{"version":2,"relationshipURLBase":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p18.html","relationshipResolution":"Resolve relative relationship URLs against relationshipURLBase; complete records and absolute URLs are in blocks.","title":"方法不是一条封闭的家谱","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p18.html","markdown":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p18.md","pages":[{"id":"h-i-methods-as-components","title":"把方法拆成可组合的设计选择","unit":"怎样增加候选，同时让历史表示仍可比较？","unitId":"i-dictionary-and-updates","chapter":"I","number":1},{"id":"h-i-moco-transfer-evidence","title":"在固定检测协议下检验预训练起点","unit":"怎样增加候选，同时让历史表示仍可比较？","unitId":"i-dictionary-and-updates","chapter":"I","number":2},{"id":"h-i-teacher-targets","title":"选读：模型提供软目标，为什么仍须检查坍塌","unit":"表示学习的选读问题","unitId":"i-extensions","chapter":"I","number":3}],"aliases":[{"oldId":"explanation","targetId":"h-i-methods-as-components","operation":"split-explanation-range"},{"oldId":"b-e6f84f19714c597952bb2bd8","targetId":"b-e6f84f19714c597952bb2bd8","page":"h-i-methods-as-components","operation":"move","beforeSHA256":"d71cf15e707c8a3674b4c90a599efee783cb84da5b8c8236a1d01c3add7b4958","afterSHA256":"d71cf15e707c8a3674b4c90a599efee783cb84da5b8c8236a1d01c3add7b4958","reason":"Same source body in authored semantic page"},{"oldId":"b-6a384b36fbe1edccd87de2ef","targetId":"b-6a384b36fbe1edccd87de2ef","page":"h-i-methods-as-components","operation":"move","beforeSHA256":"055007616c5125d629b5206876dd680a7f57a26012f2fc8aad2519c418272be5","afterSHA256":"055007616c5125d629b5206876dd680a7f57a26012f2fc8aad2519c418272be5","reason":"Same source body in authored semantic page"},{"oldId":"b-871a13db86edaf0aa9656afd","targetId":"b-871a13db86edaf0aa9656afd","page":"h-i-methods-as-components","operation":"move","beforeSHA256":"e979cb7e38da16e85a09f363cea872cd5135c89b0569bb15333ba4cfdc0bc7eb","afterSHA256":"e979cb7e38da16e85a09f363cea872cd5135c89b0569bb15333ba4cfdc0bc7eb","reason":"Same source body in authored semantic page"},{"oldId":"b-aea5eca95ece4096f9aad088","targetId":"b-aea5eca95ece4096f9aad088","page":"h-i-methods-as-components","operation":"rewrite","beforeSHA256":"3958970881c55809c0c3c85416f464e9d1e9bae72ef29804ea3dc98e760e88eb","afterSHA256":"b5ee2f62bac89712804a829bcc09decf26d46e48ac9ff0f45287afe5282d577a","reason":"Same source identity with figure semantics or reading context enriched"},{"oldId":"b-8e3b7665263f0c198e0bcf9d","targetId":"b-8e3b7665263f0c198e0bcf9d","page":"h-i-methods-as-components","operation":"move","beforeSHA256":"786ce10a6b32a492259a3dccd0a54e61cd41d28a34382fb7b5386c5f03b7ad69","afterSHA256":"786ce10a6b32a492259a3dccd0a54e61cd41d28a34382fb7b5386c5f03b7ad69","reason":"Same source body in authored semantic page"},{"oldId":"b-d70496fb9866e47ada5b5ee5","targetId":"b-d70496fb9866e47ada5b5ee5","page":"h-i-methods-as-components","operation":"move","beforeSHA256":"e33db0c7f5dc9b9db36b8f4aa0f4c22def2e09ca664350a8790d1429fefb18de","afterSHA256":"e33db0c7f5dc9b9db36b8f4aa0f4c22def2e09ca664350a8790d1429fefb18de","reason":"Same source body in authored semantic page"},{"oldId":"b-fe41d016342e8e76e32651ba","targetId":"b-fe41d016342e8e76e32651ba","page":"h-i-methods-as-components","operation":"move","beforeSHA256":"13ccb5fbeb689c6affb0b977506168ac110f6d7a364f9ebaaf1cb05b7a066556","afterSHA256":"13ccb5fbeb689c6affb0b977506168ac110f6d7a364f9ebaaf1cb05b7a066556","reason":"Same source body in authored semantic page"},{"oldId":"b-63ea6332eb6e03ff7a446cff","targetId":"b-63ea6332eb6e03ff7a446cff","page":"h-i-methods-as-components","operation":"move","beforeSHA256":"f26440f021666ee5697fc4009129450550a397b3dfbfade52b4eb71cd4dd059e","afterSHA256":"f26440f021666ee5697fc4009129450550a397b3dfbfade52b4eb71cd4dd059e","reason":"Same source body in authored semantic page"},{"oldId":"b-90ace3936d543b6c6d4aafcc","targetId":"b-90ace3936d543b6c6d4aafcc","page":"h-i-methods-as-components","operation":"move","beforeSHA256":"d2c0f7b895003028241fd58efc5b66f744ac79c984346a1d034cf151ea15e5c8","afterSHA256":"d2c0f7b895003028241fd58efc5b66f744ac79c984346a1d034cf151ea15e5c8","reason":"Same source body in authored semantic page"},{"oldId":"b-33e0bac7df6fd0582f92ad00","targetId":"b-33e0bac7df6fd0582f92ad00","page":"h-i-methods-as-components","operation":"move","beforeSHA256":"4298e64901af0b0ca244807d2e394e8d9182f3939d366cac30b4befe5ad99468","afterSHA256":"4298e64901af0b0ca244807d2e394e8d9182f3939d366cac30b4befe5ad99468","reason":"Same source body in authored semantic page"},{"oldId":"b-ff7fd4d940ae32455bdfea5f","targetId":"b-ff7fd4d940ae32455bdfea5f","page":"h-i-methods-as-components","operation":"move","beforeSHA256":"55846b29249acb632e0183a288f48e49566b6efcac196bd218ba45670ddef347","afterSHA256":"55846b29249acb632e0183a288f48e49566b6efcac196bd218ba45670ddef347","reason":"Same source body in authored semantic page"},{"oldId":"b-089b472aa4a6126178b158bf","targetId":"b-089b472aa4a6126178b158bf","page":"h-i-moco-transfer-evidence","operation":"move","beforeSHA256":"228000ca64ef7fe3e416181616a81dd94d5c61c1fe99b6f008847339fdee4d48","afterSHA256":"228000ca64ef7fe3e416181616a81dd94d5c61c1fe99b6f008847339fdee4d48","reason":"Same source body in authored semantic page"},{"oldId":"b-fc2f917a928127f3b80cfd89","targetId":"b-fc2f917a928127f3b80cfd89","page":"h-i-moco-transfer-evidence","operation":"move","beforeSHA256":"38a51d52dbda974eaf4fab05b73bdc5677b90f775950b000d2ec7cec0518fdc2","afterSHA256":"38a51d52dbda974eaf4fab05b73bdc5677b90f775950b000d2ec7cec0518fdc2","reason":"Same source body in authored semantic page"},{"oldId":"b-95c9a5a5f7f6b8fb49ca7b7d","targetId":"b-95c9a5a5f7f6b8fb49ca7b7d","page":"h-i-moco-transfer-evidence","operation":"rewrite","beforeSHA256":"8c3082b7ba11acc17d24679b827e9d7af44b0d940b49637eb47b882e1a6bde20","afterSHA256":"05be0d423c5b39f2132c94108a586c9d62c94be605786e964a2ae2455d16feca","reason":"Same source identity with figure semantics or reading context enriched"},{"oldId":"b-16468308b00322e7bc30a8d9","targetId":"b-16468308b00322e7bc30a8d9","page":"h-i-moco-transfer-evidence","operation":"move","beforeSHA256":"e7b926bf7cbc1fa25056b061f25eefaf0bba37b86c7983c3d39257ed11e0a513","afterSHA256":"e7b926bf7cbc1fa25056b061f25eefaf0bba37b86c7983c3d39257ed11e0a513","reason":"Same source body in authored semantic page"},{"oldId":"b-9d4317928eca9835ed8dd679","targetId":"b-9d4317928eca9835ed8dd679","page":"h-i-moco-transfer-evidence","operation":"move","beforeSHA256":"53fd3e352a51589d7970016133506a18b796dbfa05c222caa0740be98bbf51d8","afterSHA256":"53fd3e352a51589d7970016133506a18b796dbfa05c222caa0740be98bbf51d8","reason":"Same source body in authored semantic page"},{"oldId":"b-b7c33acf755c9e6d94c614a9","targetId":"b-b7c33acf755c9e6d94c614a9","page":"h-i-teacher-targets","operation":"move","beforeSHA256":"5bc51a0035347385ce8779fee8965d69c21f49b09a5aa430efbabc6411d7d2b8","afterSHA256":"5bc51a0035347385ce8779fee8965d69c21f49b09a5aa430efbabc6411d7d2b8","reason":"Same source body in authored semantic page"},{"oldId":"b-47201491ff9c40d647825106","targetId":"b-47201491ff9c40d647825106","page":"h-i-teacher-targets","operation":"move","beforeSHA256":"9b40607f74227610df31c60a34afc2bb8e212dabce62a2af04fcbed1aae8c1b1","afterSHA256":"9b40607f74227610df31c60a34afc2bb8e212dabce62a2af04fcbed1aae8c1b1","reason":"Same source body in authored semantic page"},{"oldId":"b-0c21dd015f4dc5c0981fbf9f","targetId":"b-0c21dd015f4dc5c0981fbf9f","page":"h-i-teacher-targets","operation":"move","beforeSHA256":"2144c511d94c649203a4096a60898b9104ed9600b5c126c135f958212ed1182a","afterSHA256":"2144c511d94c649203a4096a60898b9104ed9600b5c126c135f958212ed1182a","reason":"Same source body in authored semantic page"},{"oldId":"b-c240f92805ebb2be103063ee","targetId":"b-c240f92805ebb2be103063ee","page":"h-i-teacher-targets","operation":"move","beforeSHA256":"2d86e6daf5520a7166809b03cc90d62ffda26ba3b362d7fbcd5e33bb0bbd9cc3","afterSHA256":"2d86e6daf5520a7166809b03cc90d62ffda26ba3b362d7fbcd5e33bb0bbd9cc3","reason":"Same source body in authored semantic page"},{"oldId":"b-aa715798dd77c91da84ec96c","targetId":"b-aa715798dd77c91da84ec96c","page":"h-i-teacher-targets","operation":"move","beforeSHA256":"cf44fcc62782fb4f6bdc42c16c5187d57556a03cb7ce6065e7bd9c36d8e7d733","afterSHA256":"cf44fcc62782fb4f6bdc42c16c5187d57556a03cb7ce6065e7bd9c36d8e7d733","reason":"Same source body in authored semantic page"},{"oldId":"b-7e7b4e3a21073adb4a3bc2e0","targetId":"b-7e7b4e3a21073adb4a3bc2e0","page":"h-i-teacher-targets","operation":"move","beforeSHA256":"85efc50dc30437f045e5bd3ecfedda789345f9520e583339f496448663547d5a","afterSHA256":"85efc50dc30437f045e5bd3ecfedda789345f9520e583339f496448663547d5a","reason":"Same source body in authored semantic page"},{"oldId":"h-i-methods-as-components-title","targetId":"b-5e63df92f83c2f0bbed2ed47","operation":"retain-source-identity"},{"oldId":"h-i-moco-transfer-evidence-title","targetId":"b-98c9b52ee9ba2898e354fa52","operation":"retain-source-identity"},{"oldId":"h-i-moco-transfer-evidence-bridge","targetId":"b-f24247a8c7ac81a233d3ef91","operation":"retain-source-identity"},{"oldId":"h-i-teacher-targets-title","targetId":"b-b13747ce754b434cc5ad6d6f","operation":"retain-source-identity"},{"oldId":"h-i-teacher-targets-bridge","targetId":"b-f77ded2c0c3a92a77e3f97d1","operation":"retain-source-identity"}],"blocks":[{"id":"b-5e63df92f83c2f0bbed2ed47","type":"h2","scope":"h-i-methods-as-components","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p18.html#b-5e63df92f83c2f0bbed2ed47","content":"## 把方法拆成可组合的设计选择","sha256":"3d039df48d8dcfbc7a3b6da037cf344dd22524e9273d9631eb953d5576a6b877","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-dictionary-and-updates","title":"怎样增加候选，同时让历史表示仍可比较？"},"page":{"id":"h-i-methods-as-components","title":"把方法拆成可组合的设计选择","number":1},"sectionTitle":"把方法拆成可组合的设计选择","currentQuestion":"怎样增加候选，同时让历史表示仍可比较？","sourceSlides":["l3-r2-p18"],"knowledgeBlock":null,"formulas":[],"previousBlock":null,"nextBlock":"b-e6f84f19714c597952bb2bd8","prerequisites":[{"id":"b-871a13db86edaf0aa9656afd","url":"#b-871a13db86edaf0aa9656afd","role":"problem-context"}],"relatedBlocks":[{"id":"b-aea5eca95ece4096f9aad088","url":"#b-aea5eca95ece4096f9aad088","relation":"figure-and-explanation"}]},{"id":"b-e6f84f19714c597952bb2bd8","type":"h2","scope":"explanation","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p18.html#b-e6f84f19714c597952bb2bd8","content":"## 解释、算例与证据","sha256":"d71cf15e707c8a3674b4c90a599efee783cb84da5b8c8236a1d01c3add7b4958","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-dictionary-and-updates","title":"怎样增加候选，同时让历史表示仍可比较？"},"page":{"id":"h-i-methods-as-components","title":"把方法拆成可组合的设计选择","number":1},"sectionTitle":"把方法拆成可组合的设计选择","currentQuestion":"怎样增加候选，同时让历史表示仍可比较？","sourceSlides":["l3-r2-p18"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-5e63df92f83c2f0bbed2ed47","nextBlock":"b-6a384b36fbe1edccd87de2ef","prerequisites":[{"id":"b-5e63df92f83c2f0bbed2ed47","url":"#b-5e63df92f83c2f0bbed2ed47","role":"section-title"},{"id":"b-871a13db86edaf0aa9656afd","url":"#b-871a13db86edaf0aa9656afd","role":"problem-context"}],"relatedBlocks":[{"id":"b-aea5eca95ece4096f9aad088","url":"#b-aea5eca95ece4096f9aad088","relation":"figure-and-explanation"}]},{"id":"b-6a384b36fbe1edccd87de2ef","type":"h1","scope":"document","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p18.html#b-6a384b36fbe1edccd87de2ef","content":"# P18 · 方法不是一条封闭的家谱","sha256":"055007616c5125d629b5206876dd680a7f57a26012f2fc8aad2519c418272be5","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-dictionary-and-updates","title":"怎样增加候选，同时让历史表示仍可比较？"},"page":{"id":"h-i-methods-as-components","title":"把方法拆成可组合的设计选择","number":1},"sectionTitle":"把方法拆成可组合的设计选择","currentQuestion":"怎样增加候选，同时让历史表示仍可比较？","sourceSlides":["l3-r2-p18"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-e6f84f19714c597952bb2bd8","nextBlock":"b-871a13db86edaf0aa9656afd","prerequisites":[{"id":"b-e6f84f19714c597952bb2bd8","url":"#b-e6f84f19714c597952bb2bd8","role":"section-title"},{"id":"b-871a13db86edaf0aa9656afd","url":"#b-871a13db86edaf0aa9656afd","role":"problem-context"}],"relatedBlocks":[{"id":"b-aea5eca95ece4096f9aad088","url":"#b-aea5eca95ece4096f9aad088","relation":"figure-and-explanation"}]},{"id":"b-871a13db86edaf0aa9656afd","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p18.html#b-871a13db86edaf0aa9656afd","content":"收束对比路线：方法组件可以组合，是否有用还要看下游证据。由此转向不同目标的重建路线。","sha256":"e979cb7e38da16e85a09f363cea872cd5135c89b0569bb15333ba4cfdc0bc7eb","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-dictionary-and-updates","title":"怎样增加候选，同时让历史表示仍可比较？"},"page":{"id":"h-i-methods-as-components","title":"把方法拆成可组合的设计选择","number":1},"sectionTitle":"把方法拆成可组合的设计选择","currentQuestion":"怎样增加候选，同时让历史表示仍可比较？","sourceSlides":["l3-r2-p18"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-6a384b36fbe1edccd87de2ef","nextBlock":"b-aea5eca95ece4096f9aad088","prerequisites":[{"id":"b-6a384b36fbe1edccd87de2ef","url":"#b-6a384b36fbe1edccd87de2ef","role":"section-title"}],"relatedBlocks":[{"id":"b-aea5eca95ece4096f9aad088","url":"#b-aea5eca95ece4096f9aad088","relation":"figure-and-explanation"}]},{"id":"b-aea5eca95ece4096f9aad088","type":"figure","scope":"document","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p18.html#b-aea5eca95ece4096f9aad088","content":"图内标签与数值：P18 学习表示机制图，状态 1；2019.11；MoCo；2020.02；SimCLR；2020.03；MoCo v2；队列 + 慢编码器；非线性投影头 + 增强；设计汇合；比较的不只是名字：目标、增强、头、字典、评价协议。；课堂先讲 SimCLR 便于理解；发表顺序与教学顺序不同。\n\nMoCo 2019.11、SimCLR 2020.02、MoCo v2 2020.03时间带；队列/EMA与投影头/增强两路汇合。\n\n- 2019.11\n- MoCo\n- 2020.02\n- SimCLR\n- 2020.03\n- MoCo v2\n- 队列 + 慢编码器\n- 非线性投影头 + 增强\n- 设计汇合\n- 比较的不只是名字：目标、增强、头、字典、评价协议。\n- 课堂先讲 SimCLR 便于理解；发表顺序与教学顺序不同。MoCo 2019.11、SimCLR 2020.02、MoCo v2 2020.03时间带；队列/EMA与投影头/增强两路汇合。 [查看对应图解状态](/#/l3-r2-p18/0)[查看完整图解](/course/figures/l3-r2-p18-0.html)","sha256":"b5ee2f62bac89712804a829bcc09decf26d46e48ac9ff0f45287afe5282d577a","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-dictionary-and-updates","title":"怎样增加候选，同时让历史表示仍可比较？"},"page":{"id":"h-i-methods-as-components","title":"把方法拆成可组合的设计选择","number":1},"sectionTitle":"把方法拆成可组合的设计选择","currentQuestion":"怎样增加候选，同时让历史表示仍可比较？","sourceSlides":["l3-r2-p18"],"knowledgeBlock":null,"formulas":[],"figure":{"caption":"MoCo 2019.11、SimCLR 2020.02、MoCo v2 2020.03时间带；队列/EMA与投影头/增强两路汇合。 查看对应图解状态","labels":["2019.11","MoCo","2020.02","SimCLR","2020.03","MoCo v2","队列 + 慢编码器","非线性投影头 + 增强","设计汇合","比较的不只是名字：目标、增强、头、字典、评价协议。","课堂先讲 SimCLR 便于理解；发表顺序与教学顺序不同。"],"description":"MoCo 2019.11、SimCLR 2020.02、MoCo v2 2020.03时间带；队列/EMA与投影头/增强两路汇合。2019.11MoCo2020.02SimCLR2020.03MoCo v2队列 + 慢编码器非线性投影头 + 增强设计汇合比较的不只是名字：目标、增强、头、字典、评价协议。课堂先讲 SimCLR 便于理解；发表顺序与教学顺序不同。","evidenceIdentity":"MoCo 2019.11、SimCLR 2020.02、MoCo v2 2020.03时间带；队列/EMA与投影头/增强两路汇合。 查看对应图解状态","images":[]},"previousBlock":"b-871a13db86edaf0aa9656afd","nextBlock":"b-8e3b7665263f0c198e0bcf9d","prerequisites":[{"id":"b-6a384b36fbe1edccd87de2ef","url":"#b-6a384b36fbe1edccd87de2ef","role":"section-title"},{"id":"b-871a13db86edaf0aa9656afd","url":"#b-871a13db86edaf0aa9656afd","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-8e3b7665263f0c198e0bcf9d","type":"h2","scope":"narrative","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p18.html#b-8e3b7665263f0c198e0bcf9d","content":"## 课堂讲述","sha256":"786ce10a6b32a492259a3dccd0a54e61cd41d28a34382fb7b5386c5f03b7ad69","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-dictionary-and-updates","title":"怎样增加候选，同时让历史表示仍可比较？"},"page":{"id":"h-i-methods-as-components","title":"把方法拆成可组合的设计选择","number":1},"sectionTitle":"把方法拆成可组合的设计选择","currentQuestion":"怎样增加候选，同时让历史表示仍可比较？","sourceSlides":["l3-r2-p18"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-aea5eca95ece4096f9aad088","nextBlock":"b-d70496fb9866e47ada5b5ee5","prerequisites":[{"id":"b-6a384b36fbe1edccd87de2ef","url":"#b-6a384b36fbe1edccd87de2ef","role":"section-title"},{"id":"b-871a13db86edaf0aa9656afd","url":"#b-871a13db86edaf0aa9656afd","role":"problem-context"}],"relatedBlocks":[{"id":"b-aea5eca95ece4096f9aad088","url":"#b-aea5eca95ece4096f9aad088","relation":"figure-and-explanation"}]},{"id":"b-d70496fb9866e47ada5b5ee5","type":"p","scope":"narrative","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p18.html#b-d70496fb9866e47ada5b5ee5","content":"刚才为了讲明白计算取舍，我们先用了 SimCLR 的图。这不是历史发表顺序：MoCo 的预印本在 2019 年，SimCLR 在 2020 年，之后的 MoCo v2 明确吸收了 SimCLR 中的非线性投影头和增强设计。\nMoCo 原文还比较了多项检测与分割迁移任务，而不只报分类读出的结果。这使“不依赖逐图类别标签也能获得可用视觉表示”有了下游证据。\n这是一个很好的研究案例。持续的问题是怎样学习可迁移表示，但一个框架不必只沿着自己的名字演化；别的工作发现的有效设计，可以被吸收进去，再看它在新的组织方式下是否成立。\n所以读论文时，不只看方法缩写。把目标、增强、投影头、字典组织和评价协议拆开，才能看见到底换了什么。何恺明及合作者在这里提供的是持续追问问题的线索，不是一份把所有方法归到同一人名下的作品目录。","sha256":"e33db0c7f5dc9b9db36b8f4aa0f4c22def2e09ca664350a8790d1429fefb18de","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-dictionary-and-updates","title":"怎样增加候选，同时让历史表示仍可比较？"},"page":{"id":"h-i-methods-as-components","title":"把方法拆成可组合的设计选择","number":1},"sectionTitle":"把方法拆成可组合的设计选择","currentQuestion":"怎样增加候选，同时让历史表示仍可比较？","sourceSlides":["l3-r2-p18"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-8e3b7665263f0c198e0bcf9d","nextBlock":"b-fe41d016342e8e76e32651ba","prerequisites":[{"id":"b-8e3b7665263f0c198e0bcf9d","url":"#b-8e3b7665263f0c198e0bcf9d","role":"section-title"},{"id":"b-871a13db86edaf0aa9656afd","url":"#b-871a13db86edaf0aa9656afd","role":"problem-context"}],"relatedBlocks":[{"id":"b-aea5eca95ece4096f9aad088","url":"#b-aea5eca95ece4096f9aad088","relation":"figure-and-explanation"}]},{"id":"b-fe41d016342e8e76e32651ba","type":"h3","scope":"contrastive-designs-can-recombine","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p18.html#b-fe41d016342e8e76e32651ba","content":"### 方法是可组合的设计，不是封闭家谱","sha256":"13ccb5fbeb689c6affb0b977506168ac110f6d7a364f9ebaaf1cb05b7a066556","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-dictionary-and-updates","title":"怎样增加候选，同时让历史表示仍可比较？"},"page":{"id":"h-i-methods-as-components","title":"把方法拆成可组合的设计选择","number":1},"sectionTitle":"方法是可组合的设计，不是封闭家谱","currentQuestion":"怎样增加候选，同时让历史表示仍可比较？","sourceSlides":["l3-r2-p18"],"knowledgeBlock":"contrastive-designs-can-recombine","formulas":[],"previousBlock":"b-d70496fb9866e47ada5b5ee5","nextBlock":"b-63ea6332eb6e03ff7a446cff","prerequisites":[{"id":"b-8e3b7665263f0c198e0bcf9d","url":"#b-8e3b7665263f0c198e0bcf9d","role":"section-title"},{"id":"b-871a13db86edaf0aa9656afd","url":"#b-871a13db86edaf0aa9656afd","role":"problem-context"}],"relatedBlocks":[{"id":"b-aea5eca95ece4096f9aad088","url":"#b-aea5eca95ece4096f9aad088","relation":"figure-and-explanation"},{"id":"b-63ea6332eb6e03ff7a446cff","url":"#b-63ea6332eb6e03ff7a446cff","relation":"same-reasoning"},{"id":"b-90ace3936d543b6c6d4aafcc","url":"#b-90ace3936d543b6c6d4aafcc","relation":"same-reasoning"},{"id":"b-33e0bac7df6fd0582f92ad00","url":"#b-33e0bac7df6fd0582f92ad00","relation":"same-reasoning"}]},{"id":"b-63ea6332eb6e03ff7a446cff","type":"p","scope":"contrastive-designs-can-recombine","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p18.html#b-63ea6332eb6e03ff7a446cff","content":"MoCo突出历史字典与动量编码器，SimCLR突出批内两视图比较及投影头等设计。它们不是互相排斥的整套身份：后续方法可以把一方有效的组件放入另一种计算组织中。","sha256":"f26440f021666ee5697fc4009129450550a397b3dfbfade52b4eb71cd4dd059e","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-dictionary-and-updates","title":"怎样增加候选，同时让历史表示仍可比较？"},"page":{"id":"h-i-methods-as-components","title":"把方法拆成可组合的设计选择","number":1},"sectionTitle":"方法是可组合的设计，不是封闭家谱","currentQuestion":"怎样增加候选，同时让历史表示仍可比较？","sourceSlides":["l3-r2-p18"],"knowledgeBlock":"contrastive-designs-can-recombine","formulas":[],"previousBlock":"b-fe41d016342e8e76e32651ba","nextBlock":"b-90ace3936d543b6c6d4aafcc","prerequisites":[{"id":"b-fe41d016342e8e76e32651ba","url":"#b-fe41d016342e8e76e32651ba","role":"section-title"},{"id":"b-871a13db86edaf0aa9656afd","url":"#b-871a13db86edaf0aa9656afd","role":"problem-context"}],"relatedBlocks":[{"id":"b-aea5eca95ece4096f9aad088","url":"#b-aea5eca95ece4096f9aad088","relation":"figure-and-explanation"},{"id":"b-fe41d016342e8e76e32651ba","url":"#b-fe41d016342e8e76e32651ba","relation":"same-reasoning"},{"id":"b-90ace3936d543b6c6d4aafcc","url":"#b-90ace3936d543b6c6d4aafcc","relation":"same-reasoning"},{"id":"b-33e0bac7df6fd0582f92ad00","url":"#b-33e0bac7df6fd0582f92ad00","relation":"same-reasoning"}]},{"id":"b-90ace3936d543b6c6d4aafcc","type":"p","scope":"contrastive-designs-can-recombine","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p18.html#b-90ace3936d543b6c6d4aafcc","content":"这种看法更有助于阅读论文。把一套方法拆成数据与增强、表示网络、目标、候选组织、更新规则、评价协议六个部分，再问新论文究竟改了哪一部分。若同时改了增强、训练时长、模型和loss，最终提高可以是有效工程组合，但不能直接把全部提高归给一个名字。","sha256":"d2c0f7b895003028241fd58efc5b66f744ac79c984346a1d034cf151ea15e5c8","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-dictionary-and-updates","title":"怎样增加候选，同时让历史表示仍可比较？"},"page":{"id":"h-i-methods-as-components","title":"把方法拆成可组合的设计选择","number":1},"sectionTitle":"方法是可组合的设计，不是封闭家谱","currentQuestion":"怎样增加候选，同时让历史表示仍可比较？","sourceSlides":["l3-r2-p18"],"knowledgeBlock":"contrastive-designs-can-recombine","formulas":[],"previousBlock":"b-63ea6332eb6e03ff7a446cff","nextBlock":"b-33e0bac7df6fd0582f92ad00","prerequisites":[{"id":"b-fe41d016342e8e76e32651ba","url":"#b-fe41d016342e8e76e32651ba","role":"section-title"},{"id":"b-871a13db86edaf0aa9656afd","url":"#b-871a13db86edaf0aa9656afd","role":"problem-context"}],"relatedBlocks":[{"id":"b-aea5eca95ece4096f9aad088","url":"#b-aea5eca95ece4096f9aad088","relation":"figure-and-explanation"},{"id":"b-fe41d016342e8e76e32651ba","url":"#b-fe41d016342e8e76e32651ba","relation":"same-reasoning"},{"id":"b-63ea6332eb6e03ff7a446cff","url":"#b-63ea6332eb6e03ff7a446cff","relation":"same-reasoning"},{"id":"b-33e0bac7df6fd0582f92ad00","url":"#b-33e0bac7df6fd0582f92ad00","relation":"same-reasoning"}]},{"id":"b-33e0bac7df6fd0582f92ad00","type":"p","scope":"contrastive-designs-can-recombine","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p18.html#b-33e0bac7df6fd0582f92ad00","content":"正面进展也应看清：这些设计使没有人工类别目标的图像学习成为可迁移起点。下一块用MoCo原论文的检测微调结果展示这件事，随后转到另一条不同目标路线——预测缺失像素。","sha256":"4298e64901af0b0ca244807d2e394e8d9182f3939d366cac30b4befe5ad99468","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-dictionary-and-updates","title":"怎样增加候选，同时让历史表示仍可比较？"},"page":{"id":"h-i-methods-as-components","title":"把方法拆成可组合的设计选择","number":1},"sectionTitle":"方法是可组合的设计，不是封闭家谱","currentQuestion":"怎样增加候选，同时让历史表示仍可比较？","sourceSlides":["l3-r2-p18"],"knowledgeBlock":"contrastive-designs-can-recombine","formulas":[],"previousBlock":"b-90ace3936d543b6c6d4aafcc","nextBlock":"b-23be66d5807658710c312c08","prerequisites":[{"id":"b-fe41d016342e8e76e32651ba","url":"#b-fe41d016342e8e76e32651ba","role":"section-title"},{"id":"b-871a13db86edaf0aa9656afd","url":"#b-871a13db86edaf0aa9656afd","role":"problem-context"}],"relatedBlocks":[{"id":"b-aea5eca95ece4096f9aad088","url":"#b-aea5eca95ece4096f9aad088","relation":"figure-and-explanation"},{"id":"b-fe41d016342e8e76e32651ba","url":"#b-fe41d016342e8e76e32651ba","relation":"same-reasoning"},{"id":"b-63ea6332eb6e03ff7a446cff","url":"#b-63ea6332eb6e03ff7a446cff","relation":"same-reasoning"},{"id":"b-90ace3936d543b6c6d4aafcc","url":"#b-90ace3936d543b6c6d4aafcc","relation":"same-reasoning"}]},{"id":"b-23be66d5807658710c312c08","type":"h4","scope":"h-i-methods-as-components","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p18.html#b-23be66d5807658710c312c08","content":"#### 图示条件与公式约定","sha256":"ef7f915c1ae63ed67bd23fecf050249273ba0efa8965d2c8d1c1c6b2d3178122","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-dictionary-and-updates","title":"怎样增加候选，同时让历史表示仍可比较？"},"page":{"id":"h-i-methods-as-components","title":"把方法拆成可组合的设计选择","number":1},"sectionTitle":"把方法拆成可组合的设计选择","currentQuestion":"怎样增加候选，同时让历史表示仍可比较？","sourceSlides":["l3-r2-p18"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-33e0bac7df6fd0582f92ad00","nextBlock":"b-ff7fd4d940ae32455bdfea5f","prerequisites":[{"id":"b-fe41d016342e8e76e32651ba","url":"#b-fe41d016342e8e76e32651ba","role":"section-title"},{"id":"b-871a13db86edaf0aa9656afd","url":"#b-871a13db86edaf0aa9656afd","role":"problem-context"}],"relatedBlocks":[{"id":"b-aea5eca95ece4096f9aad088","url":"#b-aea5eca95ece4096f9aad088","relation":"figure-and-explanation"}]},{"id":"b-ff7fd4d940ae32455bdfea5f","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p18.html#b-ff7fd4d940ae32455bdfea5f","content":"原始 arXiv 首次提交：MoCo 2019-11-13；SimCLR 2020-02-13；MoCo v2 2020-03-09。v2 明确采用 MLP 投影头与更多增强。本页不展示论文性能表或未核实提升数字。","sha256":"55846b29249acb632e0183a288f48e49566b6efcac196bd218ba45670ddef347","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-dictionary-and-updates","title":"怎样增加候选，同时让历史表示仍可比较？"},"page":{"id":"h-i-methods-as-components","title":"把方法拆成可组合的设计选择","number":1},"sectionTitle":"把方法拆成可组合的设计选择","currentQuestion":"怎样增加候选，同时让历史表示仍可比较？","sourceSlides":["l3-r2-p18"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-23be66d5807658710c312c08","nextBlock":null,"prerequisites":[{"id":"b-23be66d5807658710c312c08","url":"#b-23be66d5807658710c312c08","role":"section-title"},{"id":"b-871a13db86edaf0aa9656afd","url":"#b-871a13db86edaf0aa9656afd","role":"problem-context"}],"relatedBlocks":[{"id":"b-aea5eca95ece4096f9aad088","url":"#b-aea5eca95ece4096f9aad088","relation":"figure-and-explanation"}]},{"id":"b-98c9b52ee9ba2898e354fa52","type":"h2","scope":"h-i-moco-transfer-evidence","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p18.html#b-98c9b52ee9ba2898e354fa52","content":"## 在固定检测协议下检验预训练起点","sha256":"f5331b17d6cee41632d8b1b2a7fb1806fd946b54b9d71273438c5c79128ea062","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-dictionary-and-updates","title":"怎样增加候选，同时让历史表示仍可比较？"},"page":{"id":"h-i-moco-transfer-evidence","title":"在固定检测协议下检验预训练起点","number":2},"sectionTitle":"在固定检测协议下检验预训练起点","currentQuestion":"怎样增加候选，同时让历史表示仍可比较？","sourceSlides":["l3-r2-p18","l3-r2-p24"],"knowledgeBlock":null,"formulas":[],"previousBlock":null,"nextBlock":"b-f24247a8c7ac81a233d3ef91","prerequisites":[{"id":"b-f24247a8c7ac81a233d3ef91","url":"#b-f24247a8c7ac81a233d3ef91","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-f24247a8c7ac81a233d3ef91","type":"p","scope":"h-i-moco-transfer-evidence","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p18.html#b-f24247a8c7ac81a233d3ef91","content":"比较同一下游检测器的两种预训练起点，而不是要求冻结编码器直接输出框。","sha256":"87e1bd88ab4b41875dd5a9e825c28d17b9ee6c7ede6d0f697125074de4b26ecf","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-dictionary-and-updates","title":"怎样增加候选，同时让历史表示仍可比较？"},"page":{"id":"h-i-moco-transfer-evidence","title":"在固定检测协议下检验预训练起点","number":2},"sectionTitle":"在固定检测协议下检验预训练起点","currentQuestion":"怎样增加候选，同时让历史表示仍可比较？","sourceSlides":["l3-r2-p18","l3-r2-p24"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-98c9b52ee9ba2898e354fa52","nextBlock":"b-089b472aa4a6126178b158bf","prerequisites":[{"id":"b-98c9b52ee9ba2898e354fa52","url":"#b-98c9b52ee9ba2898e354fa52","role":"section-title"}],"relatedBlocks":[]},{"id":"b-089b472aa4a6126178b158bf","type":"h3","scope":"moco-voc-transfer-evidence","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p18.html#b-089b472aa4a6126178b158bf","content":"### MoCo 的表示怎样支持检测迁移？","sha256":"228000ca64ef7fe3e416181616a81dd94d5c61c1fe99b6f008847339fdee4d48","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-dictionary-and-updates","title":"怎样增加候选，同时让历史表示仍可比较？"},"page":{"id":"h-i-moco-transfer-evidence","title":"在固定检测协议下检验预训练起点","number":2},"sectionTitle":"MoCo 的表示怎样支持检测迁移？","currentQuestion":"怎样增加候选，同时让历史表示仍可比较？","sourceSlides":["l3-r2-p18","l3-r2-p24"],"knowledgeBlock":"moco-voc-transfer-evidence","formulas":[],"previousBlock":"b-f24247a8c7ac81a233d3ef91","nextBlock":"b-fc2f917a928127f3b80cfd89","prerequisites":[{"id":"b-98c9b52ee9ba2898e354fa52","url":"#b-98c9b52ee9ba2898e354fa52","role":"section-title"},{"id":"b-f24247a8c7ac81a233d3ef91","url":"#b-f24247a8c7ac81a233d3ef91","role":"problem-context"}],"relatedBlocks":[{"id":"b-fc2f917a928127f3b80cfd89","url":"#b-fc2f917a928127f3b80cfd89","relation":"same-reasoning"},{"id":"b-95c9a5a5f7f6b8fb49ca7b7d","url":"#b-95c9a5a5f7f6b8fb49ca7b7d","relation":"same-reasoning"},{"id":"b-16468308b00322e7bc30a8d9","url":"#b-16468308b00322e7bc30a8d9","relation":"same-reasoning"},{"id":"b-9d4317928eca9835ed8dd679","url":"#b-9d4317928eca9835ed8dd679","relation":"same-reasoning"}]},{"id":"b-fc2f917a928127f3b80cfd89","type":"p","scope":"moco-voc-transfer-evidence","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p18.html#b-fc2f917a928127f3b80cfd89","content":"MoCo原论文Table 2(b)在同一Faster R-CNN、ResNet-50-C4设置下比较预训练起点：","sha256":"38a51d52dbda974eaf4fab05b73bdc5677b90f775950b000d2ec7cec0518fdc2","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-dictionary-and-updates","title":"怎样增加候选，同时让历史表示仍可比较？"},"page":{"id":"h-i-moco-transfer-evidence","title":"在固定检测协议下检验预训练起点","number":2},"sectionTitle":"MoCo 的表示怎样支持检测迁移？","currentQuestion":"怎样增加候选，同时让历史表示仍可比较？","sourceSlides":["l3-r2-p18","l3-r2-p24"],"knowledgeBlock":"moco-voc-transfer-evidence","formulas":[],"previousBlock":"b-089b472aa4a6126178b158bf","nextBlock":"b-95c9a5a5f7f6b8fb49ca7b7d","prerequisites":[{"id":"b-089b472aa4a6126178b158bf","url":"#b-089b472aa4a6126178b158bf","role":"section-title"},{"id":"b-f24247a8c7ac81a233d3ef91","url":"#b-f24247a8c7ac81a233d3ef91","role":"problem-context"}],"relatedBlocks":[{"id":"b-089b472aa4a6126178b158bf","url":"#b-089b472aa4a6126178b158bf","relation":"same-reasoning"},{"id":"b-95c9a5a5f7f6b8fb49ca7b7d","url":"#b-95c9a5a5f7f6b8fb49ca7b7d","relation":"same-reasoning"},{"id":"b-16468308b00322e7bc30a8d9","url":"#b-16468308b00322e7bc30a8d9","relation":"same-reasoning"},{"id":"b-9d4317928eca9835ed8dd679","url":"#b-9d4317928eca9835ed8dd679","relation":"same-reasoning"}]},{"id":"b-95c9a5a5f7f6b8fb49ca7b7d","type":"table","scope":"moco-voc-transfer-evidence","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p18.html#b-95c9a5a5f7f6b8fb49ca7b7d","content":"| 预训练 | box AP | AP50 | AP75 |\n| --- | --- | --- | --- |\n| ImageNet-1M 人工类别监督 | 53.5 | 81.3 | 58.8 |\n| ImageNet-1M MoCo | 55.9 | 81.5 | 62.6 |","sha256":"05be0d423c5b39f2132c94108a586c9d62c94be605786e964a2ae2455d16feca","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-dictionary-and-updates","title":"怎样增加候选，同时让历史表示仍可比较？"},"page":{"id":"h-i-moco-transfer-evidence","title":"在固定检测协议下检验预训练起点","number":2},"sectionTitle":"MoCo 的表示怎样支持检测迁移？","currentQuestion":"怎样增加候选，同时让历史表示仍可比较？","sourceSlides":["l3-r2-p18","l3-r2-p24"],"knowledgeBlock":"moco-voc-transfer-evidence","formulas":[],"table":{"headers":["预训练","box AP","AP50","AP75"],"rows":[["预训练","box AP","AP50","AP75"],["ImageNet-1M 人工类别监督","53.5","81.3","58.8"],["ImageNet-1M MoCo","55.9","81.5","62.6"]]},"previousBlock":"b-fc2f917a928127f3b80cfd89","nextBlock":"b-16468308b00322e7bc30a8d9","prerequisites":[{"id":"b-089b472aa4a6126178b158bf","url":"#b-089b472aa4a6126178b158bf","role":"section-title"},{"id":"b-f24247a8c7ac81a233d3ef91","url":"#b-f24247a8c7ac81a233d3ef91","role":"problem-context"}],"relatedBlocks":[{"id":"b-089b472aa4a6126178b158bf","url":"#b-089b472aa4a6126178b158bf","relation":"same-reasoning"},{"id":"b-fc2f917a928127f3b80cfd89","url":"#b-fc2f917a928127f3b80cfd89","relation":"same-reasoning"},{"id":"b-16468308b00322e7bc30a8d9","url":"#b-16468308b00322e7bc30a8d9","relation":"same-reasoning"},{"id":"b-9d4317928eca9835ed8dd679","url":"#b-9d4317928eca9835ed8dd679","relation":"same-reasoning"}]},{"id":"b-16468308b00322e7bc30a8d9","type":"p","scope":"moco-voc-transfer-evidence","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p18.html#b-16468308b00322e7bc30a8d9","content":"两者均在VOC trainval07+12检测标注上全层微调24k次，评价为VOC test2007，取5次实验平均；AP是COCO式多IoU阈值指标，AP50/AP75是指定阈值。MoCo相对提高2.4 AP，说明视图对比学出的参数在这套下游协议下提供了有效起点。","sha256":"e7b926bf7cbc1fa25056b061f25eefaf0bba37b86c7983c3d39257ed11e0a513","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-dictionary-and-updates","title":"怎样增加候选，同时让历史表示仍可比较？"},"page":{"id":"h-i-moco-transfer-evidence","title":"在固定检测协议下检验预训练起点","number":2},"sectionTitle":"MoCo 的表示怎样支持检测迁移？","currentQuestion":"怎样增加候选，同时让历史表示仍可比较？","sourceSlides":["l3-r2-p18","l3-r2-p24"],"knowledgeBlock":"moco-voc-transfer-evidence","formulas":[],"previousBlock":"b-95c9a5a5f7f6b8fb49ca7b7d","nextBlock":"b-9d4317928eca9835ed8dd679","prerequisites":[{"id":"b-089b472aa4a6126178b158bf","url":"#b-089b472aa4a6126178b158bf","role":"section-title"},{"id":"b-f24247a8c7ac81a233d3ef91","url":"#b-f24247a8c7ac81a233d3ef91","role":"problem-context"}],"relatedBlocks":[{"id":"b-089b472aa4a6126178b158bf","url":"#b-089b472aa4a6126178b158bf","relation":"same-reasoning"},{"id":"b-fc2f917a928127f3b80cfd89","url":"#b-fc2f917a928127f3b80cfd89","relation":"same-reasoning"},{"id":"b-95c9a5a5f7f6b8fb49ca7b7d","url":"#b-95c9a5a5f7f6b8fb49ca7b7d","relation":"same-reasoning"},{"id":"b-9d4317928eca9835ed8dd679","url":"#b-9d4317928eca9835ed8dd679","relation":"same-reasoning"}]},{"id":"b-9d4317928eca9835ed8dd679","type":"p","scope":"moco-voc-transfer-evidence","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p18.html#b-9d4317928eca9835ed8dd679","content":"这不是冻结模型直接会检测：检测头与主干都用检测监督适配。它也不把所有预训练和下游总算力配平。相同论文的另一种主干结构收益不同，因此结论应保留具体模型与协议。","sha256":"53fd3e352a51589d7970016133506a18b796dbfa05c222caa0740be98bbf51d8","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-dictionary-and-updates","title":"怎样增加候选，同时让历史表示仍可比较？"},"page":{"id":"h-i-moco-transfer-evidence","title":"在固定检测协议下检验预训练起点","number":2},"sectionTitle":"MoCo 的表示怎样支持检测迁移？","currentQuestion":"怎样增加候选，同时让历史表示仍可比较？","sourceSlides":["l3-r2-p18","l3-r2-p24"],"knowledgeBlock":"moco-voc-transfer-evidence","formulas":[],"previousBlock":"b-16468308b00322e7bc30a8d9","nextBlock":null,"prerequisites":[{"id":"b-089b472aa4a6126178b158bf","url":"#b-089b472aa4a6126178b158bf","role":"section-title"},{"id":"b-f24247a8c7ac81a233d3ef91","url":"#b-f24247a8c7ac81a233d3ef91","role":"problem-context"}],"relatedBlocks":[{"id":"b-089b472aa4a6126178b158bf","url":"#b-089b472aa4a6126178b158bf","relation":"same-reasoning"},{"id":"b-fc2f917a928127f3b80cfd89","url":"#b-fc2f917a928127f3b80cfd89","relation":"same-reasoning"},{"id":"b-95c9a5a5f7f6b8fb49ca7b7d","url":"#b-95c9a5a5f7f6b8fb49ca7b7d","relation":"same-reasoning"},{"id":"b-16468308b00322e7bc30a8d9","url":"#b-16468308b00322e7bc30a8d9","relation":"same-reasoning"}]},{"id":"b-b13747ce754b434cc5ad6d6f","type":"h2","scope":"h-i-teacher-targets","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p18.html#b-b13747ce754b434cc5ad6d6f","content":"## 选读：模型提供软目标，为什么仍须检查坍塌","sha256":"c6b00bbc4328c866c122fc2c2bdbed0633ac9f5801d81b0a7f9e1d26dfe7c482","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-extensions","title":"表示学习的选读问题"},"page":{"id":"h-i-teacher-targets","title":"选读：模型提供软目标，为什么仍须检查坍塌","number":3},"sectionTitle":"选读：模型提供软目标，为什么仍须检查坍塌","currentQuestion":"哪些细节需要进一步对照？","sourceSlides":["l3-r2-p11","l3-r2-p18"],"knowledgeBlock":null,"formulas":[],"previousBlock":null,"nextBlock":"b-f77ded2c0c3a92a77e3f97d1","prerequisites":[{"id":"b-f77ded2c0c3a92a77e3f97d1","url":"#b-f77ded2c0c3a92a77e3f97d1","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-f77ded2c0c3a92a77e3f97d1","type":"p","scope":"h-i-teacher-targets","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p18.html#b-f77ded2c0c3a92a77e3f97d1","content":"把固定教师和EMA教师分开；同一组logits改变温度，先看软目标表达了什么。","sha256":"20edd91f7639ef1ddda55ffd9172c953a23346dba35bb4245277e3705ed6c432","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-extensions","title":"表示学习的选读问题"},"page":{"id":"h-i-teacher-targets","title":"选读：模型提供软目标，为什么仍须检查坍塌","number":3},"sectionTitle":"选读：模型提供软目标，为什么仍须检查坍塌","currentQuestion":"哪些细节需要进一步对照？","sourceSlides":["l3-r2-p11","l3-r2-p18"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-b13747ce754b434cc5ad6d6f","nextBlock":"b-b7c33acf755c9e6d94c614a9","prerequisites":[{"id":"b-b13747ce754b434cc5ad6d6f","url":"#b-b13747ce754b434cc5ad6d6f","role":"section-title"}],"relatedBlocks":[]},{"id":"b-b7c33acf755c9e6d94c614a9","type":"h3","scope":"teacher-target-source-and-collapse","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p18.html#b-b7c33acf755c9e6d94c614a9","content":"### 教师模型提供的是哪一种答案？","sha256":"5bc51a0035347385ce8779fee8965d69c21f49b09a5aa430efbabc6411d7d2b8","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-extensions","title":"表示学习的选读问题"},"page":{"id":"h-i-teacher-targets","title":"选读：模型提供软目标，为什么仍须检查坍塌","number":3},"sectionTitle":"教师模型提供的是哪一种答案？","currentQuestion":"哪些细节需要进一步对照？","sourceSlides":["l3-r2-p11","l3-r2-p18"],"knowledgeBlock":"teacher-target-source-and-collapse","formulas":[],"previousBlock":"b-f77ded2c0c3a92a77e3f97d1","nextBlock":"b-47201491ff9c40d647825106","prerequisites":[{"id":"b-b13747ce754b434cc5ad6d6f","url":"#b-b13747ce754b434cc5ad6d6f","role":"section-title"},{"id":"b-f77ded2c0c3a92a77e3f97d1","url":"#b-f77ded2c0c3a92a77e3f97d1","role":"problem-context"}],"relatedBlocks":[{"id":"b-47201491ff9c40d647825106","url":"#b-47201491ff9c40d647825106","relation":"same-reasoning"},{"id":"b-0c21dd015f4dc5c0981fbf9f","url":"#b-0c21dd015f4dc5c0981fbf9f","relation":"same-reasoning"},{"id":"b-c240f92805ebb2be103063ee","url":"#b-c240f92805ebb2be103063ee","relation":"same-reasoning"}]},{"id":"b-47201491ff9c40d647825106","type":"p","scope":"teacher-target-source-and-collapse","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p18.html#b-47201491ff9c40d647825106","content":"固定教师蒸馏让学生拟合一个已有模型的概率分布；EMA教师则来自学生参数的历史平均。这两种教师都不是天然正确的类别真值，也不必比学生更大。","sha256":"9b40607f74227610df31c60a34afc2bb8e212dabce62a2af04fcbed1aae8c1b1","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-extensions","title":"表示学习的选读问题"},"page":{"id":"h-i-teacher-targets","title":"选读：模型提供软目标，为什么仍须检查坍塌","number":3},"sectionTitle":"教师模型提供的是哪一种答案？","currentQuestion":"哪些细节需要进一步对照？","sourceSlides":["l3-r2-p11","l3-r2-p18"],"knowledgeBlock":"teacher-target-source-and-collapse","formulas":[],"previousBlock":"b-b7c33acf755c9e6d94c614a9","nextBlock":"b-0c21dd015f4dc5c0981fbf9f","prerequisites":[{"id":"b-b7c33acf755c9e6d94c614a9","url":"#b-b7c33acf755c9e6d94c614a9","role":"section-title"},{"id":"b-f77ded2c0c3a92a77e3f97d1","url":"#b-f77ded2c0c3a92a77e3f97d1","role":"problem-context"}],"relatedBlocks":[{"id":"b-b7c33acf755c9e6d94c614a9","url":"#b-b7c33acf755c9e6d94c614a9","relation":"same-reasoning"},{"id":"b-0c21dd015f4dc5c0981fbf9f","url":"#b-0c21dd015f4dc5c0981fbf9f","relation":"same-reasoning"},{"id":"b-c240f92805ebb2be103063ee","url":"#b-c240f92805ebb2be103063ee","relation":"same-reasoning"}]},{"id":"b-0c21dd015f4dc5c0981fbf9f","type":"p","scope":"teacher-target-source-and-collapse","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p18.html#b-0c21dd015f4dc5c0981fbf9f","content":"以logits(4,2,0)为例，温度1的概率约(0.8668,0.1173,0.0159)，温度2约(0.6652,0.2447,0.0900)。软目标除最大类外，还表达其他类别的相对接近。学生可以用交叉熵拟合它，但也可能继承教师的背景捷径。","sha256":"2144c511d94c649203a4096a60898b9104ed9600b5c126c135f958212ed1182a","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-extensions","title":"表示学习的选读问题"},"page":{"id":"h-i-teacher-targets","title":"选读：模型提供软目标，为什么仍须检查坍塌","number":3},"sectionTitle":"教师模型提供的是哪一种答案？","currentQuestion":"哪些细节需要进一步对照？","sourceSlides":["l3-r2-p11","l3-r2-p18"],"knowledgeBlock":"teacher-target-source-and-collapse","formulas":[],"previousBlock":"b-47201491ff9c40d647825106","nextBlock":"b-c240f92805ebb2be103063ee","prerequisites":[{"id":"b-b7c33acf755c9e6d94c614a9","url":"#b-b7c33acf755c9e6d94c614a9","role":"section-title"},{"id":"b-f77ded2c0c3a92a77e3f97d1","url":"#b-f77ded2c0c3a92a77e3f97d1","role":"problem-context"}],"relatedBlocks":[{"id":"b-b7c33acf755c9e6d94c614a9","url":"#b-b7c33acf755c9e6d94c614a9","relation":"same-reasoning"},{"id":"b-47201491ff9c40d647825106","url":"#b-47201491ff9c40d647825106","relation":"same-reasoning"},{"id":"b-c240f92805ebb2be103063ee","url":"#b-c240f92805ebb2be103063ee","relation":"same-reasoning"}]},{"id":"b-c240f92805ebb2be103063ee","type":"p","scope":"teacher-target-source-and-collapse","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p18.html#b-c240f92805ebb2be103063ee","content":"BYOL、DINO说明不显式使用负例也能学表示；它们采用的不对称预测、停止梯度、教师更新及输出分布处理应整体理解。当前LAB 17训练的是解析输入上的固定线性教师蒸馏，再单独算EMA，不是DINO视觉复现。这里适合作为理解“目标从模型取得”的选读分支，阅读后回到P18。","sha256":"2d86e6daf5520a7166809b03cc90d62ffda26ba3b362d7fbcd5e33bb0bbd9cc3","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-extensions","title":"表示学习的选读问题"},"page":{"id":"h-i-teacher-targets","title":"选读：模型提供软目标，为什么仍须检查坍塌","number":3},"sectionTitle":"教师模型提供的是哪一种答案？","currentQuestion":"哪些细节需要进一步对照？","sourceSlides":["l3-r2-p11","l3-r2-p18"],"knowledgeBlock":"teacher-target-source-and-collapse","formulas":[],"previousBlock":"b-0c21dd015f4dc5c0981fbf9f","nextBlock":"b-aa715798dd77c91da84ec96c","prerequisites":[{"id":"b-b7c33acf755c9e6d94c614a9","url":"#b-b7c33acf755c9e6d94c614a9","role":"section-title"},{"id":"b-f77ded2c0c3a92a77e3f97d1","url":"#b-f77ded2c0c3a92a77e3f97d1","role":"problem-context"}],"relatedBlocks":[{"id":"b-b7c33acf755c9e6d94c614a9","url":"#b-b7c33acf755c9e6d94c614a9","relation":"same-reasoning"},{"id":"b-47201491ff9c40d647825106","url":"#b-47201491ff9c40d647825106","relation":"same-reasoning"},{"id":"b-0c21dd015f4dc5c0981fbf9f","url":"#b-0c21dd015f4dc5c0981fbf9f","relation":"same-reasoning"}]},{"id":"b-aa715798dd77c91da84ec96c","type":"h2","scope":"practice","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p18.html#b-aa715798dd77c91da84ec96c","content":"## 沿这个问题继续","sha256":"cf44fcc62782fb4f6bdc42c16c5187d57556a03cb7ce6065e7bd9c36d8e7d733","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-extensions","title":"表示学习的选读问题"},"page":{"id":"h-i-teacher-targets","title":"选读：模型提供软目标，为什么仍须检查坍塌","number":3},"sectionTitle":"选读：模型提供软目标，为什么仍须检查坍塌","currentQuestion":"哪些细节需要进一步对照？","sourceSlides":["l3-r2-p11","l3-r2-p18"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-c240f92805ebb2be103063ee","nextBlock":"b-7e7b4e3a21073adb4a3bc2e0","prerequisites":[{"id":"b-b7c33acf755c9e6d94c614a9","url":"#b-b7c33acf755c9e6d94c614a9","role":"section-title"},{"id":"b-f77ded2c0c3a92a77e3f97d1","url":"#b-f77ded2c0c3a92a77e3f97d1","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-7e7b4e3a21073adb4a3bc2e0","type":"p","scope":"practice","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p18.html#b-7e7b4e3a21073adb4a3bc2e0","content":"先读MoCo检测证据表及协议；教师目标为选读，随后直接进P19。","sha256":"85efc50dc30437f045e5bd3ecfedda789345f9520e583339f496448663547d5a","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-extensions","title":"表示学习的选读问题"},"page":{"id":"h-i-teacher-targets","title":"选读：模型提供软目标，为什么仍须检查坍塌","number":3},"sectionTitle":"选读：模型提供软目标，为什么仍须检查坍塌","currentQuestion":"哪些细节需要进一步对照？","sourceSlides":["l3-r2-p11","l3-r2-p18"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-aa715798dd77c91da84ec96c","nextBlock":"b-1142c9b6e12fc3b859ac5fbf","prerequisites":[{"id":"b-aa715798dd77c91da84ec96c","url":"#b-aa715798dd77c91da84ec96c","role":"section-title"},{"id":"b-f77ded2c0c3a92a77e3f97d1","url":"#b-f77ded2c0c3a92a77e3f97d1","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-1142c9b6e12fc3b859ac5fbf","type":"h4","scope":"practice","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p18.html#b-1142c9b6e12fc3b859ac5fbf","content":"#### 按问题回查的讲解与短练习（2）","sha256":"482a56d6c16f1fcc67a7310871c9513a1202b8628f3d5497506d151a2b4d5e51","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-extensions","title":"表示学习的选读问题"},"page":{"id":"h-i-teacher-targets","title":"选读：模型提供软目标，为什么仍须检查坍塌","number":3},"sectionTitle":"选读：模型提供软目标，为什么仍须检查坍塌","currentQuestion":"哪些细节需要进一步对照？","sourceSlides":["l3-r2-p11","l3-r2-p18"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-7e7b4e3a21073adb4a3bc2e0","nextBlock":null,"prerequisites":[{"id":"b-aa715798dd77c91da84ec96c","url":"#b-aa715798dd77c91da84ec96c","role":"section-title"},{"id":"b-f77ded2c0c3a92a77e3f97d1","url":"#b-f77ded2c0c3a92a77e3f97d1","role":"problem-context"}],"relatedBlocks":[]}]}