, ESNs \u548cLSMs\u3002\u8fd9\u7279\u522b\u9002\u7528\u4e8e\u8109\u51b2\u7f51\u7edc\uff08spiking networks\uff09\uff0c\u56e0\u4e3a\u4e00\u4e2a\u795e\u7ecf\u5143\u62e5\u6709\u66f4\u591a\u7684\u8fde\u63a5\uff0c\u5b83\u5bf9\u5e94\u7684\u6743\u91cd\u5177\u6709\u7684\u80fd\u91cf\u4e5f\u5c31\u66f4\u5c11\uff0c\u8fd9\u4e5f\u5c31\u610f\u5473\u7740\u5c06\u4f1a\u6709\u66f4\u5c11\u7684\u6269\u5c55\u548c\u91cd\u590d\u6a21\u5f0f\u3002<\/li>\n<\/ul>\n<h2><a name=\"t11\"><\/a><a id=\"111_Time_delayed_connections_76\"><\/a>1.11 \u65f6\u95f4\u6ede\u540e\u8fde\u63a5\uff08Time delayed connections\uff09<\/h2>\n<p>\u662f\u6307\u76f8\u8fde\u7684\u795e\u7ecf\u5143\uff08\u901a\u5e38\u662f\u5728\u540c\u4e00\u4e2a\u795e\u7ecf\u5143\u5c42\uff0c\u751a\u81f3\u4e8e\u4e00\u4e2a\u795e\u7ecf\u5143\u81ea\u5df1\u8ddf\u81ea\u5df1\u8fde\u63a5\uff09\uff0c\u5b83\u4eec\u4e0d\u4ece\u524d\u9762\u7684\u795e\u7ecf\u5143\u5c42\u83b7\u53d6\u4fe1\u606f\uff0c\u800c\u662f\u4ece\u795e\u7ecf\u5143\u5c42\u5148\u524d\u7684\u72b6\u6001\u83b7\u53d6\u4fe1\u606f\u3002\u8fd9\u4f7f\u5f97\u6682\u65f6\uff08\u65f6\u95f4\u4e0a\u6216\u8005\u5e8f\u5217\u4e0a\uff09\u8054\u7cfb\u5728\u4e00\u8d77\u7684\u4fe1\u606f\u53ef\u4ee5\u88ab\u5b58\u50a8\u8d77\u6765\u3002\u8fd9\u4e9b\u5f62\u5f0f\u7684\u8fde\u63a5\u7ecf\u5e38\u88ab\u624b\u5de5\u91cd\u65b0\u8fdb\u884c\u8bbe\u7f6e\uff0c\u4ece\u800c\u53ef\u4ee5\u6e05\u9664\u795e\u7ecf\u7f51\u7edc\u7684\u72b6\u6001\u3002\u548c\u5e38\u89c4\u8fde\u63a5\u7684\u4e3b\u8981\u533a\u522b\u662f\uff0c\u8fd9\u79cd\u8fde\u63a5\u4f1a\u6301\u7eed\u4e0d\u65ad\u5730\u6539\u53d8\uff0c\u5373\u4fbf\u8fd9\u4e2a\u795e\u7ecf\u7f51\u7edc\u5f53\u524d\u6ca1\u6709\u5904\u4e8e\u8bad\u7ec3\u72b6\u6001\u3002<\/p>\n<p>\u4e0b\u56fe\u5c55\u793a\u4e86\u4ee5\u4e0a\u6240\u4ecb\u7ecd\u7684\u795e\u7ecf\u7f51\u7edc\u53ca\u5176\u8fde\u63a5\u65b9\u5f0f\u3002\u5f53\u6211\u5361\u5728\u54ea\u79cd\u795e\u7ecf\u5143\u4e0e\u54ea\u4e2a\u795e\u7ecf\u7ec6\u80de\u5c42\u8be5\u8fde\u5230\u4e00\u8d77\u7684\u65f6\u5019\uff0c\u5c31\u4f1a\u62ff\u8fd9\u5f20\u56fe\u51fa\u6765\u4f5c\u4e3a\u53c2\u8003\uff08\u5c24\u5176\u662f\u5728\u5904\u7406\u548c\u5206\u6790LSTM\u4e0eGRU\u795e\u7ecf\u5143\u65f6\uff09\uff1a<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/img-blog.csdnimg.cn\/img_convert\/d9e8e1c74d95bc931930160f45257041.png#pic_center\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<p>\u663e\u800c\u6613\u89c1\uff0c\u6574\u7406\u4e00\u4efd\u5b8c\u6574\u7684\u6e05\u5355\u662f\u4e0d\u5207\u5b9e\u9645\u7684\uff0c\u56e0\u4e3a\u65b0\u7684\u67b6\u6784\u6b63\u88ab\u6e90\u6e90\u4e0d\u65ad\u5730\u53d1\u660e\u51fa\u6765\u3002\u6240\u4ee5\uff0c\u63a5\u4e0b\u6765\u8fd9\u4efd\u6e05\u5355\u7684\u76ee\u7684\uff0c\u53ea\u60f3\u5e26\u4f60\u4e00\u7aa5\u4eba\u5de5\u667a\u80fd\u9886\u57df\u7684\u57fa\u7840\u8bbe\u65bd\u3002\u5bf9\u4e8e\u6bcf\u4e00\u4e2a\u753b\u6210\u8282\u70b9\u56fe\u7684\u67b6\u6784\uff0c\u6211\u90fd\u4f1a\u5199\u4e00\u4e2a\u975e\u5e38\u975e\u5e38\u7b80\u77ed\u7684\u63cf\u8ff0\u3002\u4f60\u4f1a\u53d1\u73b0\u8fd9\u4e9b\u63cf\u8ff0\u8fd8\u662f\u5f88\u6709\u7528\u7684\uff0c\u6bd5\u7adf\uff0c\u603b\u8fd8\u662f\u6709\u4e00\u4e9b\u4f60\u5e76\u4e0d\u662f\u90a3\u4e48\u719f\u6089\u7684\u67b6\u6784\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/img-blog.csdnimg.cn\/img_convert\/3048c1d6696ffd95b9ebf5bdd737b376.png#pic_center\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<p>\u503c\u5f97\u4e00\u63d0\u7684\u662f\uff0c\u867d\u8bf4\u5927\u591a\u6570\u7684\u7b80\u5199\u90fd\u5df2\u88ab\u666e\u904d\u63a5\u53d7\uff0c\u4f46\u603b\u4f1a\u51fa\u73b0\u4e00\u4e9b\u51b2\u7a81\u3002RNNs\u6709\u65f6\u8868\u793a\u9012\u5f52\u795e\u7ecf\u7f51\u7edc\uff08recursive neural networks\uff09\uff0c\u4f46\u5927\u591a\u65f6\u5019\uff0c\u5b83\u4eec\u6307\u7684\u662f\u5faa\u73af\u795e\u7ecf\u7f51\u7edc\uff08recurrent neural networks\uff09\u3002\u8fd9\u8fd8\u6ca1\u5b8c\uff0c\u5b83\u4eec\u5728\u8bb8\u591a\u5730\u65b9\u8fd8\u4f1a\u6cdb\u6307\u5404\u79cd\u5faa\u73af\u67b6\u6784\uff0c\u8fd9\u5305\u62ec\u5728LSTMs\u3001GRU\u751a\u81f3\u662f\u53cc\u5411\u53d8\u4f53\u3002AEs\u4e5f\u7ecf\u5e38\u4f1a\u9762\u4e34\u540c\u6837\u7684\u95ee\u9898\uff0cVAEs\u3001DAEs\u53ca\u5176\u76f8\u4f3c\u7ed3\u6784\u6709\u65f6\u90fd\u88ab\u7b80\u79f0\u4e3aAEs\u3002\u5f88\u591a\u7f29\u5199\u540e\u9762\u7684\u201cN\u201d\u4e5f\u5e38\u5e38\u4f1a\u6709\u6240\u53d8\u5316\uff0c\u56e0\u4e3a\u540c\u4e00\u4e2a\u67b6\u6784\u4f60\u65e2\u53ef\u79f0\u4e4b\u4e3a\u5377\u79ef\u795e\u7ecf\u7f51\u7edc\uff08convolutional neural network\uff09\uff0c\u4e5f\u53ef\u7b80\u79f0\u4e3a\u5377\u79ef\u7f51\u7edc\uff08convolutional network\uff09\uff0c\u8fd9\u6837\u5c31\u51fa\u73b0\u4e86CNN\u548cCN\u4e24\u79cd\u5f62\u5f0f\u3002<\/p>\n<h1><a name=\"t12\"><\/a><a id=\"2_FFNN_92\"><\/a>2. \u524d\u9988\u795e\u7ecf\u7f51\u7edc\uff08FFNN\uff09<\/h1>\n<p>\u524d\u9988\u795e\u7ecf\u611f\u77e5\u7f51\u7edc\u4e0e\u611f\u77e5\u673a\uff08FF or FFNN\uff1aFeed forward neural networks and P\uff1aperceptrons\uff09\u975e\u5e38\u7b80\u5355\uff0c\u4fe1\u606f\u4ece\u524d\u5f80\u540e\u6d41\u52a8\uff08\u5206\u522b\u5bf9\u5e94\u8f93\u5165\u548c\u8f93\u51fa\uff09\u3002<\/p>\n<p>\u4e00\u822c\u5728\u63cf\u8ff0\u795e\u7ecf\u7f51\u7edc\u7684\u65f6\u5019\uff0c\u90fd\u662f\u4ece\u5b83\u7684\u5c42\u8bf4\u8d77\uff0c\u5373\u76f8\u4e92\u5e73\u884c\u7684\u8f93\u5165\u5c42\u3001\u9690\u542b\u5c42\u6216\u8005\u8f93\u51fa\u5c42\u795e\u7ecf\u7ed3\u6784\u3002\u5355\u72ec\u7684\u795e\u7ecf\u7ec6\u80de\u5c42\u5185\u90e8\uff0c\u795e\u7ecf\u5143\u4e4b\u95f4\u4e92\u4e0d\u76f8\u8fde\uff1b\u800c\u4e00\u822c\u76f8\u90bb\u7684\u4e24\u4e2a\u795e\u7ecf\u7ec6\u80de\u5c42\u5219\u662f\u5168\u8fde\u63a5\uff08\u4e00\u5c42\u7684\u6bcf\u4e2a\u795e\u7ecf\u5143\u548c\u53e6\u4e00\u5c42\u7684\u6bcf\u4e00\u4e2a\u795e\u7ecf\u5143\u76f8\u8fde\uff09\u3002\u4e00\u4e2a\u6700\u7b80\u5355\u5374\u6700\u5177\u6709\u5b9e\u7528\u6027\u7684\u795e\u7ecf\u7f51\u7edc\u7531\u4e24\u4e2a\u8f93\u5165\u795e\u7ecf\u5143\u548c\u4e00\u4e2a\u8f93\u51fa\u795e\u7ecf\u5143\u6784\u6210\uff0c\u4e5f\u5c31\u662f\u4e00\u4e2a\u903b\u8f91\u95e8\u6a21\u578b\u3002\u7ed9\u795e\u7ecf\u7f51\u7edc\u4e00\u5bf9\u6570\u636e\u96c6\uff08\u5206\u522b\u662f\u201c\u8f93\u5165\u6570\u636e\u96c6\u201d\u548c\u201c\u6211\u4eec\u671f\u671b\u7684\u8f93\u51fa\u6570\u636e\u96c6\u201d\uff09\uff0c\u4e00\u822c\u901a\u8fc7\u53cd\u5411\u4f20\u64ad\u7b97\u6cd5\u6765\u8bad\u7ec3\u524d\u9988\u795e\u7ecf\u7f51\u7edc\uff08FFNNs\uff09\u3002<\/p>\n<p>\u8fd9\u5c31\u662f\u6240\u8c13\u7684\u76d1\u7763\u5f0f\u5b66\u4e60\u3002\u4e0e\u6b64\u76f8\u53cd\u7684\u662f\u65e0\u76d1\u7763\u5b66\u4e60\uff1a\u6211\u4eec\u53ea\u7ed9\u8f93\u5165\uff0c\u7136\u540e\u8ba9\u795e\u7ecf\u7f51\u7edc\u53bb\u5bfb\u627e\u6570\u636e\u5f53\u4e2d\u7684\u89c4\u5f8b\u3002\u53cd\u5411\u4f20\u64ad\u7684\u8bef\u5dee\u5f80\u5f80\u662f\u795e\u7ecf\u7f51\u7edc\u5f53\u524d\u8f93\u51fa\u548c\u7ed9\u5b9a\u8f93\u51fa\u4e4b\u95f4\u5dee\u503c\u7684\u67d0\u79cd\u53d8\u4f53\uff08\u6bd4\u5982MSE\u6216\u8005\u4ec5\u4ec5\u662f\u5dee\u503c\u7684\u7ebf\u6027\u53d8\u5316\uff09\u3002\u5982\u679c\u795e\u7ecf\u7f51\u7edc\u5177\u6709\u8db3\u591f\u7684\u9690\u5c42\u795e\u7ecf\u5143\uff0c\u90a3\u4e48\u7406\u8bba\u4e0a\u5b83\u603b\u662f\u80fd\u591f\u5efa\u7acb\u8f93\u5165\u6570\u636e\u548c\u8f93\u51fa\u6570\u636e\u4e4b\u95f4\u7684\u5173\u7cfb\u3002\u5728\u5b9e\u8df5\u4e2d\uff0cFFNN\u7684\u4f7f\u7528\u5177\u6709\u5f88\u5927\u7684\u5c40\u9650\u6027\uff0c\u4f46\u662f\uff0c\u5b83\u4eec\u901a\u5e38\u548c\u5176\u5b83\u795e\u7ecf\u7f51\u7edc\u4e00\u8d77\u7ec4\u5408\u6210\u65b0\u7684\u67b6\u6784\u3002<\/p>\n<p>\u53c2\u8003\u6587\u732e\uff1a<br \/>\n<a href=\"http:\/\/www.ling.upenn.edu\/courses\/cogs501\/Rosenblatt1958.pdf\" target=\"_blank\" rel=\"noopener\">Rosenblatt, Frank. \u201cThe perceptron: a probabilistic model for information storage and organization in the brain.\u201d Psychological review 65.6 (1958): 386.<\/a><\/p>\n<h1><a name=\"t13\"><\/a><a id=\"3_RBF_102\"><\/a>3. \u5f84\u5411\u57fa\u795e\u7ecf\u7f51\u7edc\uff08RBF\uff09<\/h1>\n<p>\u5f84\u5411\u795e\u7ecf\u7f51\u7edc\uff08RBF\uff1aRadial basis function\uff09\u662f\u4e00\u79cd\u4ee5\u5f84\u5411\u57fa\u6838\u51fd\u6570\u4f5c\u4e3a\u6fc0\u6d3b\u51fd\u6570\u7684\u524d\u9988\u795e\u7ecf\u7f51\u7edc\u3002\u6ca1\u6709\u66f4\u591a\u63cf\u8ff0\u4e86\u3002\u8fd9\u4e0d\u662f\u8bf4\u6ca1\u6709\u76f8\u5173\u7684\u5e94\u7528\uff0c\u4f46\u5927\u591a\u6570\u4ee5\u5176\u5b83\u51fd\u6570\u4f5c\u4e3a\u6fc0\u6d3b\u51fd\u6570\u7684FFNNs\u90fd\u6ca1\u6709\u5b83\u4eec\u81ea\u5df1\u7684\u540d\u5b57\u3002\u8fd9\u6216\u8bb8\u8ddf\u5b83\u4eec\u7684\u53d1\u660e\u5e74\u4ee3\u6709\u5173\u7cfb\u3002<\/p>\n<p>\u53c2\u8003\u6587\u732e\uff1a<br \/>\n<a href=\"http:\/\/www.dtic.mil\/cgi-bin\/GetTRDoc?AD=ADA196234\" target=\"_blank\" rel=\"noopener\">Broomhead, David S., and David Lowe. Radial basis functions, multi-variable functional interpolation and adaptive networks. No. RSRE-MEMO-4148. ROYAL SIGNALS AND RADAR ESTABLISHMENT MALVERN (UNITED KINGDOM), 1988.<\/a><\/p>\n<h1><a name=\"t14\"><\/a><a id=\"4_HN_109\"><\/a>4. \u970d\u666e\u83f2\u5c14\u7f51\u7edc\uff08HN\uff09<\/h1>\n<p>\u970d\u666e\u83f2\u5c14\u7f51\u7edc\uff08HN\uff1aHopfield network\uff09\u662f\u4e00\u79cd\u6bcf\u4e00\u4e2a\u795e\u7ecf\u5143\u90fd\u8ddf\u5176\u5b83\u795e\u7ecf\u5143\u76f8\u4e92\u8fde\u63a5\u7684\u7f51\u7edc\u3002<\/p>\n<p>\u8fd9\u5c31\u50cf\u4e00\u76d8\u5b8c\u5168\u6405\u5728\u4e00\u8d77\u7684\u610f\u5927\u5229\u9762\uff0c\u56e0\u4e3a\u6bcf\u4e2a\u795e\u7ecf\u5143\u90fd\u5728\u5145\u5f53\u6240\u6709\u89d2\u8272\uff1a\u8bad\u7ec3\u524d\u7684\u6bcf\u4e00\u4e2a\u8282\u70b9\u90fd\u662f\u8f93\u5165\u795e\u7ecf\u5143\uff0c\u8bad\u7ec3\u9636\u6bb5\u662f\u9690\u795e\u7ecf\u5143\uff0c\u8f93\u51fa\u9636\u6bb5\u5219\u662f\u8f93\u51fa\u795e\u7ecf\u5143\u3002<\/p>\n<p>\u8be5\u795e\u7ecf\u7f51\u7edc\u7684\u8bad\u7ec3\uff0c\u662f\u5148\u628a\u795e\u7ecf\u5143\u7684\u503c\u8bbe\u7f6e\u5230\u671f\u671b\u6a21\u5f0f\uff0c\u7136\u540e\u8ba1\u7b97\u76f8\u5e94\u7684\u6743\u91cd\u3002\u5728\u8fd9\u4ee5\u540e\uff0c\u6743\u91cd\u5c06\u4e0d\u4f1a\u518d\u6539\u53d8\u4e86\u3002\u4e00\u65e6\u7f51\u7edc\u88ab\u8bad\u7ec3\u5305\u542b\u4e00\u79cd\u6216\u8005\u591a\u79cd\u6a21\u5f0f\uff0c\u8fd9\u4e2a\u795e\u7ecf\u7f51\u7edc\u603b\u662f\u4f1a\u6536\u655b\u4e8e\u5176\u4e2d\u7684\u67d0\u4e00\u79cd\u5b66\u4e60\u5230\u7684\u6a21\u5f0f\uff0c\u56e0\u4e3a\u5b83\u53ea\u4f1a\u5728\u67d0\u4e00\u4e2a\u72b6\u6001\u624d\u4f1a\u7a33\u5b9a\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u5b83\u5e76\u4e0d\u4e00\u5b9a\u9075\u4ece\u90a3\u4e2a\u671f\u671b\u7684\u72b6\u6001\uff08\u5f88\u9057\u61be\uff0c\u5b83\u5e76\u4e0d\u662f\u90a3\u4e2a\u5177\u6709\u9b54\u6cd5\u7684\u9ed1\u76d2\u5b50\uff09\u3002\u5b83\u4e4b\u6240\u4ee5\u4f1a\u7a33\u5b9a\u4e0b\u6765\uff0c\u90e8\u5206\u8981\u5f52\u529f\u4e8e\u5728\u8bad\u7ec3\u671f\u95f4\u6574\u4e2a\u7f51\u7edc\u7684\u201c\u80fd\u91cf\uff08Energy\uff09\u201d\u6216\u201c\u6e29\u5ea6\uff08Temperature\uff09\u201d\u4f1a\u9010\u6e10\u5730\u51cf\u5c11\u3002\u6bcf\u4e00\u4e2a\u795e\u7ecf\u5143\u7684\u6fc0\u6d3b\u51fd\u6570\u9608\u503c\u90fd\u4f1a\u88ab\u8bbe\u7f6e\u6210\u8fd9\u4e2a\u6e29\u5ea6\u7684\u503c\uff0c\u4e00\u65e6\u795e\u7ecf\u5143\u8f93\u5165\u7684\u603b\u548c\u8d85\u8fc7\u4e86\u8fd9\u4e2a\u9608\u503c\uff0c\u90a3\u4e48\u5c31\u4f1a\u8ba9\u5f53\u524d\u795e\u7ecf\u5143\u9009\u62e9\u72b6\u6001\uff08\u901a\u5e38\u662f-1\u62161\uff0c\u6709\u65f6\u4e5f\u662f0\u62161\uff09\u3002<\/p>\n<p>\u53ef\u4ee5\u591a\u4e2a\u795e\u7ecf\u5143\u540c\u6b65\uff0c\u4e5f\u53ef\u4ee5\u4e00\u4e2a\u795e\u7ecf\u5143\u4e00\u4e2a\u795e\u7ecf\u5143\u5730\u5bf9\u7f51\u7edc\u8fdb\u884c\u66f4\u65b0\u3002\u4e00\u65e6\u6240\u6709\u7684\u795e\u7ecf\u5143\u90fd\u5df2\u7ecf\u88ab\u66f4\u65b0\uff0c\u5e76\u4e14\u5b83\u4eec\u518d\u4e5f\u6ca1\u6709\u6539\u53d8\uff0c\u6574\u4e2a\u7f51\u7edc\u5c31\u7b97\u7a33\u5b9a\uff08\u9000\u706b\uff09\u4e86\uff0c\u90a3\u4f60\u5c31\u53ef\u4ee5\u8bf4\u8fd9\u4e2a\u7f51\u7edc\u5df2\u7ecf\u6536\u655b\u4e86\u3002\u8fd9\u79cd\u7c7b\u578b\u7684\u7f51\u7edc\u88ab\u79f0\u4e3a\u201c\u8054\u60f3\u8bb0\u5fc6\uff08associative memory\uff09\u201d\uff0c\u56e0\u4e3a\u5b83\u4eec\u4f1a\u6536\u655b\u5230\u548c\u8f93\u5165\u6700\u76f8\u4f3c\u7684\u72b6\u6001\uff1b\u6bd4\u5982\uff0c\u4eba\u7c7b\u770b\u5230\u684c\u5b50\u7684\u4e00\u534a\u5c31\u53ef\u4ee5\u60f3\u8c61\u51fa\u53e6\u5916\u4e00\u534a\uff1b\u4e0e\u4e4b\u76f8\u4f3c\uff0c\u5982\u679c\u8f93\u5165\u4e00\u534a\u566a\u97f3+\u4e00\u534a\u684c\u5b50\uff0c\u8fd9\u4e2a\u7f51\u7edc\u5c31\u80fd\u6536\u655b\u5230\u6574\u5f20\u684c\u5b50\u3002<\/p>\n<p>\u53c2\u8003\u6587\u732e\uff1a<br \/>\n<a href=\"https:\/\/bi.snu.ac.kr\/Courses\/g-ai09-2\/hopfield82.pdf\" target=\"_blank\" rel=\"noopener\">Hopfield, John J. \u201cNeural networks and physical systems with emergent collective computational abilities.\u201d Proceedings of the national academy of sciences 79.8 (1982): 2554-2558.<\/a><\/p>\n<h1><a name=\"t15\"><\/a><a id=\"5_MC_122\"><\/a>5. \u9a6c\u5c14\u53ef\u592b\u94fe\uff08MC\uff09<\/h1>\n<p>\u9a6c\u5c14\u53ef\u592b\u94fe\uff08MC\uff1aMarkov Chain\uff09\u6216\u79bb\u6563\u65f6\u95f4\u9a6c\u5c14\u53ef\u592b\u94fe\uff08DTMC\uff1aMC or discrete time Markov Chain\uff09\u5728\u67d0\u79cd\u610f\u4e49\u4e0a\u662fBMs\u548cHNs\u7684\u524d\u8eab\u3002\u53ef\u4ee5\u8fd9\u6837\u6765\u7406\u89e3\uff1a\u4ece\u4ece\u6211\u5f53\u524d\u6240\u5904\u7684\u8282\u70b9\u5f00\u59cb\uff0c\u8d70\u5230\u4efb\u610f\u76f8\u90bb\u8282\u70b9\u7684\u6982\u7387\u662f\u591a\u5c11\u5462\uff1f\u5b83\u4eec\u6ca1\u6709\u8bb0\u5fc6\uff08\u6240\u8c13\u7684\u9a6c\u5c14\u53ef\u592b\u7279\u6027\uff09\uff1a\u4f60\u6240\u5f97\u5230\u7684\u6bcf\u4e00\u4e2a\u72b6\u6001\u90fd\u5b8c\u5168\u4f9d\u8d56\u4e8e\u524d\u4e00\u4e2a\u72b6\u6001\u3002\u5c3d\u7ba1\u7b97\u4e0d\u4e0a\u795e\u7ecf\u7f51\u7edc\uff0c\u4f46\u5b83\u5374\u8ddf\u795e\u7ecf\u7f51\u7edc\u7c7b\u4f3c\uff0c\u5e76\u4e14\u5960\u5b9a\u4e86BM\u548cHN\u7684\u7406\u8bba\u57fa\u7840\u3002\u8ddfBM\u3001RBM\u3001HN\u4e00\u6837\uff0cMC\u5e76\u4e0d\u603b\u88ab\u8ba4\u4e3a\u662f\u795e\u7ecf\u7f51\u7edc\u3002\u6b64\u5916\uff0c\u5b83\u4e5f\u5e76\u4e0d\u603b\u662f\u5168\u8fde\u63a5\u7684\u3002<\/p>\n<p>\u53c2\u8003\u6587\u732e\uff1a<br \/>\n<a href=\"http:\/\/www.americanscientist.org\/libraries\/documents\/201321152149545-2013-03Hayes.pdf\" target=\"_blank\" rel=\"noopener\">Hayes, Brian. \u201cFirst links in the Markov chain.\u201d American Scientist 101.2 (2013): 252.<\/a><\/p>\n<h1><a name=\"t16\"><\/a><a id=\"6_BM_129\"><\/a>6. \u73bb\u5c14\u5179\u66fc\u673a\uff08BM\uff09<\/h1>\n<p>\u73bb\u5c14\u5179\u66fc\u673a\uff08BM\uff1aBoltzmann machines\uff09\u548c\u970d\u666e\u83f2\u5c14\u7f51\u7edc\u5f88\u63a5\u8fd1\uff0c\u5dee\u522b\u53ea\u662f\uff1a\u4e00\u4e9b\u795e\u7ecf\u5143\u4f5c\u4e3a\u8f93\u5165\u795e\u7ecf\u5143\uff0c\u5269\u4f59\u7684\u5219\u662f\u4f5c\u4e3a\u9690\u795e\u7ecf\u5143\u3002<\/p>\n<p>\u5728\u6574\u4e2a\u795e\u7ecf\u7f51\u7edc\u66f4\u65b0\u8fc7\u540e\uff0c\u8f93\u5165\u795e\u7ecf\u5143\u6210\u4e3a\u8f93\u51fa\u795e\u7ecf\u5143\u3002\u521a\u5f00\u59cb\u795e\u7ecf\u5143\u7684\u6743\u91cd\u90fd\u662f\u968f\u673a\u7684\uff0c\u901a\u8fc7\u53cd\u5411\u4f20\u64ad\uff08back-propagation\uff09\u7b97\u6cd5\u8fdb\u884c\u5b66\u4e60\uff0c\u6216\u662f\u6700\u8fd1\u5e38\u7528\u7684\u5bf9\u6bd4\u6563\u5ea6\uff08contrastive divergence\uff09\u7b97\u6cd5\uff08\u9a6c\u5c14\u53ef\u592b\u94fe\u7528\u4e8e\u8ba1\u7b97\u4e24\u4e2a\u4fe1\u606f\u589e\u76ca\u4e4b\u95f4\u7684\u68af\u5ea6\uff09\u3002<\/p>\n<p>\u76f8\u6bd4HN\uff0c\u5927\u591a\u6570BM\u7684\u795e\u7ecf\u5143\u6fc0\u6d3b\u6a21\u5f0f\u90fd\u662f\u4e8c\u5143\u7684\u3002BM\u7531MC\u8bad\u7ec3\u83b7\u5f97\uff0c\u56e0\u800c\u662f\u4e00\u4e2a\u968f\u673a\u7f51\u7edc\u3002BM\u7684\u8bad\u7ec3\u548c\u8fd0\u884c\u8fc7\u7a0b\uff0c\u8ddfHN\u5927\u540c\u5c0f\u5f02\uff1a\u4e3a\u8f93\u5165\u795e\u7ecf\u5143\u8bbe\u597d\u94b3\u4f4d\u503c\uff0c\u800c\u540e\u8ba9\u795e\u7ecf\u7f51\u7edc\u81ea\u884c\u5b66\u4e60\u3002\u56e0\u4e3a\u8fd9\u4e9b\u795e\u7ecf\u5143\u53ef\u80fd\u4f1a\u5f97\u5230\u4efb\u610f\u7684\u503c\uff0c\u6211\u4eec\u53cd\u590d\u5730\u5728\u8f93\u5165\u548c\u8f93\u51fa\u795e\u7ecf\u5143\u4e4b\u95f4\u6765\u56de\u5730\u8fdb\u884c\u8ba1\u7b97\u3002\u6fc0\u6d3b\u51fd\u6570\u7684\u6fc0\u6d3b\u53d7\u5168\u5c40\u6e29\u5ea6\u7684\u63a7\u5236\uff0c\u5982\u679c\u5168\u5c40\u6e29\u5ea6\u964d\u4f4e\u4e86\uff0c\u90a3\u4e48\u795e\u7ecf\u5143\u7684\u80fd\u91cf\u4e5f\u4f1a\u76f8\u5e94\u5730\u964d\u4f4e\u3002\u8fd9\u4e2a\u80fd\u91cf\u4e0a\u7684\u964d\u4f4e\u5bfc\u81f4\u4e86\u5b83\u4eec\u6fc0\u6d3b\u6a21\u5f0f\u7684\u7a33\u5b9a\u3002\u5728\u6b63\u786e\u7684\u6e29\u5ea6\u4e0b\uff0c\u8fd9\u4e2a\u7f51\u7edc\u4f1a\u62b5\u8fbe\u4e00\u4e2a\u5e73\u8861\u72b6\u6001\u3002<br \/>\n\u53c2\u8003\u6587\u732e\uff1a<br \/>\n<a href=\"https:\/\/www.researchgate.net\/profile\/Terrence_Sejnowski\/publication\/242509302_Learning_and_relearning_in_Boltzmann_machines\/links\/54a4b00f0cf256bf8bb327cc.pdf\" target=\"_blank\" rel=\"noopener\">Hinton, Geoffrey E., and Terrence J. Sejnowski. \u201cLearning and releaming in Boltzmann machines.\u201d Parallel distributed processing: Explorations in the microstructure of cognition 1 (1986): 282-317.<\/a><\/p>\n<h1><a name=\"t17\"><\/a><a id=\"7_RBM_139\"><\/a>7. \u53d7\u9650\u73bb\u5c14\u5179\u66fc\u673a\uff08RBM\uff09<\/h1>\n<p>[\u5916\u94fe\u56fe\u7247\u8f6c\u5b58\u5931\u8d25,\u6e90\u7ad9\u53ef\u80fd\u6709\u9632\u76d7\u94fe\u673a\u5236,\u5efa\u8bae\u5c06\u56fe\u7247\u4fdd\u5b58\u4e0b\u6765\u76f4\u63a5\u4e0a\u4f20(img-4bQYq9pp-1644720505581)(http:\/\/p9.pstatp.com\/large\/288d000065e0cc074b0b)]<\/p>\n<p>\u53d7\u9650\u73bb\u5c14\u5179\u66fc\u673a\uff08RBM\uff1aRestricted Boltzmann machines\uff09\u4e0eBM\u51fa\u5947\u5730\u76f8\u4f3c\uff0c\u56e0\u800c\u4e5f\u540cHN\u76f8\u4f3c\u3002<\/p>\n<p>\u5b83\u4eec\u7684\u6700\u5927\u533a\u522b\u5728\u4e8e\uff1aRBM\u66f4\u5177\u5b9e\u7528\u4ef7\u503c\uff0c\u56e0\u4e3a\u5b83\u4eec\u53d7\u5230\u4e86\u66f4\u591a\u7684\u9650\u5236\u3002\u5b83\u4eec\u4e0d\u4f1a\u968f\u610f\u5728\u6240\u6709\u795e\u7ecf\u5143\u95f4\u5efa\u7acb\u8fde\u63a5\uff0c\u800c\u53ea\u5728\u4e0d\u540c\u795e\u7ecf\u5143\u7fa4\u4e4b\u95f4\u5efa\u7acb\u8fde\u63a5\uff0c\u56e0\u6b64\u4efb\u4f55\u8f93\u5165\u795e\u7ecf\u5143\u90fd\u4e0d\u4f1a\u540c\u5176\u4ed6\u8f93\u5165\u795e\u7ecf\u5143\u76f8\u8fde\uff0c\u4efb\u4f55\u9690\u795e\u7ecf\u5143\u4e5f\u4e0d\u4f1a\u540c\u5176\u4ed6\u9690\u795e\u7ecf\u5143\u76f8\u8fde\u3002<\/p>\n<p>RBM\u7684\u8bad\u7ec3\u65b9\u5f0f\u5c31\u50cf\u7a0d\u5fae\u4fee\u6539\u8fc7\u7684FFNN\uff1a\u524d\u5411\u901a\u8fc7\u6570\u636e\u4e4b\u540e\u518d\u5c06\u8fd9\u4e9b\u6570\u636e\u53cd\u5411\u4f20\u56de\uff08\u56de\u5230\u7b2c\u4e00\u5c42\uff09\uff0c\u800c\u975e\u524d\u5411\u901a\u8fc7\u6570\u636e\u7136\u540e\u53cd\u5411\u4f20\u64ad\u8bef\u5dee\u3002\u4e4b\u540e\uff0c\u518d\u4f7f\u7528\u524d\u5411\u548c\u53cd\u5411\u4f20\u64ad\u8fdb\u884c\u8bad\u7ec3\u3002<\/p>\n<p>\u53c2\u8003\u6587\u732e\uff1a<br \/>\n<a href=\"http:\/\/www.dtic.mil\/cgi-bin\/GetTRDoc?Location=U2&#038;doc=GetTRDoc.pdf&#038;AD=ADA620727\" target=\"_blank\" rel=\"noopener\">Smolensky, Paul. Information processing in dynamical systems: Foundations of harmony theory. No. CU-CS-321-86. COLORADO UNIV AT BOULDER DEPT OF COMPUTER SCIENCE, 1986.<\/a><\/p>\n<h1><a name=\"t18\"><\/a><a id=\"8_AE_152\"><\/a>8. \u81ea\u7f16\u7801\u673a\uff08AE\uff09<\/h1>\n<p><img decoding=\"async\" src=\"https:\/\/img-blog.csdnimg.cn\/img_convert\/eb534eca29b5a5a841423fd661f72702.png#pic_center\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<p>\u81ea\u7f16\u7801\u673a\uff08AE\uff1aAutoencoders\uff09\u548cFFNN\u6709\u4e9b\u76f8\u8fd1\uff0c\u56e0\u4e3a\u5b83\u66f4\u50cf\u662fFFNN\u7684\u53e6\u4e00\u79cd\u7528\u6cd5\uff0c\u800c\u975e\u672c\u8d28\u4e0a\u5b8c\u5168\u4e0d\u540c\u7684\u53e6\u4e00\u79cd\u67b6\u6784\u3002<\/p>\n<p>\u81ea\u7f16\u7801\u673a\u7684\u57fa\u672c\u601d\u60f3\u662f\u81ea\u52a8\u5bf9\u4fe1\u606f\u8fdb\u884c\u7f16\u7801\uff08\u50cf\u538b\u7f29\u4e00\u6837\uff0c\u800c\u975e\u52a0\u5bc6\uff09\uff0c\u5b83\u4e5f\u56e0\u6b64\u800c\u5f97\u540d\u3002\u6574\u4e2a\u7f51\u7edc\u7684\u5f62\u72b6\u9177\u4f3c\u4e00\u4e2a\u6c99\u6f0f\u8ba1\u65f6\u5668\uff0c\u4e2d\u95f4\u7684\u9690\u542b\u5c42\u8f83\u5c0f\uff0c\u4e24\u8fb9\u7684\u8f93\u5165\u5c42\u3001\u8f93\u51fa\u5c42\u8f83\u5927\u3002\u81ea\u7f16\u7801\u673a\u603b\u662f\u5bf9\u79f0\u7684\uff0c\u4ee5\u4e2d\u95f4\u5c42\uff08\u4e00\u5c42\u8fd8\u662f\u4e24\u5c42\u53d6\u51b3\u4e8e\u795e\u7ecf\u7f51\u7edc\u5c42\u6570\u7684\u5947\u5076\uff09\u4e3a\u8f74\u3002\u6700\u5c0f\u7684\u5c42\uff08\u4e00\u5c42\u6216\u8005\u591a\u5c42\uff09\u603b\u662f\u5728\u4e2d\u95f4\uff0c\u5728\u8fd9\u91cc\u4fe1\u606f\u538b\u7f29\u7a0b\u5ea6\u6700\u5927\uff08\u6574\u4e2a\u7f51\u7edc\u7684\u5173\u9698\u53e3\uff09\u3002\u5728\u4e2d\u95f4\u5c42\u4e4b\u524d\u4e3a\u7f16\u7801\u90e8\u5206\uff0c\u4e2d\u95f4\u5c42\u4e4b\u540e\u4e3a\u89e3\u7801\u90e8\u5206\uff0c\u4e2d\u95f4\u5c42\u5219\u662f\u7f16\u7801\u90e8\u5206\u3002<\/p>\n<p>\u81ea\u7f16\u7801\u673a\u53ef\u7528\u53cd\u5411\u4f20\u64ad\u7b97\u6cd5\u8fdb\u884c\u8bad\u7ec3\uff0c\u7ed9\u5b9a\u8f93\u5165\uff0c\u5c06\u8bef\u5dee\u8bbe\u4e3a\u8f93\u5165\u548c\u8f93\u51fa\u4e4b\u5dee\u3002\u81ea\u7f16\u7801\u673a\u7684\u6743\u91cd\u4e5f\u662f\u5bf9\u79f0\u7684\uff0c\u56e0\u6b64\u7f16\u7801\u90e8\u5206\u6743\u91cd\u4e0e\u89e3\u7801\u90e8\u5206\u6743\u91cd\u5b8c\u5168\u4e00\u6837\u3002<\/p>\n<p>\u53c2\u8003\u6587\u732e:<br \/>\n<a href=\"https:\/\/pdfs.semanticscholar.org\/f582\/1548720901c89b3b7481f7500d7cd64e99bd.pdf\" target=\"_blank\" rel=\"noopener\">Bourlard, Herv\u00e9, and Yves Kamp. \u201cAuto-association by multilayer perceptrons and singular value decomposition.\u201d Biological cybernetics 59.4-5 (1988): 291-294.<\/a><\/p>\n<h1><a name=\"t19\"><\/a><a id=\"9_SAE_167\"><\/a>9. \u7a00\u758f\u81ea\u7f16\u7801\u673a\uff08SAE\uff09<\/h1>\n<p><img decoding=\"async\" src=\"https:\/\/img-blog.csdnimg.cn\/img_convert\/3c87939990c7c63443542c5fee5a5885.png#pic_center\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<p>\u7a00\u758f\u81ea\u7f16\u7801\u673a\uff08SAE\uff1aSparse autoencoders\uff09\u67d0\u79cd\u7a0b\u5ea6\u4e0a\u540c\u81ea\u7f16\u7801\u673a\u76f8\u53cd\u3002\u7a00\u758f\u81ea\u7f16\u7801\u673a\u4e0d\u662f\u7528\u66f4\u5c0f\u7684\u7a7a\u95f4\u8868\u5f81\u5927\u91cf\u4fe1\u606f\uff0c\u800c\u662f\u628a\u539f\u672c\u7684\u4fe1\u606f\u7f16\u7801\u5230\u66f4\u5927\u7684\u7a7a\u95f4\u5185\u3002\u56e0\u6b64\uff0c\u4e2d\u95f4\u5c42\u4e0d\u662f\u6536\u655b\uff0c\u800c\u662f\u6269\u5f20\uff0c\u7136\u540e\u518d\u8fd8\u539f\u5230\u8f93\u5165\u5927\u5c0f\u3002\u5b83\u53ef\u4ee5\u7528\u4e8e\u63d0\u53d6\u6570\u636e\u96c6\u5185\u7684\u5c0f\u7279\u5f81\u3002<\/p>\n<p>\u5982\u679c\u7528\u8bad\u7ec3\u81ea\u7f16\u7801\u673a\u7684\u65b9\u5f0f\u6765\u8bad\u7ec3\u7a00\u758f\u81ea\u7f16\u7801\u673a\uff0c\u51e0\u4e4e\u6240\u6709\u7684\u60c5\u51b5\uff0c\u90fd\u662f\u5f97\u5230\u6beb\u65e0\u7528\u5904\u7684\u6052\u7b49\u7f51\u7edc\uff08\u8f93\u5165=\u8f93\u51fa\uff0c\u6ca1\u6709\u4efb\u4f55\u5f62\u5f0f\u7684\u53d8\u6362\u6216\u5206\u89e3\uff09\u3002\u4e3a\u907f\u514d\u8fd9\u79cd\u60c5\u51b5\uff0c\u9700\u8981\u5728\u53cd\u9988\u8f93\u5165\u4e2d\u52a0\u4e0a\u7a00\u758f\u9a71\u52a8\u6570\u636e\u3002\u7a00\u758f\u9a71\u52a8\u7684\u5f62\u5f0f\u53ef\u4ee5\u662f\u9608\u503c\u8fc7\u6ee4\uff0c\u8fd9\u6837\u5c31\u53ea\u6709\u7279\u5b9a\u7684\u8bef\u5dee\u624d\u4f1a\u53cd\u5411\u4f20\u64ad\u7528\u4e8e\u8bad\u7ec3\uff0c\u800c\u5176\u5b83\u7684\u8bef\u5dee\u5219\u88ab\u5ffd\u7565\u4e3a0\uff0c\u4e0d\u4f1a\u7528\u4e8e\u53cd\u5411\u4f20\u64ad\u3002\u8fd9\u5f88\u50cf\u8109\u51b2\u795e\u7ecf\u7f51\u7edc\uff08\u5e76\u4e0d\u662f\u6240\u6709\u7684\u795e\u7ecf\u5143\u4e00\u76f4\u90fd\u4f1a\u8f93\u51fa\uff09\u3002<\/p>\n<p><a href=\"https:\/\/papers.nips.cc\/paper\/3112-efficient-learning-of-sparse-representations-with-an-energy-based-model.pdf\" target=\"_blank\" rel=\"noopener\">Marc\u2019Aurelio Ranzato, Christopher Poultney, Sumit Chopra, and Yann LeCun. \u201cEfficient learning of sparse representations with an energy-based model.\u201d Proceedings of NIPS. 2007.<\/a><\/p>\n<h1><a name=\"t20\"><\/a><a id=\"10VAE_178\"><\/a>10.\u53d8\u5206\u81ea\u7f16\u7801\u673a\uff08VAE\uff09<\/h1>\n<p><img decoding=\"async\" src=\"https:\/\/img-blog.csdnimg.cn\/img_convert\/41a3886408d01d4548714b153666af85.png#pic_center\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<p>\u53d8\u5206\u81ea\u7f16\u7801\u673a\uff08VAE\uff1aVariational autoencoders\uff09\u548cAE\u6709\u7740\u76f8\u540c\u7684\u67b6\u6784\uff0c\u5374\u88ab\u6559\u4f1a\u4e86\u4e0d\u540c\u7684\u4e8b\u60c5\uff1a\u8f93\u5165\u6837\u672c\u7684\u4e00\u4e2a\u8fd1\u4f3c\u6982\u7387\u5206\u5e03\uff0c\u8fd9\u8ba9\u5b83\u8ddfBM\u3001RBM\u66f4\u76f8\u8fd1\u3002<\/p>\n<p>\u4e0d\u8fc7\uff0cVAE\u5374\u4f9d\u8d56\u4e8e\u8d1d\u53f6\u65af\u7406\u8bba\u6765\u5904\u7406\u6982\u7387\u63a8\u65ad\u548c\u72ec\u7acb\uff08probabilistic inference and independence\uff09\uff0c\u4ee5\u53ca\u91cd\u65b0\u53c2\u6570\u5316\uff08re-parametrisation\uff09\u6765\u8fdb\u884c\u4e0d\u540c\u7684\u8868\u5f81\u3002\u63a8\u65ad\u548c\u72ec\u7acb\u975e\u5e38\u76f4\u89c2\uff0c\u4f46\u5374\u4f9d\u8d56\u4e8e\u590d\u6742\u7684\u6570\u5b66\u7406\u8bba\u3002\u57fa\u672c\u539f\u7406\u662f\uff1a\u628a\u5f71\u54cd\u7eb3\u5165\u8003\u8651\u3002\u5982\u679c\u5728\u4e00\u4e2a\u5730\u65b9\u53d1\u751f\u4e86\u4e00\u4ef6\u4e8b\u60c5\uff0c\u53e6\u5916\u4e00\u4ef6\u4e8b\u60c5\u5728\u5176\u5b83\u5730\u65b9\u53d1\u751f\u4e86\uff0c\u5b83\u4eec\u4e0d\u4e00\u5b9a\u5c31\u662f\u5173\u8054\u5728\u4e00\u8d77\u7684\u3002\u5982\u679c\u5b83\u4eec\u4e0d\u76f8\u5173\uff0c\u90a3\u4e48\u8bef\u5dee\u4f20\u64ad\u5e94\u8be5\u8003\u8651\u8fd9\u4e2a\u56e0\u7d20\u3002\u8fd9\u662f\u4e00\u4e2a\u6709\u7528\u7684\u65b9\u6cd5\uff0c\u56e0\u4e3a\u795e\u7ecf\u7f51\u7edc\u662f\u4e00\u4e2a\u975e\u5e38\u5927\u7684\u56fe\u8868\uff0c\u5982\u679c\u4f60\u80fd\u5728\u67d0\u4e9b\u8282\u70b9\u6392\u9664\u4e00\u4e9b\u6765\u81ea\u4e8e\u5176\u5b83\u8282\u70b9\u7684\u5f71\u54cd\uff0c\u968f\u7740\u7f51\u7edc\u6df1\u5ea6\u5730\u589e\u52a0\uff0c\u8fd9\u5c06\u4f1a\u975e\u5e38\u6709\u7528\u3002<\/p>\n<p><a href=\"https:\/\/arxiv.org\/pdf\/1312.6114v10.pdf\" target=\"_blank\" rel=\"noopener\">Kingma, Diederik P., and Max Welling. \u201cAuto-encoding variational bayes.\u201d arXiv preprint arXiv:1312.6114 (2013).<\/a><\/p>\n<h1><a name=\"t21\"><\/a><a id=\"11_DAE_189\"><\/a>11. \u53bb\u566a\u81ea\u7f16\u7801\u673a\uff08DAE\uff09<\/h1>\n<p>\u53bb\u566a\u81ea\u7f16\u7801\u673a\uff08DAE\uff1aDenoising autoencoders\uff09\u662f\u4e00\u79cd\u81ea\u7f16\u7801\u673a\uff0c\u5b83\u7684\u8bad\u7ec3\u8fc7\u7a0b\uff0c\u4e0d\u4ec5\u8981\u8f93\u5165\u6570\u636e\uff0c\u8fd8\u6709\u518d\u52a0\u4e0a\u566a\u97f3\u6570\u636e\uff08\u5c31\u597d\u50cf\u8ba9\u56fe\u50cf\u53d8\u5f97\u66f4\u52a0\u6a21\u7cca\u4e00\u6837\uff09\u3002<\/p>\n<p>\u4f46\u5728\u8ba1\u7b97\u8bef\u5dee\u7684\u65f6\u5019\u8ddf\u81ea\u52a8\u7f16\u7801\u673a\u4e00\u6837\uff0c\u964d\u566a\u81ea\u52a8\u7f16\u7801\u673a\u7684\u8f93\u51fa\u4e5f\u662f\u548c\u539f\u59cb\u7684\u8f93\u5165\u6570\u636e\u8fdb\u884c\u5bf9\u6bd4\u3002\u8fd9\u79cd\u5f62\u5f0f\u7684\u8bad\u7ec3\u65e8\u5728\u9f13\u52b1\u964d\u566a\u81ea\u7f16\u7801\u673a\u4e0d\u8981\u53bb\u5b66\u4e60\u7ec6\u8282\uff0c\u800c\u662f\u4e00\u4e9b\u66f4\u52a0\u5b8f\u89c2\u7684\u7279\u5f81\uff0c\u56e0\u4e3a\u7ec6\u5fae\u7279\u5f81\u53d7\u5230\u566a\u97f3\u7684\u5f71\u54cd\uff0c\u5b66\u4e60\u7ec6\u5fae\u7279\u5f81\u5f97\u5230\u7684\u6a21\u578b\u6700\u7ec8\u8868\u73b0\u51fa\u6765\u7684\u6027\u80fd\u603b\u662f\u5f88\u5dee\u3002<\/p>\n<p><a href=\"http:\/\/machinelearning.org\/archive\/icml2008\/papers\/592.pdf\" target=\"_blank\" rel=\"noopener\">Vincent, Pascal, et al. \u201cExtracting and composing robust features with denoising autoencoders.\u201d Proceedings of the 25th international conference on Machine learning. ACM, 2008.<\/a><\/p>\n<h1><a name=\"t22\"><\/a><a id=\"12_DBN_197\"><\/a>12. \u6df1\u5ea6\u4fe1\u5ff5\u7f51\u7edc\uff08DBN\uff09<\/h1>\n<p><img decoding=\"async\" src=\"https:\/\/img-blog.csdnimg.cn\/img_convert\/ffdf7fc4f34d7620ce700ced664f5dea.png#pic_center\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<p>\u6df1\u5ea6\u4fe1\u5ff5\u7f51\u7edc\uff08DBN\uff1aDeep belief networks\uff09\u4e4b\u6240\u4ee5\u53d6\u8fd9\u4e2a\u540d\u5b57\uff0c\u662f\u7531\u4e8e\u5b83\u672c\u8eab\u51e0\u4e4e\u662f\u7531\u591a\u4e2a\u53d7\u9650\u73bb\u5c14\u5179\u66fc\u673a\u6216\u8005\u53d8\u5206\u81ea\u7f16\u7801\u673a\u5806\u780c\u800c\u6210\u3002<\/p>\n<p>\u5b9e\u8df5\u8868\u660e\u4e00\u5c42\u4e00\u5c42\u5730\u5bf9\u8fd9\u79cd\u7c7b\u578b\u7684\u795e\u7ecf\u7f51\u7edc\u8fdb\u884c\u8bad\u7ec3\u975e\u5e38\u6709\u6548\uff0c\u8fd9\u6837\u6bcf\u4e00\u4e2a\u81ea\u7f16\u7801\u673a\u6216\u8005\u53d7\u9650\u73bb\u5c14\u5179\u66fc\u673a\u53ea\u9700\u8981\u5b66\u4e60\u5982\u4f55\u7f16\u7801\u524d\u4e00\u795e\u7ecf\u5143\u5c42\u7684\u8f93\u51fa\u3002\u8fd9\u79cd\u8bad\u7ec3\u6280\u672f\u4e5f\u88ab\u79f0\u4e3a\u8d2a\u5a6a\u8bad\u7ec3\uff0c\u8fd9\u91cc\u8d2a\u5a6a\u7684\u610f\u601d\u662f\u901a\u8fc7\u4e0d\u65ad\u5730\u83b7\u53d6\u5c40\u90e8\u6700\u4f18\u89e3\uff0c\u6700\u7ec8\u5f97\u5230\u4e00\u4e2a\u76f8\u5f53\u4e0d\u9519\u89e3\uff08\u4f46\u53ef\u80fd\u4e0d\u662f\u5168\u5c40\u6700\u4f18\u7684\uff09\u3002\u53ef\u4ee5\u901a\u8fc7\u5bf9\u6bd4\u6563\u5ea6\u7b97\u6cd5\u6216\u8005\u53cd\u5411\u4f20\u64ad\u7b97\u6cd5\u8fdb\u884c\u8bad\u7ec3\uff0c\u5b83\u4f1a\u6162\u6162\u5b66\u7740\u4ee5\u4e00\u79cd\u6982\u7387\u6a21\u578b\u6765\u8868\u5f81\u6570\u636e\uff0c\u5c31\u597d\u50cf\u5e38\u89c4\u7684\u81ea\u7f16\u7801\u673a\u6216\u8005\u53d7\u9650\u73bb\u5c14\u5179\u66fc\u673a\u3002\u4e00\u65e6\u7ecf\u8fc7\u975e\u76d1\u7763\u5f0f\u5b66\u4e60\u65b9\u5f0f\uff0c\u8bad\u7ec3\u6216\u8005\u6536\u655b\u5230\u4e86\u4e00\u4e2a\u7a33\u5b9a\u7684\u72b6\u6001\uff0c\u90a3\u4e48\u8fd9\u4e2a\u6a21\u578b\u5c31\u53ef\u4ee5\u7528\u6765\u4ea7\u751f\u65b0\u7684\u6570\u636e\u3002\u5982\u679c\u4ee5\u5bf9\u6bd4\u6563\u5ea6\u7b97\u6cd5\u8fdb\u884c\u8bad\u7ec3\uff0c\u90a3\u4e48\u5b83\u751a\u81f3\u53ef\u4ee5\u7528\u4e8e\u533a\u5206\u73b0\u6709\u7684\u6570\u636e\uff0c\u56e0\u4e3a\u90a3\u4e9b\u795e\u7ecf\u5143\u5df2\u7ecf\u88ab\u5f15\u5bfc\u6765\u83b7\u53d6\u6570\u636e\u7684\u4e0d\u540c\u7279\u5b9a\u3002<\/p>\n<p><a href=\"https:\/\/papers.nips.cc\/paper\/3048-greedy-layer-wise-training-of-deep-networks.pdf\" target=\"_blank\" rel=\"noopener\">Bengio, Yoshua, et al. \u201cGreedy layer-wise training of deep networks.\u201d Advances in neural information processing systems 19 (2007): 153.<\/a><\/p>\n<h1><a name=\"t23\"><\/a><a id=\"13_CNN_208\"><\/a>13. \u5377\u79ef\u795e\u7ecf\u7f51\u7edc\uff08CNN\uff09<\/h1>\n<p><img decoding=\"async\" src=\"https:\/\/img-blog.csdnimg.cn\/img_convert\/93c964b03013ea01f46ee9bfec165025.png#pic_center\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<p>\u5377\u79ef\u795e\u7ecf\u7f51\u7edc\uff08CNN\uff1aConvolutional neural networks\uff09\u6216\u6df1\u5ea6\u5377\u79ef\u795e\u7ecf\u7f51\u7edc\uff08DCNN\uff1adeep convolutional neural networks\uff09\u8ddf\u5176\u5b83\u7c7b\u578b\u7684\u795e\u7ecf\u7f51\u7edc\u5927\u6709\u4e0d\u540c\u3002\u5b83\u4eec\u4e3b\u8981\u7528\u4e8e\u5904\u7406\u56fe\u50cf\u6570\u636e\uff0c\u4f46\u53ef\u7528\u4e8e\u5176\u5b83\u5f62\u5f0f\u6570\u636e\u7684\u5904\u7406\uff0c\u5982\u8bed\u97f3\u6570\u636e\u3002\u5bf9\u4e8e\u5377\u79ef\u795e\u7ecf\u7f51\u7edc\u6765\u8bf4\uff0c\u4e00\u4e2a\u5178\u578b\u7684\u5e94\u7528\u5c31\u662f\u7ed9\u5b83\u8f93\u5165\u4e00\u4e2a\u56fe\u50cf\uff0c\u800c\u540e\u5b83\u4f1a\u7ed9\u51fa\u4e00\u4e2a\u5206\u7c7b\u7ed3\u679c\u3002\u4e5f\u5c31\u662f\u8bf4\uff0c\u5982\u679c\u4f60\u7ed9\u5b83\u4e00\u5f20\u732b\u7684\u56fe\u50cf\uff0c\u5b83\u5c31\u8f93\u51fa\u201c\u732b\u201d\uff1b\u5982\u679c\u4f60\u7ed9\u4e00\u5f20\u72d7\u7684\u56fe\u50cf\uff0c\u5b83\u5c31\u8f93\u51fa\u201c\u72d7\u201d\u3002<\/p>\n<p>\u5377\u79ef\u795e\u7ecf\u7f51\u7edc\u662f\u4ece\u4e00\u4e2a\u6570\u636e\u626b\u63cf\u5c42\u5f00\u59cb\uff0c\u8fd9\u79cd\u5f62\u5f0f\u7684\u5904\u7406\u5e76\u6ca1\u6709\u5c1d\u8bd5\u5728\u4e00\u5f00\u59cb\u5c31\u89e3\u6790\u6574\u4e2a\u8bad\u7ec3\u6570\u636e\u3002\u6bd4\u5982\uff1a\u5bf9\u4e8e\u4e00\u4e2a\u5927\u5c0f\u4e3a200X200\u50cf\u7d20\u7684\u56fe\u50cf\uff0c\u4f60\u4e0d\u4f1a\u60f3\u6784\u5efa\u4e00\u4e2a40000\u4e2a\u8282\u70b9\u7684\u795e\u7ecf\u5143\u5c42\u3002\u800c\u662f\uff0c\u6784\u5efa\u4e00\u4e2a20X20\u50cf\u7d20\u7684\u8f93\u5165\u626b\u63cf\u5c42\uff0c\u7136\u540e\uff0c\u628a\u539f\u59cb\u56fe\u50cf\u7b2c\u4e00\u90e8\u5206\u768420X20\u50cf\u7d20\u56fe\u50cf\uff08\u901a\u5e38\u662f\u4ece\u56fe\u50cf\u7684\u5de6\u4e0a\u65b9\u5f00\u59cb\uff09\u8f93\u5165\u5230\u8fd9\u4e2a\u626b\u63cf\u5c42\u3002\u5f53\u8fd9\u90e8\u5206\u56fe\u50cf\uff08\u53ef\u80fd\u662f\u7528\u4e8e\u8fdb\u884c\u5377\u79ef\u795e\u7ecf\u7f51\u7edc\u7684\u8bad\u7ec3\uff09\u5904\u7406\u5b8c\uff0c\u4f60\u4f1a\u63a5\u7740\u5904\u7406\u4e0b\u4e00\u90e8\u5206\u768420X20\u50cf\u7d20\u56fe\u50cf\uff1a\u9010\u6e10\uff08\u901a\u5e38\u60c5\u51b5\u4e0b\u662f\u79fb\u52a8\u4e00\u4e2a\u50cf\u7d20\uff0c\u4f46\u662f\uff0c\u79fb\u52a8\u7684\u6b65\u957f\u662f\u53ef\u4ee5\u8bbe\u7f6e\u7684\uff09\u79fb\u52a8\u626b\u63cf\u5c42\uff0c\u6765\u5904\u7406\u539f\u59cb\u6570\u636e\u3002<\/p>\n<p>\u6ce8\u610f\uff0c\u4f60\u4e0d\u662f\u4e00\u6b21\u6027\u79fb\u52a8\u626b\u63cf\u5c4220\u4e2a\u50cf\u7d20\uff08\u6216\u5176\u5b83\u4efb\u4f55\u626b\u63cf\u5c42\u5927\u5c0f\u7684\u5c3a\u5ea6\uff09\uff0c\u4e5f\u4e0d\u662f\u628a\u539f\u59cb\u56fe\u50cf\u5207\u5206\u621020X20\u50cf\u7d20\u7684\u56fe\u50cf\u5757\uff0c\u800c\u662f\u7528\u626b\u63cf\u5c42\u5728\u539f\u59cb\u56fe\u50cf\u4e0a\u6ed1\u8fc7\u3002\u8fd9\u4e2a\u8f93\u5165\u6570\u636e\uff0820X20\u50cf\u7d20\u7684\u56fe\u50cf\u5757\uff09\u7d27\u63a5\u7740\u88ab\u8f93\u5165\u5230\u5377\u79ef\u5c42\uff0c\u800c\u975e\u5e38\u89c4\u7684\u795e\u7ecf\u7ec6\u80de\u5c42\u2014\u2014\u5377\u79ef\u5c42\u7684\u8282\u70b9\u4e0d\u662f\u5168\u8fde\u63a5\u3002\u6bcf\u4e00\u4e2a\u8f93\u5165\u8282\u70b9\u53ea\u4f1a\u548c\u6700\u8fd1\u7684\u90a3\u4e2a\u795e\u7ecf\u5143\u8282\u70b9\u8fde\u63a5\uff08\u81f3\u4e8e\u591a\u8fd1\u8981\u53d6\u51b3\u4e8e\u5177\u4f53\u7684\u5b9e\u73b0\uff0c\u4f46\u901a\u5e38\u4e0d\u4f1a\u8d85\u8fc7\u51e0\u4e2a\uff09\u3002<\/p>\n<p>\u8fd9\u4e9b\u5377\u79ef\u5c42\u4f1a\u968f\u7740\u6df1\u5ea6\u7684\u589e\u52a0\u800c\u9010\u6e10\u53d8\u5c0f\uff1a\u5927\u591a\u6570\u60c5\u51b5\u4e0b\uff0c\u4f1a\u6309\u7167\u8f93\u5165\u5c42\u6570\u91cf\u7684\u67d0\u4e2a\u56e0\u5b50\u7f29\u5c0f\uff08\u6bd4\u5982\uff1a20\u4e2a\u795e\u7ecf\u5143\u7684\u5377\u79ef\u5c42\uff0c\u540e\u9762\u662f10\u4e2a\u795e\u7ecf\u5143\u7684\u5377\u79ef\u5c42\uff0c\u518d\u540e\u9762\u5c31\u662f5\u4e2a\u795e\u7ecf\u5143\u7684\u5377\u79ef\u5c42\uff09\u30022\u7684n\u6b21\u65b9\uff0832, 16, 8, 4, 2, 1\uff09\u4e5f\u662f\u4e00\u4e2a\u975e\u5e38\u5e38\u7528\u7684\u56e0\u5b50\uff0c\u56e0\u4e3a\u5b83\u4eec\u5728\u5b9a\u4e49\u4e0a\u53ef\u4ee5\u7b80\u6d01\u4e14\u5b8c\u6574\u5730\u9664\u5c3d\u3002\u9664\u4e86\u5377\u79ef\u5c42\uff0c\u6c60\u5316\u5c42\uff08pooling layers\uff09\u4e5f\u975e\u5e38\u91cd\u8981\u3002<\/p>\n<p>\u6c60\u5316\u662f\u4e00\u79cd\u8fc7\u6ee4\u6389\u7ec6\u8282\u7684\u65b9\u5f0f\uff1a\u4e00\u79cd\u5e38\u7528\u7684\u6c60\u5316\u65b9\u5f0f\u662f\u6700\u5927\u6c60\u5316\uff0c\u6bd4\u5982\u75282X2\u7684\u50cf\u7d20\uff0c\u7136\u540e\u53d6\u56db\u4e2a\u50cf\u7d20\u4e2d\u503c\u6700\u5927\u7684\u90a3\u4e2a\u4f20\u9012\u3002\u4e3a\u4e86\u8ba9\u5377\u79ef\u795e\u7ecf\u7f51\u7edc\u5904\u7406\u8bed\u97f3\u6570\u636e\uff0c\u9700\u8981\u628a\u8bed\u97f3\u6570\u636e\u5207\u5206\uff0c\u4e00\u6bb5\u4e00\u6bb5\u8f93\u5165\u3002\u5728\u5b9e\u9645\u5e94\u7528\u4e2d\uff0c\u901a\u5e38\u4f1a\u5728\u5377\u79ef\u795e\u7ecf\u7f51\u7edc\u540e\u9762\u52a0\u4e00\u4e2a\u524d\u9988\u795e\u7ecf\u7f51\u7edc\uff0c\u4ee5\u8fdb\u4e00\u6b65\u5904\u7406\u6570\u636e\uff0c\u4ece\u800c\u5bf9\u6570\u636e\u8fdb\u884c\u66f4\u9ad8\u6c34\u5e73\u7684\u975e\u7ebf\u6027\u62bd\u8c61\u3002<\/p>\n<p><a href=\"http:\/\/yann.lecun.com\/exdb\/publis\/pdf\/lecun-98.pdf\" target=\"_blank\" rel=\"noopener\">LeCun, Yann, et al. \u201cGradient-based learning applied to document recognition.\u201d Proceedings of the IEEE 86.11 (1998): 2278-2324.<\/a><\/p>\n<h1><a name=\"t24\"><\/a><a id=\"14_DN_225\"><\/a>14. \u89e3\u5377\u79ef\u7f51\u7edc\uff08DN\uff09<\/h1>\n<p><img decoding=\"async\" src=\"https:\/\/img-blog.csdnimg.cn\/img_convert\/cb9db2c8be36bf2012279ddb950ffcd4.png#pic_center\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<p>\u89e3\u5377\u79ef\u7f51\u7edc\uff08DN\uff1aDeconvolutional networks\uff09\uff0c\u53c8\u79f0\u4e3a\u9006\u56fe\u5f62\u7f51\u7edc\uff08IGNs\uff1ainverse graphics networks\uff09\uff0c\u662f\u9006\u5411\u7684\u5377\u79ef\u795e\u7ecf\u7f51\u7edc\u3002<\/p>\n<p>\u60f3\u8c61\u4e00\u4e0b\uff0c\u7ed9\u4e00\u4e2a\u795e\u7ecf\u7f51\u7edc\u8f93\u5165\u4e00\u4e2a\u201c\u732b\u201d\u7684\u8bcd\uff0c\u5c31\u53ef\u4ee5\u751f\u6210\u4e00\u4e2a\u50cf\u732b\u4e00\u6837\u7684\u56fe\u50cf\uff0c\u901a\u8fc7\u6bd4\u5bf9\u5b83\u548c\u771f\u5b9e\u7684\u732b\u7684\u56fe\u7247\u6765\u8fdb\u884c\u8bad\u7ec3\u3002\u8ddf\u5e38\u89c4CNN\u4e00\u6837\uff0cDN\u4e5f\u53ef\u4ee5\u7ed3\u5408FFNN\u4f7f\u7528\uff0c\u4f46\u6ca1\u5fc5\u8981\u4e3a\u8fd9\u4e2a\u65b0\u7684\u7f29\u5199\u91cd\u65b0\u505a\u56fe\u89e3\u91ca\u3002\u5b83\u4eec\u53ef\u88ab\u79f0\u4e3a\u6df1\u5ea6\u89e3\u5377\u79ef\u7f51\u7edc\uff0c\u4f46\u628aFFNN\u653e\u5230DNN\u524d\u9762\u548c\u540e\u9762\u662f\u4e0d\u540c\u7684\uff0c\u90a3\u662f\u4e24\u79cd\u67b6\u6784\uff08\u4e5f\u5c31\u9700\u8981\u4e24\u4e2a\u540d\u5b57\uff09\uff0c\u5bf9\u4e8e\u662f\u5426\u9700\u8981\u4e24\u4e2a\u4e0d\u540c\u7684\u540d\u5b57\u4f60\u4eec\u53ef\u80fd\u4f1a\u6709\u4e89\u8bba\u3002\u9700\u8981\u6ce8\u610f\u7684\u662f\uff0c\u7edd\u5927\u591a\u6570\u5e94\u7528\u90fd\u4e0d\u4f1a\u628a\u6587\u672c\u6570\u636e\u76f4\u63a5\u8f93\u5165\u5230\u795e\u7ecf\u7f51\u7edc\uff0c\u800c\u662f\u7528\u4e8c\u5143\u8f93\u5165\u5411\u91cf\u3002\u6bd4\u5982<0,1>\u4ee3\u8868\u732b\uff0c<1,0>\u4ee3\u8868\u72d7\uff0c<1,1>\u4ee3\u8868\u732b\u548c\u72d7\u3002<\/p>\n<p>CNN\u7684\u6c60\u5316\u5c42\u5f80\u5f80\u4e5f\u662f\u88ab\u5bf9\u5e94\u7684\u9006\u5411\u64cd\u4f5c\u66ff\u6362\u4e86\uff0c\u4e3b\u8981\u662f\u63d2\u503c\u548c\u5916\u63a8\uff08\u57fa\u4e8e\u4e00\u4e2a\u57fa\u672c\u7684\u5047\u8bbe\uff1a\u5982\u679c\u4e00\u4e2a\u6c60\u5316\u5c42\u4f7f\u7528\u4e86\u6700\u5927\u6c60\u5316\uff0c\u4f60\u53ef\u4ee5\u5728\u9006\u64cd\u4f5c\u7684\u65f6\u5019\u751f\u6210\u4e00\u4e9b\u76f8\u5bf9\u4e8e\u6700\u5927\u503c\u66f4\u5c0f\u7684\u6570\u636e\uff09\u3002<\/p>\n<p><a href=\"http:\/\/www.matthewzeiler.com\/pubs\/cvpr2010\/cvpr2010.pdf\" target=\"_blank\" rel=\"noopener\">Zeiler, Matthew D., et al. \u201cDeconvolutional networks.\u201d Computer Vision and Pattern Recognition (CVPR), 2010 IEEE Conference on. IEEE, 2010.<\/a><\/p>\n<h1><a name=\"t25\"><\/a><a id=\"15__DCIGN_238\"><\/a>15. \u6df1\u5ea6\u5377\u79ef\u9006\u5411\u56fe\u7f51\u7edc\uff08DCIGN\uff09<\/h1>\n<p><img decoding=\"async\" src=\"https:\/\/img-blog.csdnimg.cn\/img_convert\/46ef8a5af19daba608be370e100d5b15.png#pic_center\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<p>\u6df1\u5ea6\u5377\u79ef\u9006\u5411\u56fe\u7f51\u7edc\uff08DCIGN\uff1aDeep convolutional inverse graphics networks\uff09\uff0c\u8fd9\u4e2a\u540d\u5b57\u5177\u6709\u8bef\u5bfc\u6027\uff0c\u56e0\u4e3a\u5b83\u4eec\u5b9e\u9645\u4e0a\u662fVAE\uff0c\u4f46\u5206\u522b\u7528CNN\u3001DNN\u6765\u4f5c\u7f16\u7801\u548c\u89e3\u7801\u7684\u90e8\u5206\u3002<\/p>\n<p>\u8fd9\u4e9b\u7f51\u7edc\u5c1d\u8bd5\u5728\u7f16\u7801\u8fc7\u7a0b\u4e2d\u5bf9\u201c\u7279\u5f81\u201c\u8fdb\u884c\u6982\u7387\u5efa\u6a21\uff0c\u8fd9\u6837\u4e00\u6765\uff0c\u4f60\u53ea\u8981\u7528\u732b\u548c\u72d7\u7684\u72ec\u7167\uff0c\u5c31\u80fd\u8ba9\u5b83\u4eec\u751f\u6210\u4e00\u5f20\u732b\u548c\u72d7\u7684\u5408\u7167\u3002\u540c\u7406\uff0c\u4f60\u53ef\u4ee5\u8f93\u5165\u4e00\u5f20\u732b\u7684\u7167\u7247\uff0c\u5982\u679c\u732b\u65c1\u8fb9\u6709\u4e00\u53ea\u607c\u4eba\u7684\u90bb\u5bb6\u72d7\uff0c\u4f60\u53ef\u4ee5\u8ba9\u5b83\u4eec\u628a\u72d7\u53bb\u6389\u3002\u5f88\u591a\u6f14\u793a\u8868\u660e\uff0c\u8fd9\u79cd\u7c7b\u578b\u7684\u7f51\u7edc\u80fd\u5b66\u4f1a\u57fa\u4e8e\u56fe\u50cf\u7684\u590d\u6742\u53d8\u6362\uff0c\u6bd4\u5982\u706f\u5149\u5f3a\u5f31\u7684\u53d8\u5316\u30013D\u7269\u4f53\u7684\u65cb\u8f6c\u3002\u4e00\u822c\u4e5f\u662f\u7528\u53cd\u5411\u4f20\u64ad\u7b97\u6cd5\u6765\u8bad\u7ec3\u6b64\u7c7b\u7f51\u7edc\u3002<\/p>\n<p><a href=\"https:\/\/arxiv.org\/pdf\/1503.03167v4.pdf\" target=\"_blank\" rel=\"noopener\">Kulkarni, Tejas D., et al. \u201cDeep convolutional inverse graphics network.\u201d Advances in Neural Information Processing Systems. 2015.<\/a><\/p>\n<h1><a name=\"t26\"><\/a><a id=\"16_GAN_249\"><\/a>16. \u751f\u6210\u5f0f\u5bf9\u6297\u7f51\u7edc\uff08GAN\uff09<\/h1>\n<p><img decoding=\"async\" src=\"https:\/\/img-blog.csdnimg.cn\/img_convert\/49fd52a7a26bd76d8e5f55152c0ce9e3.png#pic_center\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<p>\u751f\u6210\u5f0f\u5bf9\u6297\u7f51\u7edc\uff08GAN\uff1aGenerative adversarial networks\uff09\u662f\u4e00\u7c7b\u4e0d\u540c\u7684\u7f51\u7edc\uff0c\u5b83\u4eec\u6709\u4e00\u5bf9\u201c\u53cc\u80de\u80ce\u201d\uff1a\u4e24\u4e2a\u7f51\u7edc\u534f\u540c\u5de5\u4f5c\u3002<\/p>\n<p>GAN\u53ef\u7531\u4efb\u610f\u4e24\u79cd\u7f51\u7edc\u7ec4\u6210\uff08\u4f46\u901a\u5e38\u662fFF\u548cCNN\uff09\uff0c\u5176\u4e2d\u4e00\u4e2a\u7528\u4e8e\u751f\u6210\u5185\u5bb9\uff0c\u53e6\u4e00\u4e2a\u5219\u7528\u4e8e\u9274\u522b\u751f\u6210\u7684\u5185\u5bb9\u3002<\/p>\n<p>\u9274\u522b\u7f51\u7edc\uff08discriminating network\uff09\u540c\u65f6\u63a5\u6536\u8bad\u7ec3\u6570\u636e\u548c\u751f\u6210\u7f51\u7edc\uff08generative network\uff09\u751f\u6210\u7684\u6570\u636e\u3002\u9274\u522b\u7f51\u7edc\u7684\u51c6\u786e\u7387\uff0c\u88ab\u7528\u4f5c\u751f\u6210\u7f51\u7edc\u8bef\u5dee\u7684\u4e00\u90e8\u5206\u3002\u8fd9\u5c31\u5f62\u6210\u4e86\u4e00\u79cd\u7ade\u4e89\uff1a\u9274\u522b\u7f51\u7edc\u8d8a\u6765\u8d8a\u64c5\u957f\u4e8e\u533a\u5206\u771f\u5b9e\u7684\u6570\u636e\u548c\u751f\u6210\u6570\u636e\uff0c\u800c\u751f\u6210\u7f51\u7edc\u4e5f\u8d8a\u6765\u8d8a\u5584\u4e8e\u751f\u6210\u96be\u4ee5\u9884\u6d4b\u7684\u6570\u636e\u3002\u8fd9\u79cd\u65b9\u5f0f\u975e\u5e38\u6709\u6548\uff0c\u90e8\u5206\u662f\u56e0\u4e3a\uff1a\u5373\u4fbf\u76f8\u5f53\u590d\u6742\u7684\u7c7b\u566a\u97f3\u6a21\u5f0f\u6700\u7ec8\u90fd\u662f\u53ef\u9884\u6d4b\u7684\uff0c\u4f46\u8ddf\u8f93\u5165\u6570\u636e\u6709\u7740\u6781\u4e3a\u76f8\u4f3c\u7279\u5f81\u7684\u751f\u6210\u6570\u636e\uff0c\u5219\u5f88\u96be\u533a\u5206\u3002<\/p>\n<p>\u8bad\u7ec3GAN\u6781\u5177\u6311\u6218\u6027\uff0c\u56e0\u4e3a\u4f60\u4e0d\u4ec5\u8981\u8bad\u7ec3\u4e24\u4e2a\u795e\u7ecf\u7f51\u7edc\uff08\u5176\u4e2d\u7684\u4efb\u4f55\u4e00\u4e2a\u90fd\u4f1a\u51fa\u73b0\u5b83\u81ea\u5df1\u7684\u95ee\u9898\uff09\uff0c\u540c\u65f6\u8fd8\u8981\u5e73\u8861\u4e24\u8005\u7684\u8fd0\u884c\u673a\u5236\u3002\u5982\u679c\u9884\u6d4b\u6216\u751f\u6210\u76f8\u6bd4\u5bf9\u65b9\u8868\u73b0\u5f97\u8fc7\u597d\uff0c\u8fd9\u4e2aGAN\u5c31\u4e0d\u4f1a\u6536\u655b\uff0c\u56e0\u4e3a\u5b83\u4f1a\u5185\u90e8\u53d1\u6563\u3002<\/p>\n<p><a href=\"https:\/\/arxiv.org\/pdf\/1406.2661v1.pdf\" target=\"_blank\" rel=\"noopener\">Goodfellow, Ian, et al. \u201cGenerative adversarial nets.\u201d Advances in Neural Information Processing Systems. 2014.<\/a><\/p>\n<h1><a name=\"t27\"><\/a><a id=\"17_RNN_264\"><\/a>17. \u5faa\u73af\u795e\u7ecf\u7f51\u7edc\uff08RNN\uff09<\/h1>\n<p><img decoding=\"async\" src=\"https:\/\/img-blog.csdnimg.cn\/img_convert\/670480dac0b69982cdd4dce9a399ed73.png#pic_center\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<p>\u5faa\u73af\u795e\u7ecf\u7f51\u7edc\uff08RNN\uff1aRecurrent neural networks\uff09\u662f\u5177\u6709\u65f6\u95f4\u8054\u7ed3\u7684\u524d\u9988\u795e\u7ecf\u7f51\u7edc\uff1a\u5b83\u4eec\u6709\u4e86\u72b6\u6001\uff0c\u901a\u9053\u4e0e\u901a\u9053\u4e4b\u95f4\u6709\u4e86\u65f6\u95f4\u4e0a\u7684\u8054\u7cfb\u3002\u795e\u7ecf\u5143\u7684\u8f93\u5165\u4fe1\u606f\uff0c\u4e0d\u4ec5\u5305\u62ec\u524d\u4e00\u795e\u7ecf\u7ec6\u80de\u5c42\u7684\u8f93\u51fa\uff0c\u8fd8\u5305\u62ec\u5b83\u81ea\u8eab\u5728\u5148\u524d\u901a\u9053\u7684\u72b6\u6001\u3002<\/p>\n<p>\u8fd9\u5c31\u610f\u5473\u7740\uff1a\u4f60\u7684\u8f93\u5165\u987a\u5e8f\u5c06\u4f1a\u5f71\u54cd\u795e\u7ecf\u7f51\u7edc\u7684\u8bad\u7ec3\u7ed3\u679c\uff1a\u76f8\u6bd4\u5148\u8f93\u5165\u201c\u66f2\u5947\u997c\u201d\u518d\u8f93\u5165\u201c\u725b\u5976\u201d\uff0c\u5148\u8f93\u5165\u201c\u725b\u5976\u201d\u518d\u8f93\u5165\u201c\u66f2\u5947\u997c\u201d\u540e\uff0c\u6216\u8bb8\u4f1a\u4ea7\u751f\u4e0d\u540c\u7684\u7ed3\u679c\u3002RNN\u5b58\u5728\u4e00\u5927\u95ee\u9898\uff1a\u68af\u5ea6\u6d88\u5931\uff08\u6216\u68af\u5ea6\u7206\u70b8\uff0c\u8fd9\u53d6\u51b3\u4e8e\u6240\u7528\u7684\u6fc0\u6d3b\u51fd\u6570\uff09\uff0c\u4fe1\u606f\u4f1a\u968f\u65f6\u95f4\u8fc5\u901f\u6d88\u5931\uff0c\u6b63\u5982FFNN\u4f1a\u968f\u7740\u6df1\u5ea6\u7684\u589e\u52a0\u800c\u5931\u53bb\u4fe1\u606f\u4e00\u6837\u3002<\/p>\n<p>\u76f4\u89c9\u4e0a\uff0c\u8fd9\u4e0d\u7b97\u4ec0\u4e48\u5927\u95ee\u9898\uff0c\u56e0\u4e3a\u8fd9\u4e9b\u90fd\u53ea\u662f\u6743\u91cd\uff0c\u800c\u975e\u795e\u7ecf\u5143\u7684\u72b6\u6001\uff0c\u4f46\u968f\u65f6\u95f4\u53d8\u5316\u7684\u6743\u91cd\u6b63\u662f\u6765\u81ea\u8fc7\u53bb\u4fe1\u606f\u7684\u5b58\u50a8\uff1b\u5982\u679c\u6743\u91cd\u662f0\u62161000000\uff0c\u90a3\u4e4b\u524d\u7684\u72b6\u6001\u5c31\u4e0d\u518d\u6709\u4fe1\u606f\u4ef7\u503c\u3002<\/p>\n<p>\u539f\u5219\u4e0a\uff0cRNN\u53ef\u4ee5\u5728\u5f88\u591a\u9886\u57df\u4f7f\u7528\uff0c\u56e0\u4e3a\u5927\u90e8\u5206\u6570\u636e\u5728\u5f62\u5f0f\u4e0a\u4e0d\u5b58\u5728\u65f6\u95f4\u7ebf\u7684\u53d8\u5316\uff0c\uff08\u4e0d\u50cf\u8bed\u97f3\u6216\u89c6\u9891\uff09\uff0c\u5b83\u4eec\u80fd\u4ee5\u67d0\u79cd\u5e8f\u5217\u7684\u5f62\u5f0f\u5448\u73b0\u51fa\u6765\u3002\u4e00\u5f20\u56fe\u7247\u6216\u4e00\u6bb5\u6587\u5b57\u53ef\u4ee5\u4e00\u4e2a\u50cf\u7d20\u6216\u8005\u4e00\u4e2a\u6587\u5b57\u5730\u8fdb\u884c\u8f93\u5165\uff0c\u56e0\u6b64\uff0c\u4e0e\u65f6\u95f4\u76f8\u5173\u7684\u6743\u91cd\u63cf\u8ff0\u4e86\u8be5\u5e8f\u5217\u524d\u4e00\u6b65\u53d1\u751f\u4e86\u4ec0\u4e48\uff0c\u800c\u4e0d\u662f\u591a\u5c11\u79d2\u4e4b\u524d\u53d1\u751f\u4e86\u4ec0\u4e48\u3002\u4e00\u822c\u6765\u8bf4\uff0c\u5faa\u73af\u795e\u7ecf\u7f51\u7edc\u662f\u63a8\u6d4b\u6216\u8865\u5168\u4fe1\u606f\u5f88\u597d\u7684\u9009\u62e9\uff0c\u6bd4\u5982\u81ea\u52a8\u8865\u5168\u3002<\/p>\n<p><a href=\"https:\/\/crl.ucsd.edu\/~elman\/Papers\/fsit.pdf\" target=\"_blank\" rel=\"noopener\">Elman, Jeffrey L. \u201cFinding structure in time.\u201d Cognitive science 14.2 (1990): 179-211.<\/a><\/p>\n<h1><a name=\"t28\"><\/a><a id=\"18_LSTM_279\"><\/a>18. \u957f\u77ed\u671f\u8bb0\u5fc6\uff08LSTM\uff09<\/h1>\n<p><img decoding=\"async\" src=\"https:\/\/img-blog.csdnimg.cn\/img_convert\/365d21235b5c206b13b19eaa223fc5dc.png#pic_center\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<p>\u957f\u77ed\u671f\u8bb0\u5fc6\uff08LSTM\uff1aLong \/ short term memory\uff09\u7f51\u7edc\u8bd5\u56fe\u901a\u8fc7\u5f15\u5165\u95e8\u7ed3\u6784\u4e0e\u660e\u786e\u5b9a\u4e49\u7684\u8bb0\u5fc6\u5355\u5143\u6765\u89e3\u51b3\u68af\u5ea6\u6d88\u5931\/\u7206\u70b8\u7684\u95ee\u9898\u3002<\/p>\n<p>\u8fd9\u66f4\u591a\u7684\u662f\u53d7\u7535\u8def\u56fe\u8bbe\u8ba1\u7684\u542f\u53d1\uff0c\u800c\u975e\u751f\u7269\u5b66\u4e0a\u67d0\u79cd\u548c\u8bb0\u5fc6\u76f8\u5173\u673a\u5236\u3002\u6bcf\u4e2a\u795e\u7ecf\u5143\u90fd\u6709\u4e00\u4e2a\u8bb0\u5fc6\u5355\u5143\u548c\u4e09\u4e2a\u95e8\uff1a\u8f93\u5165\u95e8\u3001\u8f93\u51fa\u95e8\u3001\u9057\u5fd8\u95e8\u3002\u8fd9\u4e09\u4e2a\u95e8\u7684\u529f\u80fd\u5c31\u662f\u901a\u8fc7\u7981\u6b62\u6216\u5141\u8bb8\u4fe1\u606f\u6d41\u52a8\u6765\u4fdd\u62a4\u4fe1\u606f\u3002<\/p>\n<p>\u8f93\u5165\u95e8\u51b3\u5b9a\u4e86\u6709\u591a\u5c11\u524d\u4e00\u795e\u7ecf\u7ec6\u80de\u5c42\u7684\u4fe1\u606f\u53ef\u7559\u5728\u5f53\u524d\u8bb0\u5fc6\u5355\u5143\uff0c\u8f93\u51fa\u5c42\u5728\u53e6\u4e00\u7aef\u51b3\u5b9a\u4e0b\u4e00\u795e\u7ecf\u7ec6\u80de\u5c42\u80fd\u4ece\u5f53\u524d\u795e\u7ecf\u5143\u83b7\u53d6\u591a\u5c11\u4fe1\u606f\u3002\u9057\u5fd8\u95e8\u4e4d\u770b\u5f88\u5947\u602a\uff0c\u4f46\u6709\u65f6\u5019\u9057\u5fd8\u90e8\u5206\u4fe1\u606f\u662f\u5f88\u6709\u7528\u7684\uff1a\u6bd4\u5982\u8bf4\u5b83\u5728\u5b66\u4e60\u4e00\u672c\u4e66\uff0c\u5e76\u5f00\u59cb\u5b66\u4e00\u4e2a\u65b0\u7684\u7ae0\u8282\uff0c\u90a3\u9057\u5fd8\u524d\u9762\u7ae0\u8282\u7684\u90e8\u5206\u89d2\u8272\u5c31\u5f88\u6709\u5fc5\u8981\u4e86\u3002<\/p>\n<p>\u5b9e\u8df5\u8bc1\u660e\uff0cLSTM\u53ef\u7528\u6765\u5b66\u4e60\u590d\u6742\u7684\u5e8f\u5217\uff0c\u6bd4\u5982\u50cf\u838e\u58eb\u6bd4\u4e9a\u4e00\u6837\u5199\u4f5c\uff0c\u6216\u521b\u4f5c\u5168\u65b0\u7684\u97f3\u4e50\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u6bcf\u4e00\u4e2a\u95e8\u90fd\u5bf9\u524d\u4e00\u795e\u7ecf\u5143\u7684\u8bb0\u5fc6\u5355\u5143\u8d4b\u6709\u4e00\u4e2a\u6743\u91cd\uff0c\u56e0\u6b64\u4f1a\u9700\u8981\u66f4\u591a\u7684\u8ba1\u7b97\u8d44\u6e90\u3002<\/p>\n<p><a href=\"http:\/\/deeplearning.cs.cmu.edu\/pdfs\/Hochreiter97_lstm.pdf\" target=\"_blank\" rel=\"noopener\">Hochreiter, Sepp, and J\u00fcrgen Schmidhuber. \u201cLong short-term memory.\u201d Neural computation 9.8 (1997): 1735-1780.<\/a><\/p>\n<h1><a name=\"t29\"><\/a><a id=\"19_GRU_294\"><\/a>19. \u95e8\u5faa\u73af\u5355\u5143\uff08GRU\uff09<\/h1>\n<p><img decoding=\"async\" src=\"https:\/\/img-blog.csdnimg.cn\/img_convert\/b70c3770529c3def6e4c5851fc4d6201.png#pic_center\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<p>\u95e8\u5faa\u73af\u5355\u5143\uff08GRU : Gated recurrent units\uff09\u662fLSTM\u7684\u4e00\u79cd\u8f7b\u91cf\u7ea7\u53d8\u4f53\u3002\u5b83\u4eec\u5c11\u4e86\u4e00\u4e2a\u95e8\uff0c\u540c\u65f6\u8fde\u63a5\u65b9\u5f0f\u4e5f\u7a0d\u6709\u4e0d\u540c\uff1a\u5b83\u4eec\u91c7\u7528\u4e86\u4e00\u4e2a\u66f4\u65b0\u95e8\uff08update gate\uff09\uff0c\u800c\u975eLSTM\u6240\u7528\u7684\u8f93\u5165\u95e8\u3001\u8f93\u51fa\u95e8\u3001\u9057\u5fd8\u95e8\u3002<\/p>\n<p>\u66f4\u65b0\u95e8\u51b3\u5b9a\u4e86\u4fdd\u7559\u591a\u5c11\u4e0a\u4e00\u4e2a\u72b6\u6001\u7684\u4fe1\u606f\uff0c\u8fd8\u51b3\u5b9a\u4e86\u6536\u53d6\u591a\u5c11\u6765\u81ea\u524d\u4e00\u795e\u7ecf\u7ec6\u80de\u5c42\u7684\u4fe1\u606f\u3002\u91cd\u7f6e\u95e8\uff08reset gate\uff09\u8ddfLSTM\u9057\u5fd8\u95e8\u7684\u529f\u80fd\u5f88\u76f8\u4f3c\uff0c\u4f46\u5b83\u5b58\u5728\u7684\u4f4d\u7f6e\u5374\u7a0d\u6709\u4e0d\u540c\u3002\u5b83\u4eec\u603b\u662f\u8f93\u51fa\u5b8c\u6574\u7684\u72b6\u6001\uff0c\u6ca1\u6709\u8f93\u51fa\u95e8\u3002\u591a\u6570\u60c5\u51b5\u4e0b\uff0c\u5b83\u4eec\u8ddfLSTM\u7c7b\u4f3c\uff0c\u4f46\u6700\u5927\u7684\u4e0d\u540c\u662f\uff1aGRU\u901f\u5ea6\u66f4\u5feb\u3001\u8fd0\u884c\u66f4\u5bb9\u6613\uff08\u4f46\u51fd\u6570\u8868\u8fbe\u529b\u7a0d\u5f31\uff09\u3002<\/p>\n<p>\u5728\u5b9e\u8df5\u4e2d\uff0c\u8fd9\u91cc\u7684\u4f18\u52bf\u548c\u52a3\u52bf\u4f1a\u76f8\u4e92\u62b5\u6d88\uff1a\u5f53\u4f60\u4f60\u9700\u8981\u66f4\u5927\u7684\u7f51\u7edc\u6765\u83b7\u53d6\u51fd\u6570\u8868\u8fbe\u529b\u65f6\uff0c\u8fd9\u6837\u53cd\u8fc7\u6765\uff0c\u6027\u80fd\u4f18\u52bf\u5c31\u88ab\u62b5\u6d88\u4e86\u3002\u5728\u4e0d\u9700\u8981\u989d\u5916\u7684\u51fd\u6570\u8868\u8fbe\u529b\u65f6\uff0cGRU\u7684\u7efc\u5408\u6027\u80fd\u8981\u597d\u4e8eLSTM\u3002<\/p>\n<p><a href=\"https:\/\/arxiv.org\/pdf\/1412.3555v1.pdf\" target=\"_blank\" rel=\"noopener\">Chung, Junyoung, et al. \u201cEmpirical evaluation of gated recurrent neural networks on sequence modeling.\u201d arXiv preprint arXiv:1412.3555 (2014).<\/a><\/p>\n<h1><a name=\"t30\"><\/a><a id=\"20_NTM_307\"><\/a>20. \u795e\u7ecf\u56fe\u7075\u673a\uff08NTM\uff09<\/h1>\n<p><img decoding=\"async\" src=\"https:\/\/img-blog.csdnimg.cn\/img_convert\/e405a01a895fc36bb542897ff9c03ea3.png#pic_center\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<p>\u795e\u7ecf\u56fe\u7075\u673a\uff08NTM: Neural Turing machines\uff09\u53ef\u4ee5\u7406\u89e3\u4e3a\u5bf9LSTM\u7684\u62bd\u8c61\uff0c\u5b83\u8bd5\u56fe\u628a\u795e\u7ecf\u7f51\u7edc\u53bb\u9ed1\u7bb1\u5316\uff08\u4ee5\u7aa5\u63a2\u5176\u5185\u90e8\u53d1\u751f\u7684\u7ec6\u8282\uff09\u3002<\/p>\n<p>NTM\u4e0d\u662f\u628a\u8bb0\u5fc6\u5355\u5143\u8bbe\u8ba1\u5728\u795e\u7ecf\u5143\u5185\uff0c\u800c\u662f\u5206\u79bb\u51fa\u6765\u3002NTM\u8bd5\u56fe\u7ed3\u5408\u5e38\u89c4\u6570\u5b57\u4fe1\u606f\u5b58\u50a8\u7684\u9ad8\u6548\u6027\u3001\u6c38\u4e45\u6027\u4e0e\u795e\u7ecf\u7f51\u7edc\u7684\u6548\u7387\u53ca\u51fd\u6570\u8868\u8fbe\u80fd\u529b\u3002\u5b83\u7684\u60f3\u6cd5\u662f\u8bbe\u8ba1\u4e00\u4e2a\u53ef\u4f5c\u5185\u5bb9\u5bfb\u5740\u7684\u8bb0\u5fc6\u5e93\uff0c\u5e76\u8ba9\u795e\u7ecf\u7f51\u7edc\u5bf9\u5176\u8fdb\u884c\u8bfb\u5199\u64cd\u4f5c\u3002NTM\u540d\u5b57\u4e2d\u7684\u201c\u56fe\u7075\uff08Turing\uff09\u201d\u662f\u8868\u660e\uff0c\u5b83\u662f\u56fe\u7075\u5b8c\u5907\uff08Turing complete\uff09\u7684\uff0c\u5373\u5177\u5907\u57fa\u4e8e\u5b83\u6240\u8bfb\u53d6\u7684\u5185\u5bb9\u6765\u8bfb\u53d6\u3001\u5199\u5165\u3001\u4fee\u6539\u72b6\u6001\u7684\u80fd\u529b\uff0c\u4e5f\u5c31\u662f\u80fd\u8868\u8fbe\u4e00\u4e2a\u901a\u7528\u56fe\u7075\u673a\u6240\u80fd\u8868\u8fbe\u7684\u4e00\u5207\u3002<\/p>\n<p><a href=\"https:\/\/arxiv.org\/pdf\/1410.5401v2.pdf\" target=\"_blank\" rel=\"noopener\">Graves, Alex, Greg Wayne, and Ivo Danihelka. \u201cNeural turing machines.\u201d arXiv preprint arXiv:1410.5401 (2014).<\/a><\/p>\n<h1><a name=\"t31\"><\/a><a id=\"22_BiRNNBiLSTMBiGRU_318\"><\/a>22. BiRNN\u3001BiLSTM\u3001BiGRU<\/h1>\n<p>\u53cc\u5411\u5faa\u73af\u795e\u7ecf\u7f51\u7edc\uff08BiRNN\uff1aBidirectional recurrent neural networks\uff09\u3001\u53cc\u5411\u957f\u77ed\u671f\u8bb0\u5fc6\u7f51\u7edc\uff08BiLSTM\uff1abidirectional long \/ short term memory networks \uff09\u548c\u53cc\u5411\u95e8\u63a7\u5faa\u73af\u5355\u5143\uff08BiGRU\uff1abidirectional gated recurrent units\uff09\u5728\u56fe\u8868\u4e2d\u5e76\u672a\u5448\u73b0\u51fa\u6765\uff0c\u56e0\u4e3a\u5b83\u4eec\u770b\u8d77\u6765\u4e0e\u5176\u5bf9\u5e94\u7684\u5355\u5411\u795e\u7ecf\u7f51\u7edc\u7ed3\u6784\u4e00\u6837\u3002<\/p>\n<p>\u6240\u4e0d\u540c\u7684\u662f\uff0c\u8fd9\u4e9b\u7f51\u7edc\u4e0d\u4ec5\u4e0e\u8fc7\u53bb\u7684\u72b6\u6001\u6709\u8fde\u63a5\uff0c\u800c\u4e14\u4e0e\u672a\u6765\u7684\u72b6\u6001\u4e5f\u6709\u8fde\u63a5\u3002\u6bd4\u5982\uff0c\u901a\u8fc7\u4e00\u4e2a\u4e00\u4e2a\u5730\u8f93\u5165\u5b57\u6bcd\uff0c\u8bad\u7ec3\u5355\u5411\u7684LSTM\u9884\u6d4b\u201c\u9c7c\uff08fish\uff09\u201d\uff08\u5728\u65f6\u95f4\u8f74\u4e0a\u7684\u5faa\u73af\u8fde\u63a5\u8bb0\u4f4f\u4e86\u8fc7\u53bb\u7684\u72b6\u6001\u503c\uff09\u3002\u5728BiLSTM\u7684\u53cd\u9988\u901a\u8def\u8f93\u5165\u5e8f\u5217\u4e2d\u7684\u4e0b\u4e00\u4e2a\u5b57\u6bcd\uff0c\u8fd9\u4f7f\u5f97\u5b83\u53ef\u4ee5\u4e86\u89e3\u672a\u6765\u7684\u4fe1\u606f\u662f\u4ec0\u4e48\u3002\u8fd9\u79cd\u5f62\u5f0f\u7684\u8bad\u7ec3\u4f7f\u5f97\u8be5\u7f51\u7edc\u53ef\u4ee5\u586b\u5145\u4fe1\u606f\u4e4b\u95f4\u7684\u7a7a\u767d\uff0c\u800c\u4e0d\u662f\u9884\u6d4b\u4fe1\u606f\u3002\u56e0\u6b64\uff0c\u5b83\u5728\u5904\u7406\u56fe\u50cf\u65f6\u4e0d\u662f\u6269\u5c55\u56fe\u50cf\u7684\u8fb9\u754c\uff0c\u800c\u662f\u586b\u8865\u4e00\u5f20\u56fe\u7247\u4e2d\u7684\u7f3a\u5931\u3002<\/p>\n<p><a href=\"http:\/\/www.di.ufpe.br\/~fnj\/RNA\/bibliografia\/BRNN.pdf\" target=\"_blank\" rel=\"noopener\">Schuster, Mike, and Kuldip K. Paliwal. \u201cBidirectional recurrent neural networks.\u201d IEEE Transactions on Signal Processing 45.11 (1997): 2673-2681.<\/a><\/p>\n<h1><a name=\"t32\"><\/a><a id=\"23_DRN_326\"><\/a>23. \u6df1\u5ea6\u6b8b\u5dee\u7f51\u7edc\uff08DRN\uff09<\/h1>\n<p><img decoding=\"async\" src=\"https:\/\/img-blog.csdnimg.cn\/img_convert\/c9399813de32e8b93e93e33a0603ab88.png#pic_center\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<p>\u6df1\u5ea6\u6b8b\u5dee\u7f51\u7edc\uff08DRN: Deep residual networks\uff09\u662f\u975e\u5e38\u6df1\u7684FFNN\u7f51\u7edc\uff0c\u5b83\u6709\u4e00\u79cd\u7279\u6b8a\u7684\u8fde\u63a5\uff0c\u53ef\u4ee5\u628a\u4fe1\u606f\u4ece\u67d0\u4e00\u795e\u7ecf\u7ec6\u80de\u5c42\u4f20\u81f3\u540e\u9762\u51e0\u5c42\uff08\u901a\u5e38\u662f2\u52305\u5c42\uff09\u3002<\/p>\n<p>\u8be5\u7f51\u7edc\u7684\u76ee\u7684\u4e0d\u662f\u8981\u627e\u8f93\u5165\u6570\u636e\u4e0e\u8f93\u51fa\u6570\u636e\u4e4b\u95f4\u7684\u6620\u5c04\uff0c\u800c\u662f\u81f4\u529b\u4e8e\u6784\u5efa\u8f93\u5165\u6570\u636e\u4e0e\u8f93\u51fa\u6570\u636e+\u8f93\u5165\u6570\u636e\u4e4b\u95f4\u7684\u6620\u5c04\u51fd\u6570\u3002\u672c\u8d28\u4e0a\uff0c\u5b83\u5728\u7ed3\u679c\u4e2d\u589e\u52a0\u4e00\u4e2a\u6052\u7b49\u51fd\u6570\uff0c\u5e76\u8ddf\u524d\u9762\u7684\u8f93\u5165\u4e00\u8d77\u4f5c\u4e3a\u540e\u4e00\u5c42\u7684\u65b0\u8f93\u5165\u3002\u7ed3\u679c\u8868\u660e\uff0c\u5f53\u5c42\u6570\u8d85\u8fc7150\u540e\uff0c\u8fd9\u4e00\u7f51\u7edc\u5c06\u975e\u5e38\u64c5\u4e8e\u5b66\u4e60\u6a21\u5f0f\uff0c\u8fd9\u6bd4\u5e38\u89c4\u76842\u52305\u5c42\u8981\u591a\u5f97\u591a\u3002\u7136\u800c\uff0c\u6709\u8bc1\u636e\u8868\u660e\u8fd9\u4e9b\u7f51\u7edc\u672c\u8d28\u4e0a\u53ea\u662f\u6ca1\u6709\u65f6\u95f4\u7ed3\u6784\u7684RNN\uff0c\u5b83\u4eec\u603b\u662f\u4e0e\u6ca1\u6709\u95e8\u7ed3\u6784\u7684LSTM\u76f8\u63d0\u5e76\u8bba\u3002<\/p>\n<p><a href=\"https:\/\/arxiv.org\/pdf\/1512.03385v1.pdf\" target=\"_blank\" rel=\"noopener\">He, Kaiming, et al. \u201cDeep residual learning for image recognition.\u201d arXiv preprint arXiv:1512.03385 (2015).<\/a><\/p>\n<h1><a name=\"t33\"><\/a><a id=\"24_ESN_337\"><\/a>24. \u56de\u58f0\u72b6\u6001\u7f51\u7edc\uff08ESN\uff09<\/h1>\n<p><img decoding=\"async\" src=\"https:\/\/img-blog.csdnimg.cn\/img_convert\/366321241ca7bd637aca9c0c310b9996.png#pic_center\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<p>\u56de\u58f0\u72b6\u6001\u7f51\u7edc\uff08ESN\uff1aEcho state networks\uff09\u662f\u53e6\u4e00\u79cd\u4e0d\u540c\u7c7b\u578b\u7684\uff08\u5faa\u73af\uff09\u7f51\u7edc\u3002<\/p>\n<p>\u5b83\u7684\u4e0d\u540c\u4e4b\u5904\u5728\u4e8e\uff1a\u795e\u7ecf\u5143\u4e4b\u95f4\u7684\u8fde\u63a5\u662f\u968f\u673a\u7684\uff08\u6ca1\u6709\u6574\u9f50\u5212\u4e00\u7684\u795e\u7ecf\u7ec6\u80de\u5c42\uff09\uff0c\u5176\u8bad\u7ec3\u8fc7\u7a0b\u4e5f\u6709\u6240\u4e0d\u540c\u3002\u4e0d\u540c\u4e8e\u8f93\u5165\u6570\u636e\u540e\u53cd\u5411\u4f20\u64ad\u8bef\u5dee\uff0cESN\u5148\u8f93\u5165\u6570\u636e\u3001\u524d\u9988\u3001\u800c\u540e\u66f4\u65b0\u795e\u7ecf\u5143\u72b6\u6001\uff0c\u6700\u540e\u6765\u89c2\u5bdf\u7ed3\u679c\u3002\u5b83\u7684\u8f93\u5165\u5c42\u548c\u8f93\u51fa\u5c42\u5728\u8fd9\u91cc\u626e\u6f14\u7684\u89d2\u8272\u4e0d\u592a\u5e38\u89c4\uff0c\u8f93\u5165\u5c42\u7528\u6765\u4e3b\u5bfc\u7f51\u7edc\uff0c\u8f93\u51fa\u5c42\u4f5c\u4e3a\u6fc0\u6d3b\u6a21\u5f0f\u7684\u89c2\u6d4b\u5668\u968f\u65f6\u95f4\u5c55\u5f00\u3002\u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\uff0c\u53ea\u6709\u89c2\u6d4b\u548c\u9690\u85cf\u5355\u5143\u4e4b\u95f4\u8fde\u63a5\u4f1a\u88ab\u6539\u53d8\u3002<\/p>\n<p><a href=\"https:\/\/pdfs.semanticscholar.org\/8922\/17bb82c11e6e2263178ed20ac23db6279c7a.pdf\" target=\"_blank\" rel=\"noopener\">Jaeger, Herbert, and Harald Haas. \u201cHarnessing nonlinearity: Predicting chaotic systems and saving energy in wireless communication.\u201d science 304.5667 (2004): 78-80.<\/a><\/p>\n<h1><a name=\"t34\"><\/a><a id=\"25_ELM_348\"><\/a>25. \u6781\u9650\u5b66\u4e60\u673a\uff08ELM\uff09<\/h1>\n<p><img decoding=\"async\" src=\"https:\/\/img-blog.csdnimg.cn\/img_convert\/1f33cc63ecbff95ba1e413b3de82adbb.png#pic_center\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<p>\u6781\u9650\u5b66\u4e60\u673a\uff08ELM\uff1aExtreme learning machines\uff09\u672c\u8d28\u4e0a\u662f\u62e5\u6709\u968f\u673a\u8fde\u63a5\u7684FFNN\u3002<\/p>\n<p>\u5b83\u4eec\u4e0eLSM\u3001ESN\u6781\u4e3a\u76f8\u4f3c\uff0c\u9664\u4e86\u5faa\u73af\u7279\u5f81\u548c\u8109\u51b2\u6027\u8d28\uff0c\u5b83\u4eec\u8fd8\u4e0d\u4f7f\u7528\u53cd\u5411\u4f20\u64ad\u3002\u76f8\u53cd\uff0c\u5b83\u4eec\u5148\u7ed9\u6743\u91cd\u8bbe\u5b9a\u968f\u673a\u503c\uff0c\u7136\u540e\u6839\u636e\u6700\u5c0f\u4e8c\u4e58\u6cd5\u62df\u5408\u6765\u4e00\u6b21\u6027\u8bad\u7ec3\u6743\u91cd\uff08\u5728\u6240\u6709\u51fd\u6570\u4e2d\u8bef\u5dee\u6700\u5c0f\uff09\u3002\u8fd9\u4f7fELM\u7684\u51fd\u6570\u62df\u5408\u80fd\u529b\u8f83\u5f31\uff0c\u4f46\u5176\u8fd0\u884c\u901f\u5ea6\u6bd4\u53cd\u5411\u4f20\u64ad\u5feb\u591a\u4e86\u3002<\/p>\n<p><a href=\"http:\/\/www.ntu.edu.sg\/home\/egbhuang\/pdf\/ieee-is-elm.pdf\" target=\"_blank\" rel=\"noopener\">Cambria, Erik, et al. \u201cExtreme learning machines [trends &#038; controversies].\u201d IEEE Intelligent Systems 28.6 (2013): 30-59.<\/a><\/p>\n<h1><a name=\"t35\"><\/a><a id=\"25_LSM_359\"><\/a>25. \u6db2\u6001\u673a\uff08LSM\uff09<\/h1>\n<p><img decoding=\"async\" src=\"https:\/\/img-blog.csdnimg.cn\/img_convert\/4572a6223cfea401025b8196e6ce59cd.png#pic_center\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<p>\u6db2\u6001\u673a\uff08LSM\uff1aLiquid state machines\uff09\u6362\u6c64\u4e0d\u6362\u836f\uff0c\u8ddfESN\u540c\u6837\u76f8\u8fd1\u3002<\/p>\n<p>\u533a\u522b\u5728\u4e8e\uff0cLSM\u662f\u4e00\u79cd\u8109\u51b2\u795e\u7ecf\u7f51\u7edc\uff08spiking neural networks\uff09\uff0c\u7528\u9608\u503c\u6fc0\u6d3b\u51fd\u6570\uff08threshold functions\uff09\u53d6\u4ee3\u4e86sigmoid\u6fc0\u6d3b\u51fd\u6570\uff0c\u6bcf\u4e2a\u795e\u7ecf\u5143\u540c\u65f6\u4e5f\u662f\u5177\u6709\u7d2f\u52a0\u6027\u8d28\u7684\u8bb0\u5fc6\u5355\u5143\u3002\u56e0\u6b64\uff0c\u5f53\u795e\u7ecf\u5143\u72b6\u6001\u66f4\u65b0\u65f6\uff0c\u5176\u503c\u4e0d\u662f\u76f8\u90bb\u795e\u7ecf\u5143\u7684\u7d2f\u52a0\u503c\uff0c\u800c\u662f\u5b83\u81ea\u8eab\u72b6\u6001\u503c\u7684\u7d2f\u52a0\u3002\u4e00\u65e6\u7d2f\u52a0\u5230\u9608\u503c\uff0c\u5b83\u5c31\u91ca\u653e\u80fd\u91cf\u81f3\u5176\u5b83\u795e\u7ecf\u5143\u3002\u8fd9\u5c31\u5f62\u6210\u4e86\u4e00\u79cd\u7c7b\u4f3c\u4e8e\u8109\u51b2\u7684\u6a21\u5f0f\uff1a\u795e\u7ecf\u5143\u4e0d\u4f1a\u8fdb\u884c\u4efb\u4f55\u64cd\u4f5c\uff0c\u76f4\u81f3\u5230\u8fbe\u9608\u503c\u7684\u90a3\u4e00\u523b\u3002<\/p>\n<p><a href=\"https:\/\/web.archive.org\/web\/20120222154641\/http:\/\/ramsesii.upf.es\/seminar\/Maass_et_al_2002.pdf\" target=\"_blank\" rel=\"noopener\">Maass, Wolfgang, Thomas Natschl\u00e4ger, and Henry Markram. \u201cReal-time computing without stable states: A new framework for neural computation based on perturbations.\u201d Neural computation 14.11 (2002): 2531-2560.<\/a><\/p>\n<h1><a name=\"t36\"><\/a><a id=\"26_SVM_370\"><\/a>26. \u652f\u6301\u5411\u91cf\u673a\uff08SVM\uff09<\/h1>\n<p><img decoding=\"async\" src=\"https:\/\/img-blog.csdnimg.cn\/img_convert\/54dd0ffdba08e1b268aacd6f7b2f62c2.png#pic_center\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<p>\u652f\u6301\u5411\u91cf\u673a\uff08SVM\uff1aSupport vector machines\uff09\u80fd\u4e3a\u5206\u7c7b\u95ee\u9898\u627e\u51fa\u6700\u4f18\u65b9\u6848\u3002<\/p>\n<p>\u4f20\u7edf\u610f\u4e49\u4e0a\uff0c\u5b83\u4eec\u53ea\u80fd\u5904\u7406\u7ebf\u6027\u53ef\u5206\u7684\u6570\u636e\uff1b\u6bd4\u5982\u627e\u51fa\u54ea\u5f20\u56fe\u7247\u662f\u52a0\u83f2\u732b\u3001\u54ea\u5f20\u662f\u53f2\u52aa\u6bd4\uff0c\u6b64\u5916\u5c31\u65e0\u6cd5\u505a\u5176\u5b83\u8f93\u51fa\u4e86\u3002<\/p>\n<p>\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\uff0cSVM\u53ef\u4ee5\u7406\u89e3\u4e3a\uff1a\u5148\u5728\u5e73\u9762\u56fe\u8868\u4e0a\u6807\u7ed8\u6240\u6709\u6570\u636e\uff08\u52a0\u83f2\u732b\u3001\u53f2\u52aa\u6bd4\uff09\uff0c\u7136\u540e\u627e\u51fa\u5230\u90a3\u6761\u80fd\u591f\u6700\u597d\u533a\u5206\u8fd9\u4e24\u7c7b\u6570\u636e\u70b9\u7684\u7ebf\u3002\u8fd9\u6761\u7ebf\u80fd\u628a\u6570\u636e\u5206\u4e3a\u4e24\u90e8\u5206\uff0c\u7ebf\u7684\u8fd9\u8fb9\u5168\u662f\u53f2\u52aa\u6bd4\uff0c\u7ebf\u7684\u90a3\u8fb9\u5168\u662f\u52a0\u83f2\u732b\u3002\u800c\u540e\u79fb\u52a8\u5e76\u4f18\u5316\u8be5\u76f4\u7ebf\uff0c\u4ee4\u4e24\u8fb9\u6570\u636e\u70b9\u5230\u76f4\u7ebf\u7684\u8ddd\u79bb\u6700\u5927\u5316\u3002\u5206\u7c7b\u65b0\u7684\u6570\u636e\uff0c\u5219\u5c06\u8be5\u6570\u636e\u70b9\u753b\u5728\u8fd9\u4e2a\u56fe\u8868\u4e0a\uff0c\u7136\u540e\u5bdf\u770b\u8fd9\u4e2a\u6570\u636e\u70b9\u5728\u5206\u9694\u7ebf\u7684\u54ea\u4e00\u8fb9\uff08\u53f2\u52aa\u6bd4\u4e00\u4fa7\uff0c\u8fd8\u662f\u52a0\u83f2\u732b\u4e00\u4fa7\uff09\u3002<\/p>\n<p>\u901a\u8fc7\u4f7f\u7528\u6838\u65b9\u6cd5\uff0cSVM\u4fbf\u53ef\u7528\u6765\u5206\u7c7bn\u7ef4\u7a7a\u95f4\u7684\u6570\u636e\u3002\u8fd9\u5c31\u5f15\u51fa\u4e86\u57283\u7ef4\u7a7a\u95f4\u4e2d\u6807\u7ed8\u6570\u636e\u70b9\uff0c\u4ece\u800c\u8ba9SVM\u53ef\u4ee5\u533a\u5206\u53f2\u52aa\u6bd4\u3001\u52a0\u83f2\u732b\u4e0e\u897f\u8499\uff0c\u751a\u81f3\u5728\u66f4\u9ad8\u7684\u7ef4\u5ea6\u5bf9\u66f4\u591a\u5361\u901a\u4eba\u7269\u8fdb\u884c\u5206\u7c7b\u3002SVM\u5e76\u4e0d\u603b\u88ab\u89c6\u4e3a\u795e\u7ecf\u7f51\u7edc\u3002<\/p>\n<p><a href=\"http:\/\/image.diku.dk\/imagecanon\/material\/cortes_vapnik95.pdf\" target=\"_blank\" rel=\"noopener\">Cortes, Corinna, and Vladimir Vapnik. \u201cSupport-vector networks.\u201d Machine learning 20.3 (1995): 273-297.<\/a><\/p>\n<h1><a name=\"t37\"><\/a><a id=\"27_Kohonen__385\"><\/a>27. Kohonen \u7f51\u7edc<\/h1>\n<p><img decoding=\"async\" src=\"https:\/\/img-blog.csdnimg.cn\/img_convert\/f469275180ca1bbe661db987b6f801cd.png#pic_center\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<p>\u6700\u540e\uff0c\u6211\u4eec\u6765\u4ecb\u7ecd\u4e00\u4e0bKohonen\u7f51\u7edc\uff08KN\uff0c\u4e5f\u79f0\u4e4b\u4e3a\u81ea\u7ec4\u7ec7\uff08\u7279\u5f81\uff09\u6620\u5c04\uff08SOM\/SOFM\uff1aself organising (feature) map\uff09\uff09\u3002<\/p>\n<p>KN\u5229\u7528\u7ade\u4e89\u5b66\u4e60\u6765\u5bf9\u6570\u636e\u8fdb\u884c\u5206\u7c7b\uff0c\u4e0d\u9700\u8981\u76d1\u7763\u3002\u5148\u7ed9\u795e\u7ecf\u7f51\u7edc\u4e00\u4e2a\u8f93\u5165\uff0c\u800c\u540e\u5b83\u4f1a\u8bc4\u4f30\u54ea\u4e2a\u795e\u7ecf\u5143\u6700\u5339\u914d\u8be5\u8f93\u5165\u3002\u7136\u540e\u8fd9\u4e2a\u795e\u7ecf\u5143\u4f1a\u7ee7\u7eed\u8c03\u6574\u4ee5\u66f4\u597d\u5730\u5339\u914d\u8f93\u5165\u6570\u636e\uff0c\u540c\u65f6\u5e26\u52a8\u76f8\u90bb\u7684\u795e\u7ecf\u5143\u3002\u76f8\u90bb\u795e\u7ecf\u5143\u79fb\u52a8\u7684\u8ddd\u79bb\uff0c\u53d6\u51b3\u4e8e\u5b83\u4eec\u4e0e\u6700\u4f73\u5339\u914d\u5355\u5143\u4e4b\u95f4\u7684\u8ddd\u79bb\u3002KN\u6709\u65f6\u4e5f\u4e0d\u88ab\u8ba4\u4e3a\u662f\u795e\u7ecf\u7f51\u7edc\u3002<\/p>\n<p>\u53c2\u8003\u6587\u732e\uff1a<br \/>\n<a href=\"http:\/\/cioslab.vcu.edu\/alg\/Visualize\/kohonen-82.pdf\" target=\"_blank\" rel=\"noopener\">Kohonen, Teuvo. \u201cSelf-organized formation of topologically correct feature maps.\u201d Biological cybernetics 43.1 (1982): 59-69.<\/a><\/p>\n<p>\u539f\u6587\u94fe\u63a5\uff1a<\/p>\n<ul>\n<li><a href=\"http:\/\/www.toutiao.com\/i6432188985530909186\/\" target=\"_blank\" rel=\"noopener\">\u4e2d\u6587\u94fe\u63a5<\/a><\/li>\n<li><a href=\"http:\/\/www.asimovinstitute.org\/neural-network-zoo\/\" target=\"_blank\" rel=\"noopener\">\u82f1\u6587\u94fe\u63a51<\/a><\/li>\n<li><a href=\"http:\/\/www.asimovinstitute.org\/neural-network-zoo-prequel-cells-layers\/\" target=\"_blank\" rel=\"noopener\">\u82f1\u6587\u94fe\u63a52<\/a><\/li>\n<\/ul>\n<\/div>\n<\/div>\n<\/article>\n","protected":false},"excerpt":{"rendered":"<p>\u539f\u6587\u5730\u5740 \u7f6e\u9876\u5218\u70ab320\u5df2\u4e8e\u00a02022-<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"colormag_page_container_layout":"default_layout","colormag_page_sidebar_layout":"default_layout","footnotes":""},"categories":[24],"tags":[121],"class_list":["post-6420","post","type-post","status-publish","format-standard","hentry","category-value_docs","tag-05-"],"_links":{"self":[{"href":"https:\/\/i007.cc\/wordpress\/wp-json\/wp\/v2\/posts\/6420","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/i007.cc\/wordpress\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/i007.cc\/wordpress\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/i007.cc\/wordpress\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/i007.cc\/wordpress\/wp-json\/wp\/v2\/comments?post=6420"}],"version-history":[{"count":0,"href":"https:\/\/i007.cc\/wordpress\/wp-json\/wp\/v2\/posts\/6420\/revisions"}],"wp:attachment":[{"href":"https:\/\/i007.cc\/wordpress\/wp-json\/wp\/v2\/media?parent=6420"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/i007.cc\/wordpress\/wp-json\/wp\/v2\/categories?post=6420"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/i007.cc\/wordpress\/wp-json\/wp\/v2\/tags?post=6420"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}