{"id":123044,"date":"2026-11-06T08:40:48","date_gmt":"2026-11-06T00:40:48","guid":{"rendered":"https:\/\/www.wsisp.com\/helps\/123044.html"},"modified":"2026-11-06T08:40:48","modified_gmt":"2026-11-06T00:40:48","slug":"%e8%ae%a1%e7%ae%97%e6%9c%ba%e8%a7%86%e8%a7%89%e7%9a%84%e5%8f%91%e5%b1%95%e5%8e%86%e7%a8%8b%ef%bc%9a%e4%bb%8elenet%e5%88%b0gpt-4v%e7%9a%84%e4%b8%89%e5%8d%81%e5%b9%b4","status":"publish","type":"post","link":"https:\/\/www.wsisp.com\/helps\/123044.html","title":{"rendered":"\u8ba1\u7b97\u673a\u89c6\u89c9\u7684\u53d1\u5c55\u5386\u7a0b\uff1a\u4eceLeNet\u5230GPT-4V\u7684\u4e09\u5341\u5e74"},"content":{"rendered":"

\/* MD \/ \u5bcc\u6587\u672c\u4e2d\u7684 .toc\uff08\u542b\u535a\u5ba2\u56ed\u642c\u5bb6\u7b49\u5d4c\u5957\u7ed3\u6784\uff09\uff1b.toc-box \u5728\u4fa7\u680f\uff0c\u4e0d\u53d7\u5f71\u54cd *\/
\n #content_views .toc,
\n \/* \u7f16\u8f91\u5668\u5e38\u5728\u76ee\u5f55\u524d\u540e\u63d2\u5165\u7a7a p\uff08:empty \u4ecd\u5360 20px\uff09\uff0c\u4e00\u5e76\u53bb\u6389\u907f\u514d\u9876\u7a7a\u9699 *\/
\n #content_views.markdown_views > p:empty:has(+ .toc),
\n #content_views.markdown_views > .toc + p:empty,
\n \/* \u5bcc\u6587\u672c\u65e7\u7248\u76ee\u5f55\u6807\u8bb0 *\/
\n #content_views.htmledit_views #main-toc,
\n #content_views.htmledit_views #hr-toc,
\n #content_views.htmledit_views p[id*=\\”-toc\\”] {
\n display: none !important;
\n }
\n \/* \u76ee\u5f55\u53bb\u6389\u540e\uff0c\u7d27\u8ddf\u7684\u9996\u4e2a\u6807\u9898\u4e0d\u518d\u591a\u51fa\u4e00\u5757\u4e0a\u8fb9\u8ddd *\/
\n #content_views.markdown_views > .toc + h1,
\n #content_views.markdown_views > .toc + h2,
\n #content_views.markdown_views > .toc + h3,
\n #content_views.markdown_views > .toc + h4,
\n #content_views.markdown_views > .toc + p:empty + h1,
\n #content_views.markdown_views > .toc + p:empty + h2,
\n #content_views.markdown_views > .toc + p:empty + h3,
\n #content_views.markdown_views > .toc + p:empty + h4 {
\n margin-top: 0 !important;
\n }<\/p>\n

\u4e00\u3001\u5f15\u8a00：\u4e00\u573a\u8de8\u8d8a\u4e09\u5341\u5e74\u7684\u89c6\u89c9\u9769\u547d<\/h3>\n

\u8ba1\u7b97\u673a\u89c6\u89c9\u662f\u4eba\u5de5\u667a\u80fd\u9886\u57df\u6700\u5177\u6d3b\u529b\u4e0e\u5b9e\u7528\u4ef7\u503c\u7684\u5206\u652f\u4e4b\u4e00\u3002\u5b83\u8bd5\u56fe\u56de\u7b54\u4e00\u4e2a\u770b\u4f3c\u7b80\u5355\u5374\u6781\u5176\u6df1\u523b\u7684\u95ee\u9898：\u673a\u5668\u80fd\u5426\u50cf\u4eba\u7c7b\u4e00\u6837"\u770b\u89c1"\u5e76\u7406\u89e3\u4e16\u754c？\u4ece1989\u5e74LeCun\u5728\u8d1d\u5c14\u5b9e\u9a8c\u5ba4\u5199\u51fa\u7b2c\u4e00\u4e2a\u5377\u79ef\u795e\u7ecf\u7f51\u7edc\u624b\u5199\u6570\u5b57\u8bc6\u522b\u7cfb\u7edf，\u52302024\u5e74Sora\u6a21\u578b\u80fd\u591f\u6839\u636e\u6587\u5b57\u63cf\u8ff0\u751f\u6210\u957f\u8fbe\u4e00\u5206\u949f\u7684\u9ad8\u8d28\u91cf\u89c6\u9891，\u8fd9\u573a\u8de8\u8d8a\u4e09\u5341\u4e94\u5e74\u7684\u6280\u672f\u5f81\u7a0b，\u89c1\u8bc1\u4e86\u6df1\u5ea6\u5b66\u4e60\u4ece\u6d93\u6d93\u7ec6\u6d41\u6c47\u805a\u6210\u5954\u817e\u6c5f\u6cb3\u7684\u5168\u8fc7\u7a0b\u3002<\/p>\n

\u56de\u987e\u8fd9\u6bb5\u5386\u53f2，\u6211\u4eec\u770b\u5230\u7684\u4e0d\u4ec5\u662f\u7b97\u6cd5\u7cbe\u5ea6\u7684\u8282\u8282\u6500\u5347，\u66f4\u662f\u4eba\u7c7b\u5bf9"\u89c6\u89c9\u667a\u80fd"\u672c\u8d28\u7406\u89e3\u7684\u4e0d\u65ad\u6df1\u5316\u3002\u65e9\u671f\u7684\u8ba1\u7b97\u673a\u89c6\u89c9\u4f9d\u8d56\u624b\u5de5\u8bbe\u8ba1\u7684\u7279\u5f81\u63d0\u53d6\u5668，\u5b9b\u5982\u4e00\u4e2a\u6234\u7740\u9563\u94d0\u7684\u821e\u8005，\u88ab\u9650\u5236\u5728\u4eba\u7c7b\u5de5\u7a0b\u5e08\u6709\u9650\u7684\u60f3\u8c61\u529b\u4e4b\u5185\u3002\u6df1\u5ea6\u5b66\u4e60\u7684\u51fa\u73b0\u6253\u7834\u4e86\u8fd9\u4e00\u684e\u688f，\u8ba9\u673a\u5668\u80fd\u591f\u4ece\u6d77\u91cf\u6570\u636e\u4e2d\u81ea\u52a8\u5b66\u4e60\u89c6\u89c9\u8868\u5f81，\u5176\u80fd\u529b\u8fb9\u754c\u88ab\u5f7b\u5e95\u6253\u5f00\u3002<\/p>\n

\u672c\u6587\u5c06\u4ee5\u65f6\u95f4\u7ebf\u4e3a\u7ecf，\u4ee5\u5173\u952e\u6a21\u578b\u4e3a\u7eac，\u7cfb\u7edf\u68b3\u7406\u4ece1989\u5e74\u81f32024\u5e74\u8ba1\u7b97\u673a\u89c6\u89c9\u9886\u57df\u6700\u5177\u91cc\u7a0b\u7891\u610f\u4e49\u7684\u7a81\u7834\u3002\u6211\u4eec\u5c06\u6df1\u5165\u6bcf\u4e2a\u6a21\u578b\u7684\u6280\u672f\u7ec6\u8282，\u7406\u89e3\u5b83\u4eec\u5728\u90a3\u4e2a\u65f6\u4ee3\u6240\u9762\u4e34\u7684\u6311\u6218，\u4ee5\u53ca\u5b83\u4eec\u5982\u4f55\u4e3a\u540e\u7ee7\u8005\u94fa\u5e73\u9053\u8def\u3002\u8fd9\u4e0d\u4ec5\u662f\u4e00\u6bb5\u6280\u672f\u53f2，\u66f4\u662f\u4e00\u90e8\u5173\u4e8e\u4eba\u7c7b\u667a\u6167\u5982\u4f55\u63a8\u52a8\u673a\u5668\u89c6\u89c9\u80fd\u529b\u4e0d\u65ad\u8d85\u8d8a\u6781\u9650\u7684\u53f2\u8bd7\u3002<\/p>\n

#mermaid-svg-mTqzlfXZAVaipEn8{font-family:\\”trebuchet ms\\”,verdana,arial,sans-serif;font-size:16px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-mTqzlfXZAVaipEn8 .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-mTqzlfXZAVaipEn8 .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-mTqzlfXZAVaipEn8 .error-icon{fill:#552222;}#mermaid-svg-mTqzlfXZAVaipEn8 .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-mTqzlfXZAVaipEn8 .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-mTqzlfXZAVaipEn8 .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-mTqzlfXZAVaipEn8 .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-mTqzlfXZAVaipEn8 .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-mTqzlfXZAVaipEn8 .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-mTqzlfXZAVaipEn8 .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-mTqzlfXZAVaipEn8 .marker{fill:#333333;stroke:#333333;}#mermaid-svg-mTqzlfXZAVaipEn8 .marker.cross{stroke:#333333;}#mermaid-svg-mTqzlfXZAVaipEn8 svg{font-family:\\”trebuchet ms\\”,verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-mTqzlfXZAVaipEn8 p{margin:0;}#mermaid-svg-mTqzlfXZAVaipEn8 .edge{stroke-width:3;}#mermaid-svg-mTqzlfXZAVaipEn8 .section–1 rect,#mermaid-svg-mTqzlfXZAVaipEn8 .section–1 path,#mermaid-svg-mTqzlfXZAVaipEn8 .section–1 circle,#mermaid-svg-mTqzlfXZAVaipEn8 .section–1 path{fill:hsl(240, 100%, 76.2745098039%);}#mermaid-svg-mTqzlfXZAVaipEn8 .section–1 text{fill:#ffffff;}#mermaid-svg-mTqzlfXZAVaipEn8 .node-icon–1{font-size:40px;color:#ffffff;}#mermaid-svg-mTqzlfXZAVaipEn8 .section-edge–1{stroke:hsl(240, 100%, 76.2745098039%);}#mermaid-svg-mTqzlfXZAVaipEn8 .edge-depth–1{stroke-width:17;}#mermaid-svg-mTqzlfXZAVaipEn8 .section–1 line{stroke:hsl(60, 100%, 86.2745098039%);stroke-width:3;}#mermaid-svg-mTqzlfXZAVaipEn8 .lineWrapper line{stroke:#ffffff;}#mermaid-svg-mTqzlfXZAVaipEn8 .disabled,#mermaid-svg-mTqzlfXZAVaipEn8 .disabled circle,#mermaid-svg-mTqzlfXZAVaipEn8 .disabled text{fill:lightgray;}#mermaid-svg-mTqzlfXZAVaipEn8 .disabled text{fill:#efefef;}#mermaid-svg-mTqzlfXZAVaipEn8 .section-0 rect,#mermaid-svg-mTqzlfXZAVaipEn8 .section-0 path,#mermaid-svg-mTqzlfXZAVaipEn8 .section-0 circle,#mermaid-svg-mTqzlfXZAVaipEn8 .section-0 path{fill:hsl(60, 100%, 73.5294117647%);}#mermaid-svg-mTqzlfXZAVaipEn8 .section-0 text{fill:black;}#mermaid-svg-mTqzlfXZAVaipEn8 .node-icon-0{font-size:40px;color:black;}#mermaid-svg-mTqzlfXZAVaipEn8 .section-edge-0{stroke:hsl(60, 100%, 73.5294117647%);}#mermaid-svg-mTqzlfXZAVaipEn8 .edge-depth-0{stroke-width:14;}#mermaid-svg-mTqzlfXZAVaipEn8 .section-0 line{stroke:hsl(240, 100%, 83.5294117647%);stroke-width:3;}#mermaid-svg-mTqzlfXZAVaipEn8 .lineWrapper line{stroke:black;}#mermaid-svg-mTqzlfXZAVaipEn8 .disabled,#mermaid-svg-mTqzlfXZAVaipEn8 .disabled circle,#mermaid-svg-mTqzlfXZAVaipEn8 .disabled text{fill:lightgray;}#mermaid-svg-mTqzlfXZAVaipEn8 .disabled text{fill:#efefef;}#mermaid-svg-mTqzlfXZAVaipEn8 .section-1 rect,#mermaid-svg-mTqzlfXZAVaipEn8 .section-1 path,#mermaid-svg-mTqzlfXZAVaipEn8 .section-1 circle,#mermaid-svg-mTqzlfXZAVaipEn8 .section-1 path{fill:hsl(80, 100%, 76.2745098039%);}#mermaid-svg-mTqzlfXZAVaipEn8 .section-1 text{fill:black;}#mermaid-svg-mTqzlfXZAVaipEn8 .node-icon-1{font-size:40px;color:black;}#mermaid-svg-mTqzlfXZAVaipEn8 .section-edge-1{stroke:hsl(80, 100%, 76.2745098039%);}#mermaid-svg-mTqzlfXZAVaipEn8 .edge-depth-1{stroke-width:11;}#mermaid-svg-mTqzlfXZAVaipEn8 .section-1 line{stroke:hsl(260, 100%, 86.2745098039%);stroke-width:3;}#mermaid-svg-mTqzlfXZAVaipEn8 .lineWrapper line{stroke:black;}#mermaid-svg-mTqzlfXZAVaipEn8 .disabled,#mermaid-svg-mTqzlfXZAVaipEn8 .disabled circle,#mermaid-svg-mTqzlfXZAVaipEn8 .disabled text{fill:lightgray;}#mermaid-svg-mTqzlfXZAVaipEn8 .disabled text{fill:#efefef;}#mermaid-svg-mTqzlfXZAVaipEn8 .section-2 rect,#mermaid-svg-mTqzlfXZAVaipEn8 .section-2 path,#mermaid-svg-mTqzlfXZAVaipEn8 .section-2 circle,#mermaid-svg-mTqzlfXZAVaipEn8 .section-2 path{fill:hsl(270, 100%, 76.2745098039%);}#mermaid-svg-mTqzlfXZAVaipEn8 .section-2 text{fill:#ffffff;}#mermaid-svg-mTqzlfXZAVaipEn8 .node-icon-2{font-size:40px;color:#ffffff;}#mermaid-svg-mTqzlfXZAVaipEn8 .section-edge-2{stroke:hsl(270, 100%, 76.2745098039%);}#mermaid-svg-mTqzlfXZAVaipEn8 .edge-depth-2{stroke-width:8;}#mermaid-svg-mTqzlfXZAVaipEn8 .section-2 line{stroke:hsl(90, 100%, 86.2745098039%);stroke-width:3;}#mermaid-svg-mTqzlfXZAVaipEn8 .lineWrapper line{stroke:#ffffff;}#mermaid-svg-mTqzlfXZAVaipEn8 .disabled,#mermaid-svg-mTqzlfXZAVaipEn8 .disabled circle,#mermaid-svg-mTqzlfXZAVaipEn8 .disabled text{fill:lightgray;}#mermaid-svg-mTqzlfXZAVaipEn8 .disabled text{fill:#efefef;}#mermaid-svg-mTqzlfXZAVaipEn8 .section-3 rect,#mermaid-svg-mTqzlfXZAVaipEn8 .section-3 path,#mermaid-svg-mTqzlfXZAVaipEn8 .section-3 circle,#mermaid-svg-mTqzlfXZAVaipEn8 .section-3 path{fill:hsl(300, 100%, 76.2745098039%);}#mermaid-svg-mTqzlfXZAVaipEn8 .section-3 text{fill:black;}#mermaid-svg-mTqzlfXZAVaipEn8 .node-icon-3{font-size:40px;color:black;}#mermaid-svg-mTqzlfXZAVaipEn8 .section-edge-3{stroke:hsl(300, 100%, 76.2745098039%);}#mermaid-svg-mTqzlfXZAVaipEn8 .edge-depth-3{stroke-width:5;}#mermaid-svg-mTqzlfXZAVaipEn8 .section-3 line{stroke:hsl(120, 100%, 86.2745098039%);stroke-width:3;}#mermaid-svg-mTqzlfXZAVaipEn8 .lineWrapper line{stroke:black;}#mermaid-svg-mTqzlfXZAVaipEn8 .disabled,#mermaid-svg-mTqzlfXZAVaipEn8 .disabled circle,#mermaid-svg-mTqzlfXZAVaipEn8 .disabled text{fill:lightgray;}#mermaid-svg-mTqzlfXZAVaipEn8 .disabled text{fill:#efefef;}#mermaid-svg-mTqzlfXZAVaipEn8 .section-4 rect,#mermaid-svg-mTqzlfXZAVaipEn8 .section-4 path,#mermaid-svg-mTqzlfXZAVaipEn8 .section-4 circle,#mermaid-svg-mTqzlfXZAVaipEn8 .section-4 path{fill:hsl(330, 100%, 76.2745098039%);}#mermaid-svg-mTqzlfXZAVaipEn8 .section-4 text{fill:black;}#mermaid-svg-mTqzlfXZAVaipEn8 .node-icon-4{font-size:40px;color:black;}#mermaid-svg-mTqzlfXZAVaipEn8 .section-edge-4{stroke:hsl(330, 100%, 76.2745098039%);}#mermaid-svg-mTqzlfXZAVaipEn8 .edge-depth-4{stroke-width:2;}#mermaid-svg-mTqzlfXZAVaipEn8 .section-4 line{stroke:hsl(150, 100%, 86.2745098039%);stroke-width:3;}#mermaid-svg-mTqzlfXZAVaipEn8 .lineWrapper line{stroke:black;}#mermaid-svg-mTqzlfXZAVaipEn8 .disabled,#mermaid-svg-mTqzlfXZAVaipEn8 .disabled circle,#mermaid-svg-mTqzlfXZAVaipEn8 .disabled text{fill:lightgray;}#mermaid-svg-mTqzlfXZAVaipEn8 .disabled text{fill:#efefef;}#mermaid-svg-mTqzlfXZAVaipEn8 .section-5 rect,#mermaid-svg-mTqzlfXZAVaipEn8 .section-5 path,#mermaid-svg-mTqzlfXZAVaipEn8 .section-5 circle,#mermaid-svg-mTqzlfXZAVaipEn8 .section-5 path{fill:hsl(0, 100%, 76.2745098039%);}#mermaid-svg-mTqzlfXZAVaipEn8 .section-5 text{fill:black;}#mermaid-svg-mTqzlfXZAVaipEn8 .node-icon-5{font-size:40px;color:black;}#mermaid-svg-mTqzlfXZAVaipEn8 .section-edge-5{stroke:hsl(0, 100%, 76.2745098039%);}#mermaid-svg-mTqzlfXZAVaipEn8 .edge-depth-5{stroke-width:-1;}#mermaid-svg-mTqzlfXZAVaipEn8 .section-5 line{stroke:hsl(180, 100%, 86.2745098039%);stroke-width:3;}#mermaid-svg-mTqzlfXZAVaipEn8 .lineWrapper line{stroke:black;}#mermaid-svg-mTqzlfXZAVaipEn8 .disabled,#mermaid-svg-mTqzlfXZAVaipEn8 .disabled circle,#mermaid-svg-mTqzlfXZAVaipEn8 .disabled text{fill:lightgray;}#mermaid-svg-mTqzlfXZAVaipEn8 .disabled text{fill:#efefef;}#mermaid-svg-mTqzlfXZAVaipEn8 .section-6 rect,#mermaid-svg-mTqzlfXZAVaipEn8 .section-6 path,#mermaid-svg-mTqzlfXZAVaipEn8 .section-6 circle,#mermaid-svg-mTqzlfXZAVaipEn8 .section-6 path{fill:hsl(30, 100%, 76.2745098039%);}#mermaid-svg-mTqzlfXZAVaipEn8 .section-6 text{fill:black;}#mermaid-svg-mTqzlfXZAVaipEn8 .node-icon-6{font-size:40px;color:black;}#mermaid-svg-mTqzlfXZAVaipEn8 .section-edge-6{stroke:hsl(30, 100%, 76.2745098039%);}#mermaid-svg-mTqzlfXZAVaipEn8 .edge-depth-6{stroke-width:-4;}#mermaid-svg-mTqzlfXZAVaipEn8 .section-6 line{stroke:hsl(210, 100%, 86.2745098039%);stroke-width:3;}#mermaid-svg-mTqzlfXZAVaipEn8 .lineWrapper line{stroke:black;}#mermaid-svg-mTqzlfXZAVaipEn8 .disabled,#mermaid-svg-mTqzlfXZAVaipEn8 .disabled circle,#mermaid-svg-mTqzlfXZAVaipEn8 .disabled text{fill:lightgray;}#mermaid-svg-mTqzlfXZAVaipEn8 .disabled text{fill:#efefef;}#mermaid-svg-mTqzlfXZAVaipEn8 .section-7 rect,#mermaid-svg-mTqzlfXZAVaipEn8 .section-7 path,#mermaid-svg-mTqzlfXZAVaipEn8 .section-7 circle,#mermaid-svg-mTqzlfXZAVaipEn8 .section-7 path{fill:hsl(90, 100%, 76.2745098039%);}#mermaid-svg-mTqzlfXZAVaipEn8 .section-7 text{fill:black;}#mermaid-svg-mTqzlfXZAVaipEn8 .node-icon-7{font-size:40px;color:black;}#mermaid-svg-mTqzlfXZAVaipEn8 .section-edge-7{stroke:hsl(90, 100%, 76.2745098039%);}#mermaid-svg-mTqzlfXZAVaipEn8 .edge-depth-7{stroke-width:-7;}#mermaid-svg-mTqzlfXZAVaipEn8 .section-7 line{stroke:hsl(270, 100%, 86.2745098039%);stroke-width:3;}#mermaid-svg-mTqzlfXZAVaipEn8 .lineWrapper line{stroke:black;}#mermaid-svg-mTqzlfXZAVaipEn8 .disabled,#mermaid-svg-mTqzlfXZAVaipEn8 .disabled circle,#mermaid-svg-mTqzlfXZAVaipEn8 .disabled text{fill:lightgray;}#mermaid-svg-mTqzlfXZAVaipEn8 .disabled text{fill:#efefef;}#mermaid-svg-mTqzlfXZAVaipEn8 .section-8 rect,#mermaid-svg-mTqzlfXZAVaipEn8 .section-8 path,#mermaid-svg-mTqzlfXZAVaipEn8 .section-8 circle,#mermaid-svg-mTqzlfXZAVaipEn8 .section-8 path{fill:hsl(150, 100%, 76.2745098039%);}#mermaid-svg-mTqzlfXZAVaipEn8 .section-8 text{fill:black;}#mermaid-svg-mTqzlfXZAVaipEn8 .node-icon-8{font-size:40px;color:black;}#mermaid-svg-mTqzlfXZAVaipEn8 .section-edge-8{stroke:hsl(150, 100%, 76.2745098039%);}#mermaid-svg-mTqzlfXZAVaipEn8 .edge-depth-8{stroke-width:-10;}#mermaid-svg-mTqzlfXZAVaipEn8 .section-8 line{stroke:hsl(330, 100%, 86.2745098039%);stroke-width:3;}#mermaid-svg-mTqzlfXZAVaipEn8 .lineWrapper line{stroke:black;}#mermaid-svg-mTqzlfXZAVaipEn8 .disabled,#mermaid-svg-mTqzlfXZAVaipEn8 .disabled circle,#mermaid-svg-mTqzlfXZAVaipEn8 .disabled text{fill:lightgray;}#mermaid-svg-mTqzlfXZAVaipEn8 .disabled text{fill:#efefef;}#mermaid-svg-mTqzlfXZAVaipEn8 .section-9 rect,#mermaid-svg-mTqzlfXZAVaipEn8 .section-9 path,#mermaid-svg-mTqzlfXZAVaipEn8 .section-9 circle,#mermaid-svg-mTqzlfXZAVaipEn8 .section-9 path{fill:hsl(180, 100%, 76.2745098039%);}#mermaid-svg-mTqzlfXZAVaipEn8 .section-9 text{fill:black;}#mermaid-svg-mTqzlfXZAVaipEn8 .node-icon-9{font-size:40px;color:black;}#mermaid-svg-mTqzlfXZAVaipEn8 .section-edge-9{stroke:hsl(180, 100%, 76.2745098039%);}#mermaid-svg-mTqzlfXZAVaipEn8 .edge-depth-9{stroke-width:-13;}#mermaid-svg-mTqzlfXZAVaipEn8 .section-9 line{stroke:hsl(0, 100%, 86.2745098039%);stroke-width:3;}#mermaid-svg-mTqzlfXZAVaipEn8 .lineWrapper line{stroke:black;}#mermaid-svg-mTqzlfXZAVaipEn8 .disabled,#mermaid-svg-mTqzlfXZAVaipEn8 .disabled circle,#mermaid-svg-mTqzlfXZAVaipEn8 .disabled text{fill:lightgray;}#mermaid-svg-mTqzlfXZAVaipEn8 .disabled text{fill:#efefef;}#mermaid-svg-mTqzlfXZAVaipEn8 .section-10 rect,#mermaid-svg-mTqzlfXZAVaipEn8 .section-10 path,#mermaid-svg-mTqzlfXZAVaipEn8 .section-10 circle,#mermaid-svg-mTqzlfXZAVaipEn8 .section-10 path{fill:hsl(210, 100%, 76.2745098039%);}#mermaid-svg-mTqzlfXZAVaipEn8 .section-10 text{fill:black;}#mermaid-svg-mTqzlfXZAVaipEn8 .node-icon-10{font-size:40px;color:black;}#mermaid-svg-mTqzlfXZAVaipEn8 .section-edge-10{stroke:hsl(210, 100%, 76.2745098039%);}#mermaid-svg-mTqzlfXZAVaipEn8 .edge-depth-10{stroke-width:-16;}#mermaid-svg-mTqzlfXZAVaipEn8 .section-10 line{stroke:hsl(30, 100%, 86.2745098039%);stroke-width:3;}#mermaid-svg-mTqzlfXZAVaipEn8 .lineWrapper line{stroke:black;}#mermaid-svg-mTqzlfXZAVaipEn8 .disabled,#mermaid-svg-mTqzlfXZAVaipEn8 .disabled circle,#mermaid-svg-mTqzlfXZAVaipEn8 .disabled text{fill:lightgray;}#mermaid-svg-mTqzlfXZAVaipEn8 .disabled text{fill:#efefef;}#mermaid-svg-mTqzlfXZAVaipEn8 .section-root rect,#mermaid-svg-mTqzlfXZAVaipEn8 .section-root path,#mermaid-svg-mTqzlfXZAVaipEn8 .section-root circle{fill:hsl(240, 100%, 46.2745098039%);}#mermaid-svg-mTqzlfXZAVaipEn8 .section-root text{fill:#ffffff;}#mermaid-svg-mTqzlfXZAVaipEn8 .icon-container{height:100%;display:flex;justify-content:center;align-items:center;}#mermaid-svg-mTqzlfXZAVaipEn8 .edge{fill:none;}#mermaid-svg-mTqzlfXZAVaipEn8 .eventWrapper{filter:brightness(120%);}#mermaid-svg-mTqzlfXZAVaipEn8 :root{–mermaid-font-family:\\”trebuchet ms\\”,verdana,arial,sans-serif;}1989LeNet\u8bde\u751f\u5377\u79ef\u795e\u7ecf\u7f51\u7edc\u5960\u57fa1998LeNet-5\u624b\u5199\u4f53\u8bc6\u522b\u5546\u75282012AlexNet\u6df1\u5ea6\u5b66\u4e60\u9769\u547d\u5e8f\u5e552014VGG\/GoogLeNet\u66f4\u6df1\u66f4\u5bbd\u7684\u7f51\u7edc2015ResNet\u6b8b\u5dee\u8fde\u63a5\u7a81\u7834\u6df1\u5ea6\u74f6\u98882017Transformer\u6ce8\u610f\u529b\u673a\u5236\u767b\u573a2020ViTTransformer\u5165\u4e3bCV2021CLIP\u89c6\u89c9-\u8bed\u8a00\u9884\u8bad\u7ec32022Stable Diffusion\u751f\u6210\u5f0fAI\u7206\u53d12023GPT-4V\/SAM\u591a\u6a21\u6001\u7406\u89e3\u4e0e\u5206\u52722024Sora\u89c6\u9891\u751f\u6210\u65b0\u7eaa\u5143\u8ba1\u7b97\u673a\u89c6\u89c9\u4e09\u5341\u5e74：\u5173\u952e\u91cc\u7a0b\u7891<\/p>\n

\u4e8c\u3001\u9ece\u660e\u65f6\u671f：\u5377\u79ef\u795e\u7ecf\u7f51\u7edc\u7684\u8bde\u751f（1989-2011）<\/h3>\n

2.1 LeNet(1989)：\u4e00\u5207\u5f00\u59cb\u7684\u5730\u65b9<\/h4>\n

1989\u5e74，Yann LeCun\u5728\u8d1d\u5c14\u5b9e\u9a8c\u5ba4\u53d1\u8868\u4e86\u4e00\u7bc7\u8bba\u6587，\u63cf\u8ff0\u4e86\u4e00\u4e2a\u5177\u6709\u5377\u79ef\u5c42\u548c\u53cd\u5411\u4f20\u64ad\u8bad\u7ec3\u7684\u795e\u7ecf\u7f51\u7edc，\u7528\u4e8e\u8bc6\u522b\u624b\u5199\u90ae\u653f\u7f16\u7801\u6570\u5b57\u3002\u8fd9\u4e2a\u6a21\u578b\u540e\u6765\u88ab\u79f0\u4e3aLeNet，\u5b83\u5960\u5b9a\u4e86\u73b0\u4ee3\u5377\u79ef\u795e\u7ecf\u7f51\u7edc\u7684\u57fa\u672c\u67b6\u6784\u8303\u5f0f\u3002<\/p>\n

LeNet\u7684\u6838\u5fc3\u601d\u60f3\u6765\u6e90\u4e8e\u5bf9\u751f\u7269\u89c6\u89c9\u7cfb\u7edf\u7684\u6a21\u4eff\u3002Hubel\u548cWiesel\u572820\u4e16\u7eaa60\u5e74\u4ee3\u5bf9\u732b\u7684\u89c6\u89c9\u76ae\u5c42\u7814\u7a76\u63ed\u793a\u4e86\u4e24\u4e2a\u5173\u952e\u53d1\u73b0：\u89c6\u89c9\u76ae\u5c42\u4e2d\u7684\u795e\u7ecf\u5143\u5177\u6709\u5c40\u90e8\u611f\u53d7\u91ce，\u53ea\u5bf9\u89c6\u91ce\u4e2d\u7279\u5b9a\u533a\u57df\u7684\u523a\u6fc0\u505a\u51fa\u54cd\u5e94；\u4e0d\u540c\u795e\u7ecf\u5143\u5bf9\u4e0d\u540c\u7c7b\u578b\u7684\u89c6\u89c9\u6a21\u5f0f（\u8fb9\u7f18\u3001\u65b9\u5411\u3001\u8fd0\u52a8）\u5177\u6709\u9009\u62e9\u6027\u3002\u8fd9\u4e9b\u53d1\u73b0\u76f4\u63a5\u542f\u53d1\u4e86\u5377\u79ef\u795e\u7ecf\u7f51\u7edc\u7684\u8bbe\u8ba1\u2014\u2014\u5377\u79ef\u6838\u6a21\u62df\u5c40\u90e8\u611f\u53d7\u91ce，\u591a\u4e2a\u5377\u79ef\u6838\u6a21\u62df\u4e0d\u540c\u7c7b\u578b\u7684\u7279\u5f81\u68c0\u6d4b\u5668\u3002<\/p>\n

LeNet\u7684\u67b6\u6784\u5305\u542b\u4ee5\u4e0b\u5173\u952e\u7ec4\u4ef6：\u5377\u79ef\u5c42\u8d1f\u8d23\u63d0\u53d6\u5c40\u90e8\u7279\u5f81，\u901a\u8fc7\u5171\u4eab\u6743\u91cd\u5927\u5e45\u51cf\u5c11\u53c2\u6570\u6570\u91cf；\u4e0b\u91c7\u6837\u5c42（\u6c60\u5316\u5c42\u7684\u524d\u8eab）\u964d\u4f4e\u7279\u5f81\u56fe\u7684\u7a7a\u95f4\u5206\u8fa8\u7387，\u63d0\u4f9b\u4e00\u5b9a\u7a0b\u5ea6\u7684\u5e73\u79fb\u4e0d\u53d8\u6027；\u5168\u8fde\u63a5\u5c42\u5c06\u63d0\u53d6\u7684\u7279\u5f81\u6620\u5c04\u5230\u7c7b\u522b\u7a7a\u95f4\u3002\u8fd9\u79cd"\u5377\u79ef-\u4e0b\u91c7\u6837-\u5168\u8fde\u63a5"\u7684\u4e09\u6bb5\u5f0f\u7ed3\u6784，\u81f3\u4eca\u4ecd\u7136\u662f\u5927\u591a\u6570CNN\u7684\u57fa\u672c\u9aa8\u67b6\u3002<\/p>\n

LeNet\u5728MNIST\u624b\u5199\u6570\u5b57\u6570\u636e\u96c6\u4e0a\u53d6\u5f97\u4e86\u5f53\u65f6\u4ee4\u4eba\u77a9\u76ee\u7684\u6210\u7ee9，\u5176\u8bc6\u522b\u7684\u90ae\u653f\u7f16\u7801\u88ab\u7f8e\u56fd\u90ae\u653f\u670d\u52a1\u5b9e\u9645\u90e8\u7f72\u4f7f\u7528，\u6210\u4e3a\u6df1\u5ea6\u5b66\u4e60\u6700\u65e9\u7684\u6210\u529f\u5546\u4e1a\u5e94\u7528\u4e4b\u4e00\u3002\u7136\u800c，\u53d7\u9650\u4e8e\u5f53\u65f6\u8ba1\u7b97\u8d44\u6e90（\u8bad\u7ec3\u4e00\u4e2aLeNet\u53ef\u80fd\u9700\u8981\u6570\u5929\u65f6\u95f4）\u548c\u6570\u636e\u89c4\u6a21，CNN\u5728\u63a5\u4e0b\u6765\u8fd1\u4e8c\u5341\u5e74\u95f4\u5e76\u672a\u6210\u4e3a\u4e3b\u6d41\u3002<\/p>\n

2.2 LeNet-5(1998)：\u7ecf\u5178\u67b6\u6784\u7684\u5b9a\u578b<\/h4>\n

1998\u5e74，LeCun\u56e2\u961f\u53d1\u5e03\u4e86LeNet-5，\u8fd9\u662fLeNet\u7cfb\u5217\u4e2d\u6700\u8457\u540d\u7684\u7248\u672c，\u4e5f\u662f\u73b0\u4ee3CNN\u6559\u79d1\u4e66\u7684\u7ecf\u5178\u6848\u4f8b\u3002LeNet-5\u5305\u542b7\u5c42（\u4e0d\u5305\u62ec\u8f93\u5165\u5c42）：3\u4e2a\u5377\u79ef\u5c42\u30012\u4e2a\u4e0b\u91c7\u6837\u5c42\u30011\u4e2a\u5168\u8fde\u63a5\u5c42\u548c1\u4e2a\u8f93\u51fa\u5c42\u3002\u5176\u6700\u91cd\u8981\u7684\u6539\u8fdb\u5728\u4e8e\u5f15\u5165\u4e86\u66f4\u591a\u7684\u5377\u79ef\u5c42，\u5f62\u6210\u4e86"\u5377\u79ef-\u6c60\u5316-\u5377\u79ef-\u6c60\u5316-\u5377\u79ef-\u5168\u8fde\u63a5"\u7684\u5c42\u6b21\u5316\u7ed3\u6784\u3002<\/p>\n

LeNet-5\u5c55\u793a\u4e86\u4e00\u4e2a\u91cd\u8981\u7684\u8bbe\u8ba1\u54f2\u5b66：\u968f\u7740\u7f51\u7edc\u5c42\u6570\u7684\u589e\u52a0，\u7279\u5f81\u56fe\u7684\u7a7a\u95f4\u7ef4\u5ea6\u9010\u6e10\u51cf\u5c0f（\u4ece32×32\u52301×1），\u800c\u901a\u9053\u6570\u9010\u6e10\u589e\u52a0（\u4ece1\u5230120）\u3002\u8fd9\u79cd"\u7a7a\u95f4\u6536\u7f29\u3001\u901a\u9053\u6269\u5f20"\u7684\u6a21\u5f0f，\u4f7f\u5f97\u7f51\u7edc\u80fd\u591f\u4ece\u4f4e\u7ea7\u7684\u8fb9\u7f18\u7279\u5f81\u9010\u6b65\u62bd\u8c61\u5230\u9ad8\u7ea7\u7684\u8bed\u4e49\u7279\u5f81\u3002\u8fd9\u4e00\u601d\u60f3\u5728\u540e\u7eed\u51e0\u4e4e\u6240\u6709CNN\u67b6\u6784\u4e2d\u5f97\u5230\u4e86\u7ee7\u627f\u548c\u53d1\u5c55\u3002<\/p>\n

\u5728MNIST\u6570\u636e\u96c6\u4e0a，LeNet-5\u7684\u9519\u8bef\u7387\u964d\u5230\u4e86\u7ea60.8%，\u8fd9\u4e2a\u6210\u7ee9\u5728\u5f53\u65f6\u8fdc\u8d85\u5176\u4ed6\u65b9\u6cd5\u3002\u7136\u800c，\u5f53\u9762\u5bf9\u66f4\u5927\u89c4\u6a21\u3001\u66f4\u590d\u6742\u7684\u81ea\u7136\u56fe\u50cf\u6570\u636e\u96c6\u65f6，LeNet\u7684\u6027\u80fd\u4ecd\u7136\u4e0d\u8db3\u3002\u8fd9\u4e3b\u8981\u53d7\u9650\u4e8e\u4e24\u4e2a\u56e0\u7d20：\u4e00\u662f\u7f51\u7edc\u7684\u6df1\u5ea6\u548c\u53c2\u6570\u91cf\u4ecd\u7136\u6709\u9650，\u65e0\u6cd5\u5145\u5206\u5efa\u6a21\u590d\u6742\u7684\u89c6\u89c9\u6a21\u5f0f；\u4e8c\u662f\u7f3a\u4e4f\u6709\u6548\u7684\u6b63\u5219\u5316\u624b\u6bb5，\u66f4\u6df1\u66f4\u5927\u7684\u7f51\u7edc\u5bb9\u6613\u8fc7\u62df\u5408\u3002<\/p>\n

2.3 \u5bd2\u51ac\u4e0e\u575a\u6301：2000-2011\u5e74\u7684\u63a2\u7d22<\/h4>\n

\u57282000\u5e74\u52302011\u5e74\u8fd9\u6bb5\u65f6\u671f，\u8ba1\u7b97\u673a\u89c6\u89c9\u7684\u4e3b\u6d41\u8303\u5f0f\u662f"\u624b\u5de5\u7279\u5f81+\u6d45\u5c42\u5206\u7c7b\u5668"\u3002SIFT（\u5c3a\u5ea6\u4e0d\u53d8\u7279\u5f81\u53d8\u6362，1999\u5e74）\u3001HOG（\u65b9\u5411\u68af\u5ea6\u76f4\u65b9\u56fe，2005\u5e74）\u7b49\u624b\u5de5\u8bbe\u8ba1\u7684\u7279\u5f81\u63cf\u8ff0\u5b50\u914d\u5408SVM（\u652f\u6301\u5411\u91cf\u673a）\u6210\u4e3a\u56fe\u50cf\u5206\u7c7b\u548c\u7269\u4f53\u68c0\u6d4b\u7684\u6807\u51c6\u89e3\u51b3\u65b9\u6848\u3002\u8fd9\u4e9b\u65b9\u6cd5\u867d\u7136\u5728\u67d0\u4e9b\u4efb\u52a1\u4e0a\u53d6\u5f97\u4e86\u4e0d\u9519\u7684\u6548\u679c，\u4f46\u5176\u6838\u5fc3\u74f6\u9888\u5728\u4e8e\u7279\u5f81\u8bbe\u8ba1\u5b8c\u5168\u4f9d\u8d56\u4eba\u7c7b\u5de5\u7a0b\u5e08\u7684\u9886\u57df\u77e5\u8bc6\u548c\u7ecf\u9a8c，\u96be\u4ee5\u63a8\u5e7f\u5230\u590d\u6742\u591a\u53d8\u7684\u89c6\u89c9\u573a\u666f\u3002<\/p>\n

\u5728\u8fd9\u6bb5"AI\u5bd2\u51ac"\u4e2d，\u4ecd\u6709\u5c11\u6570\u7814\u7a76\u8005\u575a\u6301\u795e\u7ecf\u7f51\u7edc\u65b9\u5411\u3002Geoffrey Hinton\u57282006\u5e74\u63d0\u51fa\u7684\u6df1\u5ea6\u4fe1\u5ff5\u7f51\u7edc（DBN）\u548c\u9010\u5c42\u9884\u8bad\u7ec3\u7b56\u7565，\u4e3a\u8bad\u7ec3\u6df1\u5c42\u7f51\u7edc\u63d0\u4f9b\u4e86\u65b0\u7684\u601d\u8def\u3002\u867d\u7136DBN\u672c\u8eab\u5e76\u975eCNN，\u4f46\u5b83\u8bc1\u660e\u4e86\u6df1\u5ea6\u7f51\u7edc\u662f\u53ef\u4ee5\u88ab\u6709\u6548\u8bad\u7ec3\u7684，\u4e3a\u540e\u7eedCNN\u7684\u590d\u5174\u505a\u4e86\u91cd\u8981\u7684\u7406\u8bba\u94fa\u57ab\u3002<\/p>\n

\u5434\u6069\u8fbe（Andrew Ng）\u57282011\u5e74\u4f7f\u7528\u5927\u89c4\u6a21\u65e0\u76d1\u7763\u5b66\u4e60\u8bad\u7ec3\u4e86\u4e00\u4e2a\u6df1\u5ea6\u81ea\u7f16\u7801\u5668，\u80fd\u591f\u4eceYouTube\u89c6\u9891\u5e27\u4e2d\u81ea\u52a8\u5b66\u4e60\u5230\u732b\u8138\u68c0\u6d4b\u5668\u3002\u8fd9\u9879\u5de5\u4f5c\u867d\u7136\u7cbe\u5ea6\u6709\u9650，\u4f46\u5c55\u793a\u4e86\u65e0\u76d1\u7763\u6df1\u5ea6\u5b66\u4e60\u5728\u5927\u89c4\u6a21\u6570\u636e\u4e0a\u7684\u6f5c\u529b，\u88ab\u89c6\u4e3a\u6df1\u5ea6\u5b66\u4e60\u590d\u5174\u7684\u91cd\u8981\u524d\u594f\u3002<\/p>\n

\u4e09\u3001\u6df1\u5ea6\u5b66\u4e60\u9769\u547d：CNN\u7684\u9ec4\u91d1\u65f6\u4ee3（2012-2017）<\/h3>\n

3.1 AlexNet(2012)：\u5f15\u7206\u9769\u547d\u7684\u6838\u5f39<\/h4>\n

2012\u5e74\u7684ImageNet\u5927\u89c4\u6a21\u89c6\u89c9\u8bc6\u522b\u6311\u6218\u8d5b（ILSVRC）\u662f\u8ba1\u7b97\u673a\u89c6\u89c9\u5386\u53f2\u4e0a\u7684\u5206\u6c34\u5cad\u3002Alex Krizhevsky\u3001Ilya Sutskever\u548cGeoffrey Hinton\u63d0\u4ea4\u4e86\u4e00\u4e2a\u540d\u4e3aAlexNet\u7684\u6df1\u5ea6\u5377\u79ef\u795e\u7ecf\u7f51\u7edc，\u4ee5top-5\u9519\u8bef\u738715.3%\u7684\u6210\u7ee9\u593a\u51a0，\u6bd4\u7b2c\u4e8c\u540d（\u57fa\u4e8e\u624b\u5de5\u7279\u5f81\u7684SIFT+FV\u65b9\u6cd5）\u4f4e\u4e86\u8d85\u8fc710\u4e2a\u767e\u5206\u70b9\u3002\u8fd9\u4e00\u5de8\u5927\u5dee\u8ddd\u4ee4\u6574\u4e2a\u9886\u57df\u4e3a\u4e4b\u9707\u60ca，\u6807\u5fd7\u7740\u6df1\u5ea6\u5b66\u4e60\u6b63\u5f0f\u63a5\u7ba1\u8ba1\u7b97\u673a\u89c6\u89c9\u3002<\/p>\n

AlexNet\u7684\u6210\u529f\u5e76\u975e\u6e90\u81ea\u5355\u4e00\u7684\u6280\u672f\u7a81\u7834，\u800c\u662f\u591a\u9879\u5173\u952e\u6280\u672f\u7684\u5de7\u5999\u7ed3\u5408\u3002\u9996\u5148，\u5b83\u4f7f\u7528\u4e86ReLU（\u4fee\u6b63\u7ebf\u6027\u5355\u5143）\u6fc0\u6d3b\u51fd\u6570\u66ff\u4ee3\u4f20\u7edf\u7684sigmoid\u6216tanh\u3002ReLU\u7684\u68af\u5ea6\u5728\u6b63\u533a\u95f4\u6052\u4e3a1，\u6709\u6548\u7f13\u89e3\u4e86\u6df1\u5c42\u7f51\u7edc\u7684\u68af\u5ea6\u6d88\u5931\u95ee\u9898，\u540c\u65f6\u8ba1\u7b97\u7b80\u5355，\u52a0\u901f\u4e86\u8bad\u7ec3\u8fc7\u7a0b\u3002\u5176\u6b21，AlexNet\u5229\u7528\u4e86GPU（NVIDIA GTX 580）\u8fdb\u884c\u5e76\u884c\u8ba1\u7b97，\u5c06\u8bad\u7ec3\u65f6\u95f4\u4ece\u6570\u6708\u7f29\u77ed\u5230\u7ea6\u4e00\u5468\u3002\u8fd9\u5728\u5f53\u65f6\u662f\u9769\u547d\u6027\u7684\u2014\u2014GPU\u4ece\u6b64\u6210\u4e3a\u6df1\u5ea6\u5b66\u4e60\u8bad\u7ec3\u7684\u6807\u914d\u786c\u4ef6\u3002<\/p>\n

AlexNet\u8fd8\u5f15\u5165\u4e86\u591a\u79cd\u6b63\u5219\u5316\u6280\u672f\u6765\u9632\u6b62\u8fc7\u62df\u5408\u3002Dropout\u5728\u8bad\u7ec3\u65f6\u4ee5\u4e00\u5b9a\u6982\u7387\u968f\u673a\u4e22\u5f03\u795e\u7ecf\u5143，\u8feb\u4f7f\u7f51\u7edc\u5b66\u4e60\u5197\u4f59\u7684\u9c81\u68d2\u7279\u5f81，\u76f8\u5f53\u4e8e\u4e00\u79cd\u9690\u5f0f\u7684\u6a21\u578b\u96c6\u6210\u3002\u6570\u636e\u589e\u5f3a（\u968f\u673a\u88c1\u526a\u3001\u6c34\u5e73\u7ffb\u8f6c\u3001\u989c\u8272\u6296\u52a8）\u4eba\u4e3a\u6269\u5145\u4e86\u8bad\u7ec3\u96c6\u89c4\u6a21，\u63d0\u9ad8\u4e86\u6a21\u578b\u7684\u6cdb\u5316\u80fd\u529b\u3002\u5c40\u90e8\u54cd\u5e94\u5f52\u4e00\u5316（LRN）\u867d\u7136\u540e\u7eed\u88ab\u8bc1\u660e\u5e76\u975e\u5fc5\u8981，\u4f46\u53cd\u6620\u4e86\u5f53\u65f6\u5bf9\u7f51\u7edc\u7a33\u5b9a\u6027\u7684\u63a2\u7d22\u3002<\/p>\n

AlexNet\u7684\u67b6\u6784\u5305\u542b5\u4e2a\u5377\u79ef\u5c42\u548c3\u4e2a\u5168\u8fde\u63a5\u5c42，\u603b\u8ba1\u7ea66000\u4e07\u4e2a\u53c2\u6570\u3002\u7531\u4e8e\u5f53\u65f6GPU\u663e\u5b58\u9650\u5236（\u5355\u5361\u4ec53GB），AlexNet\u91c7\u7528\u4e86\u53ccGPU\u5e76\u884c\u8bad\u7ec3\u7684\u7b56\u7565，\u5c06\u7f51\u7edc\u5206\u6210\u4e24\u90e8\u5206\u5206\u522b\u653e\u5728\u4e24\u4e2aGPU\u4e0a，\u53ea\u6709\u5728\u7279\u5b9a\u5c42\u4e4b\u95f4\u8fdb\u884c\u901a\u4fe1\u3002\u8fd9\u79cd\u5de5\u7a0b\u4e0a\u7684\u6298\u4e2d\u65b9\u6848，\u610f\u5916\u5730\u50ac\u751f\u4e86\u5206\u7ec4\u5377\u79ef\u7684\u601d\u60f3，\u4e3a\u540e\u7eed\u9ad8\u6548\u7f51\u7edc\u8bbe\u8ba1\u63d0\u4f9b\u4e86\u7075\u611f\u3002<\/p>\n

3.2 VGGNet(2014)：\u6df1\u5ea6\u5c31\u662f\u529b\u91cf<\/h4>\n

\u725b\u6d25\u5927\u5b66\u89c6\u89c9\u51e0\u4f55\u7ec4（Visual Geometry Group）\u57282014\u5e74\u63d0\u51fa\u7684VGGNet，\u4ee5\u5176\u7b80\u6d01\u4f18\u96c5\u7684\u67b6\u6784\u8bbe\u8ba1\u800c\u95fb\u540d\u3002VGGNet\u7684\u6838\u5fc3\u8bbe\u8ba1\u54f2\u5b66\u662f：\u4f7f\u7528\u66f4\u6df1\u7684\u7f51\u7edc，\u4f46\u6240\u6709\u5377\u79ef\u5c42\u90fd\u4f7f\u7528\u6700\u5c0f\u76843×3\u5377\u79ef\u6838\u3002\u4e24\u4e2a3×3\u5377\u79ef\u6838\u7684\u5806\u53e0\u7b49\u6548\u4e8e\u4e00\u4e2a5×5\u5377\u79ef\u6838\u7684\u611f\u53d7\u91ce，\u4e09\u4e2a3×3\u5377\u79ef\u6838\u7684\u5806\u53e0\u7b49\u6548\u4e8e\u4e00\u4e2a7×7\u5377\u79ef\u6838\u7684\u611f\u53d7\u91ce，\u4f46\u53c2\u6570\u91cf\u66f4\u5c11，\u4e14\u5f15\u5165\u4e86\u66f4\u591a\u7684\u975e\u7ebf\u6027\u53d8\u6362\u3002<\/p>\n

VGG-16\u548cVGG-19（\u5206\u522b\u5305\u542b16\u548c19\u4e2a\u6743\u91cd\u5c42）\u662fVGGNet\u6700\u5e38\u7528\u7684\u4e24\u4e2a\u7248\u672c\u3002\u5b83\u4eec\u90fd\u9075\u5faa\u76f8\u540c\u7684\u7ed3\u6784\u6a21\u5f0f：\u591a\u4e2a\u5377\u79ef\u5c42\u5806\u53e0\u540e\u63a5\u4e00\u4e2a2×2\u6700\u5927\u6c60\u5316\u5c42，\u7279\u5f81\u56fe\u5206\u8fa8\u7387\u51cf\u534a，\u901a\u9053\u6570\u52a0\u500d\u3002\u8fd9\u79cd\u7b80\u5355\u7684\u6a21\u5757\u5316\u8bbe\u8ba1\u4f7f\u5f97VGGNet\u975e\u5e38\u5bb9\u6613\u7406\u89e3\u548c\u5b9e\u73b0，\u81f3\u4eca\u4ecd\u7136\u662f\u8bb8\u591a\u89c6\u89c9\u4efb\u52a1（\u5c24\u5176\u662f\u98ce\u683c\u8fc1\u79fb）\u7684\u9996\u9009\u9aa8\u5e72\u7f51\u7edc\u3002<\/p>\n

VGGNet\u5728ImageNet 2014\u6bd4\u8d5b\u4e2d\u53d6\u5f97\u4e86top-5\u9519\u8bef\u73877.3%\u7684\u6210\u7ee9，\u8f83AlexNet\u6709\u4e86\u663e\u8457\u63d0\u5347\u3002\u7136\u800c，VGGNet\u7684\u7f3a\u70b9\u4e5f\u5f88\u660e\u663e：\u5168\u8fde\u63a5\u5c42\u5360\u636e\u4e86\u5927\u91cf\u53c2\u6570（VGG-16\u7684138M\u53c2\u6570\u4e2d，\u5168\u8fde\u63a5\u5c42\u7ea6\u536090%），\u5bfc\u81f4\u6a21\u578b\u4f53\u79ef\u5e9e\u5927\u3001\u8ba1\u7b97\u5f00\u9500\u9ad8\u3002\u8fd9\u4e00\u7f3a\u9677\u4fc3\u4f7f\u7814\u7a76\u8005\u5f00\u59cb\u601d\u8003\u5982\u4f55\u8bbe\u8ba1\u66f4\u9ad8\u6548\u7684\u7f51\u7edc\u7ed3\u6784\u3002<\/p>\n

3.3 GoogLeNet\/Inception(2014)：\u62d3\u5bbd\u800c\u975e\u52a0\u6df1<\/h4>\n

\u4e0eVGGNet\u540c\u5e74\u8bde\u751f\u7684GoogLeNet（\u53c8\u540dInception v1），\u51fa\u81eaGoogle\u7684Szegedy\u7b49\u4eba\u4e4b\u624b，\u91c7\u53d6\u4e86\u4e00\u79cd\u4e0eVGGNet\u622a\u7136\u4e0d\u540c\u7684\u601d\u8def\u3002GoogLeNet\u63d0\u51fa\u4e86Inception\u6a21\u5757，\u5728\u540c\u4e00\u5c42\u5185\u5e76\u884c\u4f7f\u7528\u4e0d\u540c\u5c3a\u5ea6\u7684\u5377\u79ef\u6838（1×1\u30013×3\u30015×5）\u548c\u6c60\u5316\u64cd\u4f5c，\u8ba9\u7f51\u7edc\u81ea\u52a8\u5b66\u4e60\u6700\u4f18\u7684\u7279\u5f81\u7ec4\u5408\u3002\u8fd9\u79cd"\u62d3\u5bbd"\u800c\u975e"\u52a0\u6df1"\u7684\u7b56\u7565，\u4f7f\u5f97\u7f51\u7edc\u80fd\u591f\u6355\u6349\u591a\u5c3a\u5ea6\u7684\u89c6\u89c9\u4fe1\u606f\u3002<\/p>\n

Inception\u6a21\u5757\u7684\u4e00\u4e2a\u5173\u952e\u521b\u65b0\u662f\u4f7f\u75281×1\u5377\u79ef\u8fdb\u884c\u964d\u7ef4\u3002\u5728\u5927\u5377\u79ef\u6838（3×3\u548c5×5）\u4e4b\u524d\u5148\u75281×1\u5377\u79ef\u51cf\u5c11\u901a\u9053\u6570，\u53ef\u4ee5\u5927\u5e45\u964d\u4f4e\u8ba1\u7b97\u91cf\u3002\u4f8b\u5982，\u5bf928x28x256\u7684\u8f93\u5165，\u76f4\u63a5\u4f7f\u75285×5\u5377\u79ef\u8f93\u51fa28x28x128\u9700\u8981\u7ea62.3\u4ebf\u6b21\u4e58\u6cd5\u8fd0\u7b97；\u800c\u5148\u75281×1\u5377\u79ef\u5c06\u901a\u9053\u6570\u964d\u523064，\u518d\u8fdb\u884c5×5\u5377\u79ef，\u4ec5\u9700\u7ea61.1\u4ebf\u6b21\u8fd0\u7b97，\u8282\u7701\u4e86\u8d85\u8fc7\u4e00\u534a\u7684\u8ba1\u7b97\u91cf\u3002<\/p>\n

GoogLeNet\u6700\u6df1\u7684\u90e8\u5206\u5305\u542b22\u4e2a\u6743\u91cd\u5c42，\u4f46\u901a\u8fc7\u5927\u91cf\u4f7f\u75281×1\u5377\u79ef\u548c\u5168\u5c40\u5e73\u5747\u6c60\u5316\u66ff\u4ee3\u5168\u8fde\u63a5\u5c42，\u5176\u53c2\u6570\u603b\u91cf\u4ec5\u7ea6500\u4e07，\u662fAlexNet\u76841\/12，VGG-16\u76841\/28\u3002\u8fd9\u79cd\u9ad8\u6548\u7684\u8bbe\u8ba1\u4f7f\u5176\u5728\u79fb\u52a8\u7aef\u548c\u5d4c\u5165\u5f0f\u8bbe\u5907\u4e0a\u5177\u6709\u6781\u5927\u7684\u5b9e\u7528\u4ef7\u503c\u3002<\/p>\n

GoogLeNet\u5728ILSVRC 2014\u5206\u7c7b\u4efb\u52a1\u4e2d\u4ee56.7%\u7684top-5\u9519\u8bef\u7387\u593a\u51a0，\u8d85\u8d8a\u4e86VGGNet\u3002\u6b64\u540e，Inception\u7cfb\u5217\u4e0d\u65ad\u8fed\u4ee3（v2\u5f15\u5165Batch Normalization，v3\u5f15\u5165\u5206\u89e3\u5377\u79ef，v4\u7ed3\u5408ResNet\u601d\u60f3），\u5728\u7cbe\u5ea6\u548c\u6548\u7387\u4e0a\u6301\u7eed\u6539\u8fdb\u3002<\/p>\n

3.4 ResNet(2015)：\u8de8\u8d8a\u6df1\u5ea6\u7684\u9e3f\u6c9f<\/h4>\n

\u968f\u7740\u7f51\u7edc\u6df1\u5ea6\u7684\u589e\u52a0，\u7814\u7a76\u8005\u89c2\u5bdf\u5230\u4e00\u4e2a\u53cd\u76f4\u89c9\u7684\u73b0\u8c61：\u66f4\u6df1\u7684\u7f51\u7edc\u4e0d\u4ec5\u6ca1\u6709\u5e26\u6765\u66f4\u9ad8\u7684\u7cbe\u5ea6，\u53cd\u800c\u51fa\u73b0\u4e86\u9000\u5316\u2014\u2014\u8bad\u7ec3\u8bef\u5dee\u548c\u6d4b\u8bd5\u8bef\u5dee\u90fd\u9ad8\u4e8e\u8f83\u6d45\u7684\u7f51\u7edc\u3002\u8fd9\u4e0d\u662f\u8fc7\u62df\u5408\u95ee\u9898，\u800c\u662f\u4f18\u5316\u56f0\u96be\u3002\u7406\u8bba\u4e0a，\u6df1\u5c42\u7f51\u7edc\u53ef\u4ee5\u5b66\u4e60\u6052\u7b49\u6620\u5c04\u6765\u5b9e\u73b0\u4e0e\u6d45\u5c42\u7f51\u7edc\u76f8\u540c\u7684\u6027\u80fd，\u4f46\u68af\u5ea6\u4e0b\u964d\u7b97\u6cd5\u96be\u4ee5\u627e\u5230\u8fd9\u6837\u7684\u89e3\u3002<\/p>\n

\u4f55\u607a\u660e\u7b49\u4eba\u57282015\u5e74\u63d0\u51fa\u7684\u6b8b\u5dee\u7f51\u7edc（ResNet），\u4ee5\u4f18\u96c5\u7684\u65b9\u5f0f\u89e3\u51b3\u4e86\u8fd9\u4e00\u95ee\u9898\u3002ResNet\u7684\u6838\u5fc3\u601d\u60f3\u662f\u6b8b\u5dee\u5b66\u4e60：\u8ba9\u7f51\u7edc\u5c42\u5b66\u4e60\u8f93\u5165\u4e0e\u8f93\u51fa\u4e4b\u95f4\u7684\u6b8b\u5dee\u6620\u5c04F(x)=H(x)-x，\u800c\u975e\u76f4\u63a5\u5b66\u4e60\u671f\u671b\u7684\u5e95\u5c42\u6620\u5c04H(x)\u3002\u901a\u8fc7\u5f15\u5165"\u8df3\u8dc3\u8fde\u63a5"（skip connection），\u4fe1\u606f\u53ef\u4ee5\u7ed5\u8fc7\u82e5\u5e72\u5c42\u76f4\u63a5\u4f20\u9012，\u4f7f\u5f97\u7f51\u7edc\u53ef\u4ee5\u8f7b\u677e\u5b66\u4e60\u6052\u7b49\u6620\u5c04（\u53ea\u9700\u5c06\u6b8b\u5dee\u903c\u8fd1\u96f6）\u3002<\/p>\n

\u8fd9\u4e00\u8bbe\u8ba1\u4f7f\u5f97\u8bad\u7ec3\u6781\u6df1\u7f51\u7edc\u6210\u4e3a\u53ef\u80fd\u3002ResNet-50\u3001ResNet-101\u3001ResNet-152\u4e43\u81f3ResNet-1001\u90fd\u53d6\u5f97\u4e86\u6210\u529f\u3002ResNet-152\u5728ILSVRC 2015\u4e2d\u4ee53.57%\u7684top-5\u9519\u8bef\u7387\u593a\u51a0，\u9996\u6b21\u8d85\u8d8a\u4e86\u4eba\u7c7b\u6c34\u5e73（\u7ea65.1%）\u3002\u8fd9\u662f\u4e00\u4e2a\u91cc\u7a0b\u7891\u5f0f\u7684\u65f6\u523b\u2014\u2014\u5728ImageNet\u8fd9\u4e2a\u7279\u5b9a\u4efb\u52a1\u4e0a，\u673a\u5668\u89c6\u89c9\u7684\u8bc6\u522b\u80fd\u529b\u5df2\u7ecf\u8d85\u8d8a\u4e86\u4eba\u7c7b\u3002<\/p>\n

ResNet\u7684\u6b8b\u5dee\u8fde\u63a5\u601d\u60f3\u6df1\u523b\u5730\u5f71\u54cd\u4e86\u540e\u7eed\u7f51\u7edc\u67b6\u6784\u7684\u8bbe\u8ba1\u3002DenseNet（2017）\u5c06\u8fd9\u4e00\u601d\u60f3\u63a8\u5411\u6781\u81f4，\u6bcf\u4e00\u5c42\u90fd\u63a5\u6536\u524d\u9762\u6240\u6709\u5c42\u7684\u8f93\u51fa\u4f5c\u4e3a\u8f93\u5165\u3002ResNeXt（2017）\u7ed3\u5408\u4e86ResNet\u7684\u6b8b\u5dee\u8fde\u63a5\u548cInception\u7684\u5206\u7ec4\u5377\u79ef\u601d\u60f3\u3002SE-Net（2018）\u5f15\u5165\u4e86\u901a\u9053\u6ce8\u610f\u529b\u673a\u5236，\u4e3a\u6bcf\u4e2a\u901a\u9053\u5b66\u4e60\u4e00\u4e2a\u6743\u91cd，\u5b9e\u73b0\u4e86\u901a\u9053\u7ef4\u5ea6\u7684\u81ea\u9002\u5e94\u7279\u5f81\u91cd\u6807\u5b9a\u3002<\/p>\n

3.5 Batch Normalization\u4e0e\u4f18\u5316\u6280\u672f<\/h4>\n

\u5728CNN\u67b6\u6784\u6f14\u8fdb\u7684\u540c\u65f6，\u8bad\u7ec3\u6280\u672f\u7684\u8fdb\u6b65\u540c\u6837\u81f3\u5173\u91cd\u8981\u30022015\u5e74，Sergey Ioffe\u548cChristian Szegedy\u63d0\u51fa\u7684\u6279\u91cf\u5f52\u4e00\u5316（Batch Normalization，BN），\u662f\u7ee7ReLU\u4e4b\u540e\u6700\u91cd\u8981\u7684\u8bad\u7ec3\u6280\u672f\u7a81\u7834\u3002<\/p>\n

BN\u7684\u6838\u5fc3\u64cd\u4f5c\u662f\u5728\u6bcf\u4e2amini-batch\u5185\u5bf9\u6bcf\u4e00\u5c42\u7684\u6fc0\u6d3b\u503c\u8fdb\u884c\u6807\u51c6\u5316，\u4f7f\u5176\u5747\u503c\u4e3a0\u3001\u65b9\u5dee\u4e3a1，\u7136\u540e\u901a\u8fc7\u53ef\u5b66\u4e60\u7684\u7f29\u653e\u548c\u5e73\u79fb\u53c2\u6570\u6062\u590d\u7f51\u7edc\u7684\u8868\u8fbe\u80fd\u529b\u3002BN\u5e26\u6765\u7684\u597d\u5904\u662f\u591a\u65b9\u9762\u7684：\u5b83\u5141\u8bb8\u4f7f\u7528\u66f4\u5927\u7684\u5b66\u4e60\u7387，\u52a0\u901f\u8bad\u7ec3\u6536\u655b；\u5b83\u964d\u4f4e\u4e86\u5bf9\u6743\u91cd\u521d\u59cb\u5316\u7684\u654f\u611f\u6027；\u5b83\u5177\u6709\u4e00\u5b9a\u7684\u6b63\u5219\u5316\u6548\u679c，\u53ef\u4ee5\u51cf\u5c11\u5bf9Dropout\u7684\u4f9d\u8d56；\u6700\u91cd\u8981\u7684\u662f，\u5b83\u6709\u6548\u7f13\u89e3\u4e86\u5185\u90e8\u534f\u53d8\u91cf\u504f\u79fb\u95ee\u9898，\u4f7f\u5f97\u8bad\u7ec3\u975e\u5e38\u6df1\u7684\u7f51\u7edc\u6210\u4e3a\u53ef\u80fd\u3002<\/p>\n

\u5728\u4f18\u5316\u5668\u65b9\u9762，Adam（2014）\u7ed3\u5408\u4e86\u52a8\u91cf\u6cd5\u548cRMSProp\u7684\u4f18\u70b9，\u901a\u8fc7\u8ba1\u7b97\u68af\u5ea6\u7684\u4e00\u9636\u77e9\u548c\u4e8c\u9636\u77e9\u4f30\u8ba1\u6765\u4e3a\u4e0d\u540c\u53c2\u6570\u8bbe\u8ba1\u72ec\u7acb\u7684\u81ea\u9002\u5e94\u5b66\u4e60\u7387\u3002Adam\u6210\u4e3a\u6df1\u5ea6\u5b66\u4e60\u4e2d\u6700\u5e38\u7528\u7684\u4f18\u5316\u5668\u4e4b\u4e00，\u5176\u9ed8\u8ba4\u8d85\u53c2\u6570（\u5b66\u4e60\u73870.001，beta1=0.9，beta2=0.999）\u5728\u5927\u591a\u6570\u60c5\u51b5\u4e0b\u90fd\u80fd\u53d6\u5f97\u4e0d\u9519\u7684\u6548\u679c\u3002<\/p>\n

\u56db\u3001\u6ce8\u610f\u529b\u9769\u547d：Transformer\u5165\u4e3b\u89c6\u89c9（2017-2021）<\/h3>\n

4.1 Transformer(2017)：NLP\u7684\u8303\u5f0f\u8f6c\u53d8<\/h4>\n

2017\u5e74，Google\u5728\u8bba\u6587"Attention Is All You Need"\u4e2d\u63d0\u51fa\u4e86Transformer\u67b6\u6784，\u6700\u521d\u662f\u4e3a\u673a\u5668\u7ffb\u8bd1\u4efb\u52a1\u8bbe\u8ba1\u7684\u3002Transformer\u5b8c\u5168\u6452\u5f03\u4e86\u4f20\u7edf\u7684\u5faa\u73af\u795e\u7ecf\u7f51\u7edc（RNN）\u548c\u5377\u79ef\u795e\u7ecf\u7f51\u7edc，\u4ec5\u4f9d\u8d56\u81ea\u6ce8\u610f\u529b（self-attention）\u673a\u5236\u6765\u5efa\u6a21\u5e8f\u5217\u4e2d\u7684\u4f9d\u8d56\u5173\u7cfb\u3002<\/p>\n

\u81ea\u6ce8\u610f\u529b\u673a\u5236\u7684\u6838\u5fc3\u8ba1\u7b97\u53ef\u4ee5\u6982\u62ec\u4e3a：\u5c06\u8f93\u5165\u5e8f\u5217\u4e2d\u7684\u6bcf\u4e2a\u5143\u7d20\u6620\u5c04\u4e3a\u67e5\u8be2（Query）\u3001\u952e（Key）\u548c\u503c（Value）\u4e09\u4e2a\u5411\u91cf，\u7136\u540e\u901a\u8fc7\u8ba1\u7b97\u67e5\u8be2\u4e0e\u6240\u6709\u952e\u7684\u76f8\u4f3c\u5ea6，\u52a0\u6743\u805a\u5408\u6240\u6709\u503c，\u5f97\u5230\u6bcf\u4e2a\u4f4d\u7f6e\u7684\u8f93\u51fa\u8868\u793a\u3002\u8fd9\u79cd\u673a\u5236\u4f7f\u5f97\u6bcf\u4e2a\u4f4d\u7f6e\u90fd\u80fd\u76f4\u63a5\u5173\u6ce8\u5e8f\u5217\u4e2d\u7684\u4efb\u610f\u5176\u4ed6\u4f4d\u7f6e，\u65e0\u9700\u50cfRNN\u90a3\u6837\u9010\u6b65\u4f20\u9012\u4fe1\u606f，\u6709\u6548\u89e3\u51b3\u4e86\u957f\u8ddd\u79bb\u4f9d\u8d56\u95ee\u9898\u3002<\/p>\n

Transformer\u5f15\u5165\u4e86\u591a\u5934\u6ce8\u610f\u529b（Multi-Head Attention），\u901a\u8fc7\u5e76\u884c\u8ba1\u7b97\u591a\u4e2a\u6ce8\u610f\u529b\u5934，\u6bcf\u4e2a\u5934\u5173\u6ce8\u4e0d\u540c\u7684\u8868\u793a\u5b50\u7a7a\u95f4，\u589e\u5f3a\u4e86\u6a21\u578b\u7684\u8868\u8fbe\u80fd\u529b\u3002\u4f4d\u7f6e\u7f16\u7801（Positional Encoding）\u4e3a\u6a21\u578b\u63d0\u4f9b\u4e86\u5e8f\u5217\u4e2d\u5143\u7d20\u7684\u4f4d\u7f6e\u4fe1\u606f，\u5f25\u8865\u4e86\u81ea\u6ce8\u610f\u529b\u673a\u5236\u672c\u8eab\u4e0d\u5177\u5907\u4f4d\u7f6e\u611f\u77e5\u80fd\u529b\u7684\u7f3a\u9677\u3002<\/p>\n

Transformer\u5728NLP\u9886\u57df\u53d6\u5f97\u4e86\u5de8\u5927\u6210\u529f，\u50ac\u751f\u4e86BERT\u3001GPT\u7b49\u4e00\u7cfb\u5217\u91cc\u7a0b\u7891\u5f0f\u6a21\u578b\u3002\u7136\u800c，\u5c06Transformer\u5e94\u7528\u4e8e\u8ba1\u7b97\u673a\u89c6\u89c9\u9762\u4e34\u4e00\u4e2a\u76f4\u63a5\u6311\u6218：\u81ea\u6ce8\u610f\u529b\u7684\u8ba1\u7b97\u590d\u6742\u5ea6\u4e0e\u5e8f\u5217\u957f\u5ea6\u7684\u5e73\u65b9\u6210\u6b63\u6bd4\u3002\u5bf9\u4e8e\u4e00\u5f20224×224\u7684\u56fe\u50cf，\u5982\u679c\u6bcf\u4e2a\u50cf\u7d20\u90fd\u4f5c\u4e3a\u4e00\u4e2atoken，\u5e8f\u5217\u957f\u5ea6\u5c06\u8fbe\u523050176，\u8ba1\u7b97\u91cf\u8d85\u51fa\u4e86\u73b0\u6709\u786c\u4ef6\u7684\u627f\u53d7\u80fd\u529b\u3002<\/p>\n

4.2 ViT(2020)：\u56fe\u50cf\u5373\u5e8f\u5217<\/h4>\n

2020\u5e74，Google\u63d0\u51fa\u7684Vision Transformer（ViT）\u4ee5\u4e00\u79cd\u7b80\u5355\u800c\u5de7\u5999\u7684\u65b9\u5f0f\u89e3\u51b3\u4e86\u4e0a\u8ff0\u95ee\u9898：\u5c06\u56fe\u50cf\u5206\u5272\u6210\u56fa\u5b9a\u5927\u5c0f\u7684patch（\u4f8b\u598216×16），\u5c06\u6bcf\u4e2apatch\u7ebf\u6027\u6295\u5f71\u4e3a\u4e00\u4e2a\u5411\u91cftoken，\u7136\u540e\u5c06\u8fd9\u4e9btoken\u5e8f\u5217\u8f93\u5165\u6807\u51c6\u7684Transformer\u7f16\u7801\u5668\u3002<\/p>\n

ViT\u7684\u8fd9\u4e00\u8bbe\u8ba1\u975e\u5e38\u5927\u80c6\u2014\u2014\u5b83\u51e0\u4e4e\u6ca1\u6709\u4efb\u4f55\u56fe\u50cf\u7279\u5b9a\u7684\u5f52\u7eb3\u504f\u7f6e（inductive bias），\u5b8c\u5168\u4f9d\u8d56Transformer\u7684\u81ea\u6ce8\u610f\u529b\u673a\u5236\u6765\u5b66\u4e60\u56fe\u50cf\u7684\u7a7a\u95f4\u7ed3\u6784\u3002CNN\u4e2d\u7684\u5c40\u90e8\u8fde\u63a5\u3001\u5e73\u79fb\u4e0d\u53d8\u6027\u7b49\u5bf9\u56fe\u50cf\u5904\u7406\u975e\u5e38\u6709\u6548\u7684\u5148\u9a8c\u77e5\u8bc6，\u5728ViT\u4e2d\u90fd\u4e0d\u5b58\u5728\u3002\u8fd9\u610f\u5473\u7740ViT\u9700\u8981\u4ece\u6570\u636e\u4e2d\u5b66\u4e60\u4e00\u5207，\u56e0\u6b64\u5bf9\u9884\u8bad\u7ec3\u6570\u636e\u91cf\u6709\u6781\u9ad8\u7684\u8981\u6c42\u3002<\/p>\n

\u5b9e\u9a8c\u7ed3\u679c\u8868\u660e，\u5f53\u5728ImageNet-21K\u6216JFT-300M\u7b49\u5927\u89c4\u6a21\u6570\u636e\u96c6\u4e0a\u9884\u8bad\u7ec3\u540e，ViT\u80fd\u591f\u8d85\u8d8a\u540c\u7b49\u89c4\u6a21\u7684CNN\u6a21\u578b\u3002\u7136\u800c，\u5728ImageNet-1K（\u7ea6130\u4e07\u5f20\u56fe\u7247）\u4e0a\u4ece\u5934\u8bad\u7ec3\u65f6，ViT\u7684\u6027\u80fd\u4e0d\u5982CNN\u3002\u8fd9\u63ed\u793a\u4e86ViT\u7684\u4e00\u4e2a\u91cd\u8981\u7279\u6027：\u5b83\u7684\u5f52\u7eb3\u504f\u7f6e\u66f4\u5f31，\u4f46\u8868\u793a\u80fd\u529b\u66f4\u5f3a，\u5173\u952e\u5728\u4e8e\u662f\u5426\u6709\u8db3\u591f\u7684\u6570\u636e\u6765\u91ca\u653e\u8fd9\u79cd\u80fd\u529b\u3002<\/p>\n

ViT\u7684\u51fa\u73b0\u6807\u5fd7\u7740\u8ba1\u7b97\u673a\u89c6\u89c9\u9886\u57df\u8fdb\u5165\u4e86\u4e00\u4e2a\u65b0\u7684\u8303\u5f0f\u2014\u2014\u5728\u6b64\u4e4b\u524d，CNN\u662f\u89c6\u89c9\u4efb\u52a1\u7684\u7edd\u5bf9\u738b\u8005；\u5728\u6b64\u4e4b\u540e，Transformer\u5f00\u59cb\u4e0eCNN\u5206\u5ead\u6297\u793c，\u5e76\u9010\u6e10\u5c55\u73b0\u51fa\u66f4\u5f3a\u7684\u6f5c\u529b\u3002<\/p>\n

4.3 DeiT\u3001Swin Transformer\u7b49\u540e\u7eed\u5de5\u4f5c<\/h4>\n

ViT\u7684\u6210\u529f\u6fc0\u53d1\u4e86\u5927\u91cf\u540e\u7eed\u5de5\u4f5c，\u8fd9\u4e9b\u5de5\u4f5c\u4ece\u4e0d\u540c\u89d2\u5ea6\u6539\u8fdb\u4e86\u89c6\u89c9Transformer\u3002DeiT（Data-efficient Image Transformers，2020）\u901a\u8fc7\u77e5\u8bc6\u84b8\u998f（\u4f7f\u7528CNN\u6559\u5e08\u6a21\u578b\u6307\u5bfcViT\u8bad\u7ec3）\u548c\u66f4\u5f3a\u7684\u6570\u636e\u589e\u5f3a\u7b56\u7565，\u4f7f\u5f97ViT\u5728ImageNet-1K\u4e0a\u4e5f\u80fd\u53d6\u5f97\u6709\u7ade\u4e89\u529b\u7684\u7ed3\u679c，\u964d\u4f4e\u4e86\u5bf9\u5927\u89c4\u6a21\u9884\u8bad\u7ec3\u6570\u636e\u7684\u4f9d\u8d56\u3002<\/p>\n

Swin Transformer（2021）\u662f\u89c6\u89c9Transformer\u53d1\u5c55\u7684\u53c8\u4e00\u4e2a\u91cd\u8981\u91cc\u7a0b\u7891\u3002\u5b83\u5f15\u5165\u4e86\u4e24\u4e2a\u5173\u952e\u521b\u65b0：\u5c42\u6b21\u5316\u7279\u5f81\u56fe\u6784\u5efa（\u7c7b\u4f3cCNN\u7684\u91d1\u5b57\u5854\u7ed3\u6784）\u548c\u79fb\u4f4d\u7a97\u53e3\u81ea\u6ce8\u610f\u529b（Shifted Window Self-Attention）\u3002\u524d\u8005\u4f7f\u5f97Swin Transformer\u80fd\u591f\u4f5c\u4e3a\u901a\u7528\u9aa8\u5e72\u7f51\u7edc\u7528\u4e8e\u5404\u79cd\u89c6\u89c9\u4efb\u52a1（\u5206\u7c7b\u3001\u68c0\u6d4b\u3001\u5206\u5272），\u540e\u8005\u5728\u4fdd\u6301\u8ba1\u7b97\u6548\u7387\u7684\u540c\u65f6\u5b9e\u73b0\u4e86\u8de8\u7a97\u53e3\u7684\u4fe1\u606f\u4ea4\u4e92\u3002<\/p>\n

Swin Transformer\u5728ImageNet-22K\u4e0a\u9884\u8bad\u7ec3\u540e\u5728ImageNet-1K\u4e0a\u8fbe\u5230\u4e8687.3%\u7684top-1\u51c6\u786e\u7387，\u5e76\u5728COCO\u76ee\u6807\u68c0\u6d4b\u548cADE20K\u8bed\u4e49\u5206\u5272\u7b49\u4efb\u52a1\u4e0a\u5237\u65b0\u4e86\u5f53\u65f6\u7684\u8bb0\u5f55\u3002\u5b83\u8bc1\u660e\u4e86\u7ecf\u8fc7\u7cbe\u5fc3\u8bbe\u8ba1\u7684Transformer\u67b6\u6784\u53ef\u4ee5\u5168\u9762\u8d85\u8d8aCNN，\u6210\u4e3a\u65b0\u7684\u89c6\u89c9\u9aa8\u5e72\u7f51\u7edc\u8303\u5f0f\u3002<\/p>\n

4.4 CLIP(2021)：\u89c6\u89c9-\u8bed\u8a00\u7684\u6865\u6881<\/h4>\n

OpenAI\u57282021\u5e74\u53d1\u5e03\u7684CLIP（Contrastive Language-Image Pre-training），\u5f00\u542f\u4e86\u89c6\u89c9-\u8bed\u8a00\u8054\u5408\u5b66\u4e60\u7684\u65b0\u7eaa\u5143\u3002CLIP\u7684\u6838\u5fc3\u601d\u60f3\u662f\u7528\u81ea\u7136\u8bed\u8a00\u4f5c\u4e3a\u76d1\u7763\u4fe1\u53f7\u6765\u8bad\u7ec3\u89c6\u89c9\u6a21\u578b，\u800c\u975e\u4f20\u7edf\u7684\u56fa\u5b9a\u7c7b\u522b\u6807\u7b7e\u3002<\/p>\n

CLIP\u7684\u8bad\u7ec3\u76ee\u6807\u662f：\u5bf9\u4e8e\u4e00\u6279（\u56fe\u50cf，\u6587\u672c）\u914d\u5bf9\u6570\u636e，\u8ba9\u914d\u5bf9\u7684\u56fe\u50cf\u548c\u6587\u672c\u7684\u5d4c\u5165\u5411\u91cf\u5c3d\u53ef\u80fd\u76f8\u4f3c，\u540c\u65f6\u8ba9\u975e\u914d\u5bf9\u7684\u5c3d\u53ef\u80fd\u4e0d\u76f8\u4f3c\u3002\u8fd9\u79cd\u5bf9\u6bd4\u5b66\u4e60\u76ee\u6807\u4f7f\u5f97\u6a21\u578b\u5b66\u4e60\u5230\u7684\u89c6\u89c9\u8868\u793a\u4e0e\u8bed\u8a00\u8868\u793a\u5bf9\u9f50，\u4ece\u800c\u5177\u5907\u4e86\u96f6\u6837\u672c（zero-shot）\u5206\u7c7b\u80fd\u529b\u2014\u2014\u6a21\u578b\u53ef\u4ee5\u76f4\u63a5\u6839\u636e\u7c7b\u522b\u540d\u79f0\u7684\u6587\u672c\u63cf\u8ff0\u6765\u8bc6\u522b\u672a\u89c1\u8fc7\u7684\u7c7b\u522b，\u65e0\u9700\u4efb\u4f55\u5fae\u8c03\u3002<\/p>\n

CLIP\u7684\u8bad\u7ec3\u6570\u636e\u89c4\u6a21\u4ee4\u4eba\u5370\u8c61\u6df1\u523b：4\u4ebf\u5bf9（\u56fe\u50cf，\u6587\u672c）\u4ece\u4e92\u8054\u7f51\u4e0a\u6536\u96c6\u3002\u5728\u5982\u6b64\u5927\u89c4\u6a21\u7684\u6570\u636e\u4e0a\u8bad\u7ec3\u540e，CLIP\u5c55\u73b0\u51fa\u4e86\u60ca\u4eba\u7684\u6cdb\u5316\u80fd\u529b，\u53ef\u4ee5\u572830\u591a\u4e2a\u4e0d\u540c\u7684\u89c6\u89c9\u6570\u636e\u96c6\u4e0a\u8fdb\u884c\u96f6\u6837\u672c\u5206\u7c7b，\u4e14\u6548\u679c\u4e0e\u76d1\u7763\u8bad\u7ec3\u7684ResNet-50\u76f8\u5f53\u751a\u81f3\u66f4\u597d\u3002<\/p>\n

CLIP\u7684\u5f71\u54cd\u8fdc\u8d85\u5206\u7c7b\u4efb\u52a1\u672c\u8eab\u3002\u5b83\u4e3a\u540e\u7eed\u7684\u6587\u672c\u5230\u56fe\u50cf\u751f\u6210\u6a21\u578b（\u5982DALL-E 2\u3001Stable Diffusion）\u63d0\u4f9b\u4e86\u6838\u5fc3\u7ec4\u4ef6，\u4e5f\u542f\u53d1\u4e86\u5927\u91cf\u89c6\u89c9-\u8bed\u8a00\u6a21\u578b\u7684\u7814\u7a76\u3002CLIP\u8bc1\u660e\u4e86\u8bed\u8a00\u53ef\u4ee5\u6210\u4e3a\u89c6\u89c9\u5b66\u4e60\u7684\u5f3a\u5927\u76d1\u7763\u4fe1\u53f7，\u8fd9\u4e00\u53d1\u73b0\u6df1\u523b\u6539\u53d8\u4e86\u8ba1\u7b97\u673a\u89c6\u89c9\u7684\u7814\u7a76\u8303\u5f0f\u3002<\/p>\n

\u4e94\u3001\u751f\u6210\u5f0fAI\u7206\u53d1：\u4ece\u56fe\u50cf\u5230\u89c6\u9891（2022-2024）<\/h3>\n

5.1 Stable Diffusion(2022)：\u56fe\u50cf\u751f\u6210\u7684\u6c11\u4e3b\u5316<\/h4>\n

2022\u5e74，Stability AI\u53d1\u5e03\u7684Stable Diffusion\u6a21\u578b，\u5c06\u6587\u672c\u5230\u56fe\u50cf\u751f\u6210\u6280\u672f\u5e26\u5165\u4e86\u5927\u4f17\u89c6\u91ce\u3002\u4e0e\u4e4b\u524d\u7684\u6269\u6563\u6a21\u578b（\u5982DDPM\u3001DDIM）\u4e0d\u540c，Stable Diffusion\u5728\u6f5c\u5728\u7a7a\u95f4（latent space）\u800c\u975e\u50cf\u7d20\u7a7a\u95f4\u8fdb\u884c\u6269\u6563\u548c\u53bb\u566a\u8fc7\u7a0b，\u5927\u5e45\u964d\u4f4e\u4e86\u8ba1\u7b97\u5f00\u9500，\u4f7f\u5f97\u5728\u6d88\u8d39\u7ea7GPU\u4e0a\u8fd0\u884c\u56fe\u50cf\u751f\u6210\u6210\u4e3a\u53ef\u80fd\u3002<\/p>\n

Stable Diffusion\u7684\u6280\u672f\u67b6\u6784\u5305\u542b\u4e09\u4e2a\u6838\u5fc3\u7ec4\u4ef6：\u4e00\u4e2a\u53d8\u5206\u81ea\u7f16\u7801\u5668（VAE），\u8d1f\u8d23\u5c06\u56fe\u50cf\u538b\u7f29\u5230\u6f5c\u5728\u7a7a\u95f4\u5e76\u5728\u751f\u6210\u540e\u89e3\u7801\u56de\u50cf\u7d20\u7a7a\u95f4；\u4e00\u4e2aU-Net\u53bb\u566a\u7f51\u7edc，\u5728\u6f5c\u5728\u7a7a\u95f4\u4e2d\u9010\u6b65\u9884\u6d4b\u5e76\u53bb\u9664\u566a\u58f0；\u4e00\u4e2aCLIP\u6587\u672c\u7f16\u7801\u5668，\u5c06\u6587\u672c\u63cf\u8ff0\u7f16\u7801\u4e3a\u6761\u4ef6\u5411\u91cf，\u5f15\u5bfc\u751f\u6210\u8fc7\u7a0b\u3002<\/p>\n

\u6269\u6563\u6a21\u578b\u7684\u57fa\u672c\u539f\u7406\u662f：\u524d\u5411\u8fc7\u7a0b\u9010\u6b65\u5411\u6570\u636e\u6dfb\u52a0\u9ad8\u65af\u566a\u58f0，\u76f4\u5230\u6570\u636e\u5b8c\u5168\u53d8\u6210\u7eaf\u566a\u58f0；\u53cd\u5411\u8fc7\u7a0b\u5b66\u4e60\u4ece\u566a\u58f0\u4e2d\u6062\u590d\u6570\u636e，\u901a\u8fc7\u9010\u6b65\u53bb\u566a\u751f\u6210\u65b0\u7684\u6837\u672c\u3002\u8bad\u7ec3\u65f6，\u6a21\u578b\u5b66\u4e60\u9884\u6d4b\u7ed9\u5b9a\u566a\u58f0\u56fe\u50cf\u548c\u566a\u58f0\u6c34\u5e73\u4e0b\u7684\u566a\u58f0\u5206\u91cf；\u63a8\u7406\u65f6，\u4ece\u968f\u673a\u566a\u58f0\u5f00\u59cb，\u9010\u6b65\u53bb\u566a，\u6700\u7ec8\u5f97\u5230\u751f\u6210\u56fe\u50cf\u3002<\/p>\n

Stable Diffusion\u7684\u5f00\u6e90\u53d1\u5e03\u6781\u5927\u5730\u63a8\u52a8\u4e86AI\u56fe\u50cf\u751f\u6210\u6280\u672f\u7684\u666e\u53ca\u3002\u793e\u533a\u8d21\u732e\u4e86\u6570\u5343\u4e2a\u5fae\u8c03\u6a21\u578b（\u5982LoRA\u3001DreamBooth）\u3001\u63d2\u4ef6\u548c\u5de5\u5177，\u5f62\u6210\u4e86\u7e41\u8363\u7684\u751f\u6001\u7cfb\u7edf\u3002ControlNet\u3001IP-Adapter\u7b49\u540e\u7eed\u5de5\u4f5c\u8fdb\u4e00\u6b65\u589e\u5f3a\u4e86\u56fe\u50cf\u751f\u6210\u7684\u53ef\u63a7\u6027\u548c\u7075\u6d3b\u6027\u3002<\/p>\n

5.2 SAM(2023)：\u5206\u5272\u4e00\u5207\u6a21\u578b<\/h4>\n

Meta\u57282023\u5e74\u53d1\u5e03\u7684SAM（Segment Anything Model），\u662f\u56fe\u50cf\u5206\u5272\u9886\u57df\u7684\u91cc\u7a0b\u7891\u5f0f\u5de5\u4f5c\u3002SAM\u7684\u76ee\u6807\u662f\u6784\u5efa\u4e00\u4e2a\u901a\u7528\u7684\u56fe\u50cf\u5206\u5272\u57fa\u7840\u6a21\u578b，\u80fd\u591f\u5728\u4e0d\u8fdb\u884c\u7279\u5b9a\u4efb\u52a1\u5fae\u8c03\u7684\u60c5\u51b5\u4e0b，\u5bf9\u4efb\u610f\u56fe\u50cf\u4e2d\u7684\u4efb\u610f\u7269\u4f53\u8fdb\u884c\u5206\u5272\u3002<\/p>\n

SAM\u7684\u67b6\u6784\u5305\u542b\u4e09\u4e2a\u90e8\u5206：\u4e00\u4e2a\u56fe\u50cf\u7f16\u7801\u5668（\u57fa\u4e8eViT），\u7528\u4e8e\u751f\u6210\u56fe\u50cf\u7684\u5d4c\u5165\u8868\u793a；\u4e00\u4e2a\u63d0\u793a\u7f16\u7801\u5668，\u7528\u4e8e\u7f16\u7801\u7528\u6237\u63d0\u4f9b\u7684\u63d0\u793a（\u70b9\u3001\u6846\u3001\u63a9\u7801\u6216\u6587\u672c）；\u4e00\u4e2a\u63a9\u7801\u89e3\u7801\u5668，\u5c06\u56fe\u50cf\u5d4c\u5165\u548c\u63d0\u793a\u5d4c\u5165\u7ed3\u5408，\u8f93\u51fa\u5206\u5272\u63a9\u7801\u3002\u8fd9\u79cd\u8bbe\u8ba1\u4f7f\u5f97SAM\u80fd\u591f\u7075\u6d3b\u5730\u54cd\u5e94\u5404\u79cd\u5f62\u5f0f\u7684\u63d0\u793a，\u4ece\u5355\u70b9\u70b9\u51fb\u5230\u5bc6\u96c6\u7684\u5168\u56fe\u5206\u5272\u3002<\/p>\n

SAM\u7684\u8bad\u7ec3\u6570\u636e\u89c4\u6a21\u4ee4\u4eba\u77a9\u76ee：\u901a\u8fc7"\u6570\u636e\u5f15\u64ce"\u7684\u4e09\u9636\u6bb5\u8fed\u4ee3\u6d41\u7a0b（\u8f85\u52a9\u4eba\u5de5\u6807\u6ce8\u3001\u534a\u81ea\u52a8\u6807\u6ce8\u3001\u5168\u81ea\u52a8\u6807\u6ce8），\u6700\u7ec8\u6784\u5efa\u4e86\u5305\u542b1100\u4e07\u5f20\u56fe\u50cf\u548c11\u4ebf\u4e2a\u63a9\u7801\u7684SA-1B\u6570\u636e\u96c6\u3002\u8fd9\u662f\u8fc4\u4eca\u4e3a\u6b62\u6700\u5927\u7684\u56fe\u50cf\u5206\u5272\u6570\u636e\u96c6，\u6bd4\u4e4b\u524d\u6700\u5927\u7684\u6570\u636e\u96c6\u5927\u4e86\u4e24\u4e2a\u6570\u91cf\u7ea7\u3002<\/p>\n

SAM\u7684\u51fa\u73b0\u6807\u5fd7\u7740\u8ba1\u7b97\u673a\u89c6\u89c9\u5411"\u57fa\u7840\u6a21\u578b"\u8303\u5f0f\u8fc8\u51fa\u4e86\u91cd\u8981\u4e00\u6b65\u3002\u5b83\u8bc1\u660e\u4e86\u901a\u8fc7\u5927\u89c4\u6a21\u6570\u636e\u548c\u6a21\u578b\u5bb9\u91cf，\u53ef\u4ee5\u8bad\u7ec3\u51fa\u5177\u6709\u5f3a\u5927\u6cdb\u5316\u80fd\u529b\u7684\u5206\u5272\u6a21\u578b，\u8fd9\u4e3a\u4e0b\u6e38\u4efb\u52a1（\u5982\u533b\u5b66\u56fe\u50cf\u5206\u5272\u3001\u9065\u611f\u56fe\u50cf\u5206\u6790\u3001\u89c6\u9891\u7f16\u8f91\u7b49）\u63d0\u4f9b\u4e86\u5f3a\u5927\u7684\u57fa\u7840\u5de5\u5177\u3002<\/p>\n

5.3 GPT-4V(2023)：\u591a\u6a21\u6001\u7406\u89e3\u7684\u96c6\u5927\u6210\u8005<\/h4>\n

2023\u5e749\u6708，OpenAI\u53d1\u5e03\u4e86GPT-4V（GPT-4 with Vision），\u5c06GPT-4\u7684\u8bed\u8a00\u7406\u89e3\u80fd\u529b\u6269\u5c55\u5230\u4e86\u89c6\u89c9\u9886\u57df\u3002GPT-4V\u80fd\u591f\u7406\u89e3\u56fe\u50cf\u5185\u5bb9，\u56de\u7b54\u5173\u4e8e\u56fe\u50cf\u7684\u95ee\u9898，\u8fdb\u884c\u89c6\u89c9\u63a8\u7406，\u751a\u81f3\u6839\u636e\u56fe\u50cf\u8fdb\u884c\u521b\u610f\u5199\u4f5c\u548c\u5206\u6790\u3002<\/p>\n

GPT-4V\u7684\u591a\u6a21\u6001\u80fd\u529b\u5e76\u975e\u7b80\u5355\u7684\u56fe\u6587\u5339\u914d，\u800c\u662f\u5c55\u73b0\u4e86\u6df1\u5c42\u6b21\u7684\u89c6\u89c9\u7406\u89e3\u3002\u5b83\u80fd\u591f\u89e3\u8bfb\u56fe\u8868\u548c\u6570\u636e\u53ef\u89c6\u5316，\u7406\u89e3\u6f2b\u753b\u548c\u8868\u60c5\u5305\u4e2d\u7684\u5e7d\u9ed8，\u8bc6\u522b\u624b\u5199\u6587\u5b57，\u5206\u6790\u533b\u5b66\u5f71\u50cf，\u751a\u81f3\u80fd\u591f\u7406\u89e3\u590d\u6742\u7684\u591a\u6b65\u9aa4\u89c6\u89c9\u63a8\u7406\u4efb\u52a1\u3002\u8fd9\u79cd\u80fd\u529b\u4f7f\u5b83\u6210\u4e3a\u591a\u6a21\u6001\u5927\u8bed\u8a00\u6a21\u578b\u7684\u6807\u6746\u3002<\/p>\n

GPT-4V\u7684\u53d1\u5e03\u6807\u5fd7\u7740\u8ba1\u7b97\u673a\u89c6\u89c9\u8fdb\u5165\u4e86\u4e00\u4e2a\u65b0\u9636\u6bb5：\u4ece"\u8bc6\u522b"\u5230"\u7406\u89e3"\u3002\u4f20\u7edf\u7684\u89c6\u89c9\u6a21\u578b\u56de\u7b54\u7684\u662f"\u8fd9\u5f20\u56fe\u7247\u91cc\u6709\u4ec0\u4e48"\u7684\u95ee\u9898，\u800cGPT-4V\u80fd\u591f\u56de\u7b54"\u8fd9\u5f20\u56fe\u7247\u610f\u5473\u7740\u4ec0\u4e48"\u3001"\u6839\u636e\u8fd9\u5f20\u56fe\u7247\u5e94\u8be5\u600e\u4e48\u505a"\u7b49\u66f4\u590d\u6742\u7684\u95ee\u9898\u3002\u8fd9\u79cd\u4ece\u611f\u77e5\u5230\u8ba4\u77e5\u7684\u8de8\u8d8a，\u4e3a\u8ba1\u7b97\u673a\u89c6\u89c9\u7684\u5e94\u7528\u5f00\u8f9f\u4e86\u5168\u65b0\u7684\u53ef\u80fd\u6027\u3002<\/p>\n

5.4 Sora(2024)：\u89c6\u9891\u751f\u6210\u7684\u65b0\u7eaa\u5143<\/h4>\n

2024\u5e742\u6708，OpenAI\u53d1\u5e03\u4e86Sora，\u4e00\u4e2a\u80fd\u591f\u6839\u636e\u6587\u672c\u63cf\u8ff0\u751f\u6210\u957f\u8fbe\u4e00\u5206\u949f\u9ad8\u8d28\u91cf\u89c6\u9891\u7684\u6a21\u578b\u3002Sora\u7684\u51fa\u73b0\u9707\u60ca\u4e86\u5168\u4e16\u754c，\u56e0\u4e3a\u5b83\u751f\u6210\u7684\u89c6\u9891\u4e0d\u4ec5\u5728\u89c6\u89c9\u8d28\u91cf\u4e0a\u8fbe\u5230\u4e86\u524d\u6240\u672a\u6709\u7684\u6c34\u5e73，\u66f4\u91cd\u8981\u7684\u662f\u5c55\u73b0\u4e86\u6df1\u5c42\u7684\u7269\u7406\u4e16\u754c\u7406\u89e3\u80fd\u529b\u2014\u2014\u7269\u4f53\u5728\u8fd0\u52a8\u3001\u906e\u6321\u3001\u5149\u7167\u53d8\u5316\u4e2d\u7684\u8868\u73b0\u9ad8\u5ea6\u7b26\u5408\u7269\u7406\u89c4\u5f8b\u3002<\/p>\n

Sora\u7684\u6280\u672f\u8def\u7ebf\u57fa\u4e8e\u6269\u6563\u53d8\u6362\u5668（Diffusion Transformer，DiT），\u5c06\u6269\u6563\u6a21\u578b\u4e0eTransformer\u67b6\u6784\u7ed3\u5408\u3002\u4e0e\u4f20\u7edf\u7684\u89c6\u9891\u751f\u6210\u6a21\u578b\u4e0d\u540c，Sora\u4e0d\u5c06\u89c6\u9891\u89c6\u4e3a\u56fa\u5b9a\u5e27\u7387\u7684\u65f6\u95f4\u5e8f\u5217，\u800c\u662f\u4f7f\u7528"\u65f6\u7a7apatch"\u7684\u8868\u793a\u65b9\u5f0f\u2014\u2014\u5c06\u89c6\u9891\u5728\u65f6\u95f4\u548c\u7a7a\u95f4\u7ef4\u5ea6\u4e0a\u540c\u65f6\u5206\u5272\u6210\u5c0f\u5757，\u7c7b\u4f3c\u4e8eViT\u5c06\u56fe\u50cf\u5206\u5272\u6210\u7a7a\u95f4patch\u3002\u8fd9\u79cd\u7edf\u4e00\u7684\u8868\u793a\u65b9\u5f0f\u4f7f\u5f97Sora\u80fd\u591f\u5904\u7406\u4efb\u610f\u5206\u8fa8\u7387\u3001\u4efb\u610f\u957f\u5bbd\u6bd4\u548c\u4efb\u610f\u65f6\u957f\u7684\u89c6\u9891\u3002<\/p>\n

Sora\u7684\u53e6\u4e00\u4e2a\u91cd\u8981\u7279\u6027\u662f\u5b83\u7684\u6d8c\u73b0\u80fd\u529b\u3002\u901a\u8fc7\u5927\u89c4\u6a21\u8bad\u7ec3，Sora\u5c55\u73b0\u51fa\u4e86\u5bf9\u4e09\u7ef4\u7a7a\u95f4\u3001\u7269\u4f53\u6301\u4e45\u6027\u3001\u957f\u7a0b\u65f6\u95f4\u4e00\u81f4\u6027\u7b49\u7269\u7406\u4e16\u754c\u5c5e\u6027\u7684\u7406\u89e3，\u8fd9\u4e9b\u80fd\u529b\u5e76\u975e\u88ab\u663e\u5f0f\u7f16\u7a0b，\u800c\u662f\u4ece\u6570\u636e\u4e2d\u81ea\u53d1\u6d8c\u73b0\u51fa\u6765\u7684\u3002\u8fd9\u6697\u793a\u4e86\u5927\u6a21\u578b\u5728\u89c6\u89c9\u9886\u57df\u53ef\u80fd\u5177\u6709\u4e0e\u8bed\u8a00\u9886\u57df\u7c7b\u4f3c\u7684\u6d8c\u73b0\u80fd\u529b\u3002<\/p>\n

Sora\u76ee\u524d\u5c1a\u672a\u516c\u5f00\u53d1\u5e03，\u4f46\u5176\u6280\u672f\u62a5\u544a\u548c\u6f14\u793a\u5df2\u7ecf\u5f15\u53d1\u4e86\u5173\u4e8e\u89c6\u9891\u751f\u6210\u3001\u4e16\u754c\u6a21\u578b\u548c\u901a\u7528\u4eba\u5de5\u667a\u80fd\u7684\u5e7f\u6cdb\u8ba8\u8bba\u3002\u5b83\u4ee3\u8868\u4e86\u8ba1\u7b97\u673a\u89c6\u89c9\u57282024\u5e74\u6240\u8fbe\u5230\u7684\u6700\u9ad8\u6c34\u5e73\u2014\u2014\u4ece\u8bc6\u522b\u9759\u6001\u56fe\u50cf，\u5230\u7406\u89e3\u56fe\u50cf\u5185\u5bb9，\u518d\u5230\u751f\u6210\u52a8\u6001\u7684\u4e16\u754c\u6a21\u62df\u3002<\/p>\n

\u516d\u3001\u5173\u952e\u57fa\u7840\u8bbe\u65bd：\u6570\u636e\u96c6\u3001\u57fa\u51c6\u4e0e\u8bc4\u4f30\u4f53\u7cfb\u7684\u6f14\u8fdb<\/h3>\n

6.1 \u6570\u636e\u96c6：\u89c6\u89c9AI\u7684\u71c3\u6599<\/h4>\n

\u8ba1\u7b97\u673a\u89c6\u89c9\u7684\u53d1\u5c55\u53f2，\u5728\u5f88\u5927\u7a0b\u5ea6\u4e0a\u4e5f\u662f\u4e00\u90e8\u6570\u636e\u96c6\u7684\u6f14\u8fdb\u53f2\u3002\u6ca1\u6709\u9ad8\u8d28\u91cf\u7684\u5927\u89c4\u6a21\u6807\u6ce8\u6570\u636e，\u518d\u7cbe\u5999\u7684\u7b97\u6cd5\u4e5f\u53ea\u662f\u7a7a\u4e2d\u697c\u9601\u3002\u4eceMNIST\u5230ImageNet\u518d\u5230SA-1B，\u6bcf\u4e00\u4e2a\u91cc\u7a0b\u7891\u5f0f\u6570\u636e\u96c6\u7684\u8bde\u751f，\u90fd\u63a8\u52a8\u4e86\u7b97\u6cd5\u80fd\u529b\u7684\u8d28\u53d8\u3002<\/p>\n

MNIST（1998\u5e74）\u662f\u8ba1\u7b97\u673a\u89c6\u89c9\u9886\u57df\u6700\u8457\u540d\u7684\u6570\u636e\u96c6\u4e4b\u4e00，\u5305\u542b60000\u5f20\u8bad\u7ec3\u56fe\u50cf\u548c10000\u5f20\u6d4b\u8bd5\u56fe\u50cf\u7684\u624b\u5199\u6570\u5b57（0-9）\u3002\u867d\u7136\u4ee5\u4eca\u5929\u7684\u6807\u51c6\u6765\u770b，MNIST\u7684\u89c4\u6a21\u5f88\u5c0f\u3001\u4efb\u52a1\u5f88\u7b80\u5355，\u4f46\u5b83\u5728CNN\u53d1\u5c55\u7684\u65e9\u671f\u9636\u6bb5\u63d0\u4f9b\u4e86\u4e0d\u53ef\u66ff\u4ee3\u7684\u57fa\u51c6\u6d4b\u8bd5\u5e73\u53f0\u3002\u76f4\u5230\u4eca\u5929，MNIST\u4ecd\u7136\u662f\u521d\u5b66\u8005\u5b66\u4e60\u6df1\u5ea6\u5b66\u4e60\u7684\u7b2c\u4e00\u4e2a\u6570\u636e\u96c6，LeNet-5\u5728MNIST\u4e0a\u7684\u6210\u529f\u4e3aCNN\u7684\u540e\u7eed\u53d1\u5c55\u5960\u5b9a\u4e86\u4fe1\u5fc3\u57fa\u7840\u3002<\/p>\n

CIFAR-10\u548cCIFAR-100（2009\u5e74）\u7531Alex Krizhevsky\u548cGeoffrey Hinton\u6574\u7406\u53d1\u5e03，\u5c06\u56fe\u50cf\u5206\u7c7b\u4ece\u7070\u5ea6\u624b\u5199\u6570\u5b57\u63a8\u8fdb\u5230\u4e86\u5f69\u8272\u81ea\u7136\u56fe\u50cf\u3002CIFAR-10\u5305\u542b10\u4e2a\u7c7b\u522b，\u6bcf\u4e2a\u7c7b\u522b6000\u5f2032×32\u7684\u5c0f\u56fe\u50cf，\u603b\u8ba160000\u5f20\u3002CIFAR-100\u5305\u542b100\u4e2a\u7c7b\u522b，\u6bcf\u4e2a\u7c7b\u522b600\u5f20\u56fe\u50cf\u3002\u867d\u7136\u56fe\u50cf\u5206\u8fa8\u7387\u4f4e，\u4f46\u7c7b\u522b\u591a\u6837\u6027\u548c\u56fe\u50cf\u590d\u6742\u6027\u8fdc\u9ad8\u4e8eMNIST，\u6210\u4e3a\u9a8c\u8bc1\u65b0\u578bCNN\u67b6\u6784\u7684\u6807\u51c6\u6d4b\u8bd5\u96c6\u3002<\/p>\n

ImageNet（2009\u5e74\u53d1\u5e03，2010\u5e74\u8d77\u4e3e\u529eILSVRC\u7ade\u8d5b）\u662f\u8ba1\u7b97\u673a\u89c6\u89c9\u5386\u53f2\u4e0a\u6700\u5177\u5f71\u54cd\u529b\u7684\u6570\u636e\u96c6\u3002\u5b83\u7531\u65af\u5766\u798f\u5927\u5b66\u7684\u674e\u98de\u98de\u56e2\u961f\u53d1\u8d77，\u6700\u521d\u5305\u542b\u8d85\u8fc71400\u4e07\u5f20\u56fe\u50cf，\u8986\u76d621841\u4e2a\u7c7b\u522b，\u6784\u5efa\u4e86\u4e00\u4e2a\u6309\u7167WordNet\u8bed\u4e49\u5c42\u6b21\u7ed3\u6784\u7ec4\u7ec7\u7684\u6d77\u91cf\u56fe\u50cf\u6570\u636e\u5e93\u3002ILSVRC（ImageNet\u5927\u89c4\u6a21\u89c6\u89c9\u8bc6\u522b\u6311\u6218\u8d5b）\u4f7f\u7528\u7684\u5b50\u96c6\u5305\u542b1000\u4e2a\u7c7b\u522b，\u7ea6120\u4e07\u5f20\u8bad\u7ec3\u56fe\u50cf\u30015\u4e07\u5f20\u9a8c\u8bc1\u56fe\u50cf\u548c10\u4e07\u5f20\u6d4b\u8bd5\u56fe\u50cf\u3002ImageNet\u7684\u89c4\u6a21\u5728\u5f53\u65f6\u662f\u524d\u6240\u672a\u6709\u7684\u2014\u2014\u6b64\u524d\u6700\u5927\u7684\u56fe\u50cf\u5206\u7c7b\u6570\u636e\u96c6\u4e5f\u4ec5\u6709\u6570\u4e07\u5f20\u56fe\u50cf\u3002\u6b63\u662f\u8fd9\u79cd\u5927\u89c4\u6a21\u6570\u636e，\u4e3aAlexNet\u7b49\u6df1\u5ea6\u6a21\u578b\u63d0\u4f9b\u4e86\u5145\u8db3\u7684\u8bad\u7ec3\u4fe1\u53f7，\u4f7f\u7aef\u5230\u7aef\u6df1\u5ea6\u5b66\u4e60\u6210\u4e3a\u53ef\u80fd\u3002<\/p>\n

COCO（Common Objects in Context，2014\u5e74）\u7531\u5fae\u8f6f\u53d1\u5e03，\u5c06\u8ba1\u7b97\u673a\u89c6\u89c9\u7684\u8bc4\u4f30\u4ece\u7b80\u5355\u7684\u5206\u7c7b\u6269\u5c55\u5230\u4e86\u66f4\u590d\u6742\u7684\u573a\u666f\u7406\u89e3\u3002COCO\u5305\u542b\u8d85\u8fc733\u4e07\u5f20\u56fe\u50cf，\u6807\u6ce8\u4e86\u8d85\u8fc7250\u4e07\u4e2a\u76ee\u6807\u5b9e\u4f8b，\u6db5\u76d680\u4e2a\u7269\u4f53\u7c7b\u522b\u3002COCO\u7684\u72ec\u7279\u4e4b\u5904\u5728\u4e8e\u5b83\u5f3a\u8c03"\u4e0a\u4e0b\u6587\u4e2d\u7684\u7269\u4f53"\u2014\u2014\u56fe\u50cf\u4e2d\u7684\u7269\u4f53\u901a\u5e38\u51fa\u73b0\u5728\u81ea\u7136\u573a\u666f\u4e2d，\u800c\u975e\u7ecf\u8fc7\u88c1\u526a\u548c\u5c45\u4e2d\u5904\u7406\u7684\u5b64\u7acb\u7269\u4f53\u3002\u8fd9\u4f7f\u5f97COCO\u6210\u4e3a\u76ee\u6807\u68c0\u6d4b\u3001\u5b9e\u4f8b\u5206\u5272\u548c\u56fe\u50cf\u63cf\u8ff0\u751f\u6210\u7684\u6807\u51c6\u57fa\u51c6\u3002COCO\u7684\u8bc4\u4f30\u6307\u6807（mAP@[0.5:0.95]）\u6bd4ImageNet\u7684top-5\u51c6\u786e\u7387\u66f4\u52a0\u4e25\u683c，\u63a8\u52a8\u4e86\u68c0\u6d4b\u548c\u5206\u5272\u7b97\u6cd5\u7684\u6301\u7eed\u6539\u8fdb\u3002<\/p>\n

SA-1B（Segment Anything 1 Billion，2023\u5e74）\u662fMeta\u5728\u53d1\u5e03SAM\u6a21\u578b\u65f6\u540c\u65f6\u516c\u5f00\u7684\u6570\u636e\u96c6，\u5305\u542b1100\u4e07\u5f20\u56fe\u50cf\u548c\u8d85\u8fc711\u4ebf\u4e2a\u5206\u5272\u63a9\u7801\u3002\u8fd9\u662f\u8fc4\u4eca\u4e3a\u6b62\u6700\u5927\u7684\u56fe\u50cf\u5206\u5272\u6570\u636e\u96c6，\u5176\u89c4\u6a21\u6bd4\u4e4b\u524d\u6700\u5927\u7684\u5206\u5272\u6570\u636e\u96c6\u5927\u4e86\u4e24\u4e2a\u6570\u91cf\u7ea7\u3002SA-1B\u7684\u6784\u5efa\u91c7\u7528\u4e86\u521b\u65b0\u7684"\u6570\u636e\u5f15\u64ce"\u4e09\u9636\u6bb5\u6d41\u7a0b\u2014\u2014\u4ece\u4eba\u5de5\u6807\u6ce8\u5230\u534a\u81ea\u52a8\u6807\u6ce8\u518d\u5230\u5168\u81ea\u52a8\u6807\u6ce8\u2014\u2014\u5c55\u793a\u4e86\u5982\u4f55\u5229\u7528\u6a21\u578b\u81ea\u8eab\u7684\u80fd\u529b\u6765\u9ad8\u6548\u6784\u5efa\u5927\u89c4\u6a21\u6570\u636e\u96c6\u3002<\/p>\n

6.2 \u8bc4\u4f30\u57fa\u51c6：\u4ece\u5355\u4e00\u6307\u6807\u5230\u5168\u9762\u8861\u91cf<\/h4>\n

\u8bc4\u4f30\u57fa\u51c6\u7684\u6f14\u53d8\u53cd\u6620\u4e86\u9886\u57df\u5bf9\u6a21\u578b\u80fd\u529b\u671f\u671b\u7684\u4e0d\u65ad\u63d0\u5347\u3002\u65e9\u671f\u7684\u8bc4\u4f30\u4f9d\u8d56\u7b80\u5355\u7684\u51c6\u786e\u7387\u6307\u6807，\u5982\u4eca\u7684\u8bc4\u4f30\u4f53\u7cfb\u5219\u6db5\u76d6\u4e86\u7cbe\u5ea6\u3001\u9c81\u68d2\u6027\u3001\u516c\u5e73\u6027\u3001\u6548\u7387\u7b49\u591a\u4e2a\u7ef4\u5ea6\u3002<\/p>\n

ImageNet\u5206\u7c7b\u4ece\u6700\u521d\u7684top-5\u9519\u8bef\u7387\u8bc4\u4f30，\u53d1\u5c55\u5230\u5305\u62ectop-1\u51c6\u786e\u7387\u3001\u7c7b\u522b\u5e73\u5747\u51c6\u786e\u7387\u7b49\u591a\u7ef4\u6307\u6807\u30022019\u5e74，ImageNet\u8fd8\u5f15\u5165\u4e86ImageNet-A（\u81ea\u7136\u5bf9\u6297\u6837\u672c）\u3001ImageNet-R（\u827a\u672f\u6e32\u67d3\u56fe\u50cf）\u3001ImageNet-Sketch（\u7d20\u63cf\u56fe\u50cf）\u7b49\u53d8\u4f53，\u7528\u4e8e\u8bc4\u4f30\u6a21\u578b\u5728\u5206\u5e03\u5916\u6570\u636e\u4e0a\u7684\u9c81\u68d2\u6027\u3002\u8fd9\u4e9b\u53d8\u4f53\u63ed\u793a\u4e86\u6807\u51c6ImageNet\u8bad\u7ec3\u7684\u6a21\u578b\u5728\u975e\u81ea\u7136\u573a\u666f\u4e0b\u7684\u6027\u80fd\u9000\u5316，\u63a8\u52a8\u4e86\u9c81\u68d2\u6027\u7814\u7a76\u7684\u8fdb\u5c55\u3002<\/p>\n

COCO\u7684\u8bc4\u4f30\u4f53\u7cfb\u662f\u76ee\u6807\u68c0\u6d4b\u548c\u5206\u5272\u9886\u57df\u6700\u5168\u9762\u7684\u57fa\u51c6\u4e4b\u4e00\u3002\u9664\u4e86mAP（\u5e73\u5747\u7cbe\u5ea6\u5747\u503c），COCO\u8fd8\u62a5\u544aAP50\u3001AP75（\u4e0d\u540cIoU\u9608\u503c\u4e0b\u7684\u7cbe\u5ea6）\u3001AP_small\u3001AP_medium\u3001AP_large（\u4e0d\u540c\u5c3a\u5bf8\u76ee\u6807\u7684\u7cbe\u5ea6）\u4ee5\u53caAR（\u5e73\u5747\u53ec\u56de\u7387）\u3002\u8fd9\u79cd\u591a\u7ef4\u5ea6\u8bc4\u4f30\u53ef\u4ee5\u5168\u9762\u53cd\u6620\u68c0\u6d4b\u5668\u5728\u4e0d\u540c\u573a\u666f\u4e0b\u7684\u8868\u73b0，\u907f\u514d\u4e86\u5355\u4e00\u6307\u6807\u53ef\u80fd\u5e26\u6765\u7684\u8bef\u5bfc\u3002<\/p>\n

6.3 \u7ade\u8d5b\u5e73\u53f0：\u63a8\u52a8\u8fdb\u6b65\u7684\u50ac\u5316\u5242<\/h4>\n

ILSVRC（2010-2017）\u662f\u6700\u5177\u4f20\u5947\u8272\u5f69\u7684\u8ba1\u7b97\u673a\u89c6\u89c9\u7ade\u8d5b\u30022012\u5e74AlexNet\u7684\u593a\u51a0\u5ba3\u544a\u4e86\u6df1\u5ea6\u5b66\u4e60\u65f6\u4ee3\u7684\u5230\u6765，2015\u5e74ResNet\u7684\u8d85\u8d8a\u4eba\u7c7b\u8868\u73b0\u6807\u5fd7\u7740\u6df1\u5ea6\u5b66\u4e60\u5728\u7279\u5b9a\u4efb\u52a1\u4e0a\u7684\u6210\u719f\u3002ILSVRC\u4e8e2017\u5e74\u6b63\u5f0f\u7ed3\u675f，\u4f46\u5176\u5f71\u54cd\u529b\u5ef6\u7eed\u81f3\u4eca\u2014\u2014\u51e0\u4e4e\u6240\u6709\u7684\u7ecf\u5178CNN\u67b6\u6784\u90fd\u5728\u8fd9\u4e2a\u821e\u53f0\u4e0a\u9996\u6b21\u4eae\u76f8\u3002<\/p>\n

Kaggle\u548cPapers With Code\u7b49\u5e73\u53f0\u5728ILSVRC\u4e4b\u540e\u63a5\u8fc7\u4e86\u63a8\u52a8\u7b97\u6cd5\u8fdb\u6b65\u7684\u5927\u65d7\u3002Kaggle\u63d0\u4f9b\u4e86\u4e30\u5bcc\u7684\u6570\u636e\u79d1\u5b66\u7ade\u8d5b，\u6db5\u76d6\u56fe\u50cf\u5206\u7c7b\u3001\u5206\u5272\u3001\u751f\u6210\u7b49\u591a\u79cdCV\u4efb\u52a1，\u5438\u5f15\u4e86\u5168\u7403\u6570\u5341\u4e07\u6570\u636e\u79d1\u5b66\u5bb6\u53c2\u4e0e\u3002Papers With Code\u5219\u901a\u8fc7\u5c06\u8bba\u6587\u3001\u4ee3\u7801\u548c\u6392\u884c\u699c\u5173\u8054\u8d77\u6765，\u5efa\u7acb\u4e86\u4e00\u4e2a\u900f\u660e\u3001\u53ef\u590d\u73b0\u7684\u7814\u7a76\u8bc4\u4f30\u4f53\u7cfb，\u4efb\u4f55\u7814\u7a76\u8005\u90fd\u53ef\u4ee5\u63d0\u4ea4\u81ea\u5df1\u7684\u7ed3\u679c\u4e0e\u5b66\u754c\u6700\u4f73\u6c34\u5e73\u8fdb\u884c\u5bf9\u6bd4\u3002<\/p>\n

\u4e03\u3001\u4e09\u5341\u5e74\u7684\u542f\u793a\u4e0e\u672a\u6765\u5c55\u671b<\/h3>\n

\u56de\u987e\u8ba1\u7b97\u673a\u89c6\u89c9\u4e09\u5341\u5e74\u7684\u53d1\u5c55\u5386\u7a0b，\u6211\u4eec\u53ef\u4ee5\u603b\u7ed3\u51fa\u51e0\u6761\u6838\u5fc3\u89c4\u5f8b\u3002\u7b2c\u4e00，\u6570\u636e\u89c4\u6a21\u662f\u63a8\u52a8\u8fdb\u6b65\u7684\u5173\u952e\u529b\u91cf\u3002\u4eceMNIST\u76846\u4e07\u5f20\u624b\u5199\u6570\u5b57，\u5230ImageNet\u76841400\u4e07\u5f20\u81ea\u7136\u56fe\u50cf，\u518d\u5230SA-1B\u768411\u4ebf\u4e2a\u5206\u5272\u63a9\u7801，\u6bcf\u4e00\u6b21\u6570\u636e\u89c4\u6a21\u7684\u8dc3\u5347\u90fd\u5e26\u6765\u4e86\u80fd\u529b\u7684\u8d28\u53d8\u3002\u7b2c\u4e8c，\u67b6\u6784\u521b\u65b0\u4e0e\u8ba1\u7b97\u80fd\u529b\u7684\u63d0\u5347\u76f8\u8f85\u76f8\u6210\u3002CNN\u7684\u590d\u5174\u5f97\u76ca\u4e8eGPU\u5e76\u884c\u8ba1\u7b97，Transformer\u7684\u5d1b\u8d77\u79bb\u4e0d\u5f00\u5927\u89c4\u6a21\u5206\u5e03\u5f0f\u8bad\u7ec3\u57fa\u7840\u8bbe\u65bd\u3002\u7b2c\u4e09，\u4ece\u4e13\u7528\u5230\u901a\u7528\u7684\u8d8b\u52bf\u4e0d\u53ef\u9006\u8f6c\u3002\u65e9\u671f\u7684CNN\u53ea\u80fd\u505a\u5206\u7c7b，\u540e\u6765\u7684\u6a21\u578b\u53ef\u4ee5\u540c\u65f6\u505a\u68c0\u6d4b\u548c\u5206\u5272，\u5982\u4eca\u7684GPT-4V\u548cSora\u5219\u5411\u7740\u901a\u7528\u89c6\u89c9\u7406\u89e3\u8fc8\u8fdb\u3002\u7b2c\u56db，\u5f00\u6e90\u6587\u5316\u5bf9CV\u9886\u57df\u7684\u63a8\u52a8\u4f5c\u7528\u4e0d\u53ef\u4f30\u91cf\u3002PyTorch\u7684\u751f\u6001\u3001Hugging Face\u7684\u6a21\u578b\u5e93\u3001GitHub\u4e0a\u7684\u5f00\u6e90\u4ee3\u7801，\u4f7f\u5f97\u5168\u4e16\u754c\u7684\u7814\u7a76\u8005\u53ef\u4ee5\u5728\u524d\u4eba\u5de5\u4f5c\u7684\u57fa\u7840\u4e0a\u5feb\u901f\u8fed\u4ee3，\u8fd9\u79cd\u5f00\u653e\u534f\u4f5c\u7684\u6a21\u5f0f\u662fCV\u9886\u57df\u521b\u65b0\u901f\u5ea6\u8fdc\u8d85\u5176\u4ed6\u9886\u57df\u7684\u91cd\u8981\u539f\u56e0\u3002<\/p>\n

\u5c55\u671b\u672a\u6765，\u8ba1\u7b97\u673a\u89c6\u89c9\u7684\u53d1\u5c55\u65b9\u5411\u5305\u62ec\u4f46\u4e0d\u9650\u4e8e：\u66f4\u9ad8\u6548\u7684\u89c6\u89c9\u8868\u793a，\u63a2\u7d22\u5982\u4f55\u5728\u4fdd\u6301\u6027\u80fd\u7684\u540c\u65f6\u51cf\u5c11\u8ba1\u7b97\u5f00\u9500，\u4f7f\u89c6\u89c9\u6a21\u578b\u80fd\u591f\u5728\u79fb\u52a8\u7aef\u548c\u5d4c\u5165\u5f0f\u8bbe\u5907\u4e0a\u9ad8\u6548\u8fd0\u884c；\u66f4\u5f3a\u5927\u7684\u89c6\u89c9-\u8bed\u8a00\u6574\u5408，\u7814\u7a76\u5982\u4f55\u8ba9\u89c6\u89c9\u7406\u89e3\u4e0e\u8bed\u8a00\u63a8\u7406\u66f4\u597d\u5730\u534f\u540c，\u5b9e\u73b0\u771f\u6b63\u7684\u591a\u6a21\u6001\u667a\u80fd；\u4e16\u754c\u6a21\u578b\u7684\u6784\u5efa，\u4ece\u50cf\u7d20\u7ea7\u7406\u89e3\u8fc8\u5411\u5bf9\u7269\u7406\u89c4\u5f8b\u548c\u56e0\u679c\u5173\u7cfb\u7684\u5efa\u6a21，\u8ba9\u673a\u5668\u4e0d\u4ec5"\u770b\u5230"\u8868\u9762\u73b0\u8c61，\u8fd8\u80fd"\u7406\u89e3"\u6df1\u5c42\u539f\u56e0；\u89c6\u89c9AI\u7684\u5b89\u5168\u4e0e\u4f26\u7406，\u5728\u751f\u6210\u5f0fAI（Deepfake\u3001AI\u6362\u8138）\u65e5\u76ca\u903c\u771f\u7684\u80cc\u666f\u4e0b，\u5982\u4f55\u786e\u4fdd\u6280\u672f\u4e0d\u88ab\u6ee5\u7528，\u5982\u4f55\u6784\u5efa\u516c\u5e73\u3001\u900f\u660e\u3001\u53ef\u89e3\u91ca\u7684\u89c6\u89c9\u7cfb\u7edf，\u8fd9\u4e9b\u95ee\u9898\u53d8\u5f97\u6bd4\u4ee5\u5f80\u4efb\u4f55\u65f6\u5019\u90fd\u66f4\u52a0\u7d27\u8feb\u3002<\/p>\n

\u4e09\u5341\u5e74\u7684\u6280\u672f\u79ef\u7d2f，\u5df2\u7ecf\u8ba9\u8ba1\u7b97\u673a\u89c6\u89c9\u4ece\u5b9e\u9a8c\u5ba4\u8d70\u5411\u4e86\u5343\u5bb6\u4e07\u6237\u3002\u4ece\u6211\u4eec\u6bcf\u5929\u4f7f\u7528\u7684\u9762\u90e8\u89e3\u9501\u548c\u4eba\u50cf\u6a21\u5f0f，\u5230\u81ea\u52a8\u9a7e\u9a76\u6c7d\u8f66\u611f\u77e5\u5468\u56f4\u73af\u5883，\u518d\u5230AI\u8f85\u52a9\u533b\u7597\u8bca\u65ad，\u8ba1\u7b97\u673a\u89c6\u89c9\u6b63\u5728\u6df1\u523b\u5730\u6539\u53d8\u7740\u4eba\u7c7b\u7684\u751f\u6d3b\u3002\u8fd9\u573a\u8de8\u8d8a\u4e09\u5341\u5e74\u7684\u89c6\u89c9\u9769\u547d，\u4eceLeNet\u624b\u5199\u6570\u5b57\u8bc6\u522b\u7684\u6d93\u6d93\u7ec6\u6d41，\u6c47\u805a\u6210GPT-4V\u548cSora\u7684\u5954\u817e\u6c5f\u6cb3，\u672a\u6765\u8fd8\u5c06\u7ee7\u7eed\u5411\u7740\u901a\u7528\u89c6\u89c9\u667a\u80fd\u7684\u661f\u8fb0\u5927\u6d77\u5954\u6d8c\u800c\u53bb\u3002\u800c\u8fd9，\u624d\u521a\u521a\u5f00\u59cb\u3002<\/p>\n

\n

\u53c2\u8003\u6587\u732e<\/p>\n

  • LeCun, Y., et al. (1989). Backpropagation Applied to Handwritten Zip Code Recognition. Neural Computation.<\/li>\n
  • LeCun, Y., et al. (1998). Gradient-Based Learning Applied to Document Recognition. Proceedings of the IEEE.<\/li>\n
  • Krizhevsky, A., Sutskever, I., & Hinton, G. (2012). ImageNet Classification with Deep Convolutional Neural Networks. NeurIPS.<\/li>\n
  • Simonyan, K., & Zisserman, A. (2014). Very Deep Convolutional Networks for Large-Scale Image Recognition. ICLR.<\/li>\n
  • Szegedy, C., et al. (2015). Going Deeper with Convolutions. CVPR.<\/li>\n
  • He, K., et al. (2016). Deep Residual Learning for Image Recognition. CVPR.<\/li>\n
  • Vaswani, A., et al. (2017). Attention Is All You Need. NeurIPS.<\/li>\n
  • Dosovitskiy, A., et al. (2020). An Image is Worth 16×16 Words: Transformers for Image Recognition at Scale. ICLR.<\/li>\n
  • Radford, A., et al. (2021). Learning Transferable Visual Models From Natural Language Supervision. ICML.<\/li>\n
  • Rombach, R., et al. (2022). High-Resolution Image Synthesis with Latent Diffusion Models. CVPR.<\/li>\n
  • Kirillov, A., et al. (2023). Segment Anything. ICCV.<\/li>\n
  • OpenAI. (2023). GPT-4V(ision) System Card.<\/li>\n
  • OpenAI. (2024). Video Generation Models as World Simulators. Technical Report.<\/li>\n
  • Liu, Z., et al. (2021). Swin Transformer: Hierarchical Vision Transformer using Shifted Windows. ICCV.<\/li>\n
  • Touvron, H., et al. (2021). Training data-efficient image transformers & distillation through attention. ICML.<\/li>\n","protected":false},"excerpt":{"rendered":"

    \/* MD \/ \u5bcc\u6587\u672c\u4e2d\u7684 .toc\uff08\u542b\u535a\u5ba2\u56ed\u642c\u5bb6\u7b49\u5d4c\u5957\u7ed3\u6784\uff09\uff1b.toc-box \u5728\u4fa7\u680f\uff0c\u4e0d\u53d7\u5f71\u54cd *\/#content_views .toc,\/* \u7f16\u8f91\u5668\u5e38\u5728\u76ee\u5f55\u524d\u540e\u63d2\u5165\u7a7a p\uff08:empty \u4ecd\u5360 20px\uff09\uff0c\u4e00\u5e76\u53bb\u6389\u907f\u514d\u9876\u7a7a\u9699 *\/#content_views.markdown_views > p:empty:has(+ .toc),#content_views.markdown_views > .toc + p:empty,\/* \u5bcc\u6587\u672c\u65e7\u7248\u76ee\u5f55\u6807\u8bb0 *\/#content<\/p>\n","protected":false},"author":2,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[14056,2753,1460,1665,523],"topic":[],"class_list":["post-123044","post","type-post","status-publish","format-standard","hentry","category-server","tag-14056","tag-cv","tag-opencv","tag-1665","tag-523"],"yoast_head":"\n\u8ba1\u7b97\u673a\u89c6\u89c9\u7684\u53d1\u5c55\u5386\u7a0b\uff1a\u4eceLeNet\u5230GPT-4V\u7684\u4e09\u5341\u5e74 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/www.wsisp.com\/helps\/123044.html\" \/>\n<meta property=\"og:locale\" content=\"zh_CN\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"\u8ba1\u7b97\u673a\u89c6\u89c9\u7684\u53d1\u5c55\u5386\u7a0b\uff1a\u4eceLeNet\u5230GPT-4V\u7684\u4e09\u5341\u5e74 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"og:description\" content=\"\/* MD \/ \u5bcc\u6587\u672c\u4e2d\u7684 .toc\uff08\u542b\u535a\u5ba2\u56ed\u642c\u5bb6\u7b49\u5d4c\u5957\u7ed3\u6784\uff09\uff1b.toc-box \u5728\u4fa7\u680f\uff0c\u4e0d\u53d7\u5f71\u54cd *\/#content_views .toc,\/* \u7f16\u8f91\u5668\u5e38\u5728\u76ee\u5f55\u524d\u540e\u63d2\u5165\u7a7a p\uff08:empty \u4ecd\u5360 20px\uff09\uff0c\u4e00\u5e76\u53bb\u6389\u907f\u514d\u9876\u7a7a\u9699 *\/#content_views.markdown_views > p:empty:has(+ .toc),#content_views.markdown_views > .toc + p:empty,\/* \u5bcc\u6587\u672c\u65e7\u7248\u76ee\u5f55\u6807\u8bb0 *\/#content\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.wsisp.com\/helps\/123044.html\" \/>\n<meta property=\"og:site_name\" content=\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"article:published_time\" content=\"2026-11-06T00:40:48+00:00\" \/>\n<meta name=\"author\" content=\"admin\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:label1\" content=\"\u4f5c\u8005\" \/>\n\t<meta name=\"twitter:data1\" content=\"admin\" \/>\n\t<meta name=\"twitter:label2\" content=\"\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4\" \/>\n\t<meta name=\"twitter:data2\" content=\"13 \u5206\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/123044.html\",\"url\":\"https:\/\/www.wsisp.com\/helps\/123044.html\",\"name\":\"\u8ba1\u7b97\u673a\u89c6\u89c9\u7684\u53d1\u5c55\u5386\u7a0b\uff1a\u4eceLeNet\u5230GPT-4V\u7684\u4e09\u5341\u5e74 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"isPartOf\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\"},\"datePublished\":\"2026-11-06T00:40:48+00:00\",\"dateModified\":\"2026-11-06T00:40:48+00:00\",\"author\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\"},\"breadcrumb\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/123044.html#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/www.wsisp.com\/helps\/123044.html\"]}]},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/123044.html#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/www.wsisp.com\/helps\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u8ba1\u7b97\u673a\u89c6\u89c9\u7684\u53d1\u5c55\u5386\u7a0b\uff1a\u4eceLeNet\u5230GPT-4V\u7684\u4e09\u5341\u5e74\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\",\"url\":\"https:\/\/www.wsisp.com\/helps\/\",\"name\":\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"description\":\"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}\"},\"query-input\":\"required name=search_term_string\"}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\",\"name\":\"admin\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/\",\"url\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"contentUrl\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"caption\":\"admin\"},\"sameAs\":[\"http:\/\/wp.wsisp.com\"],\"url\":\"https:\/\/www.wsisp.com\/helps\/author\/admin\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"\u8ba1\u7b97\u673a\u89c6\u89c9\u7684\u53d1\u5c55\u5386\u7a0b\uff1a\u4eceLeNet\u5230GPT-4V\u7684\u4e09\u5341\u5e74 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/www.wsisp.com\/helps\/123044.html","og_locale":"zh_CN","og_type":"article","og_title":"\u8ba1\u7b97\u673a\u89c6\u89c9\u7684\u53d1\u5c55\u5386\u7a0b\uff1a\u4eceLeNet\u5230GPT-4V\u7684\u4e09\u5341\u5e74 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","og_description":"\/* MD \/ \u5bcc\u6587\u672c\u4e2d\u7684 .toc\uff08\u542b\u535a\u5ba2\u56ed\u642c\u5bb6\u7b49\u5d4c\u5957\u7ed3\u6784\uff09\uff1b.toc-box \u5728\u4fa7\u680f\uff0c\u4e0d\u53d7\u5f71\u54cd *\/#content_views .toc,\/* \u7f16\u8f91\u5668\u5e38\u5728\u76ee\u5f55\u524d\u540e\u63d2\u5165\u7a7a p\uff08:empty \u4ecd\u5360 20px\uff09\uff0c\u4e00\u5e76\u53bb\u6389\u907f\u514d\u9876\u7a7a\u9699 *\/#content_views.markdown_views > p:empty:has(+ .toc),#content_views.markdown_views > .toc + p:empty,\/* \u5bcc\u6587\u672c\u65e7\u7248\u76ee\u5f55\u6807\u8bb0 *\/#content","og_url":"https:\/\/www.wsisp.com\/helps\/123044.html","og_site_name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","article_published_time":"2026-11-06T00:40:48+00:00","author":"admin","twitter_card":"summary_large_image","twitter_misc":{"\u4f5c\u8005":"admin","\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4":"13 \u5206"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/www.wsisp.com\/helps\/123044.html","url":"https:\/\/www.wsisp.com\/helps\/123044.html","name":"\u8ba1\u7b97\u673a\u89c6\u89c9\u7684\u53d1\u5c55\u5386\u7a0b\uff1a\u4eceLeNet\u5230GPT-4V\u7684\u4e09\u5341\u5e74 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","isPartOf":{"@id":"https:\/\/www.wsisp.com\/helps\/#website"},"datePublished":"2026-11-06T00:40:48+00:00","dateModified":"2026-11-06T00:40:48+00:00","author":{"@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41"},"breadcrumb":{"@id":"https:\/\/www.wsisp.com\/helps\/123044.html#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.wsisp.com\/helps\/123044.html"]}]},{"@type":"BreadcrumbList","@id":"https:\/\/www.wsisp.com\/helps\/123044.html#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/www.wsisp.com\/helps"},{"@type":"ListItem","position":2,"name":"\u8ba1\u7b97\u673a\u89c6\u89c9\u7684\u53d1\u5c55\u5386\u7a0b\uff1a\u4eceLeNet\u5230GPT-4V\u7684\u4e09\u5341\u5e74"}]},{"@type":"WebSite","@id":"https:\/\/www.wsisp.com\/helps\/#website","url":"https:\/\/www.wsisp.com\/helps\/","name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","description":"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}"},"query-input":"required name=search_term_string"}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41","name":"admin","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/","url":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","contentUrl":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","caption":"admin"},"sameAs":["http:\/\/wp.wsisp.com"],"url":"https:\/\/www.wsisp.com\/helps\/author\/admin"}]}},"_links":{"self":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/123044","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/comments?post=123044"}],"version-history":[{"count":0,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/123044\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media?parent=123044"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/categories?post=123044"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/tags?post=123044"},{"taxonomy":"topic","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/topic?post=123044"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}