{"version":2,"relationshipURLBase":"https://codingai-lec03.pages.dev/course/notebook-blocks/5dc1a98878994b8989105349ef7bd0c6.html","relationshipResolution":"Resolve relative relationship URLs against relationshipURLBase; complete records and absolute URLs are in blocks.","title":"MAE 让昂贵编码器只看少量位置","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/5dc1a98878994b8989105349ef7bd0c6.html","markdown":"https://codingai-lec03.pages.dev/course/notebook-blocks/5dc1a98878994b8989105349ef7bd0c6.md","pages":[{"id":"reading-opening","title":"MAE 让昂贵编码器只看少量位置","unit":"MAE 让昂贵编码器只看少量位置","unitId":"notebook-blocks/5dc1a98878994b8989105349ef7bd0c6","chapter":"course","number":1}],"aliases":[{"oldId":"b-2458bcd11b986f3f9226a700","targetId":"b-2458bcd11b986f3f9226a700","page":"reading-opening","operation":"move","beforeSHA256":"b7d46a4de36382cd644f1e71aaeaa89f3bbda6ead2c255e2d0f3cffd9453a606","afterSHA256":"b7d46a4de36382cd644f1e71aaeaa89f3bbda6ead2c255e2d0f3cffd9453a606","reason":"Same source body in authored semantic page"},{"oldId":"b-330d23bd06e51a1a2e1d4527","targetId":"b-330d23bd06e51a1a2e1d4527","page":"reading-opening","operation":"move","beforeSHA256":"3b23c0af8215660bbfa3c749e87be57dc041af4fbf73f04b7db0703260949785","afterSHA256":"3b23c0af8215660bbfa3c749e87be57dc041af4fbf73f04b7db0703260949785","reason":"Same source body in authored semantic page"},{"oldId":"b-5fb19a87d2282865034b1d3a","targetId":"b-5fb19a87d2282865034b1d3a","page":"reading-opening","operation":"move","beforeSHA256":"e91f24e4e4ca40e2099a0d81d30db6568f4bf2074b499eeee6e3959617bd97ac","afterSHA256":"e91f24e4e4ca40e2099a0d81d30db6568f4bf2074b499eeee6e3959617bd97ac","reason":"Same source body in authored semantic page"},{"oldId":"b-267b8c42e81ade59ddf64bb7","targetId":"b-267b8c42e81ade59ddf64bb7","page":"reading-opening","operation":"move","beforeSHA256":"db36e128a2d5a5e8c2d5360cbf60671bab45cb5494bbdd81f27d6b7d95ea74e0","afterSHA256":"db36e128a2d5a5e8c2d5360cbf60671bab45cb5494bbdd81f27d6b7d95ea74e0","reason":"Same source body in authored semantic page"},{"oldId":"b-61e1c08c124f2e9847eb98ac","targetId":"b-61e1c08c124f2e9847eb98ac","page":"reading-opening","operation":"move","beforeSHA256":"94c92248638a95e30f63c78bc247c7bac3e334d71bda920fd01554a003dbda6d","afterSHA256":"94c92248638a95e30f63c78bc247c7bac3e334d71bda920fd01554a003dbda6d","reason":"Same source body in authored semantic page"},{"oldId":"b-9b958f967a4083ceff464353","targetId":"b-9b958f967a4083ceff464353","page":"reading-opening","operation":"move","beforeSHA256":"11689a6e9d1b70e1f58e6c9dc369e5b3e87fa0cb70c4c6f28d85cbea174eb60a","afterSHA256":"11689a6e9d1b70e1f58e6c9dc369e5b3e87fa0cb70c4c6f28d85cbea174eb60a","reason":"Same source body in authored semantic page"},{"oldId":"b-13f73e1ba7920053747057a9","targetId":"b-13f73e1ba7920053747057a9","page":"reading-opening","operation":"move","beforeSHA256":"5ac484a3f720ac4cf71140e2cb746a9e171f2c492b27c78bb5e433c48f6370f7","afterSHA256":"5ac484a3f720ac4cf71140e2cb746a9e171f2c492b27c78bb5e433c48f6370f7","reason":"Same source body in authored semantic page"},{"oldId":"b-01936ccb5de2597a9c4533de","targetId":"b-01936ccb5de2597a9c4533de","page":"reading-opening","operation":"move","beforeSHA256":"b5e0f378794cb8b5e98506bac2baa689d3b4fed7d73ae9b28175dfaaf27dbe4b","afterSHA256":"b5e0f378794cb8b5e98506bac2baa689d3b4fed7d73ae9b28175dfaaf27dbe4b","reason":"Same source body in authored semantic page"},{"oldId":"b-c15977e2314e8c82ce83206b","targetId":"b-c15977e2314e8c82ce83206b","page":"reading-opening","operation":"move","beforeSHA256":"3f5dc0bbd2e7534e9bd07a1f682b523e93864d9abfc9188f92d5dc37ff605a01","afterSHA256":"3f5dc0bbd2e7534e9bd07a1f682b523e93864d9abfc9188f92d5dc37ff605a01","reason":"Same source body in authored semantic page"},{"oldId":"b-33da961af4c1bf5b60a6bbae","targetId":"b-33da961af4c1bf5b60a6bbae","page":"reading-opening","operation":"move","beforeSHA256":"36259a44332912c3afa0a6cc99679f3f32046cdd68b393af2a095cb0dc8b8542","afterSHA256":"36259a44332912c3afa0a6cc99679f3f32046cdd68b393af2a095cb0dc8b8542","reason":"Same source body in authored semantic page"},{"oldId":"b-3e6369f3899fb605e2217bbd","targetId":"b-3e6369f3899fb605e2217bbd","page":"reading-opening","operation":"rewrite","beforeSHA256":"c0823b50d032b609150e0475d34c6006c1e48c8a07e376dcdbee74ebc8409eb8","afterSHA256":"762a8c5fbf71524b89e155ad28cc9e70823159c29af9b3f81e9709b17aef7982","reason":"Same source identity with figure semantics or reading context enriched"},{"oldId":"b-3c4b599d511183876d381c4d","targetId":"b-3c4b599d511183876d381c4d","page":"reading-opening","operation":"move","beforeSHA256":"bdd741431b0bfcf6895a3ec1e163d576a9a39b43e2d32f28568c43e170bf94d1","afterSHA256":"bdd741431b0bfcf6895a3ec1e163d576a9a39b43e2d32f28568c43e170bf94d1","reason":"Same source body in authored semantic page"},{"oldId":"b-28d74128e5b2fff19d16ed58","targetId":"b-28d74128e5b2fff19d16ed58","page":"reading-opening","operation":"move","beforeSHA256":"63ef714a664c8436ef49a563821c2fb3836648b0e9361da95f69018103d353e6","afterSHA256":"63ef714a664c8436ef49a563821c2fb3836648b0e9361da95f69018103d353e6","reason":"Same source body in authored semantic page"},{"oldId":"b-faee23d60e9e96cc84e55c3c","targetId":"b-faee23d60e9e96cc84e55c3c","page":"reading-opening","operation":"move","beforeSHA256":"ba893ef885b664c6949f9713db226dd1b41af3d792fd9ab9f3d46cee131de26b","afterSHA256":"ba893ef885b664c6949f9713db226dd1b41af3d792fd9ab9f3d46cee131de26b","reason":"Same source body in authored semantic page"},{"oldId":"b-63cd000f85622c97828d5bab","targetId":"b-63cd000f85622c97828d5bab","page":"reading-opening","operation":"rewrite","beforeSHA256":"3a586279318b165191ce70d3d763f8cfaa1fde74125422ac404a69584f33b5e0","afterSHA256":"abc9ce8051bfb0d16d084f87dc7c4be7312f5993868cfd50a3846998ba18fc0c","reason":"Same source identity with figure semantics or reading context enriched"},{"oldId":"b-98333d0d68335878ef6bb469","targetId":"b-98333d0d68335878ef6bb469","page":"reading-opening","operation":"move","beforeSHA256":"686b884b78320da1fd4e9f3041bd3fe80f96edebbbe915a529890ae251d76293","afterSHA256":"122ba20c7f0fe242e9c627f9ff6ce34ff9163bdabed0dd9bdccd57346b14e7aa","reason":"Same source body in authored semantic page"},{"oldId":"b-61c77784c598e2ee0ee9f14b","targetId":"b-61c77784c598e2ee0ee9f14b","page":"reading-opening","operation":"move","beforeSHA256":"6d6419fd9b8cf38b3215e50deaae56b50c8735f4332672de31f0b542104500bf","afterSHA256":"6d6419fd9b8cf38b3215e50deaae56b50c8735f4332672de31f0b542104500bf","reason":"Same source body in authored semantic page"},{"oldId":"b-21738eae9e9c97d664907c5d","targetId":"b-21738eae9e9c97d664907c5d","page":"reading-opening","operation":"move","beforeSHA256":"3299792b5c910a1b3a63b2c4677082dda6868df4222b47dc500d9c26add22fe1","afterSHA256":"3299792b5c910a1b3a63b2c4677082dda6868df4222b47dc500d9c26add22fe1","reason":"Same source body in authored semantic page"},{"oldId":"b-d8d093a0bbf4d5943d10f618","targetId":"b-d8d093a0bbf4d5943d10f618","page":"reading-opening","operation":"move","beforeSHA256":"342160d8a63afceb0d11b164fdae9c202f3e9be639da6a98adb895229b074ebe","afterSHA256":"342160d8a63afceb0d11b164fdae9c202f3e9be639da6a98adb895229b074ebe","reason":"Same source body in authored semantic page"},{"oldId":"block-5dc1a98878994b8989105349ef7bd0c6--mae-让昂贵编码器只看少量位置","targetId":"b-61e1c08c124f2e9847eb98ac","operation":"retain-source-identity"},{"oldId":"block-5dc1a98878994b8989105349ef7bd0c6--已运行的图像重建恢复更好表示评价另算","targetId":"b-28d74128e5b2fff19d16ed58","operation":"retain-source-identity"}],"blocks":[{"id":"b-2458bcd11b986f3f9226a700","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/5dc1a98878994b8989105349ef7bd0c6.html#b-2458bcd11b986f3f9226a700","content":"[CLS 04 · 重建：约束、位置与表示评价](/course/notebooks/23b0affc4b824723927bd0938fd2bbda.html)","sha256":"b7d46a4de36382cd644f1e71aaeaa89f3bbda6ead2c255e2d0f3cffd9453a606","chapter":{"id":"course","title":"MAE 让昂贵编码器只看少量位置"},"unit":{"id":"notebook-blocks/5dc1a98878994b8989105349ef7bd0c6","title":"MAE 让昂贵编码器只看少量位置"},"page":{"id":"reading-opening","title":"MAE 让昂贵编码器只看少量位置","number":1},"sectionTitle":"MAE 让昂贵编码器只看少量位置","currentQuestion":"MAE 让昂贵编码器只看少量位置","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":null,"nextBlock":"b-330d23bd06e51a1a2e1d4527","prerequisites":[],"relatedBlocks":[]},{"id":"b-330d23bd06e51a1a2e1d4527","type":"h1","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/5dc1a98878994b8989105349ef7bd0c6.html#b-330d23bd06e51a1a2e1d4527","content":"# MAE 让昂贵编码器只看少量位置","sha256":"3b23c0af8215660bbfa3c749e87be57dc041af4fbf73f04b7db0703260949785","chapter":{"id":"course","title":"MAE 让昂贵编码器只看少量位置"},"unit":{"id":"notebook-blocks/5dc1a98878994b8989105349ef7bd0c6","title":"MAE 让昂贵编码器只看少量位置"},"page":{"id":"reading-opening","title":"MAE 让昂贵编码器只看少量位置","number":1},"sectionTitle":"MAE 让昂贵编码器只看少量位置","currentQuestion":"MAE 让昂贵编码器只看少量位置","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-2458bcd11b986f3f9226a700","nextBlock":"b-5fb19a87d2282865034b1d3a","prerequisites":[{"id":"b-2458bcd11b986f3f9226a700","url":"#b-2458bcd11b986f3f9226a700","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-5fb19a87d2282865034b1d3a","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/5dc1a98878994b8989105349ef7bd0c6.html#b-5fb19a87d2282865034b1d3a","content":"按问题继续","sha256":"e91f24e4e4ca40e2099a0d81d30db6568f4bf2074b499eeee6e3959617bd97ac","chapter":{"id":"course","title":"MAE 让昂贵编码器只看少量位置"},"unit":{"id":"notebook-blocks/5dc1a98878994b8989105349ef7bd0c6","title":"MAE 让昂贵编码器只看少量位置"},"page":{"id":"reading-opening","title":"MAE 让昂贵编码器只看少量位置","number":1},"sectionTitle":"MAE 让昂贵编码器只看少量位置","currentQuestion":"MAE 让昂贵编码器只看少量位置","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-330d23bd06e51a1a2e1d4527","nextBlock":"b-267b8c42e81ade59ddf64bb7","prerequisites":[{"id":"b-330d23bd06e51a1a2e1d4527","url":"#b-330d23bd06e51a1a2e1d4527","role":"section-title"},{"id":"b-2458bcd11b986f3f9226a700","url":"#b-2458bcd11b986f3f9226a700","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-267b8c42e81ade59ddf64bb7","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/5dc1a98878994b8989105349ef7bd0c6.html#b-267b8c42e81ade59ddf64bb7","content":"词典/PCA数值、线性自编码约束、遮挡索引、masked ConvAE与MAE区别","sha256":"db36e128a2d5a5e8c2d5360cbf60671bab45cb5494bbdd81f27d6b7d95ea74e0","chapter":{"id":"course","title":"MAE 让昂贵编码器只看少量位置"},"unit":{"id":"notebook-blocks/5dc1a98878994b8989105349ef7bd0c6","title":"MAE 让昂贵编码器只看少量位置"},"page":{"id":"reading-opening","title":"MAE 让昂贵编码器只看少量位置","number":1},"sectionTitle":"MAE 让昂贵编码器只看少量位置","currentQuestion":"MAE 让昂贵编码器只看少量位置","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-5fb19a87d2282865034b1d3a","nextBlock":"b-61e1c08c124f2e9847eb98ac","prerequisites":[{"id":"b-330d23bd06e51a1a2e1d4527","url":"#b-330d23bd06e51a1a2e1d4527","role":"section-title"},{"id":"b-2458bcd11b986f3f9226a700","url":"#b-2458bcd11b986f3f9226a700","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-61e1c08c124f2e9847eb98ac","type":"h2","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/5dc1a98878994b8989105349ef7bd0c6.html#b-61e1c08c124f2e9847eb98ac","content":"## MAE 让昂贵编码器只看少量位置","sha256":"94c92248638a95e30f63c78bc247c7bac3e334d71bda920fd01554a003dbda6d","chapter":{"id":"course","title":"MAE 让昂贵编码器只看少量位置"},"unit":{"id":"notebook-blocks/5dc1a98878994b8989105349ef7bd0c6","title":"MAE 让昂贵编码器只看少量位置"},"page":{"id":"reading-opening","title":"MAE 让昂贵编码器只看少量位置","number":1},"sectionTitle":"MAE 让昂贵编码器只看少量位置","currentQuestion":"MAE 让昂贵编码器只看少量位置","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-267b8c42e81ade59ddf64bb7","nextBlock":"b-9b958f967a4083ceff464353","prerequisites":[{"id":"b-330d23bd06e51a1a2e1d4527","url":"#b-330d23bd06e51a1a2e1d4527","role":"section-title"},{"id":"b-2458bcd11b986f3f9226a700","url":"#b-2458bcd11b986f3f9226a700","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-9b958f967a4083ceff464353","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/5dc1a98878994b8989105349ef7bd0c6.html#b-9b958f967a4083ceff464353","content":"把图像切为 patch 后，MAE 随机选择可见子集，编码器只处理可见 patch 及其位置；轻量解码器补入 mask token，恢复完整位置顺序，再预测被遮挡的像素目标。原 MAE 的一个关键设计是高遮挡率与这种不对称编码—解码结构，见 [MAE §3 与图 1](https://arxiv.org/html/2111.06377v3)。这里不是先让编码器看完整图，再在输出端涂黑几块。","sha256":"11689a6e9d1b70e1f58e6c9dc369e5b3e87fa0cb70c4c6f28d85cbea174eb60a","chapter":{"id":"course","title":"MAE 让昂贵编码器只看少量位置"},"unit":{"id":"notebook-blocks/5dc1a98878994b8989105349ef7bd0c6","title":"MAE 让昂贵编码器只看少量位置"},"page":{"id":"reading-opening","title":"MAE 让昂贵编码器只看少量位置","number":1},"sectionTitle":"MAE 让昂贵编码器只看少量位置","currentQuestion":"MAE 让昂贵编码器只看少量位置","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-61e1c08c124f2e9847eb98ac","nextBlock":"b-13f73e1ba7920053747057a9","prerequisites":[{"id":"b-61e1c08c124f2e9847eb98ac","url":"#b-61e1c08c124f2e9847eb98ac","role":"section-title"},{"id":"b-2458bcd11b986f3f9226a700","url":"#b-2458bcd11b986f3f9226a700","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-13f73e1ba7920053747057a9","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/5dc1a98878994b8989105349ef7bd0c6.html#b-13f73e1ba7920053747057a9","content":"四个原始位置是 `[0,1,2,3]`。随机排列 `ids_shuffle=[2,0,3,1]`，只保留前两项，所以编码器看到原位置 2、0。解码器先得到 `[z2,z0,M,M]`，再用逆排列 `ids_restore=[1,3,0,2]` 取回 `[z0,M,z2,M]`。位置 1、3 的损失可以按","sha256":"5ac484a3f720ac4cf71140e2cb746a9e171f2c492b27c78bb5e433c48f6370f7","chapter":{"id":"course","title":"MAE 让昂贵编码器只看少量位置"},"unit":{"id":"notebook-blocks/5dc1a98878994b8989105349ef7bd0c6","title":"MAE 让昂贵编码器只看少量位置"},"page":{"id":"reading-opening","title":"MAE 让昂贵编码器只看少量位置","number":1},"sectionTitle":"MAE 让昂贵编码器只看少量位置","currentQuestion":"MAE 让昂贵编码器只看少量位置","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-9b958f967a4083ceff464353","nextBlock":"b-01936ccb5de2597a9c4533de","prerequisites":[{"id":"b-61e1c08c124f2e9847eb98ac","url":"#b-61e1c08c124f2e9847eb98ac","role":"section-title"},{"id":"b-2458bcd11b986f3f9226a700","url":"#b-2458bcd11b986f3f9226a700","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-01936ccb5de2597a9c4533de","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/5dc1a98878994b8989105349ef7bd0c6.html#b-01936ccb5de2597a9c4533de","content":"$L=\\frac{1}{|\\mathcal M|}\\sum_{i\\in\\mathcal M}\\|\\hat x_i-x_i\\|^2$","sha256":"b5e0f378794cb8b5e98506bac2baa689d3b4fed7d73ae9b28175dfaaf27dbe4b","chapter":{"id":"course","title":"MAE 让昂贵编码器只看少量位置"},"unit":{"id":"notebook-blocks/5dc1a98878994b8989105349ef7bd0c6","title":"MAE 让昂贵编码器只看少量位置"},"page":{"id":"reading-opening","title":"MAE 让昂贵编码器只看少量位置","number":1},"sectionTitle":"MAE 让昂贵编码器只看少量位置","currentQuestion":"MAE 让昂贵编码器只看少量位置","sourceSlides":[],"knowledgeBlock":null,"formulas":[{"latex":"L=\\frac{1}{|\\mathcal M|}\\sum_{i\\in\\mathcal M}\\|\\hat x_i-x_i\\|^2","definitionContext":"MAE 让昂贵编码器只看少量位置"}],"previousBlock":"b-13f73e1ba7920053747057a9","nextBlock":"b-c15977e2314e8c82ce83206b","prerequisites":[{"id":"b-61e1c08c124f2e9847eb98ac","url":"#b-61e1c08c124f2e9847eb98ac","role":"section-title"},{"id":"b-2458bcd11b986f3f9226a700","url":"#b-2458bcd11b986f3f9226a700","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-c15977e2314e8c82ce83206b","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/5dc1a98878994b8989105349ef7bd0c6.html#b-c15977e2314e8c82ce83206b","content":"计算，其中 $\\mathcal M$ 为遮挡位置集合。若补 mask 后忘记恢复顺序，即便预测数值合理，也是在与错误位置的目标比较。排列练习可以证明索引正确；它不能证明编码器已经学会图像语义。","sha256":"3f5dc0bbd2e7534e9bd07a1f682b523e93864d9abfc9188f92d5dc37ff605a01","chapter":{"id":"course","title":"MAE 让昂贵编码器只看少量位置"},"unit":{"id":"notebook-blocks/5dc1a98878994b8989105349ef7bd0c6","title":"MAE 让昂贵编码器只看少量位置"},"page":{"id":"reading-opening","title":"MAE 让昂贵编码器只看少量位置","number":1},"sectionTitle":"MAE 让昂贵编码器只看少量位置","currentQuestion":"MAE 让昂贵编码器只看少量位置","sourceSlides":[],"knowledgeBlock":null,"formulas":[{"latex":"\\mathcal M","definitionContext":"MAE 让昂贵编码器只看少量位置"}],"previousBlock":"b-01936ccb5de2597a9c4533de","nextBlock":"b-33da961af4c1bf5b60a6bbae","prerequisites":[{"id":"b-61e1c08c124f2e9847eb98ac","url":"#b-61e1c08c124f2e9847eb98ac","role":"section-title"},{"id":"b-2458bcd11b986f3f9226a700","url":"#b-2458bcd11b986f3f9226a700","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-33da961af4c1bf5b60a6bbae","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/5dc1a98878994b8989105349ef7bd0c6.html#b-33da961af4c1bf5b60a6bbae","content":"再看计算：16 个 patch 只保留 4 个，编码器注意力的成对位置数由 $16^2=256$ 变为 $4^2=16$，这是二次项的 1/16；线性层、patch embedding、解码器和实际硬件开销不按同样比例缩小。因此不能把这个比例直接称为端到端加速 16 倍。高 mask 比例同时改变任务难度与编码成本，需要一起讨论。","sha256":"36259a44332912c3afa0a6cc99679f3f32046cdd68b393af2a095cb0dc8b8542","chapter":{"id":"course","title":"MAE 让昂贵编码器只看少量位置"},"unit":{"id":"notebook-blocks/5dc1a98878994b8989105349ef7bd0c6","title":"MAE 让昂贵编码器只看少量位置"},"page":{"id":"reading-opening","title":"MAE 让昂贵编码器只看少量位置","number":1},"sectionTitle":"MAE 让昂贵编码器只看少量位置","currentQuestion":"MAE 让昂贵编码器只看少量位置","sourceSlides":[],"knowledgeBlock":null,"formulas":[{"latex":"16^2=256","definitionContext":"MAE 让昂贵编码器只看少量位置"},{"latex":"4^2=16","definitionContext":"MAE 让昂贵编码器只看少量位置"}],"previousBlock":"b-c15977e2314e8c82ce83206b","nextBlock":"b-3e6369f3899fb605e2217bbd","prerequisites":[{"id":"b-61e1c08c124f2e9847eb98ac","url":"#b-61e1c08c124f2e9847eb98ac","role":"section-title"},{"id":"b-2458bcd11b986f3f9226a700","url":"#b-2458bcd11b986f3f9226a700","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-3e6369f3899fb605e2217bbd","type":"table","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/5dc1a98878994b8989105349ef7bd0c6.html#b-3e6369f3899fb605e2217bbd","content":"| 方法 | 输入受什么限制 | 代码/表示约束 | 重建好能支持什么判断 |\n| --- | --- | --- | --- |\n| K-means 词典 | 通常看到完整 patch | 一个中心索引 | 该词典的量化误差小 |\n| PCA | 通常看到完整向量 | 低维线性子空间 | 方差被线性子空间有效保留 |\n| 去噪自编码器 | 看到明确破坏后的图 | 架构、瓶颈及正则化 | 对指定破坏能恢复保留图像 |\n| MAE | 编码器只看可见 patch | 可见位置及编码—解码设计 | 能利用上下文预测遮挡目标 |","sha256":"762a8c5fbf71524b89e155ad28cc9e70823159c29af9b3f81e9709b17aef7982","chapter":{"id":"course","title":"MAE 让昂贵编码器只看少量位置"},"unit":{"id":"notebook-blocks/5dc1a98878994b8989105349ef7bd0c6","title":"MAE 让昂贵编码器只看少量位置"},"page":{"id":"reading-opening","title":"MAE 让昂贵编码器只看少量位置","number":1},"sectionTitle":"MAE 让昂贵编码器只看少量位置","currentQuestion":"MAE 让昂贵编码器只看少量位置","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"table":{"headers":["方法","输入受什么限制","代码/表示约束","重建好能支持什么判断"],"rows":[["方法","输入受什么限制","代码/表示约束","重建好能支持什么判断"],["K-means 词典","通常看到完整 patch","一个中心索引","该词典的量化误差小"],["PCA","通常看到完整向量","低维线性子空间","方差被线性子空间有效保留"],["去噪自编码器","看到明确破坏后的图","架构、瓶颈及正则化","对指定破坏能恢复保留图像"],["MAE","编码器只看可见 patch","可见位置及编码—解码设计","能利用上下文预测遮挡目标"]]},"previousBlock":"b-33da961af4c1bf5b60a6bbae","nextBlock":"b-3c4b599d511183876d381c4d","prerequisites":[{"id":"b-61e1c08c124f2e9847eb98ac","url":"#b-61e1c08c124f2e9847eb98ac","role":"section-title"},{"id":"b-2458bcd11b986f3f9226a700","url":"#b-2458bcd11b986f3f9226a700","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-3c4b599d511183876d381c4d","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/5dc1a98878994b8989105349ef7bd0c6.html#b-3c4b599d511183876d381c4d","content":"这四行都没有自动回答“少量标签能否读出类别”“颜色是否保留”“边界是否准确”。那些是另设的评价问题。[LAB 08](https://deepnote.com/project/38755e4e-23f6-4f5f-8c68-8a2e4114b1c1/notebook/b5ba27f2c5ba4ec0922208a7d5dc80a4) 把索引检查与图像重建分开：先确认输入—mask—目标位置，再比较输入、重建和误差，最后冻结编码器作读出。划分必须先按原图进行，再生成破坏版本；否则同一原图的干净版和带噪版可能分居训练与测试，给出过于乐观的结果。","sha256":"bdd741431b0bfcf6895a3ec1e163d576a9a39b43e2d32f28568c43e170bf94d1","chapter":{"id":"course","title":"MAE 让昂贵编码器只看少量位置"},"unit":{"id":"notebook-blocks/5dc1a98878994b8989105349ef7bd0c6","title":"MAE 让昂贵编码器只看少量位置"},"page":{"id":"reading-opening","title":"MAE 让昂贵编码器只看少量位置","number":1},"sectionTitle":"MAE 让昂贵编码器只看少量位置","currentQuestion":"MAE 让昂贵编码器只看少量位置","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-3e6369f3899fb605e2217bbd","nextBlock":"b-28d74128e5b2fff19d16ed58","prerequisites":[{"id":"b-61e1c08c124f2e9847eb98ac","url":"#b-61e1c08c124f2e9847eb98ac","role":"section-title"},{"id":"b-2458bcd11b986f3f9226a700","url":"#b-2458bcd11b986f3f9226a700","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-28d74128e5b2fff19d16ed58","type":"h3","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/5dc1a98878994b8989105349ef7bd0c6.html#b-28d74128e5b2fff19d16ed58","content":"### 已运行的图像重建：恢复更好，表示评价另算","sha256":"63ef714a664c8436ef49a563821c2fb3836648b0e9361da95f69018103d353e6","chapter":{"id":"course","title":"MAE 让昂贵编码器只看少量位置"},"unit":{"id":"notebook-blocks/5dc1a98878994b8989105349ef7bd0c6","title":"MAE 让昂贵编码器只看少量位置"},"page":{"id":"reading-opening","title":"MAE 让昂贵编码器只看少量位置","number":1},"sectionTitle":"MAE 让昂贵编码器只看少量位置","currentQuestion":"MAE 让昂贵编码器只看少量位置","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-3c4b599d511183876d381c4d","nextBlock":"b-faee23d60e9e96cc84e55c3c","prerequisites":[{"id":"b-61e1c08c124f2e9847eb98ac","url":"#b-61e1c08c124f2e9847eb98ac","role":"section-title"},{"id":"b-2458bcd11b986f3f9226a700","url":"#b-2458bcd11b986f3f9226a700","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-faee23d60e9e96cc84e55c3c","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/5dc1a98878994b8989105349ef7bd0c6.html#b-faee23d60e9e96cc84e55c3c","content":"LAB 08 在 2026-09-22 本地 CPU 从头训练卷积自编码器。数据是 `sklearn load_digits` 的 1797 张 8×8 灰度图，先按原图分层随机划分 1257 训练、540 保留，再生成破坏；没有按书写者分组，因此结论是这个原图随机划分下的数字任务。标签只用于划分与读出，不进入重建损失。16 维瓶颈的去噪网络训练 40 epoch，破坏为 25% 像素丢弃加标准差 0.15 的高斯噪声并裁到 [0,1]。","sha256":"ba893ef885b664c6949f9713db226dd1b41af3d792fd9ab9f3d46cee131de26b","chapter":{"id":"course","title":"MAE 让昂贵编码器只看少量位置"},"unit":{"id":"notebook-blocks/5dc1a98878994b8989105349ef7bd0c6","title":"MAE 让昂贵编码器只看少量位置"},"page":{"id":"reading-opening","title":"MAE 让昂贵编码器只看少量位置","number":1},"sectionTitle":"MAE 让昂贵编码器只看少量位置","currentQuestion":"MAE 让昂贵编码器只看少量位置","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-28d74128e5b2fff19d16ed58","nextBlock":"b-63cd000f85622c97828d5bab","prerequisites":[{"id":"b-28d74128e5b2fff19d16ed58","url":"#b-28d74128e5b2fff19d16ed58","role":"section-title"},{"id":"b-2458bcd11b986f3f9226a700","url":"#b-2458bcd11b986f3f9226a700","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-63cd000f85622c97828d5bab","type":"table","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/5dc1a98878994b8989105349ef7bd0c6.html#b-63cd000f85622c97828d5bab","content":"| 保留集方法 | 与干净目标的逐像素 MSE |\n| --- | --- |\n| 直接使用带噪输入 | 0.06244 |\n| 训练集拟合的 rank-16 PCA 重建 | 0.04160 |\n| 学到的卷积去噪自编码器 | 0.03217 |","sha256":"abc9ce8051bfb0d16d084f87dc7c4be7312f5993868cfd50a3846998ba18fc0c","chapter":{"id":"course","title":"MAE 让昂贵编码器只看少量位置"},"unit":{"id":"notebook-blocks/5dc1a98878994b8989105349ef7bd0c6","title":"MAE 让昂贵编码器只看少量位置"},"page":{"id":"reading-opening","title":"MAE 让昂贵编码器只看少量位置","number":1},"sectionTitle":"MAE 让昂贵编码器只看少量位置","currentQuestion":"MAE 让昂贵编码器只看少量位置","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"table":{"headers":["保留集方法","与干净目标的逐像素 MSE"],"rows":[["保留集方法","与干净目标的逐像素 MSE"],["直接使用带噪输入","0.06244"],["训练集拟合的 rank-16 PCA 重建","0.04160"],["学到的卷积去噪自编码器","0.03217"]]},"previousBlock":"b-faee23d60e9e96cc84e55c3c","nextBlock":"b-98333d0d68335878ef6bb469","prerequisites":[{"id":"b-28d74128e5b2fff19d16ed58","url":"#b-28d74128e5b2fff19d16ed58","role":"section-title"},{"id":"b-2458bcd11b986f3f9226a700","url":"#b-2458bcd11b986f3f9226a700","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-98333d0d68335878ef6bb469","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/5dc1a98878994b8989105349ef7bd0c6.html#b-98333d0d68335878ef6bb469","content":"![真实输入、破坏、重建与遮挡恢复](/assets/display/cdcfe2b4d2103d5484256b71.webp)","sha256":"122ba20c7f0fe242e9c627f9ff6ce34ff9163bdabed0dd9bdccd57346b14e7aa","chapter":{"id":"course","title":"MAE 让昂贵编码器只看少量位置"},"unit":{"id":"notebook-blocks/5dc1a98878994b8989105349ef7bd0c6","title":"MAE 让昂贵编码器只看少量位置"},"page":{"id":"reading-opening","title":"MAE 让昂贵编码器只看少量位置","number":1},"sectionTitle":"MAE 让昂贵编码器只看少量位置","currentQuestion":"MAE 让昂贵编码器只看少量位置","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-63cd000f85622c97828d5bab","nextBlock":"b-61c77784c598e2ee0ee9f14b","prerequisites":[{"id":"b-28d74128e5b2fff19d16ed58","url":"#b-28d74128e5b2fff19d16ed58","role":"section-title"},{"id":"b-2458bcd11b986f3f9226a700","url":"#b-2458bcd11b986f3f9226a700","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-61c77784c598e2ee0ee9f14b","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/5dc1a98878994b8989105349ef7bd0c6.html#b-61c77784c598e2ee0ee9f14b","content":"另一个两通道卷积模型接收“被遮挡图像＋可见性 mask”，在 50% 的 2×2 patch 隐藏率下训练 50 epoch。它处理整张密集网格，**不是原论文只编码可见 token 的 ViT MAE**。固定每图同一排列、让隐藏集合嵌套，测试 25%/50%/75% 遮挡时，隐藏位置 MSE 分别为 0.03933/0.04341/0.05563；相应训练均值基线为 0.07420/0.07402/0.07390。遮挡增加后重建变难，但这个卷积模型并未获得 MAE 的可见 token 计算节省。","sha256":"6d6419fd9b8cf38b3215e50deaae56b50c8735f4332672de31f0b542104500bf","chapter":{"id":"course","title":"MAE 让昂贵编码器只看少量位置"},"unit":{"id":"notebook-blocks/5dc1a98878994b8989105349ef7bd0c6","title":"MAE 让昂贵编码器只看少量位置"},"page":{"id":"reading-opening","title":"MAE 让昂贵编码器只看少量位置","number":1},"sectionTitle":"MAE 让昂贵编码器只看少量位置","currentQuestion":"MAE 让昂贵编码器只看少量位置","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-98333d0d68335878ef6bb469","nextBlock":"b-21738eae9e9c97d664907c5d","prerequisites":[{"id":"b-28d74128e5b2fff19d16ed58","url":"#b-28d74128e5b2fff19d16ed58","role":"section-title"},{"id":"b-2458bcd11b986f3f9226a700","url":"#b-2458bcd11b986f3f9226a700","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-21738eae9e9c97d664907c5d","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/5dc1a98878994b8989105349ef7bd0c6.html#b-21738eae9e9c97d664907c5d","content":"冻结编码器、输入干净保留图像作 5NN，去噪架构由随机初始化的 84.26% 到训练后的 95.00%；遮挡架构由 88.52% 到 92.04%。两种架构输入通道不同，随机基线也应分别保留，不能把它们混成同一个起点。重建改善和读出改善在这次实验中同时出现，但它们仍是两项独立测量。配置、输入 hash 与权重均在 LAB 08 对应材料中。","sha256":"3299792b5c910a1b3a63b2c4677082dda6868df4222b47dc500d9c26add22fe1","chapter":{"id":"course","title":"MAE 让昂贵编码器只看少量位置"},"unit":{"id":"notebook-blocks/5dc1a98878994b8989105349ef7bd0c6","title":"MAE 让昂贵编码器只看少量位置"},"page":{"id":"reading-opening","title":"MAE 让昂贵编码器只看少量位置","number":1},"sectionTitle":"MAE 让昂贵编码器只看少量位置","currentQuestion":"MAE 让昂贵编码器只看少量位置","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-61c77784c598e2ee0ee9f14b","nextBlock":"b-d8d093a0bbf4d5943d10f618","prerequisites":[{"id":"b-28d74128e5b2fff19d16ed58","url":"#b-28d74128e5b2fff19d16ed58","role":"section-title"},{"id":"b-2458bcd11b986f3f9226a700","url":"#b-2458bcd11b986f3f9226a700","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-d8d093a0bbf4d5943d10f618","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/5dc1a98878994b8989105349ef7bd0c6.html#b-d8d093a0bbf4d5943d10f618","content":"重建为“保存信息”提供了可计算目标；下一册 [CLS 05](https://deepnote.com/project/38755e4e-23f6-4f5f-8c68-8a2e4114b1c1/notebook/b8320404134f4a2e858ccfe7d50a71d3) 会主动要求某些变化在表示中不产生差异。二者的张力是具体的：准确重建颜色需要留下颜色，颜色不变的对比目标却可能鼓励忽略它。好的表示要服务什么，必须由下游问题回答。","sha256":"342160d8a63afceb0d11b164fdae9c202f3e9be639da6a98adb895229b074ebe","chapter":{"id":"course","title":"MAE 让昂贵编码器只看少量位置"},"unit":{"id":"notebook-blocks/5dc1a98878994b8989105349ef7bd0c6","title":"MAE 让昂贵编码器只看少量位置"},"page":{"id":"reading-opening","title":"MAE 让昂贵编码器只看少量位置","number":1},"sectionTitle":"MAE 让昂贵编码器只看少量位置","currentQuestion":"MAE 让昂贵编码器只看少量位置","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-21738eae9e9c97d664907c5d","nextBlock":null,"prerequisites":[{"id":"b-28d74128e5b2fff19d16ed58","url":"#b-28d74128e5b2fff19d16ed58","role":"section-title"},{"id":"b-2458bcd11b986f3f9226a700","url":"#b-2458bcd11b986f3f9226a700","role":"problem-context"}],"relatedBlocks":[]}]}
