{"version":2,"relationshipURLBase":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p55.html","relationshipResolution":"Resolve relative relationship URLs against relationshipURLBase; complete records and absolute URLs are in blocks.","title":"区域表示与语言预训练，有了可连接的部件","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p55.html","markdown":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p55.md","pages":[{"id":"vilbert-reusable-joint-start","title":"复用组件以后，还要学习哪些对应？","unit":"语言怎样选择当前需要的视觉信息？","unitId":"language-local-and-question","chapter":"III","number":1}],"aliases":[{"oldId":"explanation","targetId":"vilbert-reusable-joint-start","operation":"split-explanation-range"},{"oldId":"b-67f74137b3ad9ae6412b2677","targetId":"b-67f74137b3ad9ae6412b2677","page":"vilbert-reusable-joint-start","operation":"move","beforeSHA256":"d71cf15e707c8a3674b4c90a599efee783cb84da5b8c8236a1d01c3add7b4958","afterSHA256":"d71cf15e707c8a3674b4c90a599efee783cb84da5b8c8236a1d01c3add7b4958","reason":"Same source body in authored semantic page"},{"oldId":"b-ac335ef2afc942ea48883037","targetId":"b-ac335ef2afc942ea48883037","page":"vilbert-reusable-joint-start","operation":"move","beforeSHA256":"633b31d695a69fed728c43d18876f53158dbedc79fbbb3c77f4b89ccda94271d","afterSHA256":"633b31d695a69fed728c43d18876f53158dbedc79fbbb3c77f4b89ccda94271d","reason":"Same source body in authored semantic page"},{"oldId":"b-9bf5350a1e41b65debd144c3","targetId":"b-9bf5350a1e41b65debd144c3","page":"vilbert-reusable-joint-start","operation":"move","beforeSHA256":"24e1538a82758ae32abd530853b95a844b61c3dba2d9c3c045a7eabdfd4bc917","afterSHA256":"24e1538a82758ae32abd530853b95a844b61c3dba2d9c3c045a7eabdfd4bc917","reason":"Same source body in authored semantic page"},{"oldId":"b-3e17640eb66c3ff78412d9f9","targetId":"b-3e17640eb66c3ff78412d9f9","page":"vilbert-reusable-joint-start","operation":"rewrite","beforeSHA256":"721f56a10a716be0e460a225b3cac30fff4fc0a862934e98ba0dcd7dde5eb484","afterSHA256":"b9e7995dd1b416c726938387816a646c19d07e67d24bd37e545e88d80326cb2a","reason":"Same source identity with figure semantics or reading context enriched"},{"oldId":"b-80cc5c9ce092f62b13f94296","targetId":"b-80cc5c9ce092f62b13f94296","page":"vilbert-reusable-joint-start","operation":"move","beforeSHA256":"786ce10a6b32a492259a3dccd0a54e61cd41d28a34382fb7b5386c5f03b7ad69","afterSHA256":"786ce10a6b32a492259a3dccd0a54e61cd41d28a34382fb7b5386c5f03b7ad69","reason":"Same source body in authored semantic page"},{"oldId":"b-8435e27d1af521466f6b80fa","targetId":"b-8435e27d1af521466f6b80fa","page":"vilbert-reusable-joint-start","operation":"move","beforeSHA256":"daed565d78332d73480fe384eac5c8702c6776e1eed60bd0e02b75a32546f694","afterSHA256":"daed565d78332d73480fe384eac5c8702c6776e1eed60bd0e02b75a32546f694","reason":"Same source body in authored semantic page"},{"oldId":"b-cb2bb18687e770ad7bda58d7","targetId":"b-cb2bb18687e770ad7bda58d7","page":"vilbert-reusable-joint-start","operation":"move","beforeSHA256":"d81a20bbac0b0b5fefed1504ff552211c423aec38914f1efa09e62aedd815051","afterSHA256":"d81a20bbac0b0b5fefed1504ff552211c423aec38914f1efa09e62aedd815051","reason":"Same source body in authored semantic page"},{"oldId":"b-62db299265a722784a95d09f","targetId":"b-62db299265a722784a95d09f","page":"vilbert-reusable-joint-start","operation":"move","beforeSHA256":"6b66eb7026936a6211fd1af69e04f8e1bda430ac6e49121d19342d95c3196ff2","afterSHA256":"6b66eb7026936a6211fd1af69e04f8e1bda430ac6e49121d19342d95c3196ff2","reason":"Same source body in authored semantic page"},{"oldId":"b-7a3a213cbda04a5ff4eb98b7","targetId":"b-7a3a213cbda04a5ff4eb98b7","page":"vilbert-reusable-joint-start","operation":"move","beforeSHA256":"140b7f8b1fcac0e0fb869e4d00ebf68b40c5c5c9b147974f227e50a1bd3825d6","afterSHA256":"140b7f8b1fcac0e0fb869e4d00ebf68b40c5c5c9b147974f227e50a1bd3825d6","reason":"Same source body in authored semantic page"},{"oldId":"b-ee70f124cce1be8fa76b5a00","targetId":"b-ee70f124cce1be8fa76b5a00","page":"vilbert-reusable-joint-start","operation":"move","beforeSHA256":"934be7ba01b3582844964e06b9e11c3888072baf4180f5e80ab0499fd7d72781","afterSHA256":"934be7ba01b3582844964e06b9e11c3888072baf4180f5e80ab0499fd7d72781","reason":"Same source body in authored semantic page"},{"oldId":"b-30fd10e958fadd8b4e44f53b","targetId":"b-30fd10e958fadd8b4e44f53b","page":"vilbert-reusable-joint-start","operation":"move","beforeSHA256":"81c56d5a04c99fa7229181969b48cab83009bf34a15ef233d494188dab00cf8a","afterSHA256":"81c56d5a04c99fa7229181969b48cab83009bf34a15ef233d494188dab00cf8a","reason":"Same source body in authored semantic page"},{"oldId":"b-40abdba333fde45fbc7d0734","targetId":"b-40abdba333fde45fbc7d0734","page":"vilbert-reusable-joint-start","operation":"move","beforeSHA256":"cf44fcc62782fb4f6bdc42c16c5187d57556a03cb7ce6065e7bd9c36d8e7d733","afterSHA256":"cf44fcc62782fb4f6bdc42c16c5187d57556a03cb7ce6065e7bd9c36d8e7d733","reason":"Same source body in authored semantic page"},{"oldId":"b-a133a3bb690b3bc34fd6298b","targetId":"b-a133a3bb690b3bc34fd6298b","page":"vilbert-reusable-joint-start","operation":"move","beforeSHA256":"06c4a84d89a169fa08b171e0c41bf06cc7b44322a6298066900a0a92f0412cb0","afterSHA256":"06c4a84d89a169fa08b171e0c41bf06cc7b44322a6298066900a0a92f0412cb0","reason":"Same source body in authored semantic page"},{"oldId":"b-d284ac472087de6f863a138d","targetId":"b-d284ac472087de6f863a138d","page":"vilbert-reusable-joint-start","operation":"move","beforeSHA256":"9b9d51ee29dd871d540df4d0261781112df08ee7f27650f1a175ce3e58796e03","afterSHA256":"9b9d51ee29dd871d540df4d0261781112df08ee7f27650f1a175ce3e58796e03","reason":"Same source body in authored semantic page"},{"oldId":"b-48060622e7b97db08ce87a9e","targetId":"b-48060622e7b97db08ce87a9e","page":"vilbert-reusable-joint-start","operation":"move","beforeSHA256":"7bd8c57e1e8f610ad98cab953e3840e15405cd6dd60992afe6b8d84c173b6867","afterSHA256":"7bd8c57e1e8f610ad98cab953e3840e15405cd6dd60992afe6b8d84c173b6867","reason":"Same source body in authored semantic page"},{"oldId":"b-f79ffb1115b688ed82d21834","targetId":"b-f79ffb1115b688ed82d21834","page":"vilbert-reusable-joint-start","operation":"move","beforeSHA256":"5409efce4e823ced2b94060bf818e5f476c0a0032ae15add3384007e8a08d063","afterSHA256":"5409efce4e823ced2b94060bf818e5f476c0a0032ae15add3384007e8a08d063","reason":"Same source body in authored semantic page"},{"oldId":"b-2a53618f834dc8304db71e52","targetId":"b-2a53618f834dc8304db71e52","page":"vilbert-reusable-joint-start","operation":"move","beforeSHA256":"08fc293b9ead2c6db138ace3a6118cf7a6458ed5a252f9dcecebd11dc4371bc8","afterSHA256":"08fc293b9ead2c6db138ace3a6118cf7a6458ed5a252f9dcecebd11dc4371bc8","reason":"Same source body in authored semantic page"},{"oldId":"b-f55597f40aea653725260420","targetId":"b-f55597f40aea653725260420","page":"vilbert-reusable-joint-start","operation":"move","beforeSHA256":"841759ff9eb6d168b22e5f92a41fd01d6cb50f7fb97162d8df1fed7727375264","afterSHA256":"841759ff9eb6d168b22e5f92a41fd01d6cb50f7fb97162d8df1fed7727375264","reason":"Same source body in authored semantic page"},{"oldId":"b-c4146821e63fe222c7d71cc8","targetId":"b-c4146821e63fe222c7d71cc8","page":"vilbert-reusable-joint-start","operation":"move","beforeSHA256":"d24658903a9d2440ee1547f395309286cb95a5813edb1a71ad8c32febd199825","afterSHA256":"d24658903a9d2440ee1547f395309286cb95a5813edb1a71ad8c32febd199825","reason":"Same source body in authored semantic page"},{"oldId":"vilbert-reusable-joint-start-title","targetId":"b-0bfabdbbe9af488996322868","operation":"retain-source-identity"}],"blocks":[{"id":"b-0bfabdbbe9af488996322868","type":"h2","scope":"vilbert-reusable-joint-start","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p55.html#b-0bfabdbbe9af488996322868","content":"## 复用组件以后，还要学习哪些对应？","sha256":"c6a60d2f6fdb417b06166bea6108c4288233597bab93d36db3954174445b3080","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"language-local-and-question","title":"语言怎样选择当前需要的视觉信息？"},"page":{"id":"vilbert-reusable-joint-start","title":"复用组件以后，还要学习哪些对应？","number":1},"sectionTitle":"复用组件以后，还要学习哪些对应？","currentQuestion":"局部对应、问题条件和联合预训练分别增加了哪条关系？","sourceSlides":["l3-r2-p55"],"knowledgeBlock":null,"formulas":[],"previousBlock":null,"nextBlock":"b-67f74137b3ad9ae6412b2677","prerequisites":[{"id":"b-9bf5350a1e41b65debd144c3","url":"#b-9bf5350a1e41b65debd144c3","role":"problem-context"}],"relatedBlocks":[{"id":"b-3e17640eb66c3ff78412d9f9","url":"#b-3e17640eb66c3ff78412d9f9","relation":"figure-and-explanation"}]},{"id":"b-67f74137b3ad9ae6412b2677","type":"h2","scope":"explanation","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p55.html#b-67f74137b3ad9ae6412b2677","content":"## 解释、算例与证据","sha256":"d71cf15e707c8a3674b4c90a599efee783cb84da5b8c8236a1d01c3add7b4958","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"language-local-and-question","title":"语言怎样选择当前需要的视觉信息？"},"page":{"id":"vilbert-reusable-joint-start","title":"复用组件以后，还要学习哪些对应？","number":1},"sectionTitle":"复用组件以后，还要学习哪些对应？","currentQuestion":"局部对应、问题条件和联合预训练分别增加了哪条关系？","sourceSlides":["l3-r2-p55"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-0bfabdbbe9af488996322868","nextBlock":"b-ac335ef2afc942ea48883037","prerequisites":[{"id":"b-0bfabdbbe9af488996322868","url":"#b-0bfabdbbe9af488996322868","role":"section-title"},{"id":"b-9bf5350a1e41b65debd144c3","url":"#b-9bf5350a1e41b65debd144c3","role":"problem-context"}],"relatedBlocks":[{"id":"b-3e17640eb66c3ff78412d9f9","url":"#b-3e17640eb66c3ff78412d9f9","relation":"figure-and-explanation"}]},{"id":"b-ac335ef2afc942ea48883037","type":"h1","scope":"document","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p55.html#b-ac335ef2afc942ea48883037","content":"# P55 · 区域表示与语言预训练，有了可连接的部件","sha256":"633b31d695a69fed728c43d18876f53158dbedc79fbbb3c77f4b89ccda94271d","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"language-local-and-question","title":"语言怎样选择当前需要的视觉信息？"},"page":{"id":"vilbert-reusable-joint-start","title":"复用组件以后，还要学习哪些对应？","number":1},"sectionTitle":"复用组件以后，还要学习哪些对应？","currentQuestion":"局部对应、问题条件和联合预训练分别增加了哪条关系？","sourceSlides":["l3-r2-p55"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-67f74137b3ad9ae6412b2677","nextBlock":"b-9bf5350a1e41b65debd144c3","prerequisites":[{"id":"b-67f74137b3ad9ae6412b2677","url":"#b-67f74137b3ad9ae6412b2677","role":"section-title"},{"id":"b-9bf5350a1e41b65debd144c3","url":"#b-9bf5350a1e41b65debd144c3","role":"problem-context"}],"relatedBlocks":[{"id":"b-3e17640eb66c3ff78412d9f9","url":"#b-3e17640eb66c3ff78412d9f9","relation":"figure-and-explanation"}]},{"id":"b-9bf5350a1e41b65debd144c3","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p55.html#b-9bf5350a1e41b65debd144c3","content":"ViLBERT 让前面两部分的积累变得具体：检测器提供区域表示，语言预训练提供文字表示，新的预训练目标学习二者怎样共同服务任务。读本页不需要展开 attention；重点是指出可复用组件、尚未学会的对应以及跨任务迁移的研究证据。","sha256":"24e1538a82758ae32abd530853b95a844b61c3dba2d9c3c045a7eabdfd4bc917","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"language-local-and-question","title":"语言怎样选择当前需要的视觉信息？"},"page":{"id":"vilbert-reusable-joint-start","title":"复用组件以后，还要学习哪些对应？","number":1},"sectionTitle":"复用组件以后，还要学习哪些对应？","currentQuestion":"局部对应、问题条件和联合预训练分别增加了哪条关系？","sourceSlides":["l3-r2-p55"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-ac335ef2afc942ea48883037","nextBlock":"b-3e17640eb66c3ff78412d9f9","prerequisites":[{"id":"b-ac335ef2afc942ea48883037","url":"#b-ac335ef2afc942ea48883037","role":"section-title"}],"relatedBlocks":[{"id":"b-3e17640eb66c3ff78412d9f9","url":"#b-3e17640eb66c3ff78412d9f9","relation":"figure-and-explanation"}]},{"id":"b-3e17640eb66c3ff78412d9f9","type":"figure","scope":"document","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p55.html#b-3e17640eb66c3ff78412d9f9","content":"图内标签与数值：区域视觉输入；检测器留下的区域特征；语言输入；“红衣人拿着球”；预训练的文字表示；ViLBERT · 2019；跨模态目标学习对应；图文匹配；遮挡内容预测；适配问答与检索等任务；检测器区域特征与预训练语言表示是实际组件来源；此页仅呈现模块接口。\n\n左侧检测器区域特征、右侧预训练语言表示，进入ViLBERT的跨模态目标，再适配问答或检索。\n\n- 区域视觉输入\n- 检测器留下的区域特征\n- 语言输入\n- “红衣人拿着球”\n- 预训练的文字表示\n- ViLBERT · 2019\n- 跨模态目标学习对应\n- 图文匹配\n- 遮挡内容预测\n- 适配问答与检索等任务\n- 检测器区域特征与预训练语言表示是实际组件来源；此页仅呈现模块接口。左侧检测器区域特征、右侧预训练语言表示，进入ViLBERT的跨模态目标，再适配问答或检索。 [查看对应图解状态](/#/l3-r2-p55/0)[查看完整图解](/course/figures/l3-r2-p55-0.html)","sha256":"b9e7995dd1b416c726938387816a646c19d07e67d24bd37e545e88d80326cb2a","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"language-local-and-question","title":"语言怎样选择当前需要的视觉信息？"},"page":{"id":"vilbert-reusable-joint-start","title":"复用组件以后，还要学习哪些对应？","number":1},"sectionTitle":"复用组件以后，还要学习哪些对应？","currentQuestion":"局部对应、问题条件和联合预训练分别增加了哪条关系？","sourceSlides":["l3-r2-p55"],"knowledgeBlock":null,"formulas":[],"figure":{"caption":"左侧检测器区域特征、右侧预训练语言表示，进入ViLBERT的跨模态目标，再适配问答或检索。 查看对应图解状态","labels":["区域视觉输入","检测器留下的区域特征","语言输入","“红衣人拿着球”","预训练的文字表示","ViLBERT · 2019","跨模态目标学习对应","图文匹配","遮挡内容预测","适配问答与检索等任务","检测器区域特征与预训练语言表示是实际组件来源；此页仅呈现模块接口。"],"description":"左侧检测器区域特征、右侧预训练语言表示，进入ViLBERT的跨模态目标，再适配问答或检索。区域视觉输入检测器留下的区域特征语言输入“红衣人拿着球”预训练的文字表示ViLBERT · 2019跨模态目标学习对应图文匹配遮挡内容预测适配问答与检索等任务检测器区域特征与预训练语言表示是实际组件来源；此页仅呈现模块接口。","evidenceIdentity":"左侧检测器区域特征、右侧预训练语言表示，进入ViLBERT的跨模态目标，再适配问答或检索。 查看对应图解状态","images":[{"src":"/assets/display/512610a019c560ff0384e86a.webp","alt":""}]},"previousBlock":"b-9bf5350a1e41b65debd144c3","nextBlock":"b-80cc5c9ce092f62b13f94296","prerequisites":[{"id":"b-ac335ef2afc942ea48883037","url":"#b-ac335ef2afc942ea48883037","role":"section-title"},{"id":"b-9bf5350a1e41b65debd144c3","url":"#b-9bf5350a1e41b65debd144c3","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-80cc5c9ce092f62b13f94296","type":"h2","scope":"narrative","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p55.html#b-80cc5c9ce092f62b13f94296","content":"## 课堂讲述","sha256":"786ce10a6b32a492259a3dccd0a54e61cd41d28a34382fb7b5386c5f03b7ad69","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"language-local-and-question","title":"语言怎样选择当前需要的视觉信息？"},"page":{"id":"vilbert-reusable-joint-start","title":"复用组件以后，还要学习哪些对应？","number":1},"sectionTitle":"复用组件以后，还要学习哪些对应？","currentQuestion":"局部对应、问题条件和联合预训练分别增加了哪条关系？","sourceSlides":["l3-r2-p55"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-3e17640eb66c3ff78412d9f9","nextBlock":"b-8435e27d1af521466f6b80fa","prerequisites":[{"id":"b-ac335ef2afc942ea48883037","url":"#b-ac335ef2afc942ea48883037","role":"section-title"},{"id":"b-9bf5350a1e41b65debd144c3","url":"#b-9bf5350a1e41b65debd144c3","role":"problem-context"}],"relatedBlocks":[{"id":"b-3e17640eb66c3ff78412d9f9","url":"#b-3e17640eb66c3ff78412d9f9","relation":"figure-and-explanation"}]},{"id":"b-8435e27d1af521466f6b80fa","type":"p","scope":"narrative","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p55.html#b-8435e27d1af521466f6b80fa","content":"到了区域语言预训练这一类工作，前面的积累开始以模块形式汇合。以 ViLBERT 为例，一边是检测器提取的区域表示，另一边是语言预训练，再通过跨模态目标学习它们怎样配合。\n这条连接值得看，不是因为今天必须掌握完整网络结构，而是它让技术演进变得具体：检测研究留下的区域特征，能成为视觉语言系统的输入；语言研究留下的表示，也不用每个视觉任务从零学。\n但新的对应关系仍要训练。复用组件不等于复用完整能力。这是今天不断出现的同一个模式：已有表示降低了新问题的起步成本，新任务又反过来要求不同的接口、监督和评价。","sha256":"daed565d78332d73480fe384eac5c8702c6776e1eed60bd0e02b75a32546f694","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"language-local-and-question","title":"语言怎样选择当前需要的视觉信息？"},"page":{"id":"vilbert-reusable-joint-start","title":"复用组件以后，还要学习哪些对应？","number":1},"sectionTitle":"复用组件以后，还要学习哪些对应？","currentQuestion":"局部对应、问题条件和联合预训练分别增加了哪条关系？","sourceSlides":["l3-r2-p55"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-80cc5c9ce092f62b13f94296","nextBlock":"b-cb2bb18687e770ad7bda58d7","prerequisites":[{"id":"b-80cc5c9ce092f62b13f94296","url":"#b-80cc5c9ce092f62b13f94296","role":"section-title"},{"id":"b-9bf5350a1e41b65debd144c3","url":"#b-9bf5350a1e41b65debd144c3","role":"problem-context"}],"relatedBlocks":[{"id":"b-3e17640eb66c3ff78412d9f9","url":"#b-3e17640eb66c3ff78412d9f9","relation":"figure-and-explanation"}]},{"id":"b-cb2bb18687e770ad7bda58d7","type":"h3","scope":"vilbert-components-and-transfer","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p55.html#b-cb2bb18687e770ad7bda58d7","content":"### ViLBERT：把已有区域和文字表示变成可迁移的联合起点","sha256":"d81a20bbac0b0b5fefed1504ff552211c423aec38914f1efa09e62aedd815051","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"language-local-and-question","title":"语言怎样选择当前需要的视觉信息？"},"page":{"id":"vilbert-reusable-joint-start","title":"复用组件以后，还要学习哪些对应？","number":1},"sectionTitle":"ViLBERT：把已有区域和文字表示变成可迁移的联合起点","currentQuestion":"局部对应、问题条件和联合预训练分别增加了哪条关系？","sourceSlides":["l3-r2-p55"],"knowledgeBlock":"vilbert-components-and-transfer","formulas":[],"previousBlock":"b-8435e27d1af521466f6b80fa","nextBlock":"b-62db299265a722784a95d09f","prerequisites":[{"id":"b-80cc5c9ce092f62b13f94296","url":"#b-80cc5c9ce092f62b13f94296","role":"section-title"},{"id":"b-9bf5350a1e41b65debd144c3","url":"#b-9bf5350a1e41b65debd144c3","role":"problem-context"}],"relatedBlocks":[{"id":"b-3e17640eb66c3ff78412d9f9","url":"#b-3e17640eb66c3ff78412d9f9","relation":"figure-and-explanation"},{"id":"b-62db299265a722784a95d09f","url":"#b-62db299265a722784a95d09f","relation":"same-reasoning"},{"id":"b-7a3a213cbda04a5ff4eb98b7","url":"#b-7a3a213cbda04a5ff4eb98b7","relation":"same-reasoning"},{"id":"b-ee70f124cce1be8fa76b5a00","url":"#b-ee70f124cce1be8fa76b5a00","relation":"same-reasoning"},{"id":"b-30fd10e958fadd8b4e44f53b","url":"#b-30fd10e958fadd8b4e44f53b","relation":"same-reasoning"}]},{"id":"b-62db299265a722784a95d09f","type":"p","scope":"vilbert-components-and-transfer","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p55.html#b-62db299265a722784a95d09f","content":"第二部分的检测器不只留下“这里有杯子”的类别输出，也能留下每个候选区域的特征。语言预训练同样留下文字表示。ViLBERT 将两侧的已有能力接入跨模态预训练，再适配问答、指代表达、常识推理与图文检索等任务。","sha256":"6b66eb7026936a6211fd1af69e04f8e1bda430ac6e49121d19342d95c3196ff2","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"language-local-and-question","title":"语言怎样选择当前需要的视觉信息？"},"page":{"id":"vilbert-reusable-joint-start","title":"复用组件以后，还要学习哪些对应？","number":1},"sectionTitle":"ViLBERT：把已有区域和文字表示变成可迁移的联合起点","currentQuestion":"局部对应、问题条件和联合预训练分别增加了哪条关系？","sourceSlides":["l3-r2-p55"],"knowledgeBlock":"vilbert-components-and-transfer","formulas":[],"previousBlock":"b-cb2bb18687e770ad7bda58d7","nextBlock":"b-7a3a213cbda04a5ff4eb98b7","prerequisites":[{"id":"b-cb2bb18687e770ad7bda58d7","url":"#b-cb2bb18687e770ad7bda58d7","role":"section-title"},{"id":"b-9bf5350a1e41b65debd144c3","url":"#b-9bf5350a1e41b65debd144c3","role":"problem-context"}],"relatedBlocks":[{"id":"b-3e17640eb66c3ff78412d9f9","url":"#b-3e17640eb66c3ff78412d9f9","relation":"figure-and-explanation"},{"id":"b-cb2bb18687e770ad7bda58d7","url":"#b-cb2bb18687e770ad7bda58d7","relation":"same-reasoning"},{"id":"b-7a3a213cbda04a5ff4eb98b7","url":"#b-7a3a213cbda04a5ff4eb98b7","relation":"same-reasoning"},{"id":"b-ee70f124cce1be8fa76b5a00","url":"#b-ee70f124cce1be8fa76b5a00","relation":"same-reasoning"},{"id":"b-30fd10e958fadd8b4e44f53b","url":"#b-30fd10e958fadd8b4e44f53b","relation":"same-reasoning"}]},{"id":"b-7a3a213cbda04a5ff4eb98b7","type":"p","scope":"vilbert-components-and-transfer","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p55.html#b-7a3a213cbda04a5ff4eb98b7","content":"为什么不直接拼起来就结束？设区域向量来自检测器、词向量来自文字模型。两个向量长度可以相同，但相同坐标号并不自动代表同一种语义。跨模态目标要学习哪些局部与哪些文字对应、哪些信息可以共同预测。维度匹配只让计算合法，训练才使这条关系有任务含义。","sha256":"140b7f8b1fcac0e0fb869e4d00ebf68b40c5c5c9b147974f227e50a1bd3825d6","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"language-local-and-question","title":"语言怎样选择当前需要的视觉信息？"},"page":{"id":"vilbert-reusable-joint-start","title":"复用组件以后，还要学习哪些对应？","number":1},"sectionTitle":"ViLBERT：把已有区域和文字表示变成可迁移的联合起点","currentQuestion":"局部对应、问题条件和联合预训练分别增加了哪条关系？","sourceSlides":["l3-r2-p55"],"knowledgeBlock":"vilbert-components-and-transfer","formulas":[],"previousBlock":"b-62db299265a722784a95d09f","nextBlock":"b-ee70f124cce1be8fa76b5a00","prerequisites":[{"id":"b-cb2bb18687e770ad7bda58d7","url":"#b-cb2bb18687e770ad7bda58d7","role":"section-title"},{"id":"b-9bf5350a1e41b65debd144c3","url":"#b-9bf5350a1e41b65debd144c3","role":"problem-context"}],"relatedBlocks":[{"id":"b-3e17640eb66c3ff78412d9f9","url":"#b-3e17640eb66c3ff78412d9f9","relation":"figure-and-explanation"},{"id":"b-cb2bb18687e770ad7bda58d7","url":"#b-cb2bb18687e770ad7bda58d7","relation":"same-reasoning"},{"id":"b-62db299265a722784a95d09f","url":"#b-62db299265a722784a95d09f","relation":"same-reasoning"},{"id":"b-ee70f124cce1be8fa76b5a00","url":"#b-ee70f124cce1be8fa76b5a00","relation":"same-reasoning"},{"id":"b-30fd10e958fadd8b4e44f53b","url":"#b-30fd10e958fadd8b4e44f53b","relation":"same-reasoning"}]},{"id":"b-ee70f124cce1be8fa76b5a00","type":"p","scope":"vilbert-components-and-transfer","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p55.html#b-ee70f124cce1be8fa76b5a00","content":"原论文通过代理预训练任务学习联合表示，再对四类视觉语言任务做少量结构增补与任务适配，报告了相对当时任务专用模型的改善。它支持的研究方向是：区域—语言对应可以成为预训练并迁移的能力，未必每个下游任务都从头建立全部对应关系。这里保留这一正面进展，不展开内部注意力实现。","sha256":"934be7ba01b3582844964e06b9e11c3888072baf4180f5e80ab0499fd7d72781","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"language-local-and-question","title":"语言怎样选择当前需要的视觉信息？"},"page":{"id":"vilbert-reusable-joint-start","title":"复用组件以后，还要学习哪些对应？","number":1},"sectionTitle":"ViLBERT：把已有区域和文字表示变成可迁移的联合起点","currentQuestion":"局部对应、问题条件和联合预训练分别增加了哪条关系？","sourceSlides":["l3-r2-p55"],"knowledgeBlock":"vilbert-components-and-transfer","formulas":[],"previousBlock":"b-7a3a213cbda04a5ff4eb98b7","nextBlock":"b-30fd10e958fadd8b4e44f53b","prerequisites":[{"id":"b-cb2bb18687e770ad7bda58d7","url":"#b-cb2bb18687e770ad7bda58d7","role":"section-title"},{"id":"b-9bf5350a1e41b65debd144c3","url":"#b-9bf5350a1e41b65debd144c3","role":"problem-context"}],"relatedBlocks":[{"id":"b-3e17640eb66c3ff78412d9f9","url":"#b-3e17640eb66c3ff78412d9f9","relation":"figure-and-explanation"},{"id":"b-cb2bb18687e770ad7bda58d7","url":"#b-cb2bb18687e770ad7bda58d7","relation":"same-reasoning"},{"id":"b-62db299265a722784a95d09f","url":"#b-62db299265a722784a95d09f","relation":"same-reasoning"},{"id":"b-7a3a213cbda04a5ff4eb98b7","url":"#b-7a3a213cbda04a5ff4eb98b7","relation":"same-reasoning"},{"id":"b-30fd10e958fadd8b4e44f53b","url":"#b-30fd10e958fadd8b4e44f53b","relation":"same-reasoning"}]},{"id":"b-30fd10e958fadd8b4e44f53b","type":"p","scope":"vilbert-components-and-transfer","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p55.html#b-30fd10e958fadd8b4e44f53b","content":"阅读图时注意“可复用”有两个尺度：检测器区域特征作为输入被复用，跨模态预训练后的表示又被多个任务复用。二者都不是说下游无需目标数据或评价。若关心小字、细小对象或未被候选器提出的区域，首先还要检查最前面的视觉输入是否已经丢失所需证据。","sha256":"81c56d5a04c99fa7229181969b48cab83009bf34a15ef233d494188dab00cf8a","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"language-local-and-question","title":"语言怎样选择当前需要的视觉信息？"},"page":{"id":"vilbert-reusable-joint-start","title":"复用组件以后，还要学习哪些对应？","number":1},"sectionTitle":"ViLBERT：把已有区域和文字表示变成可迁移的联合起点","currentQuestion":"局部对应、问题条件和联合预训练分别增加了哪条关系？","sourceSlides":["l3-r2-p55"],"knowledgeBlock":"vilbert-components-and-transfer","formulas":[],"previousBlock":"b-ee70f124cce1be8fa76b5a00","nextBlock":"b-40abdba333fde45fbc7d0734","prerequisites":[{"id":"b-cb2bb18687e770ad7bda58d7","url":"#b-cb2bb18687e770ad7bda58d7","role":"section-title"},{"id":"b-9bf5350a1e41b65debd144c3","url":"#b-9bf5350a1e41b65debd144c3","role":"problem-context"}],"relatedBlocks":[{"id":"b-3e17640eb66c3ff78412d9f9","url":"#b-3e17640eb66c3ff78412d9f9","relation":"figure-and-explanation"},{"id":"b-cb2bb18687e770ad7bda58d7","url":"#b-cb2bb18687e770ad7bda58d7","relation":"same-reasoning"},{"id":"b-62db299265a722784a95d09f","url":"#b-62db299265a722784a95d09f","relation":"same-reasoning"},{"id":"b-7a3a213cbda04a5ff4eb98b7","url":"#b-7a3a213cbda04a5ff4eb98b7","relation":"same-reasoning"},{"id":"b-ee70f124cce1be8fa76b5a00","url":"#b-ee70f124cce1be8fa76b5a00","relation":"same-reasoning"}]},{"id":"b-40abdba333fde45fbc7d0734","type":"h2","scope":"practice","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p55.html#b-40abdba333fde45fbc7d0734","content":"## 沿这个问题继续","sha256":"cf44fcc62782fb4f6bdc42c16c5187d57556a03cb7ce6065e7bd9c36d8e7d733","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"language-local-and-question","title":"语言怎样选择当前需要的视觉信息？"},"page":{"id":"vilbert-reusable-joint-start","title":"复用组件以后，还要学习哪些对应？","number":1},"sectionTitle":"复用组件以后，还要学习哪些对应？","currentQuestion":"局部对应、问题条件和联合预训练分别增加了哪条关系？","sourceSlides":["l3-r2-p55"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-30fd10e958fadd8b4e44f53b","nextBlock":"b-a133a3bb690b3bc34fd6298b","prerequisites":[{"id":"b-cb2bb18687e770ad7bda58d7","url":"#b-cb2bb18687e770ad7bda58d7","role":"section-title"},{"id":"b-9bf5350a1e41b65debd144c3","url":"#b-9bf5350a1e41b65debd144c3","role":"problem-context"}],"relatedBlocks":[{"id":"b-3e17640eb66c3ff78412d9f9","url":"#b-3e17640eb66c3ff78412d9f9","relation":"figure-and-explanation"}]},{"id":"b-a133a3bb690b3bc34fd6298b","type":"p","scope":"practice","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p55.html#b-a133a3bb690b3bc34fd6298b","content":"先辨认区域表示和文字表示分别来自哪个已有组件，再读图文匹配、遮挡预测和下游迁移的关系；用输入与目标对照下一页的 CLIP。","sha256":"06c4a84d89a169fa08b171e0c41bf06cc7b44322a6298066900a0a92f0412cb0","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"language-local-and-question","title":"语言怎样选择当前需要的视觉信息？"},"page":{"id":"vilbert-reusable-joint-start","title":"复用组件以后，还要学习哪些对应？","number":1},"sectionTitle":"复用组件以后，还要学习哪些对应？","currentQuestion":"局部对应、问题条件和联合预训练分别增加了哪条关系？","sourceSlides":["l3-r2-p55"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-40abdba333fde45fbc7d0734","nextBlock":"b-def7a81dd9355e09acfbb8cb","prerequisites":[{"id":"b-40abdba333fde45fbc7d0734","url":"#b-40abdba333fde45fbc7d0734","role":"section-title"},{"id":"b-9bf5350a1e41b65debd144c3","url":"#b-9bf5350a1e41b65debd144c3","role":"problem-context"}],"relatedBlocks":[{"id":"b-3e17640eb66c3ff78412d9f9","url":"#b-3e17640eb66c3ff78412d9f9","relation":"figure-and-explanation"}]},{"id":"b-def7a81dd9355e09acfbb8cb","type":"h4","scope":"vilbert-reusable-joint-start","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p55.html#b-def7a81dd9355e09acfbb8cb","content":"#### 图示条件与公式约定","sha256":"ef7f915c1ae63ed67bd23fecf050249273ba0efa8965d2c8d1c1c6b2d3178122","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"language-local-and-question","title":"语言怎样选择当前需要的视觉信息？"},"page":{"id":"vilbert-reusable-joint-start","title":"复用组件以后，还要学习哪些对应？","number":1},"sectionTitle":"复用组件以后，还要学习哪些对应？","currentQuestion":"局部对应、问题条件和联合预训练分别增加了哪条关系？","sourceSlides":["l3-r2-p55"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-a133a3bb690b3bc34fd6298b","nextBlock":"b-d284ac472087de6f863a138d","prerequisites":[{"id":"b-40abdba333fde45fbc7d0734","url":"#b-40abdba333fde45fbc7d0734","role":"section-title"},{"id":"b-9bf5350a1e41b65debd144c3","url":"#b-9bf5350a1e41b65debd144c3","role":"problem-context"}],"relatedBlocks":[{"id":"b-3e17640eb66c3ff78412d9f9","url":"#b-3e17640eb66c3ff78412d9f9","relation":"figure-and-explanation"}]},{"id":"b-d284ac472087de6f863a138d","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p55.html#b-d284ac472087de6f863a138d","content":"此页以原始 ViLBERT 为例，展示检测器区域表示和预训练语言表示进入跨模态训练的模块关系；不展开内部架构。","sha256":"9b9d51ee29dd871d540df4d0261781112df08ee7f27650f1a175ce3e58796e03","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"language-local-and-question","title":"语言怎样选择当前需要的视觉信息？"},"page":{"id":"vilbert-reusable-joint-start","title":"复用组件以后，还要学习哪些对应？","number":1},"sectionTitle":"复用组件以后，还要学习哪些对应？","currentQuestion":"局部对应、问题条件和联合预训练分别增加了哪条关系？","sourceSlides":["l3-r2-p55"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-def7a81dd9355e09acfbb8cb","nextBlock":"b-f25c5e251ea5388dd47ec28b","prerequisites":[{"id":"b-def7a81dd9355e09acfbb8cb","url":"#b-def7a81dd9355e09acfbb8cb","role":"section-title"},{"id":"b-9bf5350a1e41b65debd144c3","url":"#b-9bf5350a1e41b65debd144c3","role":"problem-context"}],"relatedBlocks":[{"id":"b-3e17640eb66c3ff78412d9f9","url":"#b-3e17640eb66c3ff78412d9f9","relation":"figure-and-explanation"}]},{"id":"b-f25c5e251ea5388dd47ec28b","type":"h4","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p55.html#b-f25c5e251ea5388dd47ec28b","content":"#### 这条线索的展开与后续阅读","sha256":"48b8229f7559b36ed4d5628416e3991e5c73293fbce43374db67386c8736f3bc","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"language-local-and-question","title":"语言怎样选择当前需要的视觉信息？"},"page":{"id":"vilbert-reusable-joint-start","title":"复用组件以后，还要学习哪些对应？","number":1},"sectionTitle":"复用组件以后，还要学习哪些对应？","currentQuestion":"局部对应、问题条件和联合预训练分别增加了哪条关系？","sourceSlides":["l3-r2-p55"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-d284ac472087de6f863a138d","nextBlock":"b-48060622e7b97db08ce87a9e","prerequisites":[{"id":"b-def7a81dd9355e09acfbb8cb","url":"#b-def7a81dd9355e09acfbb8cb","role":"section-title"},{"id":"b-9bf5350a1e41b65debd144c3","url":"#b-9bf5350a1e41b65debd144c3","role":"problem-context"}],"relatedBlocks":[{"id":"b-3e17640eb66c3ff78412d9f9","url":"#b-3e17640eb66c3ff78412d9f9","relation":"figure-and-explanation"}]},{"id":"b-48060622e7b97db08ce87a9e","type":"ul","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p55.html#b-48060622e7b97db08ce87a9e","content":"- [区域表示与语言预训练，在 ViLBERT 中接上了](/supplement.html?return=l3-r2-p55/0#/l3-vilbert/0) 本页说明\n- [区域 token 让语言预训练接入物体｜核心改动怎样工作](/supplement.html?return=l3-r2-p55/0#/l3-region-language-mechanism/0) 第五讲预备阅读\n- [区域 token 让语言预训练接入物体｜怎样检查这项解释](/supplement.html?return=l3-r2-p55/0#/l3-region-language-evidence/0) L03补充阅读\n- [区域 token 让语言预训练接入物体｜什么结论还不能推出](/supplement.html?return=l3-r2-p55/0#/l3-region-language-boundary/0) L03补充阅读","sha256":"7bd8c57e1e8f610ad98cab953e3840e15405cd6dd60992afe6b8d84c173b6867","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"language-local-and-question","title":"语言怎样选择当前需要的视觉信息？"},"page":{"id":"vilbert-reusable-joint-start","title":"复用组件以后，还要学习哪些对应？","number":1},"sectionTitle":"复用组件以后，还要学习哪些对应？","currentQuestion":"局部对应、问题条件和联合预训练分别增加了哪条关系？","sourceSlides":["l3-r2-p55"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-f25c5e251ea5388dd47ec28b","nextBlock":"b-f79ffb1115b688ed82d21834","prerequisites":[{"id":"b-f25c5e251ea5388dd47ec28b","url":"#b-f25c5e251ea5388dd47ec28b","role":"section-title"},{"id":"b-9bf5350a1e41b65debd144c3","url":"#b-9bf5350a1e41b65debd144c3","role":"problem-context"}],"relatedBlocks":[{"id":"b-3e17640eb66c3ff78412d9f9","url":"#b-3e17640eb66c3ff78412d9f9","relation":"figure-and-explanation"}]},{"id":"b-f79ffb1115b688ed82d21834","type":"li","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p55.html#b-f79ffb1115b688ed82d21834","content":"- [区域表示与语言预训练，在 ViLBERT 中接上了](/supplement.html?return=l3-r2-p55/0#/l3-vilbert/0) 本页说明","sha256":"5409efce4e823ced2b94060bf818e5f476c0a0032ae15add3384007e8a08d063","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"language-local-and-question","title":"语言怎样选择当前需要的视觉信息？"},"page":{"id":"vilbert-reusable-joint-start","title":"复用组件以后，还要学习哪些对应？","number":1},"sectionTitle":"复用组件以后，还要学习哪些对应？","currentQuestion":"局部对应、问题条件和联合预训练分别增加了哪条关系？","sourceSlides":["l3-r2-p55"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-48060622e7b97db08ce87a9e","nextBlock":"b-2a53618f834dc8304db71e52","prerequisites":[{"id":"b-f25c5e251ea5388dd47ec28b","url":"#b-f25c5e251ea5388dd47ec28b","role":"section-title"},{"id":"b-9bf5350a1e41b65debd144c3","url":"#b-9bf5350a1e41b65debd144c3","role":"problem-context"}],"relatedBlocks":[{"id":"b-3e17640eb66c3ff78412d9f9","url":"#b-3e17640eb66c3ff78412d9f9","relation":"figure-and-explanation"}]},{"id":"b-2a53618f834dc8304db71e52","type":"li","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p55.html#b-2a53618f834dc8304db71e52","content":"- [区域 token 让语言预训练接入物体｜核心改动怎样工作](/supplement.html?return=l3-r2-p55/0#/l3-region-language-mechanism/0) 第五讲预备阅读","sha256":"08fc293b9ead2c6db138ace3a6118cf7a6458ed5a252f9dcecebd11dc4371bc8","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"language-local-and-question","title":"语言怎样选择当前需要的视觉信息？"},"page":{"id":"vilbert-reusable-joint-start","title":"复用组件以后，还要学习哪些对应？","number":1},"sectionTitle":"复用组件以后，还要学习哪些对应？","currentQuestion":"局部对应、问题条件和联合预训练分别增加了哪条关系？","sourceSlides":["l3-r2-p55"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-f79ffb1115b688ed82d21834","nextBlock":"b-f55597f40aea653725260420","prerequisites":[{"id":"b-f25c5e251ea5388dd47ec28b","url":"#b-f25c5e251ea5388dd47ec28b","role":"section-title"},{"id":"b-9bf5350a1e41b65debd144c3","url":"#b-9bf5350a1e41b65debd144c3","role":"problem-context"}],"relatedBlocks":[{"id":"b-3e17640eb66c3ff78412d9f9","url":"#b-3e17640eb66c3ff78412d9f9","relation":"figure-and-explanation"}]},{"id":"b-f55597f40aea653725260420","type":"li","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p55.html#b-f55597f40aea653725260420","content":"- [区域 token 让语言预训练接入物体｜怎样检查这项解释](/supplement.html?return=l3-r2-p55/0#/l3-region-language-evidence/0) L03补充阅读","sha256":"841759ff9eb6d168b22e5f92a41fd01d6cb50f7fb97162d8df1fed7727375264","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"language-local-and-question","title":"语言怎样选择当前需要的视觉信息？"},"page":{"id":"vilbert-reusable-joint-start","title":"复用组件以后，还要学习哪些对应？","number":1},"sectionTitle":"复用组件以后，还要学习哪些对应？","currentQuestion":"局部对应、问题条件和联合预训练分别增加了哪条关系？","sourceSlides":["l3-r2-p55"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-2a53618f834dc8304db71e52","nextBlock":"b-c4146821e63fe222c7d71cc8","prerequisites":[{"id":"b-f25c5e251ea5388dd47ec28b","url":"#b-f25c5e251ea5388dd47ec28b","role":"section-title"},{"id":"b-9bf5350a1e41b65debd144c3","url":"#b-9bf5350a1e41b65debd144c3","role":"problem-context"}],"relatedBlocks":[{"id":"b-3e17640eb66c3ff78412d9f9","url":"#b-3e17640eb66c3ff78412d9f9","relation":"figure-and-explanation"}]},{"id":"b-c4146821e63fe222c7d71cc8","type":"li","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p55.html#b-c4146821e63fe222c7d71cc8","content":"- [区域 token 让语言预训练接入物体｜什么结论还不能推出](/supplement.html?return=l3-r2-p55/0#/l3-region-language-boundary/0) L03补充阅读","sha256":"d24658903a9d2440ee1547f395309286cb95a5813edb1a71ad8c32febd199825","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"language-local-and-question","title":"语言怎样选择当前需要的视觉信息？"},"page":{"id":"vilbert-reusable-joint-start","title":"复用组件以后，还要学习哪些对应？","number":1},"sectionTitle":"复用组件以后，还要学习哪些对应？","currentQuestion":"局部对应、问题条件和联合预训练分别增加了哪条关系？","sourceSlides":["l3-r2-p55"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-f55597f40aea653725260420","nextBlock":null,"prerequisites":[{"id":"b-f25c5e251ea5388dd47ec28b","url":"#b-f25c5e251ea5388dd47ec28b","role":"section-title"},{"id":"b-9bf5350a1e41b65debd144c3","url":"#b-9bf5350a1e41b65debd144c3","role":"problem-context"}],"relatedBlocks":[{"id":"b-3e17640eb66c3ff78412d9f9","url":"#b-3e17640eb66c3ff78412d9f9","relation":"figure-and-explanation"}]}]}
