{"version":2,"relationshipURLBase":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p11.html","relationshipResolution":"Resolve relative relationship URLs against relationshipURLBase; complete records and absolute URLs are in blocks.","title":"为什么不能把所有图片都映射成同一个向量？","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p11.html","markdown":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p11.md","pages":[{"id":"h-i-collapse-test","title":"代入常量表示，检查目标的平凡解","unit":"视图关系怎样成为学习目标，又会牺牲什么？","unitId":"i-view-objective","chapter":"I","number":1},{"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":2}],"aliases":[{"oldId":"explanation","targetId":"h-i-collapse-test","operation":"split-explanation-range"},{"oldId":"b-b955082e4c3acc74c92f8f49","targetId":"b-b955082e4c3acc74c92f8f49","page":"h-i-collapse-test","operation":"move","beforeSHA256":"d71cf15e707c8a3674b4c90a599efee783cb84da5b8c8236a1d01c3add7b4958","afterSHA256":"d71cf15e707c8a3674b4c90a599efee783cb84da5b8c8236a1d01c3add7b4958","reason":"Same source body in authored semantic page"},{"oldId":"b-d23921e9b8b91633f458bc92","targetId":"b-d23921e9b8b91633f458bc92","page":"h-i-collapse-test","operation":"move","beforeSHA256":"85c85fa6e632087927bf846ac9e91624e176aa8d8149b07c8e05d36ab98425e6","afterSHA256":"85c85fa6e632087927bf846ac9e91624e176aa8d8149b07c8e05d36ab98425e6","reason":"Same source body in authored semantic page"},{"oldId":"b-35e38f2ba26d22fa95c9352e","targetId":"b-35e38f2ba26d22fa95c9352e","page":"h-i-collapse-test","operation":"move","beforeSHA256":"95d0ea9de4112353631eff4a418936851b49b5b6d1914dad3ff6c767a373014e","afterSHA256":"95d0ea9de4112353631eff4a418936851b49b5b6d1914dad3ff6c767a373014e","reason":"Same source body in authored semantic page"},{"oldId":"b-c269a99c9cec06976fc6ea1b","targetId":"b-c269a99c9cec06976fc6ea1b","page":"h-i-collapse-test","operation":"rewrite","beforeSHA256":"102485b25a5918198db9a0c3b4c39ed0c78d5306933ee56b1339256ad272469b","afterSHA256":"865090b91b083851a7bc82a62847317181a618dc9a9c2529ef9e05d4a9a194d7","reason":"Same source identity with figure semantics or reading context enriched"},{"oldId":"b-ba32459b66052cda6ec4465a","targetId":"b-ba32459b66052cda6ec4465a","page":"h-i-collapse-test","operation":"move","beforeSHA256":"786ce10a6b32a492259a3dccd0a54e61cd41d28a34382fb7b5386c5f03b7ad69","afterSHA256":"786ce10a6b32a492259a3dccd0a54e61cd41d28a34382fb7b5386c5f03b7ad69","reason":"Same source body in authored semantic page"},{"oldId":"b-74d0dcf013c994e1d977b30c","targetId":"b-74d0dcf013c994e1d977b30c","page":"h-i-collapse-test","operation":"move","beforeSHA256":"8b5821b8fa18dad2d1b03d8503313d19d2558c37934d9921a0af5fe814b50bb8","afterSHA256":"8b5821b8fa18dad2d1b03d8503313d19d2558c37934d9921a0af5fe814b50bb8","reason":"Same source body in authored semantic page"},{"oldId":"b-c392bdbb506fa991e29e8a67","targetId":"b-c392bdbb506fa991e29e8a67","page":"h-i-collapse-test","operation":"move","beforeSHA256":"fdc129fe823b2d8431e19e4c47d133c7b74cd2b7d2ecc29e2a3402c96cbfa988","afterSHA256":"fdc129fe823b2d8431e19e4c47d133c7b74cd2b7d2ecc29e2a3402c96cbfa988","reason":"Same source body in authored semantic page"},{"oldId":"b-f0597232a34cb120dabaf3ba","targetId":"b-f0597232a34cb120dabaf3ba","page":"h-i-collapse-test","operation":"move","beforeSHA256":"a4f811abf8841ce547ab4a4ec5f93a9b90fcecd439bfdaadbf5361d21f2ca629","afterSHA256":"a4f811abf8841ce547ab4a4ec5f93a9b90fcecd439bfdaadbf5361d21f2ca629","reason":"Same source body in authored semantic page"},{"oldId":"b-146b10866204929c1abe6bf5","targetId":"b-146b10866204929c1abe6bf5","page":"h-i-collapse-test","operation":"move","beforeSHA256":"05781d392f17aa73a2d68d7ae87712419670fc33ffd2bc95807e2430782a6883","afterSHA256":"05781d392f17aa73a2d68d7ae87712419670fc33ffd2bc95807e2430782a6883","reason":"Same source body in authored semantic page"},{"oldId":"b-7239260bf0f1495fe5511197","targetId":"b-7239260bf0f1495fe5511197","page":"h-i-collapse-test","operation":"move","beforeSHA256":"b158e5ec2ffe03688f2f2970a6f66f02bc7b208497140d5673102c71fdfaebd2","afterSHA256":"b158e5ec2ffe03688f2f2970a6f66f02bc7b208497140d5673102c71fdfaebd2","reason":"Same source body in authored semantic page"},{"oldId":"b-b354338d144264b265d2927b","targetId":"b-b354338d144264b265d2927b","page":"h-i-collapse-test","operation":"move","beforeSHA256":"700fddd7d5f83b71f89c9a291acb588ca5e9e5e322a029cd954f682cd49771df","afterSHA256":"700fddd7d5f83b71f89c9a291acb588ca5e9e5e322a029cd954f682cd49771df","reason":"Same source body in authored semantic page"},{"oldId":"b-37edb577e4438084d033a9c3","targetId":"b-37edb577e4438084d033a9c3","page":"h-i-collapse-test","operation":"move","beforeSHA256":"91a6eaa0ce8aa5a6dd1260d5c1650290572dd8971f80d7b7fbe475772bcd8fd2","afterSHA256":"91a6eaa0ce8aa5a6dd1260d5c1650290572dd8971f80d7b7fbe475772bcd8fd2","reason":"Same source body in authored semantic page"},{"oldId":"b-408fa162767be351ce0b08e0","targetId":"b-408fa162767be351ce0b08e0","page":"h-i-collapse-test","operation":"move","beforeSHA256":"a28b760be1a5fdca6253197edc6c9c92bea0f7d6f41c086daf755daa266d291e","afterSHA256":"a28b760be1a5fdca6253197edc6c9c92bea0f7d6f41c086daf755daa266d291e","reason":"Same source body in authored semantic page"},{"oldId":"b-3949b40f1ee252351f434794","targetId":"b-3949b40f1ee252351f434794","page":"h-i-collapse-test","operation":"move","beforeSHA256":"a6b61a4f053c6cb48afc79cd2f45fe911f5ba7c0cefb1039dc96b309be1016bb","afterSHA256":"a6b61a4f053c6cb48afc79cd2f45fe911f5ba7c0cefb1039dc96b309be1016bb","reason":"Same source body in authored semantic page"},{"oldId":"b-8d79fdfc8eb8f066c0560162","targetId":"b-8d79fdfc8eb8f066c0560162","page":"h-i-collapse-test","operation":"move","beforeSHA256":"e8f006fbb69bcf5781beeca8a597a8f068087275300443b6504824f4c68092a3","afterSHA256":"e8f006fbb69bcf5781beeca8a597a8f068087275300443b6504824f4c68092a3","reason":"Same source body in authored semantic page"},{"oldId":"b-a5e8189bc33781eb0938cdc6","targetId":"b-a5e8189bc33781eb0938cdc6","page":"h-i-collapse-test","operation":"move","beforeSHA256":"3209c8ee3c635d99cb9c1a1c5892dbfebac92620b0f972464acbb43b684515f3","afterSHA256":"3209c8ee3c635d99cb9c1a1c5892dbfebac92620b0f972464acbb43b684515f3","reason":"Same source body in authored semantic page"},{"oldId":"b-1025ba3bd4594767f13268f7","targetId":"b-1025ba3bd4594767f13268f7","page":"h-i-collapse-test","operation":"move","beforeSHA256":"1fe7126c3086d263e32a424fb341240631d4ee41beadd5aa3291acc44fdb4350","afterSHA256":"1fe7126c3086d263e32a424fb341240631d4ee41beadd5aa3291acc44fdb4350","reason":"Same source body in authored semantic page"},{"oldId":"b-868173bbb07725749854fbb3","targetId":"b-868173bbb07725749854fbb3","page":"h-i-teacher-targets","operation":"move","beforeSHA256":"5bc51a0035347385ce8779fee8965d69c21f49b09a5aa430efbabc6411d7d2b8","afterSHA256":"5bc51a0035347385ce8779fee8965d69c21f49b09a5aa430efbabc6411d7d2b8","reason":"Same source body in authored semantic page"},{"oldId":"b-f8a0c93720f84f4b95c190c6","targetId":"b-f8a0c93720f84f4b95c190c6","page":"h-i-teacher-targets","operation":"move","beforeSHA256":"9b40607f74227610df31c60a34afc2bb8e212dabce62a2af04fcbed1aae8c1b1","afterSHA256":"9b40607f74227610df31c60a34afc2bb8e212dabce62a2af04fcbed1aae8c1b1","reason":"Same source body in authored semantic page"},{"oldId":"b-2d769178d26329ce9d2a436c","targetId":"b-2d769178d26329ce9d2a436c","page":"h-i-teacher-targets","operation":"move","beforeSHA256":"2144c511d94c649203a4096a60898b9104ed9600b5c126c135f958212ed1182a","afterSHA256":"2144c511d94c649203a4096a60898b9104ed9600b5c126c135f958212ed1182a","reason":"Same source body in authored semantic page"},{"oldId":"b-f1c395d374a763fcd8654a05","targetId":"b-f1c395d374a763fcd8654a05","page":"h-i-teacher-targets","operation":"move","beforeSHA256":"2d86e6daf5520a7166809b03cc90d62ffda26ba3b362d7fbcd5e33bb0bbd9cc3","afterSHA256":"2d86e6daf5520a7166809b03cc90d62ffda26ba3b362d7fbcd5e33bb0bbd9cc3","reason":"Same source body in authored semantic page"},{"oldId":"b-281c245a6e6f295bba3f321a","targetId":"b-281c245a6e6f295bba3f321a","page":"h-i-teacher-targets","operation":"move","beforeSHA256":"cf44fcc62782fb4f6bdc42c16c5187d57556a03cb7ce6065e7bd9c36d8e7d733","afterSHA256":"cf44fcc62782fb4f6bdc42c16c5187d57556a03cb7ce6065e7bd9c36d8e7d733","reason":"Same source body in authored semantic page"},{"oldId":"b-e2ef084fc9bdb61d2d725d9a","targetId":"b-e2ef084fc9bdb61d2d725d9a","page":"h-i-teacher-targets","operation":"move","beforeSHA256":"7f4d7f7d2df9d1f82ac8d3bac4995a7069224eda7e147de23540e2c839b19a4a","afterSHA256":"7f4d7f7d2df9d1f82ac8d3bac4995a7069224eda7e147de23540e2c839b19a4a","reason":"Same source body in authored semantic page"},{"oldId":"h-i-collapse-test-title","targetId":"b-3d137e626953b2062e46565b","operation":"retain-source-identity"},{"oldId":"h-i-teacher-targets-title","targetId":"b-8305b3913a6b3b281594bb8a","operation":"retain-source-identity"},{"oldId":"h-i-teacher-targets-bridge","targetId":"b-fa644fa8dc1bd6820868fbe0","operation":"retain-source-identity"}],"blocks":[{"id":"b-3d137e626953b2062e46565b","type":"h2","scope":"h-i-collapse-test","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p11.html#b-3d137e626953b2062e46565b","content":"## 代入常量表示，检查目标的平凡解","sha256":"6ade2f694feac7d52d5b912c05aeaf55cd0c75aba7c9222ae28cd7745a1f166d","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-view-objective","title":"视图关系怎样成为学习目标，又会牺牲什么？"},"page":{"id":"h-i-collapse-test","title":"代入常量表示，检查目标的平凡解","number":1},"sectionTitle":"代入常量表示，检查目标的平凡解","currentQuestion":"视图关系怎样成为学习目标，又会牺牲什么？","sourceSlides":["l3-r2-p11"],"knowledgeBlock":null,"formulas":[],"previousBlock":null,"nextBlock":"b-b955082e4c3acc74c92f8f49","prerequisites":[{"id":"b-35e38f2ba26d22fa95c9352e","url":"#b-35e38f2ba26d22fa95c9352e","role":"problem-context"}],"relatedBlocks":[{"id":"b-c269a99c9cec06976fc6ea1b","url":"#b-c269a99c9cec06976fc6ea1b","relation":"figure-and-explanation"}]},{"id":"b-b955082e4c3acc74c92f8f49","type":"h2","scope":"explanation","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p11.html#b-b955082e4c3acc74c92f8f49","content":"## 解释、算例与证据","sha256":"d71cf15e707c8a3674b4c90a599efee783cb84da5b8c8236a1d01c3add7b4958","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-view-objective","title":"视图关系怎样成为学习目标，又会牺牲什么？"},"page":{"id":"h-i-collapse-test","title":"代入常量表示，检查目标的平凡解","number":1},"sectionTitle":"代入常量表示，检查目标的平凡解","currentQuestion":"视图关系怎样成为学习目标，又会牺牲什么？","sourceSlides":["l3-r2-p11"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-3d137e626953b2062e46565b","nextBlock":"b-d23921e9b8b91633f458bc92","prerequisites":[{"id":"b-3d137e626953b2062e46565b","url":"#b-3d137e626953b2062e46565b","role":"section-title"},{"id":"b-35e38f2ba26d22fa95c9352e","url":"#b-35e38f2ba26d22fa95c9352e","role":"problem-context"}],"relatedBlocks":[{"id":"b-c269a99c9cec06976fc6ea1b","url":"#b-c269a99c9cec06976fc6ea1b","relation":"figure-and-explanation"}]},{"id":"b-d23921e9b8b91633f458bc92","type":"h1","scope":"document","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p11.html#b-d23921e9b8b91633f458bc92","content":"# P11 · 为什么不能把所有图片都映射成同一个向量？","sha256":"85c85fa6e632087927bf846ac9e91624e176aa8d8149b07c8e05d36ab98425e6","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-view-objective","title":"视图关系怎样成为学习目标，又会牺牲什么？"},"page":{"id":"h-i-collapse-test","title":"代入常量表示，检查目标的平凡解","number":1},"sectionTitle":"代入常量表示，检查目标的平凡解","currentQuestion":"视图关系怎样成为学习目标，又会牺牲什么？","sourceSlides":["l3-r2-p11"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-b955082e4c3acc74c92f8f49","nextBlock":"b-35e38f2ba26d22fa95c9352e","prerequisites":[{"id":"b-b955082e4c3acc74c92f8f49","url":"#b-b955082e4c3acc74c92f8f49","role":"section-title"},{"id":"b-35e38f2ba26d22fa95c9352e","url":"#b-35e38f2ba26d22fa95c9352e","role":"problem-context"}],"relatedBlocks":[{"id":"b-c269a99c9cec06976fc6ea1b","url":"#b-c269a99c9cec06976fc6ea1b","relation":"figure-and-explanation"}]},{"id":"b-35e38f2ba26d22fa95c9352e","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p11.html#b-35e38f2ba26d22fa95c9352e","content":"先代入常量函数检查目标的平凡解，说明仅有一致性不够，也为教师方法留下合理选读入口。","sha256":"95d0ea9de4112353631eff4a418936851b49b5b6d1914dad3ff6c767a373014e","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-view-objective","title":"视图关系怎样成为学习目标，又会牺牲什么？"},"page":{"id":"h-i-collapse-test","title":"代入常量表示，检查目标的平凡解","number":1},"sectionTitle":"代入常量表示，检查目标的平凡解","currentQuestion":"视图关系怎样成为学习目标，又会牺牲什么？","sourceSlides":["l3-r2-p11"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-d23921e9b8b91633f458bc92","nextBlock":"b-c269a99c9cec06976fc6ea1b","prerequisites":[{"id":"b-d23921e9b8b91633f458bc92","url":"#b-d23921e9b8b91633f458bc92","role":"section-title"}],"relatedBlocks":[{"id":"b-c269a99c9cec06976fc6ea1b","url":"#b-c269a99c9cec06976fc6ea1b","relation":"figure-and-explanation"}]},{"id":"b-c269a99c9cec06976fc6ea1b","type":"figure","scope":"document","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p11.html#b-c269a99c9cec06976fc6ea1b","content":"图内标签与数值：P11 学习表示机制图，状态 1；全部映成同一向量；能区分指定对应；k⁺；1/3；0.665；k₁；1/3；0.245；k₂；1/3；0.090；q 与 k⁺；其他候选；L = ln 3 ≈ 1.099；L ≈ 0.408；分数全部相同，选择任务无法完成；此图比较目标值，不是训练曲线。\n\n坍塌表示的三候选均分1/3与可区分候选并排，损失ln3约1.099对比0.408。\n\n- 全部映成同一向量\n- 能区分指定对应\n- k⁺\n- 1/3\n- 0.665\n- k₁\n- 1/3\n- 0.245\n- k₂\n- 1/3\n- 0.090\n- q 与 k⁺\n- 其他候选\n- L = ln 3 ≈ 1.099\n- L ≈ 0.408\n- 分数全部相同，选择任务无法完成；此图比较目标值，不是训练曲线。坍塌表示的三候选均分1/3与可区分候选并排，损失ln3约1.099对比0.408。 [查看对应图解状态](/#/l3-r2-p11/0)[查看完整图解](/course/figures/l3-r2-p11-0.html)","sha256":"865090b91b083851a7bc82a62847317181a618dc9a9c2529ef9e05d4a9a194d7","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-view-objective","title":"视图关系怎样成为学习目标，又会牺牲什么？"},"page":{"id":"h-i-collapse-test","title":"代入常量表示，检查目标的平凡解","number":1},"sectionTitle":"代入常量表示，检查目标的平凡解","currentQuestion":"视图关系怎样成为学习目标，又会牺牲什么？","sourceSlides":["l3-r2-p11"],"knowledgeBlock":null,"formulas":[],"figure":{"caption":"坍塌表示的三候选均分1/3与可区分候选并排，损失ln3约1.099对比0.408。 查看对应图解状态","labels":["全部映成同一向量","能区分指定对应","k⁺","1/3","0.665","k₁","1/3","0.245","k₂","1/3","0.090","q 与 k⁺","其他候选","L = ln 3 ≈ 1.099","L ≈ 0.408","分数全部相同，选择任务无法完成；此图比较目标值，不是训练曲线。"],"description":"坍塌表示的三候选均分1/3与可区分候选并排，损失ln3约1.099对比0.408。全部映成同一向量能区分指定对应k⁺1/30.665k₁1/30.245k₂1/30.090q 与 k⁺其他候选L = ln 3 ≈ 1.099L ≈ 0.408分数全部相同，选择任务无法完成；此图比较目标值，不是训练曲线。","evidenceIdentity":"坍塌表示的三候选均分1/3与可区分候选并排，损失ln3约1.099对比0.408。 查看对应图解状态","images":[]},"previousBlock":"b-35e38f2ba26d22fa95c9352e","nextBlock":"b-ba32459b66052cda6ec4465a","prerequisites":[{"id":"b-d23921e9b8b91633f458bc92","url":"#b-d23921e9b8b91633f458bc92","role":"section-title"},{"id":"b-35e38f2ba26d22fa95c9352e","url":"#b-35e38f2ba26d22fa95c9352e","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-ba32459b66052cda6ec4465a","type":"h2","scope":"narrative","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p11.html#b-ba32459b66052cda6ec4465a","content":"## 课堂讲述","sha256":"786ce10a6b32a492259a3dccd0a54e61cd41d28a34382fb7b5386c5f03b7ad69","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-view-objective","title":"视图关系怎样成为学习目标，又会牺牲什么？"},"page":{"id":"h-i-collapse-test","title":"代入常量表示，检查目标的平凡解","number":1},"sectionTitle":"代入常量表示，检查目标的平凡解","currentQuestion":"视图关系怎样成为学习目标，又会牺牲什么？","sourceSlides":["l3-r2-p11"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-c269a99c9cec06976fc6ea1b","nextBlock":"b-74d0dcf013c994e1d977b30c","prerequisites":[{"id":"b-d23921e9b8b91633f458bc92","url":"#b-d23921e9b8b91633f458bc92","role":"section-title"},{"id":"b-35e38f2ba26d22fa95c9352e","url":"#b-35e38f2ba26d22fa95c9352e","role":"problem-context"}],"relatedBlocks":[{"id":"b-c269a99c9cec06976fc6ea1b","url":"#b-c269a99c9cec06976fc6ea1b","relation":"figure-and-explanation"}]},{"id":"b-74d0dcf013c994e1d977b30c","type":"p","scope":"narrative","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p11.html#b-74d0dcf013c994e1d977b30c","content":"一个看似省事的办法是：所有输入都输出同一个向量。这样同图的两个视图当然相同，但任何其他图也相同，模型已经没有办法选择对应项。\n在这个三候选算例中，所有分数相同，正确项只有三分之一的概率，损失约为 1.099；能区分对应关系时，损失可以更低。这个比较解释了为什么单纯要求“两视图一样”不够，而带有区分任务的目标还提出了另一项要求。\n不要由此推成“只要有负例，训练就一定成功”。我们现在只是看见平凡表示为什么不能很好完成这个选择任务。没有负例的方法怎样避免坍塌，是另一组设计问题，已经保留在补充材料，今天不再展开一套教师网络。","sha256":"8b5821b8fa18dad2d1b03d8503313d19d2558c37934d9921a0af5fe814b50bb8","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-view-objective","title":"视图关系怎样成为学习目标，又会牺牲什么？"},"page":{"id":"h-i-collapse-test","title":"代入常量表示，检查目标的平凡解","number":1},"sectionTitle":"代入常量表示，检查目标的平凡解","currentQuestion":"视图关系怎样成为学习目标，又会牺牲什么？","sourceSlides":["l3-r2-p11"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-ba32459b66052cda6ec4465a","nextBlock":"b-c392bdbb506fa991e29e8a67","prerequisites":[{"id":"b-ba32459b66052cda6ec4465a","url":"#b-ba32459b66052cda6ec4465a","role":"section-title"},{"id":"b-35e38f2ba26d22fa95c9352e","url":"#b-35e38f2ba26d22fa95c9352e","role":"problem-context"}],"relatedBlocks":[{"id":"b-c269a99c9cec06976fc6ea1b","url":"#b-c269a99c9cec06976fc6ea1b","relation":"figure-and-explanation"}]},{"id":"b-c392bdbb506fa991e29e8a67","type":"h3","scope":"collapse-and-discrimination","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p11.html#b-c392bdbb506fa991e29e8a67","content":"### 常量表示为什么不解决对比任务？","sha256":"fdc129fe823b2d8431e19e4c47d133c7b74cd2b7d2ecc29e2a3402c96cbfa988","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-view-objective","title":"视图关系怎样成为学习目标，又会牺牲什么？"},"page":{"id":"h-i-collapse-test","title":"代入常量表示，检查目标的平凡解","number":1},"sectionTitle":"常量表示为什么不解决对比任务？","currentQuestion":"视图关系怎样成为学习目标，又会牺牲什么？","sourceSlides":["l3-r2-p11"],"knowledgeBlock":"collapse-and-discrimination","formulas":[],"previousBlock":"b-74d0dcf013c994e1d977b30c","nextBlock":"b-f0597232a34cb120dabaf3ba","prerequisites":[{"id":"b-ba32459b66052cda6ec4465a","url":"#b-ba32459b66052cda6ec4465a","role":"section-title"},{"id":"b-35e38f2ba26d22fa95c9352e","url":"#b-35e38f2ba26d22fa95c9352e","role":"problem-context"}],"relatedBlocks":[{"id":"b-c269a99c9cec06976fc6ea1b","url":"#b-c269a99c9cec06976fc6ea1b","relation":"figure-and-explanation"},{"id":"b-f0597232a34cb120dabaf3ba","url":"#b-f0597232a34cb120dabaf3ba","relation":"same-reasoning"},{"id":"b-146b10866204929c1abe6bf5","url":"#b-146b10866204929c1abe6bf5","relation":"same-reasoning"},{"id":"b-7239260bf0f1495fe5511197","url":"#b-7239260bf0f1495fe5511197","relation":"same-reasoning"}]},{"id":"b-f0597232a34cb120dabaf3ba","type":"p","scope":"collapse-and-discrimination","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p11.html#b-f0597232a34cb120dabaf3ba","content":"若只最小化两视图间距离，$f(x)=c$ 对任何图都输出同一向量，配对距离就是零。它满足一致性，却让杯子与书无法区分。加入候选竞争后，如果所有单位向量相同，各候选得分也相同，K个候选的正例概率只能是1/K，损失是 $\\ln K$；例如三项时约1.099。","sha256":"a4f811abf8841ce547ab4a4ec5f93a9b90fcecd439bfdaadbf5361d21f2ca629","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-view-objective","title":"视图关系怎样成为学习目标，又会牺牲什么？"},"page":{"id":"h-i-collapse-test","title":"代入常量表示，检查目标的平凡解","number":1},"sectionTitle":"常量表示为什么不解决对比任务？","currentQuestion":"视图关系怎样成为学习目标，又会牺牲什么？","sourceSlides":["l3-r2-p11"],"knowledgeBlock":"collapse-and-discrimination","formulas":[{"latex":"f(x)=c","definitionContext":"若只最小化两视图间距离，$f(x)=c$ 对任何图都输出同一向量，配对距离就是零。它满足一致性，却让杯子与书无法区分。加入候选竞争后，如果所有单位向量相同，各候选得分也相同，K个候选的正例概率只能是1/K，损失是 $\\ln K$；例如三项时约1.099。\n这说明对比目标没有把“全部一样”当成一个好的区分结果。但仅凭这个算例，不应宣布任何网络或优化过程都不可能停在退化状态。真正检查还包括样本间方差、表示秩以及下游读出。\n没有显式负例的方法也能成功：它们组合教师目标、预测器、停止梯度和分布处理等设计。下面的教师补充块解释这些组件各自负责什么，不把一条EMA箭头当成防坍塌的充分证明。"},{"latex":"\\ln K","definitionContext":"若只最小化两视图间距离，$f(x)=c$ 对任何图都输出同一向量，配对距离就是零。它满足一致性，却让杯子与书无法区分。加入候选竞争后，如果所有单位向量相同，各候选得分也相同，K个候选的正例概率只能是1/K，损失是 $\\ln K$；例如三项时约1.099。\n这说明对比目标没有把“全部一样”当成一个好的区分结果。但仅凭这个算例，不应宣布任何网络或优化过程都不可能停在退化状态。真正检查还包括样本间方差、表示秩以及下游读出。\n没有显式负例的方法也能成功：它们组合教师目标、预测器、停止梯度和分布处理等设计。下面的教师补充块解释这些组件各自负责什么，不把一条EMA箭头当成防坍塌的充分证明。"}],"previousBlock":"b-c392bdbb506fa991e29e8a67","nextBlock":"b-146b10866204929c1abe6bf5","prerequisites":[{"id":"b-c392bdbb506fa991e29e8a67","url":"#b-c392bdbb506fa991e29e8a67","role":"section-title"},{"id":"b-35e38f2ba26d22fa95c9352e","url":"#b-35e38f2ba26d22fa95c9352e","role":"problem-context"}],"relatedBlocks":[{"id":"b-c269a99c9cec06976fc6ea1b","url":"#b-c269a99c9cec06976fc6ea1b","relation":"figure-and-explanation"},{"id":"b-c392bdbb506fa991e29e8a67","url":"#b-c392bdbb506fa991e29e8a67","relation":"same-reasoning"},{"id":"b-146b10866204929c1abe6bf5","url":"#b-146b10866204929c1abe6bf5","relation":"same-reasoning"},{"id":"b-7239260bf0f1495fe5511197","url":"#b-7239260bf0f1495fe5511197","relation":"same-reasoning"}]},{"id":"b-146b10866204929c1abe6bf5","type":"p","scope":"collapse-and-discrimination","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p11.html#b-146b10866204929c1abe6bf5","content":"这说明对比目标没有把“全部一样”当成一个好的区分结果。但仅凭这个算例，不应宣布任何网络或优化过程都不可能停在退化状态。真正检查还包括样本间方差、表示秩以及下游读出。","sha256":"05781d392f17aa73a2d68d7ae87712419670fc33ffd2bc95807e2430782a6883","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-view-objective","title":"视图关系怎样成为学习目标，又会牺牲什么？"},"page":{"id":"h-i-collapse-test","title":"代入常量表示，检查目标的平凡解","number":1},"sectionTitle":"常量表示为什么不解决对比任务？","currentQuestion":"视图关系怎样成为学习目标，又会牺牲什么？","sourceSlides":["l3-r2-p11"],"knowledgeBlock":"collapse-and-discrimination","formulas":[],"previousBlock":"b-f0597232a34cb120dabaf3ba","nextBlock":"b-7239260bf0f1495fe5511197","prerequisites":[{"id":"b-c392bdbb506fa991e29e8a67","url":"#b-c392bdbb506fa991e29e8a67","role":"section-title"},{"id":"b-35e38f2ba26d22fa95c9352e","url":"#b-35e38f2ba26d22fa95c9352e","role":"problem-context"}],"relatedBlocks":[{"id":"b-c269a99c9cec06976fc6ea1b","url":"#b-c269a99c9cec06976fc6ea1b","relation":"figure-and-explanation"},{"id":"b-c392bdbb506fa991e29e8a67","url":"#b-c392bdbb506fa991e29e8a67","relation":"same-reasoning"},{"id":"b-f0597232a34cb120dabaf3ba","url":"#b-f0597232a34cb120dabaf3ba","relation":"same-reasoning"},{"id":"b-7239260bf0f1495fe5511197","url":"#b-7239260bf0f1495fe5511197","relation":"same-reasoning"}]},{"id":"b-7239260bf0f1495fe5511197","type":"p","scope":"collapse-and-discrimination","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p11.html#b-7239260bf0f1495fe5511197","content":"没有显式负例的方法也能成功：它们组合教师目标、预测器、停止梯度和分布处理等设计。下面的教师补充块解释这些组件各自负责什么，不把一条EMA箭头当成防坍塌的充分证明。","sha256":"b158e5ec2ffe03688f2f2970a6f66f02bc7b208497140d5673102c71fdfaebd2","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-view-objective","title":"视图关系怎样成为学习目标，又会牺牲什么？"},"page":{"id":"h-i-collapse-test","title":"代入常量表示，检查目标的平凡解","number":1},"sectionTitle":"常量表示为什么不解决对比任务？","currentQuestion":"视图关系怎样成为学习目标，又会牺牲什么？","sourceSlides":["l3-r2-p11"],"knowledgeBlock":"collapse-and-discrimination","formulas":[],"previousBlock":"b-146b10866204929c1abe6bf5","nextBlock":"b-9fdbe2ae461f0637ebcc7ccb","prerequisites":[{"id":"b-c392bdbb506fa991e29e8a67","url":"#b-c392bdbb506fa991e29e8a67","role":"section-title"},{"id":"b-35e38f2ba26d22fa95c9352e","url":"#b-35e38f2ba26d22fa95c9352e","role":"problem-context"}],"relatedBlocks":[{"id":"b-c269a99c9cec06976fc6ea1b","url":"#b-c269a99c9cec06976fc6ea1b","relation":"figure-and-explanation"},{"id":"b-c392bdbb506fa991e29e8a67","url":"#b-c392bdbb506fa991e29e8a67","relation":"same-reasoning"},{"id":"b-f0597232a34cb120dabaf3ba","url":"#b-f0597232a34cb120dabaf3ba","relation":"same-reasoning"},{"id":"b-146b10866204929c1abe6bf5","url":"#b-146b10866204929c1abe6bf5","relation":"same-reasoning"}]},{"id":"b-9fdbe2ae461f0637ebcc7ccb","type":"h4","scope":"h-i-collapse-test","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p11.html#b-9fdbe2ae461f0637ebcc7ccb","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-view-objective","title":"视图关系怎样成为学习目标，又会牺牲什么？"},"page":{"id":"h-i-collapse-test","title":"代入常量表示，检查目标的平凡解","number":1},"sectionTitle":"代入常量表示，检查目标的平凡解","currentQuestion":"视图关系怎样成为学习目标，又会牺牲什么？","sourceSlides":["l3-r2-p11"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-7239260bf0f1495fe5511197","nextBlock":"b-b354338d144264b265d2927b","prerequisites":[{"id":"b-c392bdbb506fa991e29e8a67","url":"#b-c392bdbb506fa991e29e8a67","role":"section-title"},{"id":"b-35e38f2ba26d22fa95c9352e","url":"#b-35e38f2ba26d22fa95c9352e","role":"problem-context"}],"relatedBlocks":[{"id":"b-c269a99c9cec06976fc6ea1b","url":"#b-c269a99c9cec06976fc6ea1b","relation":"figure-and-explanation"}]},{"id":"b-b354338d144264b265d2927b","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p11.html#b-b354338d144264b265d2927b","content":"三候选同分：p+=1/3，L=ln3=1.0986122887。图上的向量与条形为机制示意，没有训练曲线或成功保证。","sha256":"700fddd7d5f83b71f89c9a291acb588ca5e9e5e322a029cd954f682cd49771df","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-view-objective","title":"视图关系怎样成为学习目标，又会牺牲什么？"},"page":{"id":"h-i-collapse-test","title":"代入常量表示，检查目标的平凡解","number":1},"sectionTitle":"代入常量表示，检查目标的平凡解","currentQuestion":"视图关系怎样成为学习目标，又会牺牲什么？","sourceSlides":["l3-r2-p11"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-9fdbe2ae461f0637ebcc7ccb","nextBlock":"b-4762c00ffc15bb8616f8a644","prerequisites":[{"id":"b-9fdbe2ae461f0637ebcc7ccb","url":"#b-9fdbe2ae461f0637ebcc7ccb","role":"section-title"},{"id":"b-35e38f2ba26d22fa95c9352e","url":"#b-35e38f2ba26d22fa95c9352e","role":"problem-context"}],"relatedBlocks":[{"id":"b-c269a99c9cec06976fc6ea1b","url":"#b-c269a99c9cec06976fc6ea1b","relation":"figure-and-explanation"}]},{"id":"b-4762c00ffc15bb8616f8a644","type":"h4","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p11.html#b-4762c00ffc15bb8616f8a644","content":"#### 这条线索的展开与后续阅读","sha256":"48b8229f7559b36ed4d5628416e3991e5c73293fbce43374db67386c8736f3bc","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-view-objective","title":"视图关系怎样成为学习目标，又会牺牲什么？"},"page":{"id":"h-i-collapse-test","title":"代入常量表示，检查目标的平凡解","number":1},"sectionTitle":"代入常量表示，检查目标的平凡解","currentQuestion":"视图关系怎样成为学习目标，又会牺牲什么？","sourceSlides":["l3-r2-p11"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-b354338d144264b265d2927b","nextBlock":"b-37edb577e4438084d033a9c3","prerequisites":[{"id":"b-9fdbe2ae461f0637ebcc7ccb","url":"#b-9fdbe2ae461f0637ebcc7ccb","role":"section-title"},{"id":"b-35e38f2ba26d22fa95c9352e","url":"#b-35e38f2ba26d22fa95c9352e","role":"problem-context"}],"relatedBlocks":[{"id":"b-c269a99c9cec06976fc6ea1b","url":"#b-c269a99c9cec06976fc6ea1b","relation":"figure-and-explanation"}]},{"id":"b-37edb577e4438084d033a9c3","type":"ul","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p11.html#b-37edb577e4438084d033a9c3","content":"- [一致性目标的平凡解｜问题从哪里出现](/supplement.html?return=l3-r2-p11/0#/l3-collapse-teacher-problem/0) L03补充阅读\n- [一致性目标的平凡解｜核心改动怎样工作](/supplement.html?return=l3-r2-p11/0#/l3-collapse-teacher-mechanism/0) L03补充阅读\n- [一致性目标的平凡解｜一起算一个完整例子](/supplement.html?return=l3-r2-p11/0#/l3-collapse-teacher-example/0) L03补充阅读\n- [一致性目标的平凡解｜怎样检查这项解释](/supplement.html?return=l3-r2-p11/0#/l3-collapse-teacher-evidence/0) L03补充阅读\n- [一致性目标的平凡解｜什么结论还不能推出](/supplement.html?return=l3-r2-p11/0#/l3-collapse-teacher-boundary/0) L03补充阅读","sha256":"91a6eaa0ce8aa5a6dd1260d5c1650290572dd8971f80d7b7fbe475772bcd8fd2","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-view-objective","title":"视图关系怎样成为学习目标，又会牺牲什么？"},"page":{"id":"h-i-collapse-test","title":"代入常量表示，检查目标的平凡解","number":1},"sectionTitle":"代入常量表示，检查目标的平凡解","currentQuestion":"视图关系怎样成为学习目标，又会牺牲什么？","sourceSlides":["l3-r2-p11"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-4762c00ffc15bb8616f8a644","nextBlock":"b-408fa162767be351ce0b08e0","prerequisites":[{"id":"b-4762c00ffc15bb8616f8a644","url":"#b-4762c00ffc15bb8616f8a644","role":"section-title"},{"id":"b-35e38f2ba26d22fa95c9352e","url":"#b-35e38f2ba26d22fa95c9352e","role":"problem-context"}],"relatedBlocks":[{"id":"b-c269a99c9cec06976fc6ea1b","url":"#b-c269a99c9cec06976fc6ea1b","relation":"figure-and-explanation"}]},{"id":"b-408fa162767be351ce0b08e0","type":"li","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p11.html#b-408fa162767be351ce0b08e0","content":"- [一致性目标的平凡解｜问题从哪里出现](/supplement.html?return=l3-r2-p11/0#/l3-collapse-teacher-problem/0) L03补充阅读","sha256":"a28b760be1a5fdca6253197edc6c9c92bea0f7d6f41c086daf755daa266d291e","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-view-objective","title":"视图关系怎样成为学习目标，又会牺牲什么？"},"page":{"id":"h-i-collapse-test","title":"代入常量表示，检查目标的平凡解","number":1},"sectionTitle":"代入常量表示，检查目标的平凡解","currentQuestion":"视图关系怎样成为学习目标，又会牺牲什么？","sourceSlides":["l3-r2-p11"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-37edb577e4438084d033a9c3","nextBlock":"b-3949b40f1ee252351f434794","prerequisites":[{"id":"b-4762c00ffc15bb8616f8a644","url":"#b-4762c00ffc15bb8616f8a644","role":"section-title"},{"id":"b-35e38f2ba26d22fa95c9352e","url":"#b-35e38f2ba26d22fa95c9352e","role":"problem-context"}],"relatedBlocks":[{"id":"b-c269a99c9cec06976fc6ea1b","url":"#b-c269a99c9cec06976fc6ea1b","relation":"figure-and-explanation"}]},{"id":"b-3949b40f1ee252351f434794","type":"li","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p11.html#b-3949b40f1ee252351f434794","content":"- [一致性目标的平凡解｜核心改动怎样工作](/supplement.html?return=l3-r2-p11/0#/l3-collapse-teacher-mechanism/0) L03补充阅读","sha256":"a6b61a4f053c6cb48afc79cd2f45fe911f5ba7c0cefb1039dc96b309be1016bb","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-view-objective","title":"视图关系怎样成为学习目标，又会牺牲什么？"},"page":{"id":"h-i-collapse-test","title":"代入常量表示，检查目标的平凡解","number":1},"sectionTitle":"代入常量表示，检查目标的平凡解","currentQuestion":"视图关系怎样成为学习目标，又会牺牲什么？","sourceSlides":["l3-r2-p11"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-408fa162767be351ce0b08e0","nextBlock":"b-8d79fdfc8eb8f066c0560162","prerequisites":[{"id":"b-4762c00ffc15bb8616f8a644","url":"#b-4762c00ffc15bb8616f8a644","role":"section-title"},{"id":"b-35e38f2ba26d22fa95c9352e","url":"#b-35e38f2ba26d22fa95c9352e","role":"problem-context"}],"relatedBlocks":[{"id":"b-c269a99c9cec06976fc6ea1b","url":"#b-c269a99c9cec06976fc6ea1b","relation":"figure-and-explanation"}]},{"id":"b-8d79fdfc8eb8f066c0560162","type":"li","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p11.html#b-8d79fdfc8eb8f066c0560162","content":"- [一致性目标的平凡解｜一起算一个完整例子](/supplement.html?return=l3-r2-p11/0#/l3-collapse-teacher-example/0) L03补充阅读","sha256":"e8f006fbb69bcf5781beeca8a597a8f068087275300443b6504824f4c68092a3","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-view-objective","title":"视图关系怎样成为学习目标，又会牺牲什么？"},"page":{"id":"h-i-collapse-test","title":"代入常量表示，检查目标的平凡解","number":1},"sectionTitle":"代入常量表示，检查目标的平凡解","currentQuestion":"视图关系怎样成为学习目标，又会牺牲什么？","sourceSlides":["l3-r2-p11"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-3949b40f1ee252351f434794","nextBlock":"b-a5e8189bc33781eb0938cdc6","prerequisites":[{"id":"b-4762c00ffc15bb8616f8a644","url":"#b-4762c00ffc15bb8616f8a644","role":"section-title"},{"id":"b-35e38f2ba26d22fa95c9352e","url":"#b-35e38f2ba26d22fa95c9352e","role":"problem-context"}],"relatedBlocks":[{"id":"b-c269a99c9cec06976fc6ea1b","url":"#b-c269a99c9cec06976fc6ea1b","relation":"figure-and-explanation"}]},{"id":"b-a5e8189bc33781eb0938cdc6","type":"li","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p11.html#b-a5e8189bc33781eb0938cdc6","content":"- [一致性目标的平凡解｜怎样检查这项解释](/supplement.html?return=l3-r2-p11/0#/l3-collapse-teacher-evidence/0) L03补充阅读","sha256":"3209c8ee3c635d99cb9c1a1c5892dbfebac92620b0f972464acbb43b684515f3","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-view-objective","title":"视图关系怎样成为学习目标，又会牺牲什么？"},"page":{"id":"h-i-collapse-test","title":"代入常量表示，检查目标的平凡解","number":1},"sectionTitle":"代入常量表示，检查目标的平凡解","currentQuestion":"视图关系怎样成为学习目标，又会牺牲什么？","sourceSlides":["l3-r2-p11"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-8d79fdfc8eb8f066c0560162","nextBlock":"b-1025ba3bd4594767f13268f7","prerequisites":[{"id":"b-4762c00ffc15bb8616f8a644","url":"#b-4762c00ffc15bb8616f8a644","role":"section-title"},{"id":"b-35e38f2ba26d22fa95c9352e","url":"#b-35e38f2ba26d22fa95c9352e","role":"problem-context"}],"relatedBlocks":[{"id":"b-c269a99c9cec06976fc6ea1b","url":"#b-c269a99c9cec06976fc6ea1b","relation":"figure-and-explanation"}]},{"id":"b-1025ba3bd4594767f13268f7","type":"li","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p11.html#b-1025ba3bd4594767f13268f7","content":"- [一致性目标的平凡解｜什么结论还不能推出](/supplement.html?return=l3-r2-p11/0#/l3-collapse-teacher-boundary/0) L03补充阅读","sha256":"1fe7126c3086d263e32a424fb341240631d4ee41beadd5aa3291acc44fdb4350","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-view-objective","title":"视图关系怎样成为学习目标，又会牺牲什么？"},"page":{"id":"h-i-collapse-test","title":"代入常量表示，检查目标的平凡解","number":1},"sectionTitle":"代入常量表示，检查目标的平凡解","currentQuestion":"视图关系怎样成为学习目标，又会牺牲什么？","sourceSlides":["l3-r2-p11"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-a5e8189bc33781eb0938cdc6","nextBlock":null,"prerequisites":[{"id":"b-4762c00ffc15bb8616f8a644","url":"#b-4762c00ffc15bb8616f8a644","role":"section-title"},{"id":"b-35e38f2ba26d22fa95c9352e","url":"#b-35e38f2ba26d22fa95c9352e","role":"problem-context"}],"relatedBlocks":[{"id":"b-c269a99c9cec06976fc6ea1b","url":"#b-c269a99c9cec06976fc6ea1b","relation":"figure-and-explanation"}]},{"id":"b-8305b3913a6b3b281594bb8a","type":"h2","scope":"h-i-teacher-targets","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p11.html#b-8305b3913a6b3b281594bb8a","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":2},"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-fa644fa8dc1bd6820868fbe0","prerequisites":[{"id":"b-fa644fa8dc1bd6820868fbe0","url":"#b-fa644fa8dc1bd6820868fbe0","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-fa644fa8dc1bd6820868fbe0","type":"p","scope":"h-i-teacher-targets","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p11.html#b-fa644fa8dc1bd6820868fbe0","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":2},"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-8305b3913a6b3b281594bb8a","nextBlock":"b-868173bbb07725749854fbb3","prerequisites":[{"id":"b-8305b3913a6b3b281594bb8a","url":"#b-8305b3913a6b3b281594bb8a","role":"section-title"}],"relatedBlocks":[]},{"id":"b-868173bbb07725749854fbb3","type":"h3","scope":"teacher-target-source-and-collapse","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p11.html#b-868173bbb07725749854fbb3","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":2},"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-fa644fa8dc1bd6820868fbe0","nextBlock":"b-f8a0c93720f84f4b95c190c6","prerequisites":[{"id":"b-8305b3913a6b3b281594bb8a","url":"#b-8305b3913a6b3b281594bb8a","role":"section-title"},{"id":"b-fa644fa8dc1bd6820868fbe0","url":"#b-fa644fa8dc1bd6820868fbe0","role":"problem-context"}],"relatedBlocks":[{"id":"b-f8a0c93720f84f4b95c190c6","url":"#b-f8a0c93720f84f4b95c190c6","relation":"same-reasoning"},{"id":"b-2d769178d26329ce9d2a436c","url":"#b-2d769178d26329ce9d2a436c","relation":"same-reasoning"},{"id":"b-f1c395d374a763fcd8654a05","url":"#b-f1c395d374a763fcd8654a05","relation":"same-reasoning"}]},{"id":"b-f8a0c93720f84f4b95c190c6","type":"p","scope":"teacher-target-source-and-collapse","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p11.html#b-f8a0c93720f84f4b95c190c6","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":2},"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-868173bbb07725749854fbb3","nextBlock":"b-2d769178d26329ce9d2a436c","prerequisites":[{"id":"b-868173bbb07725749854fbb3","url":"#b-868173bbb07725749854fbb3","role":"section-title"},{"id":"b-fa644fa8dc1bd6820868fbe0","url":"#b-fa644fa8dc1bd6820868fbe0","role":"problem-context"}],"relatedBlocks":[{"id":"b-868173bbb07725749854fbb3","url":"#b-868173bbb07725749854fbb3","relation":"same-reasoning"},{"id":"b-2d769178d26329ce9d2a436c","url":"#b-2d769178d26329ce9d2a436c","relation":"same-reasoning"},{"id":"b-f1c395d374a763fcd8654a05","url":"#b-f1c395d374a763fcd8654a05","relation":"same-reasoning"}]},{"id":"b-2d769178d26329ce9d2a436c","type":"p","scope":"teacher-target-source-and-collapse","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p11.html#b-2d769178d26329ce9d2a436c","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":2},"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-f8a0c93720f84f4b95c190c6","nextBlock":"b-f1c395d374a763fcd8654a05","prerequisites":[{"id":"b-868173bbb07725749854fbb3","url":"#b-868173bbb07725749854fbb3","role":"section-title"},{"id":"b-fa644fa8dc1bd6820868fbe0","url":"#b-fa644fa8dc1bd6820868fbe0","role":"problem-context"}],"relatedBlocks":[{"id":"b-868173bbb07725749854fbb3","url":"#b-868173bbb07725749854fbb3","relation":"same-reasoning"},{"id":"b-f8a0c93720f84f4b95c190c6","url":"#b-f8a0c93720f84f4b95c190c6","relation":"same-reasoning"},{"id":"b-f1c395d374a763fcd8654a05","url":"#b-f1c395d374a763fcd8654a05","relation":"same-reasoning"}]},{"id":"b-f1c395d374a763fcd8654a05","type":"p","scope":"teacher-target-source-and-collapse","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p11.html#b-f1c395d374a763fcd8654a05","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":2},"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-2d769178d26329ce9d2a436c","nextBlock":"b-281c245a6e6f295bba3f321a","prerequisites":[{"id":"b-868173bbb07725749854fbb3","url":"#b-868173bbb07725749854fbb3","role":"section-title"},{"id":"b-fa644fa8dc1bd6820868fbe0","url":"#b-fa644fa8dc1bd6820868fbe0","role":"problem-context"}],"relatedBlocks":[{"id":"b-868173bbb07725749854fbb3","url":"#b-868173bbb07725749854fbb3","relation":"same-reasoning"},{"id":"b-f8a0c93720f84f4b95c190c6","url":"#b-f8a0c93720f84f4b95c190c6","relation":"same-reasoning"},{"id":"b-2d769178d26329ce9d2a436c","url":"#b-2d769178d26329ce9d2a436c","relation":"same-reasoning"}]},{"id":"b-281c245a6e6f295bba3f321a","type":"h2","scope":"practice","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p11.html#b-281c245a6e6f295bba3f321a","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":2},"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-f1c395d374a763fcd8654a05","nextBlock":"b-e2ef084fc9bdb61d2d725d9a","prerequisites":[{"id":"b-868173bbb07725749854fbb3","url":"#b-868173bbb07725749854fbb3","role":"section-title"},{"id":"b-fa644fa8dc1bd6820868fbe0","url":"#b-fa644fa8dc1bd6820868fbe0","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-e2ef084fc9bdb61d2d725d9a","type":"p","scope":"practice","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p11.html#b-e2ef084fc9bdb61d2d725d9a","content":"用K=3手算常量对比loss；想追问无负例学习，再选LAB 17，不必先学整个DINO。","sha256":"7f4d7f7d2df9d1f82ac8d3bac4995a7069224eda7e147de23540e2c839b19a4a","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":2},"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-281c245a6e6f295bba3f321a","nextBlock":null,"prerequisites":[{"id":"b-281c245a6e6f295bba3f321a","url":"#b-281c245a6e6f295bba3f321a","role":"section-title"},{"id":"b-fa644fa8dc1bd6820868fbe0","url":"#b-fa644fa8dc1bd6820868fbe0","role":"problem-context"}],"relatedBlocks":[]}]}