{"version":2,"relationshipURLBase":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p22.html","relationshipResolution":"Resolve relative relationship URLs against relationshipURLBase; complete records and absolute URLs are in blocks.","title":"能补图，不等于每项任务都读得好","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p22.html","markdown":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p22.md","pages":[{"id":"h-i-conditional-mean","title":"平方误差为什么偏好可能模糊的平均值","unit":"限制可见信息，怎样构成可验证的重建任务？","unitId":"i-masked-prediction","chapter":"I","number":1},{"id":"h-i-reconstruction-and-readout-evidence","title":"读 LAB08：重建和读出各交一份成绩","unit":"限制可见信息，怎样构成可验证的重建任务？","unitId":"i-masked-prediction","chapter":"I","number":2},{"id":"h-i-mae-evaluation-interface","title":"MAE 的正面证据，也要说清使用接口","unit":"限制可见信息，怎样构成可验证的重建任务？","unitId":"i-masked-prediction","chapter":"I","number":3},{"id":"h-i-architecture-and-target","title":"选读：结构和重建目标需要共同检验","unit":"表示学习的选读问题","unitId":"i-extensions","chapter":"I","number":4}],"aliases":[{"oldId":"explanation","targetId":"h-i-reconstruction-and-readout-evidence","operation":"split-explanation-range"},{"oldId":"b-1c8abafd47b2eb5164aa85cc","targetId":"b-1c8abafd47b2eb5164aa85cc","page":"h-i-conditional-mean","operation":"move","beforeSHA256":"d36a29e68458bc4431c86b5ebb1f5058bcc0303161d609f34d45264b9dc5f22a","afterSHA256":"d36a29e68458bc4431c86b5ebb1f5058bcc0303161d609f34d45264b9dc5f22a","reason":"Same source body in authored semantic page"},{"oldId":"b-7cda26dd3e456c00ab819f56","targetId":"b-7cda26dd3e456c00ab819f56","page":"h-i-conditional-mean","operation":"move","beforeSHA256":"de02bc9146e47128b33ca1647b6f0980b16b5f51af7fc3e9fb46d75b2ba76ea9","afterSHA256":"de02bc9146e47128b33ca1647b6f0980b16b5f51af7fc3e9fb46d75b2ba76ea9","reason":"Same source body in authored semantic page"},{"oldId":"b-1c37eccd1b21732c219d2732","targetId":"b-1c37eccd1b21732c219d2732","page":"h-i-conditional-mean","operation":"move","beforeSHA256":"4a0b4d1f376f9bb7978ec8811ca0e45417b4a451b5aebafc20a9bf1f249f542b","afterSHA256":"4a0b4d1f376f9bb7978ec8811ca0e45417b4a451b5aebafc20a9bf1f249f542b","reason":"Same source body in authored semantic page"},{"oldId":"b-d5d3717a06d232f47452624b","targetId":"b-d5d3717a06d232f47452624b","page":"h-i-conditional-mean","operation":"move","beforeSHA256":"41dbdd5dd96752401412f665b978bf02ddb76075360efa0153bbcffb00c8e182","afterSHA256":"41dbdd5dd96752401412f665b978bf02ddb76075360efa0153bbcffb00c8e182","reason":"Same source body in authored semantic page"},{"oldId":"b-911114295198745df9d53101","targetId":"b-911114295198745df9d53101","page":"h-i-reconstruction-and-readout-evidence","operation":"move","beforeSHA256":"d71cf15e707c8a3674b4c90a599efee783cb84da5b8c8236a1d01c3add7b4958","afterSHA256":"d71cf15e707c8a3674b4c90a599efee783cb84da5b8c8236a1d01c3add7b4958","reason":"Same source body in authored semantic page"},{"oldId":"b-8b59f0dd341b11507aea4473","targetId":"b-8b59f0dd341b11507aea4473","page":"h-i-reconstruction-and-readout-evidence","operation":"move","beforeSHA256":"bcf0c4cc6f9c787333c050ee1992c958a6e90af55d9cb400d2b7813c88e5af42","afterSHA256":"bcf0c4cc6f9c787333c050ee1992c958a6e90af55d9cb400d2b7813c88e5af42","reason":"Same source body in authored semantic page"},{"oldId":"b-f08cf3e632c681951f0051ce","targetId":"b-f08cf3e632c681951f0051ce","page":"h-i-reconstruction-and-readout-evidence","operation":"move","beforeSHA256":"8c276b31f27003561e22b48178ed6b725fbacc9a187aa3f629d2af3b2d3fa6b8","afterSHA256":"8c276b31f27003561e22b48178ed6b725fbacc9a187aa3f629d2af3b2d3fa6b8","reason":"Same source body in authored semantic page"},{"oldId":"b-3f3fdafaabbc82d5e3f1e116","targetId":"b-3f3fdafaabbc82d5e3f1e116","page":"h-i-reconstruction-and-readout-evidence","operation":"rewrite","beforeSHA256":"aebaf65b1e97a9cfab0c71a3376c50eda691f9814264129e01bf9d7264228d82","afterSHA256":"b52d8856e3b43ec40060c97a444157188897fefcdb79bd2ce0224b6101a47f2a","reason":"Same source identity with figure semantics or reading context enriched"},{"oldId":"b-7b9a1f32d818e86e2e753b30","targetId":"b-7b9a1f32d818e86e2e753b30","page":"h-i-reconstruction-and-readout-evidence","operation":"move","beforeSHA256":"786ce10a6b32a492259a3dccd0a54e61cd41d28a34382fb7b5386c5f03b7ad69","afterSHA256":"786ce10a6b32a492259a3dccd0a54e61cd41d28a34382fb7b5386c5f03b7ad69","reason":"Same source body in authored semantic page"},{"oldId":"b-abef257eddf6b5e1b4551b04","targetId":"b-abef257eddf6b5e1b4551b04","page":"h-i-reconstruction-and-readout-evidence","operation":"move","beforeSHA256":"8590b6339d1d319c5481671869d9d6f87ead41d0d98b610224cb808aa9c0af22","afterSHA256":"8590b6339d1d319c5481671869d9d6f87ead41d0d98b610224cb808aa9c0af22","reason":"Same source body in authored semantic page"},{"oldId":"b-6ad2c84650fe42daf0757e94","targetId":"b-6ad2c84650fe42daf0757e94","page":"h-i-reconstruction-and-readout-evidence","operation":"move","beforeSHA256":"b9200ecce5033ba762f0d64ac8b818952289e3d6e981d70f000aa4f19b70caeb","afterSHA256":"b9200ecce5033ba762f0d64ac8b818952289e3d6e981d70f000aa4f19b70caeb","reason":"Same source body in authored semantic page"},{"oldId":"b-d94d07f3dc4caec310e6442d","targetId":"b-d94d07f3dc4caec310e6442d","page":"h-i-reconstruction-and-readout-evidence","operation":"move","beforeSHA256":"5688d9ffd1e591d194670c4ea54b33a33cb66c35efcfafbe31b9c46a305139bd","afterSHA256":"5688d9ffd1e591d194670c4ea54b33a33cb66c35efcfafbe31b9c46a305139bd","reason":"Same source body in authored semantic page"},{"oldId":"b-f149c08474962de2aed78e77","targetId":"b-f149c08474962de2aed78e77","page":"h-i-reconstruction-and-readout-evidence","operation":"move","beforeSHA256":"b2ab128d3317f61761e18c7c9e0d2ec1c0ba28422fcafc6f986a276bb99a0955","afterSHA256":"b2ab128d3317f61761e18c7c9e0d2ec1c0ba28422fcafc6f986a276bb99a0955","reason":"Same source body in authored semantic page"},{"oldId":"b-f46bf2cc9d941792f221767a","targetId":"b-f46bf2cc9d941792f221767a","page":"h-i-reconstruction-and-readout-evidence","operation":"move","beforeSHA256":"1d7cb0553d5370b5f86c2bf98758fc9383e10d2917873fea007f494dc1ede516","afterSHA256":"1d7cb0553d5370b5f86c2bf98758fc9383e10d2917873fea007f494dc1ede516","reason":"Same source body in authored semantic page"},{"oldId":"b-4dcb596613f167b62fdca592","targetId":"b-4dcb596613f167b62fdca592","page":"h-i-reconstruction-and-readout-evidence","operation":"move","beforeSHA256":"cca4d3e84124f36eb656680a22b16a44449afb4b3f74a7c6ccbd479206280704","afterSHA256":"cca4d3e84124f36eb656680a22b16a44449afb4b3f74a7c6ccbd479206280704","reason":"Same source body in authored semantic page"},{"oldId":"b-078151c63112769bd21b4a8f","targetId":"b-078151c63112769bd21b4a8f","page":"h-i-reconstruction-and-readout-evidence","operation":"move","beforeSHA256":"6c0883a05332991cbfa598b3f0e397e3e0ad862432090d758bc0a0434408d15a","afterSHA256":"6c0883a05332991cbfa598b3f0e397e3e0ad862432090d758bc0a0434408d15a","reason":"Same source body in authored semantic page"},{"oldId":"b-c73406d907fc82afd6efbd5d","targetId":"b-c73406d907fc82afd6efbd5d","page":"h-i-reconstruction-and-readout-evidence","operation":"move","beforeSHA256":"c8a25dbabc3335471aa1a8f45dd97444af42929b3ab5fa6356bc36446c9dcfa9","afterSHA256":"c8a25dbabc3335471aa1a8f45dd97444af42929b3ab5fa6356bc36446c9dcfa9","reason":"Same source body in authored semantic page"},{"oldId":"b-b6f5e02415c65077622de063","targetId":"b-b6f5e02415c65077622de063","page":"h-i-reconstruction-and-readout-evidence","operation":"move","beforeSHA256":"118dc668367b10a51e704e3accb4bbaa0b20ebc3bf02733068ab51f95753ece8","afterSHA256":"118dc668367b10a51e704e3accb4bbaa0b20ebc3bf02733068ab51f95753ece8","reason":"Same source body in authored semantic page"},{"oldId":"b-188ca1b857bf2a1a67420278","targetId":"b-188ca1b857bf2a1a67420278","page":"h-i-reconstruction-and-readout-evidence","operation":"move","beforeSHA256":"c1c15b8807149bcbdab0ed47bb8dd971f501e0dc14fe351508bfd17f31ec0358","afterSHA256":"c1c15b8807149bcbdab0ed47bb8dd971f501e0dc14fe351508bfd17f31ec0358","reason":"Same source body in authored semantic page"},{"oldId":"b-0d44129b7af29f106bc8a208","targetId":"b-0d44129b7af29f106bc8a208","page":"h-i-mae-evaluation-interface","operation":"move","beforeSHA256":"8ce3eef83eb109d3b7630255be43a3fdb744e2cfc074570b15c8fd7432a1486c","afterSHA256":"8ce3eef83eb109d3b7630255be43a3fdb744e2cfc074570b15c8fd7432a1486c","reason":"Same source body in authored semantic page"},{"oldId":"b-414280c80a3926dbe0cab732","targetId":"b-414280c80a3926dbe0cab732","page":"h-i-mae-evaluation-interface","operation":"move","beforeSHA256":"8e45489df883e86d34df4691779f3cf7e9e27251094df0e7a920779e4922e03c","afterSHA256":"8e45489df883e86d34df4691779f3cf7e9e27251094df0e7a920779e4922e03c","reason":"Same source body in authored semantic page"},{"oldId":"b-6ef915ec7f990490a5f7f770","targetId":"b-6ef915ec7f990490a5f7f770","page":"h-i-mae-evaluation-interface","operation":"move","beforeSHA256":"f2f725904b17bcd54b2673d5a68bc81103ae385a1e35a3653c410868e548ec52","afterSHA256":"f2f725904b17bcd54b2673d5a68bc81103ae385a1e35a3653c410868e548ec52","reason":"Same source body in authored semantic page"},{"oldId":"b-5d70faa61f04bf4735c98958","targetId":"b-5d70faa61f04bf4735c98958","page":"h-i-mae-evaluation-interface","operation":"move","beforeSHA256":"857539d92a95c906eb9fcd172edca91e7dce5831d1fd66022c510cf90947bf93","afterSHA256":"857539d92a95c906eb9fcd172edca91e7dce5831d1fd66022c510cf90947bf93","reason":"Same source body in authored semantic page"},{"oldId":"b-4b5bb5061034104a36307b53","targetId":"b-4b5bb5061034104a36307b53","page":"h-i-architecture-and-target","operation":"move","beforeSHA256":"60344dd625f5f286022e81a326bb4283bca502621019d629ef7897d2863f6fe7","afterSHA256":"60344dd625f5f286022e81a326bb4283bca502621019d629ef7897d2863f6fe7","reason":"Same source body in authored semantic page"},{"oldId":"b-83fd46bf8aef1bcacf079a65","targetId":"b-83fd46bf8aef1bcacf079a65","page":"h-i-architecture-and-target","operation":"move","beforeSHA256":"f90a46d599e1e60c5744399504a4b565df7745bd6e5f034a33ed662df7be5117","afterSHA256":"f90a46d599e1e60c5744399504a4b565df7745bd6e5f034a33ed662df7be5117","reason":"Same source body in authored semantic page"},{"oldId":"b-c35729d284ad52e8061a7abb","targetId":"b-c35729d284ad52e8061a7abb","page":"h-i-architecture-and-target","operation":"move","beforeSHA256":"5f412f089df24a97ca57f905320ac07758cc419ed15af144a02d763a3d5b011d","afterSHA256":"5f412f089df24a97ca57f905320ac07758cc419ed15af144a02d763a3d5b011d","reason":"Same source body in authored semantic page"},{"oldId":"b-7f98852cbc75ebe13c692c9b","targetId":"b-7f98852cbc75ebe13c692c9b","page":"h-i-architecture-and-target","operation":"rewrite","beforeSHA256":"337c5ccc6ecfd699559a48b9211bf03849b063a69762709a661fff267be474ab","afterSHA256":"34076280b7dc70991b3bc8eef7d2adec72f68cfb32c548503d53852f2ee695d6","reason":"Same source identity with figure semantics or reading context enriched"},{"oldId":"b-7a68f9594175c2e15c84b528","targetId":"b-7a68f9594175c2e15c84b528","page":"h-i-architecture-and-target","operation":"move","beforeSHA256":"bc8d8b41f24cb8500fbdfecfb6289b5ef4dc3c5bcc086c466aaaa7caad8b4b15","afterSHA256":"bc8d8b41f24cb8500fbdfecfb6289b5ef4dc3c5bcc086c466aaaa7caad8b4b15","reason":"Same source body in authored semantic page"},{"oldId":"b-1bca931ab02760a029960de2","targetId":"b-1bca931ab02760a029960de2","page":"h-i-architecture-and-target","operation":"move","beforeSHA256":"cf44fcc62782fb4f6bdc42c16c5187d57556a03cb7ce6065e7bd9c36d8e7d733","afterSHA256":"cf44fcc62782fb4f6bdc42c16c5187d57556a03cb7ce6065e7bd9c36d8e7d733","reason":"Same source body in authored semantic page"},{"oldId":"b-c01907a791608f1f0f1f9bb0","targetId":"b-c01907a791608f1f0f1f9bb0","page":"h-i-architecture-and-target","operation":"move","beforeSHA256":"5bf6b9348a61224c9e47f1af545ae9404fd41534e1648bc6ed1b80d86160b455","afterSHA256":"5bf6b9348a61224c9e47f1af545ae9404fd41534e1648bc6ed1b80d86160b455","reason":"Same source body in authored semantic page"},{"oldId":"h-i-conditional-mean-title","targetId":"b-e64fbba78e7c95141085faad","operation":"retain-source-identity"},{"oldId":"h-i-conditional-mean-bridge","targetId":"b-0a276b7584301af39fb9a566","operation":"retain-source-identity"},{"oldId":"h-i-reconstruction-and-readout-evidence-title","targetId":"b-1ba94dc3f1723e1e846e7abd","operation":"retain-source-identity"},{"oldId":"h-i-mae-evaluation-interface-title","targetId":"b-bd9605313c93f4f5577a074c","operation":"retain-source-identity"},{"oldId":"h-i-mae-evaluation-interface-bridge","targetId":"b-969d668c82c7a1ced11f4ffc","operation":"retain-source-identity"},{"oldId":"h-i-architecture-and-target-title","targetId":"b-01394caa0965f81884dda6bb","operation":"retain-source-identity"},{"oldId":"h-i-architecture-and-target-bridge","targetId":"b-22a7a6fbc5e1014881f46408","operation":"retain-source-identity"}],"blocks":[{"id":"b-e64fbba78e7c95141085faad","type":"h2","scope":"h-i-conditional-mean","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p22.html#b-e64fbba78e7c95141085faad","content":"## 平方误差为什么偏好可能模糊的平均值","sha256":"0e60dc361726ddcf0733e86dab5e569b3befb5854bb69ec0feff9547be504cf6","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-masked-prediction","title":"限制可见信息，怎样构成可验证的重建任务？"},"page":{"id":"h-i-conditional-mean","title":"平方误差为什么偏好可能模糊的平均值","number":1},"sectionTitle":"平方误差为什么偏好可能模糊的平均值","currentQuestion":"限制可见信息，怎样构成可验证的重建任务？","sourceSlides":["l3-r2-p19","l3-r2-p22"],"knowledgeBlock":null,"formulas":[],"previousBlock":null,"nextBlock":"b-0a276b7584301af39fb9a566","prerequisites":[{"id":"b-0a276b7584301af39fb9a566","url":"#b-0a276b7584301af39fb9a566","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-0a276b7584301af39fb9a566","type":"p","scope":"h-i-conditional-mean","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p22.html#b-0a276b7584301af39fb9a566","content":"固定可见条件，让隐藏值有0和1两种等可能答案，再比较三个确定预测。","sha256":"e9eaf444b263dfb8448212562056e0156e928b82f842bffda8a6aecea26f0be6","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-masked-prediction","title":"限制可见信息，怎样构成可验证的重建任务？"},"page":{"id":"h-i-conditional-mean","title":"平方误差为什么偏好可能模糊的平均值","number":1},"sectionTitle":"平方误差为什么偏好可能模糊的平均值","currentQuestion":"限制可见信息，怎样构成可验证的重建任务？","sourceSlides":["l3-r2-p19","l3-r2-p22"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-e64fbba78e7c95141085faad","nextBlock":"b-1c8abafd47b2eb5164aa85cc","prerequisites":[{"id":"b-e64fbba78e7c95141085faad","url":"#b-e64fbba78e7c95141085faad","role":"section-title"}],"relatedBlocks":[]},{"id":"b-1c8abafd47b2eb5164aa85cc","type":"h3","scope":"squared-error-conditional-mean","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p22.html#b-1c8abafd47b2eb5164aa85cc","content":"### 为什么缺失处可能被补得模糊？","sha256":"d36a29e68458bc4431c86b5ebb1f5058bcc0303161d609f34d45264b9dc5f22a","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-masked-prediction","title":"限制可见信息，怎样构成可验证的重建任务？"},"page":{"id":"h-i-conditional-mean","title":"平方误差为什么偏好可能模糊的平均值","number":1},"sectionTitle":"为什么缺失处可能被补得模糊？","currentQuestion":"限制可见信息，怎样构成可验证的重建任务？","sourceSlides":["l3-r2-p19","l3-r2-p22"],"knowledgeBlock":"squared-error-conditional-mean","formulas":[],"previousBlock":"b-0a276b7584301af39fb9a566","nextBlock":"b-7cda26dd3e456c00ab819f56","prerequisites":[{"id":"b-e64fbba78e7c95141085faad","url":"#b-e64fbba78e7c95141085faad","role":"section-title"},{"id":"b-0a276b7584301af39fb9a566","url":"#b-0a276b7584301af39fb9a566","role":"problem-context"}],"relatedBlocks":[{"id":"b-7cda26dd3e456c00ab819f56","url":"#b-7cda26dd3e456c00ab819f56","relation":"same-reasoning"},{"id":"b-1c37eccd1b21732c219d2732","url":"#b-1c37eccd1b21732c219d2732","relation":"same-reasoning"},{"id":"b-d5d3717a06d232f47452624b","url":"#b-d5d3717a06d232f47452624b","relation":"same-reasoning"}]},{"id":"b-7cda26dd3e456c00ab819f56","type":"p","scope":"squared-error-conditional-mean","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p22.html#b-7cda26dd3e456c00ab819f56","content":"假设可见条件完全相同，而某个隐藏像素等概率为0或1。总输出0时，平均平方误差是0.5；总输出1也为0.5；输出0.5时，两种情况误差都是0.25，因此平均更小。","sha256":"de02bc9146e47128b33ca1647b6f0980b16b5f51af7fc3e9fb46d75b2ba76ea9","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-masked-prediction","title":"限制可见信息，怎样构成可验证的重建任务？"},"page":{"id":"h-i-conditional-mean","title":"平方误差为什么偏好可能模糊的平均值","number":1},"sectionTitle":"为什么缺失处可能被补得模糊？","currentQuestion":"限制可见信息，怎样构成可验证的重建任务？","sourceSlides":["l3-r2-p19","l3-r2-p22"],"knowledgeBlock":"squared-error-conditional-mean","formulas":[],"previousBlock":"b-1c8abafd47b2eb5164aa85cc","nextBlock":"b-1c37eccd1b21732c219d2732","prerequisites":[{"id":"b-1c8abafd47b2eb5164aa85cc","url":"#b-1c8abafd47b2eb5164aa85cc","role":"section-title"},{"id":"b-0a276b7584301af39fb9a566","url":"#b-0a276b7584301af39fb9a566","role":"problem-context"}],"relatedBlocks":[{"id":"b-1c8abafd47b2eb5164aa85cc","url":"#b-1c8abafd47b2eb5164aa85cc","relation":"same-reasoning"},{"id":"b-1c37eccd1b21732c219d2732","url":"#b-1c37eccd1b21732c219d2732","relation":"same-reasoning"},{"id":"b-d5d3717a06d232f47452624b","url":"#b-d5d3717a06d232f47452624b","relation":"same-reasoning"}]},{"id":"b-1c37eccd1b21732c219d2732","type":"p","scope":"squared-error-conditional-mean","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p22.html#b-1c37eccd1b21732c219d2732","content":"一般地，平方损失的最佳确定性预测是给定可见输入的条件均值。存在多种合理补全时，均值可能看起来模糊，即使模型已很好地优化这项损失。这与“完全没学到结构”不是同一个判断。","sha256":"4a0b4d1f376f9bb7978ec8811ca0e45417b4a451b5aebafc20a9bf1f249f542b","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-masked-prediction","title":"限制可见信息，怎样构成可验证的重建任务？"},"page":{"id":"h-i-conditional-mean","title":"平方误差为什么偏好可能模糊的平均值","number":1},"sectionTitle":"为什么缺失处可能被补得模糊？","currentQuestion":"限制可见信息，怎样构成可验证的重建任务？","sourceSlides":["l3-r2-p19","l3-r2-p22"],"knowledgeBlock":"squared-error-conditional-mean","formulas":[],"previousBlock":"b-7cda26dd3e456c00ab819f56","nextBlock":"b-d5d3717a06d232f47452624b","prerequisites":[{"id":"b-1c8abafd47b2eb5164aa85cc","url":"#b-1c8abafd47b2eb5164aa85cc","role":"section-title"},{"id":"b-0a276b7584301af39fb9a566","url":"#b-0a276b7584301af39fb9a566","role":"problem-context"}],"relatedBlocks":[{"id":"b-1c8abafd47b2eb5164aa85cc","url":"#b-1c8abafd47b2eb5164aa85cc","relation":"same-reasoning"},{"id":"b-7cda26dd3e456c00ab819f56","url":"#b-7cda26dd3e456c00ab819f56","relation":"same-reasoning"},{"id":"b-d5d3717a06d232f47452624b","url":"#b-d5d3717a06d232f47452624b","relation":"same-reasoning"}]},{"id":"b-d5d3717a06d232f47452624b","type":"p","scope":"squared-error-conditional-mean","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p22.html#b-d5d3717a06d232f47452624b","content":"LAB 08里应把恢复的主要笔画、错误补全和误差图一起读。到第四讲再问：如果我们希望产生某一种合理样本，而不是平均多个可能，训练目标和输出过程应怎样改变？图像描述也已在生成文本，后续生成课将把条件多解作为更一般的问题。","sha256":"41dbdd5dd96752401412f665b978bf02ddb76075360efa0153bbcffb00c8e182","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-masked-prediction","title":"限制可见信息，怎样构成可验证的重建任务？"},"page":{"id":"h-i-conditional-mean","title":"平方误差为什么偏好可能模糊的平均值","number":1},"sectionTitle":"为什么缺失处可能被补得模糊？","currentQuestion":"限制可见信息，怎样构成可验证的重建任务？","sourceSlides":["l3-r2-p19","l3-r2-p22"],"knowledgeBlock":"squared-error-conditional-mean","formulas":[],"previousBlock":"b-1c37eccd1b21732c219d2732","nextBlock":null,"prerequisites":[{"id":"b-1c8abafd47b2eb5164aa85cc","url":"#b-1c8abafd47b2eb5164aa85cc","role":"section-title"},{"id":"b-0a276b7584301af39fb9a566","url":"#b-0a276b7584301af39fb9a566","role":"problem-context"}],"relatedBlocks":[{"id":"b-1c8abafd47b2eb5164aa85cc","url":"#b-1c8abafd47b2eb5164aa85cc","relation":"same-reasoning"},{"id":"b-7cda26dd3e456c00ab819f56","url":"#b-7cda26dd3e456c00ab819f56","relation":"same-reasoning"},{"id":"b-1c37eccd1b21732c219d2732","url":"#b-1c37eccd1b21732c219d2732","relation":"same-reasoning"}]},{"id":"b-1ba94dc3f1723e1e846e7abd","type":"h2","scope":"h-i-reconstruction-and-readout-evidence","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p22.html#b-1ba94dc3f1723e1e846e7abd","content":"## 读 LAB08：重建和读出各交一份成绩","sha256":"8844c0d797c3d41101fe6c2c680ae884537647340ab38ac03a3854e036de3eba","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-masked-prediction","title":"限制可见信息，怎样构成可验证的重建任务？"},"page":{"id":"h-i-reconstruction-and-readout-evidence","title":"读 LAB08：重建和读出各交一份成绩","number":2},"sectionTitle":"读 LAB08：重建和读出各交一份成绩","currentQuestion":"限制可见信息，怎样构成可验证的重建任务？","sourceSlides":["l3-r2-p22"],"knowledgeBlock":null,"formulas":[],"previousBlock":null,"nextBlock":"b-911114295198745df9d53101","prerequisites":[{"id":"b-f08cf3e632c681951f0051ce","url":"#b-f08cf3e632c681951f0051ce","role":"problem-context"}],"relatedBlocks":[{"id":"b-3f3fdafaabbc82d5e3f1e116","url":"#b-3f3fdafaabbc82d5e3f1e116","relation":"figure-and-explanation"}]},{"id":"b-911114295198745df9d53101","type":"h2","scope":"explanation","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p22.html#b-911114295198745df9d53101","content":"## 解释、算例与证据","sha256":"d71cf15e707c8a3674b4c90a599efee783cb84da5b8c8236a1d01c3add7b4958","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-masked-prediction","title":"限制可见信息，怎样构成可验证的重建任务？"},"page":{"id":"h-i-reconstruction-and-readout-evidence","title":"读 LAB08：重建和读出各交一份成绩","number":2},"sectionTitle":"读 LAB08：重建和读出各交一份成绩","currentQuestion":"限制可见信息，怎样构成可验证的重建任务？","sourceSlides":["l3-r2-p22"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-1ba94dc3f1723e1e846e7abd","nextBlock":"b-8b59f0dd341b11507aea4473","prerequisites":[{"id":"b-1ba94dc3f1723e1e846e7abd","url":"#b-1ba94dc3f1723e1e846e7abd","role":"section-title"},{"id":"b-f08cf3e632c681951f0051ce","url":"#b-f08cf3e632c681951f0051ce","role":"problem-context"}],"relatedBlocks":[{"id":"b-3f3fdafaabbc82d5e3f1e116","url":"#b-3f3fdafaabbc82d5e3f1e116","relation":"figure-and-explanation"}]},{"id":"b-8b59f0dd341b11507aea4473","type":"h1","scope":"document","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p22.html#b-8b59f0dd341b11507aea4473","content":"# P22 · 能补图，不等于每项任务都读得好","sha256":"bcf0c4cc6f9c787333c050ee1992c958a6e90af55d9cb400d2b7813c88e5af42","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-masked-prediction","title":"限制可见信息，怎样构成可验证的重建任务？"},"page":{"id":"h-i-reconstruction-and-readout-evidence","title":"读 LAB08：重建和读出各交一份成绩","number":2},"sectionTitle":"读 LAB08：重建和读出各交一份成绩","currentQuestion":"限制可见信息，怎样构成可验证的重建任务？","sourceSlides":["l3-r2-p22"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-911114295198745df9d53101","nextBlock":"b-f08cf3e632c681951f0051ce","prerequisites":[{"id":"b-911114295198745df9d53101","url":"#b-911114295198745df9d53101","role":"section-title"},{"id":"b-f08cf3e632c681951f0051ce","url":"#b-f08cf3e632c681951f0051ce","role":"problem-context"}],"relatedBlocks":[{"id":"b-3f3fdafaabbc82d5e3f1e116","url":"#b-3f3fdafaabbc82d5e3f1e116","relation":"figure-and-explanation"}]},{"id":"b-f08cf3e632c681951f0051ce","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p22.html#b-f08cf3e632c681951f0051ce","content":"重建结果与表示用途分开评价，并给出真实正面研究结果和课程小模型结果各自的身份。","sha256":"8c276b31f27003561e22b48178ed6b725fbacc9a187aa3f629d2af3b2d3fa6b8","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-masked-prediction","title":"限制可见信息，怎样构成可验证的重建任务？"},"page":{"id":"h-i-reconstruction-and-readout-evidence","title":"读 LAB08：重建和读出各交一份成绩","number":2},"sectionTitle":"读 LAB08：重建和读出各交一份成绩","currentQuestion":"限制可见信息，怎样构成可验证的重建任务？","sourceSlides":["l3-r2-p22"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-8b59f0dd341b11507aea4473","nextBlock":"b-3f3fdafaabbc82d5e3f1e116","prerequisites":[{"id":"b-8b59f0dd341b11507aea4473","url":"#b-8b59f0dd341b11507aea4473","role":"section-title"}],"relatedBlocks":[{"id":"b-3f3fdafaabbc82d5e3f1e116","url":"#b-3f3fdafaabbc82d5e3f1e116","relation":"figure-and-explanation"}]},{"id":"b-3f3fdafaabbc82d5e3f1e116","type":"figure","scope":"document","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p22.html#b-3f3fdafaabbc82d5e3f1e116","content":"图内标签与数值：P22 学习表示机制图，状态 1；重建训练；预训练编码器；形状读出；能否分出零件类型？；颜色读出；红蓝是否仍可区分？；重建误差；下游任务结果；两种证据，分别测量\n\n遮挡预训练编码器分别连接形状与颜色任务，重建误差与下游成绩明确分开。\n\n- 重建训练\n- 预训练编码器\n- 形状读出\n- 能否分出零件类型？\n- 颜色读出\n- 红蓝是否仍可区分？\n- 重建误差\n- 下游任务结果\n- 两种证据，分别测量遮挡预训练编码器分别连接形状与颜色任务，重建误差与下游成绩明确分开。 [查看对应图解状态](/#/l3-r2-p22/0)[查看完整图解](/course/figures/l3-r2-p22-0.html)","sha256":"b52d8856e3b43ec40060c97a444157188897fefcdb79bd2ce0224b6101a47f2a","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-masked-prediction","title":"限制可见信息，怎样构成可验证的重建任务？"},"page":{"id":"h-i-reconstruction-and-readout-evidence","title":"读 LAB08：重建和读出各交一份成绩","number":2},"sectionTitle":"读 LAB08：重建和读出各交一份成绩","currentQuestion":"限制可见信息，怎样构成可验证的重建任务？","sourceSlides":["l3-r2-p22"],"knowledgeBlock":null,"formulas":[],"figure":{"caption":"遮挡预训练编码器分别连接形状与颜色任务，重建误差与下游成绩明确分开。 查看对应图解状态","labels":["重建训练","预训练编码器","形状读出","能否分出零件类型？","颜色读出","红蓝是否仍可区分？","重建误差","下游任务结果","两种证据，分别测量"],"description":"遮挡预训练编码器分别连接形状与颜色任务，重建误差与下游成绩明确分开。重建训练预训练编码器形状读出能否分出零件类型？颜色读出红蓝是否仍可区分？重建误差下游任务结果两种证据，分别测量","evidenceIdentity":"遮挡预训练编码器分别连接形状与颜色任务，重建误差与下游成绩明确分开。 查看对应图解状态","images":[{"src":"/assets/display/584c419ba431799ff41df983.webp","alt":""},{"src":"/assets/display/8757616f3f92bbccd2d8db9c.webp","alt":""},{"src":"/assets/display/2ebe2f0515d593c5292e6367.webp","alt":""}]},"previousBlock":"b-f08cf3e632c681951f0051ce","nextBlock":"b-7b9a1f32d818e86e2e753b30","prerequisites":[{"id":"b-8b59f0dd341b11507aea4473","url":"#b-8b59f0dd341b11507aea4473","role":"section-title"},{"id":"b-f08cf3e632c681951f0051ce","url":"#b-f08cf3e632c681951f0051ce","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-7b9a1f32d818e86e2e753b30","type":"h2","scope":"narrative","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p22.html#b-7b9a1f32d818e86e2e753b30","content":"## 课堂讲述","sha256":"786ce10a6b32a492259a3dccd0a54e61cd41d28a34382fb7b5386c5f03b7ad69","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-masked-prediction","title":"限制可见信息，怎样构成可验证的重建任务？"},"page":{"id":"h-i-reconstruction-and-readout-evidence","title":"读 LAB08：重建和读出各交一份成绩","number":2},"sectionTitle":"读 LAB08：重建和读出各交一份成绩","currentQuestion":"限制可见信息，怎样构成可验证的重建任务？","sourceSlides":["l3-r2-p22"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-3f3fdafaabbc82d5e3f1e116","nextBlock":"b-abef257eddf6b5e1b4551b04","prerequisites":[{"id":"b-8b59f0dd341b11507aea4473","url":"#b-8b59f0dd341b11507aea4473","role":"section-title"},{"id":"b-f08cf3e632c681951f0051ce","url":"#b-f08cf3e632c681951f0051ce","role":"problem-context"}],"relatedBlocks":[{"id":"b-3f3fdafaabbc82d5e3f1e116","url":"#b-3f3fdafaabbc82d5e3f1e116","relation":"figure-and-explanation"}]},{"id":"b-abef257eddf6b5e1b4551b04","type":"p","scope":"narrative","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p22.html#b-abef257eddf6b5e1b4551b04","content":"重建好，是关于重建任务的证据。它可能利用了有价值的结构，也可能将很多能力花在纹理和局部统计上；我们还不能由一幅看起来不错的补图，推出所有语义任务都已经解决。\n反过来也一样：一种表示不容易被简单分类头读出，不等于它完全没有相关信息。评价方法在限定我们允许怎样使用表示。\n因此，要把预训练成绩和下游用途分开。对比学习、遮挡重建、有监督分类，都应该送进你真正关心的任务，检查在相同数据和训练条件下能做什么。我们要的不是一个抽象的“最好表示”称号，而是清楚知道某种表示以什么方式帮助了当前问题。","sha256":"8590b6339d1d319c5481671869d9d6f87ead41d0d98b610224cb808aa9c0af22","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-masked-prediction","title":"限制可见信息，怎样构成可验证的重建任务？"},"page":{"id":"h-i-reconstruction-and-readout-evidence","title":"读 LAB08：重建和读出各交一份成绩","number":2},"sectionTitle":"读 LAB08：重建和读出各交一份成绩","currentQuestion":"限制可见信息，怎样构成可验证的重建任务？","sourceSlides":["l3-r2-p22"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-7b9a1f32d818e86e2e753b30","nextBlock":"b-6ad2c84650fe42daf0757e94","prerequisites":[{"id":"b-7b9a1f32d818e86e2e753b30","url":"#b-7b9a1f32d818e86e2e753b30","role":"section-title"},{"id":"b-f08cf3e632c681951f0051ce","url":"#b-f08cf3e632c681951f0051ce","role":"problem-context"}],"relatedBlocks":[{"id":"b-3f3fdafaabbc82d5e3f1e116","url":"#b-3f3fdafaabbc82d5e3f1e116","relation":"figure-and-explanation"}]},{"id":"b-6ad2c84650fe42daf0757e94","type":"h3","scope":"reconstruction-and-readout-two-results","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p22.html#b-6ad2c84650fe42daf0757e94","content":"### 能补图与能识别，需要两张成绩单","sha256":"b9200ecce5033ba762f0d64ac8b818952289e3d6e981d70f000aa4f19b70caeb","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-masked-prediction","title":"限制可见信息，怎样构成可验证的重建任务？"},"page":{"id":"h-i-reconstruction-and-readout-evidence","title":"读 LAB08：重建和读出各交一份成绩","number":2},"sectionTitle":"能补图与能识别，需要两张成绩单","currentQuestion":"限制可见信息，怎样构成可验证的重建任务？","sourceSlides":["l3-r2-p22"],"knowledgeBlock":"reconstruction-and-readout-two-results","formulas":[],"previousBlock":"b-abef257eddf6b5e1b4551b04","nextBlock":"b-d94d07f3dc4caec310e6442d","prerequisites":[{"id":"b-7b9a1f32d818e86e2e753b30","url":"#b-7b9a1f32d818e86e2e753b30","role":"section-title"},{"id":"b-f08cf3e632c681951f0051ce","url":"#b-f08cf3e632c681951f0051ce","role":"problem-context"}],"relatedBlocks":[{"id":"b-3f3fdafaabbc82d5e3f1e116","url":"#b-3f3fdafaabbc82d5e3f1e116","relation":"figure-and-explanation"},{"id":"b-d94d07f3dc4caec310e6442d","url":"#b-d94d07f3dc4caec310e6442d","relation":"same-reasoning"},{"id":"b-f149c08474962de2aed78e77","url":"#b-f149c08474962de2aed78e77","relation":"same-reasoning"},{"id":"b-f46bf2cc9d941792f221767a","url":"#b-f46bf2cc9d941792f221767a","relation":"same-reasoning"}]},{"id":"b-d94d07f3dc4caec310e6442d","type":"p","scope":"reconstruction-and-readout-two-results","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p22.html#b-d94d07f3dc4caec310e6442d","content":"重建评价比较预测像素与原图，冻结读出则固定编码器，用有标签参考集或新头预测类别。前者回答模型是否利用可见结构恢复缺失，后者回答这份表示能否支持指定用途。","sha256":"5688d9ffd1e591d194670c4ea54b33a33cb66c35efcfafbe31b9c46a305139bd","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-masked-prediction","title":"限制可见信息，怎样构成可验证的重建任务？"},"page":{"id":"h-i-reconstruction-and-readout-evidence","title":"读 LAB08：重建和读出各交一份成绩","number":2},"sectionTitle":"能补图与能识别，需要两张成绩单","currentQuestion":"限制可见信息，怎样构成可验证的重建任务？","sourceSlides":["l3-r2-p22"],"knowledgeBlock":"reconstruction-and-readout-two-results","formulas":[],"previousBlock":"b-6ad2c84650fe42daf0757e94","nextBlock":"b-f149c08474962de2aed78e77","prerequisites":[{"id":"b-6ad2c84650fe42daf0757e94","url":"#b-6ad2c84650fe42daf0757e94","role":"section-title"},{"id":"b-f08cf3e632c681951f0051ce","url":"#b-f08cf3e632c681951f0051ce","role":"problem-context"}],"relatedBlocks":[{"id":"b-3f3fdafaabbc82d5e3f1e116","url":"#b-3f3fdafaabbc82d5e3f1e116","relation":"figure-and-explanation"},{"id":"b-6ad2c84650fe42daf0757e94","url":"#b-6ad2c84650fe42daf0757e94","relation":"same-reasoning"},{"id":"b-f149c08474962de2aed78e77","url":"#b-f149c08474962de2aed78e77","relation":"same-reasoning"},{"id":"b-f46bf2cc9d941792f221767a","url":"#b-f46bf2cc9d941792f221767a","relation":"same-reasoning"}]},{"id":"b-f149c08474962de2aed78e77","type":"p","scope":"reconstruction-and-readout-two-results","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p22.html#b-f149c08474962de2aed78e77","content":"当前LAB 08的历史基准用1257张训练、540张保留的8×8数字图。去噪ConvAE的全像素MSE从直接使用破坏图的0.06244降至0.03217；清洁输入h的冻结5NN则从相同架构随机起点84.26%提高到95.00%。遮挡ConvAE的对应读出从88.52%到92.04%。这里是两份独立测量共同支持有用性。","sha256":"b2ab128d3317f61761e18c7c9e0d2ec1c0ba28422fcafc6f986a276bb99a0955","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-masked-prediction","title":"限制可见信息，怎样构成可验证的重建任务？"},"page":{"id":"h-i-reconstruction-and-readout-evidence","title":"读 LAB08：重建和读出各交一份成绩","number":2},"sectionTitle":"能补图与能识别，需要两张成绩单","currentQuestion":"限制可见信息，怎样构成可验证的重建任务？","sourceSlides":["l3-r2-p22"],"knowledgeBlock":"reconstruction-and-readout-two-results","formulas":[],"previousBlock":"b-d94d07f3dc4caec310e6442d","nextBlock":"b-f46bf2cc9d941792f221767a","prerequisites":[{"id":"b-6ad2c84650fe42daf0757e94","url":"#b-6ad2c84650fe42daf0757e94","role":"section-title"},{"id":"b-f08cf3e632c681951f0051ce","url":"#b-f08cf3e632c681951f0051ce","role":"problem-context"}],"relatedBlocks":[{"id":"b-3f3fdafaabbc82d5e3f1e116","url":"#b-3f3fdafaabbc82d5e3f1e116","relation":"figure-and-explanation"},{"id":"b-6ad2c84650fe42daf0757e94","url":"#b-6ad2c84650fe42daf0757e94","relation":"same-reasoning"},{"id":"b-d94d07f3dc4caec310e6442d","url":"#b-d94d07f3dc4caec310e6442d","relation":"same-reasoning"},{"id":"b-f46bf2cc9d941792f221767a","url":"#b-f46bf2cc9d941792f221767a","relation":"same-reasoning"}]},{"id":"b-f46bf2cc9d941792f221767a","type":"p","scope":"reconstruction-and-readout-two-results","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p22.html#b-f46bf2cc9d941792f221767a","content":"原图先切分，再生成噪声和mask；标签不进入重建损失。结果来自已保存运行，本轮没有重训；当前Notebook另记有旧云端复核。它不是原ViT MAE复现，也没有写作者分组，不能据此声称跨写作者泛化。","sha256":"1d7cb0553d5370b5f86c2bf98758fc9383e10d2917873fea007f494dc1ede516","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-masked-prediction","title":"限制可见信息，怎样构成可验证的重建任务？"},"page":{"id":"h-i-reconstruction-and-readout-evidence","title":"读 LAB08：重建和读出各交一份成绩","number":2},"sectionTitle":"能补图与能识别，需要两张成绩单","currentQuestion":"限制可见信息，怎样构成可验证的重建任务？","sourceSlides":["l3-r2-p22"],"knowledgeBlock":"reconstruction-and-readout-two-results","formulas":[],"previousBlock":"b-f149c08474962de2aed78e77","nextBlock":"b-7ea82b1e9ba8efbaddd2d00e","prerequisites":[{"id":"b-6ad2c84650fe42daf0757e94","url":"#b-6ad2c84650fe42daf0757e94","role":"section-title"},{"id":"b-f08cf3e632c681951f0051ce","url":"#b-f08cf3e632c681951f0051ce","role":"problem-context"}],"relatedBlocks":[{"id":"b-3f3fdafaabbc82d5e3f1e116","url":"#b-3f3fdafaabbc82d5e3f1e116","relation":"figure-and-explanation"},{"id":"b-6ad2c84650fe42daf0757e94","url":"#b-6ad2c84650fe42daf0757e94","relation":"same-reasoning"},{"id":"b-d94d07f3dc4caec310e6442d","url":"#b-d94d07f3dc4caec310e6442d","relation":"same-reasoning"},{"id":"b-f149c08474962de2aed78e77","url":"#b-f149c08474962de2aed78e77","relation":"same-reasoning"}]},{"id":"b-7ea82b1e9ba8efbaddd2d00e","type":"h4","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p22.html#b-7ea82b1e9ba8efbaddd2d00e","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-masked-prediction","title":"限制可见信息，怎样构成可验证的重建任务？"},"page":{"id":"h-i-reconstruction-and-readout-evidence","title":"读 LAB08：重建和读出各交一份成绩","number":2},"sectionTitle":"读 LAB08：重建和读出各交一份成绩","currentQuestion":"限制可见信息，怎样构成可验证的重建任务？","sourceSlides":["l3-r2-p22"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-f46bf2cc9d941792f221767a","nextBlock":"b-4dcb596613f167b62fdca592","prerequisites":[{"id":"b-6ad2c84650fe42daf0757e94","url":"#b-6ad2c84650fe42daf0757e94","role":"section-title"},{"id":"b-f08cf3e632c681951f0051ce","url":"#b-f08cf3e632c681951f0051ce","role":"problem-context"}],"relatedBlocks":[{"id":"b-3f3fdafaabbc82d5e3f1e116","url":"#b-3f3fdafaabbc82d5e3f1e116","relation":"figure-and-explanation"}]},{"id":"b-4dcb596613f167b62fdca592","type":"ul","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p22.html#b-4dcb596613f167b62fdca592","content":"- [PCA 是带约束的线性编码—解码｜怎样检查这项解释](/supplement.html?return=l3-r2-p22/0#/l3-pca-reconstruction-evidence/0) L03补充阅读\n- [PCA 是带约束的线性编码—解码｜什么结论还不能推出](/supplement.html?return=l3-r2-p22/0#/l3-pca-reconstruction-boundary/0) L03补充阅读\n- [复制输入为何不是好表示｜怎样检查这项解释](/supplement.html?return=l3-r2-p22/0#/l3-autoencoder-denoise-evidence/0) L03补充阅读\n- [复制输入为何不是好表示｜什么结论还不能推出](/supplement.html?return=l3-r2-p22/0#/l3-autoencoder-denoise-boundary/0) L03补充阅读","sha256":"cca4d3e84124f36eb656680a22b16a44449afb4b3f74a7c6ccbd479206280704","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-masked-prediction","title":"限制可见信息，怎样构成可验证的重建任务？"},"page":{"id":"h-i-reconstruction-and-readout-evidence","title":"读 LAB08：重建和读出各交一份成绩","number":2},"sectionTitle":"读 LAB08：重建和读出各交一份成绩","currentQuestion":"限制可见信息，怎样构成可验证的重建任务？","sourceSlides":["l3-r2-p22"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-7ea82b1e9ba8efbaddd2d00e","nextBlock":"b-078151c63112769bd21b4a8f","prerequisites":[{"id":"b-7ea82b1e9ba8efbaddd2d00e","url":"#b-7ea82b1e9ba8efbaddd2d00e","role":"section-title"},{"id":"b-f08cf3e632c681951f0051ce","url":"#b-f08cf3e632c681951f0051ce","role":"problem-context"}],"relatedBlocks":[{"id":"b-3f3fdafaabbc82d5e3f1e116","url":"#b-3f3fdafaabbc82d5e3f1e116","relation":"figure-and-explanation"}]},{"id":"b-078151c63112769bd21b4a8f","type":"li","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p22.html#b-078151c63112769bd21b4a8f","content":"- [PCA 是带约束的线性编码—解码｜怎样检查这项解释](/supplement.html?return=l3-r2-p22/0#/l3-pca-reconstruction-evidence/0) L03补充阅读","sha256":"6c0883a05332991cbfa598b3f0e397e3e0ad862432090d758bc0a0434408d15a","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-masked-prediction","title":"限制可见信息，怎样构成可验证的重建任务？"},"page":{"id":"h-i-reconstruction-and-readout-evidence","title":"读 LAB08：重建和读出各交一份成绩","number":2},"sectionTitle":"读 LAB08：重建和读出各交一份成绩","currentQuestion":"限制可见信息，怎样构成可验证的重建任务？","sourceSlides":["l3-r2-p22"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-4dcb596613f167b62fdca592","nextBlock":"b-c73406d907fc82afd6efbd5d","prerequisites":[{"id":"b-7ea82b1e9ba8efbaddd2d00e","url":"#b-7ea82b1e9ba8efbaddd2d00e","role":"section-title"},{"id":"b-f08cf3e632c681951f0051ce","url":"#b-f08cf3e632c681951f0051ce","role":"problem-context"}],"relatedBlocks":[{"id":"b-3f3fdafaabbc82d5e3f1e116","url":"#b-3f3fdafaabbc82d5e3f1e116","relation":"figure-and-explanation"}]},{"id":"b-c73406d907fc82afd6efbd5d","type":"li","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p22.html#b-c73406d907fc82afd6efbd5d","content":"- [PCA 是带约束的线性编码—解码｜什么结论还不能推出](/supplement.html?return=l3-r2-p22/0#/l3-pca-reconstruction-boundary/0) L03补充阅读","sha256":"c8a25dbabc3335471aa1a8f45dd97444af42929b3ab5fa6356bc36446c9dcfa9","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-masked-prediction","title":"限制可见信息，怎样构成可验证的重建任务？"},"page":{"id":"h-i-reconstruction-and-readout-evidence","title":"读 LAB08：重建和读出各交一份成绩","number":2},"sectionTitle":"读 LAB08：重建和读出各交一份成绩","currentQuestion":"限制可见信息，怎样构成可验证的重建任务？","sourceSlides":["l3-r2-p22"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-078151c63112769bd21b4a8f","nextBlock":"b-b6f5e02415c65077622de063","prerequisites":[{"id":"b-7ea82b1e9ba8efbaddd2d00e","url":"#b-7ea82b1e9ba8efbaddd2d00e","role":"section-title"},{"id":"b-f08cf3e632c681951f0051ce","url":"#b-f08cf3e632c681951f0051ce","role":"problem-context"}],"relatedBlocks":[{"id":"b-3f3fdafaabbc82d5e3f1e116","url":"#b-3f3fdafaabbc82d5e3f1e116","relation":"figure-and-explanation"}]},{"id":"b-b6f5e02415c65077622de063","type":"li","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p22.html#b-b6f5e02415c65077622de063","content":"- [复制输入为何不是好表示｜怎样检查这项解释](/supplement.html?return=l3-r2-p22/0#/l3-autoencoder-denoise-evidence/0) L03补充阅读","sha256":"118dc668367b10a51e704e3accb4bbaa0b20ebc3bf02733068ab51f95753ece8","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-masked-prediction","title":"限制可见信息，怎样构成可验证的重建任务？"},"page":{"id":"h-i-reconstruction-and-readout-evidence","title":"读 LAB08：重建和读出各交一份成绩","number":2},"sectionTitle":"读 LAB08：重建和读出各交一份成绩","currentQuestion":"限制可见信息，怎样构成可验证的重建任务？","sourceSlides":["l3-r2-p22"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-c73406d907fc82afd6efbd5d","nextBlock":"b-188ca1b857bf2a1a67420278","prerequisites":[{"id":"b-7ea82b1e9ba8efbaddd2d00e","url":"#b-7ea82b1e9ba8efbaddd2d00e","role":"section-title"},{"id":"b-f08cf3e632c681951f0051ce","url":"#b-f08cf3e632c681951f0051ce","role":"problem-context"}],"relatedBlocks":[{"id":"b-3f3fdafaabbc82d5e3f1e116","url":"#b-3f3fdafaabbc82d5e3f1e116","relation":"figure-and-explanation"}]},{"id":"b-188ca1b857bf2a1a67420278","type":"li","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p22.html#b-188ca1b857bf2a1a67420278","content":"- [复制输入为何不是好表示｜什么结论还不能推出](/supplement.html?return=l3-r2-p22/0#/l3-autoencoder-denoise-boundary/0) L03补充阅读","sha256":"c1c15b8807149bcbdab0ed47bb8dd971f501e0dc14fe351508bfd17f31ec0358","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-masked-prediction","title":"限制可见信息，怎样构成可验证的重建任务？"},"page":{"id":"h-i-reconstruction-and-readout-evidence","title":"读 LAB08：重建和读出各交一份成绩","number":2},"sectionTitle":"读 LAB08：重建和读出各交一份成绩","currentQuestion":"限制可见信息，怎样构成可验证的重建任务？","sourceSlides":["l3-r2-p22"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-b6f5e02415c65077622de063","nextBlock":null,"prerequisites":[{"id":"b-7ea82b1e9ba8efbaddd2d00e","url":"#b-7ea82b1e9ba8efbaddd2d00e","role":"section-title"},{"id":"b-f08cf3e632c681951f0051ce","url":"#b-f08cf3e632c681951f0051ce","role":"problem-context"}],"relatedBlocks":[{"id":"b-3f3fdafaabbc82d5e3f1e116","url":"#b-3f3fdafaabbc82d5e3f1e116","relation":"figure-and-explanation"}]},{"id":"b-bd9605313c93f4f5577a074c","type":"h2","scope":"h-i-mae-evaluation-interface","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p22.html#b-bd9605313c93f4f5577a074c","content":"## MAE 的正面证据，也要说清使用接口","sha256":"b3d28a0a4c169ff30e4c538ad8d1abe3726bb57eb6816619db5403228d933780","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-masked-prediction","title":"限制可见信息，怎样构成可验证的重建任务？"},"page":{"id":"h-i-mae-evaluation-interface","title":"MAE 的正面证据，也要说清使用接口","number":3},"sectionTitle":"MAE 的正面证据，也要说清使用接口","currentQuestion":"限制可见信息，怎样构成可验证的重建任务？","sourceSlides":["l3-r2-p22","l3-r2-p23"],"knowledgeBlock":null,"formulas":[],"previousBlock":null,"nextBlock":"b-969d668c82c7a1ced11f4ffc","prerequisites":[{"id":"b-969d668c82c7a1ced11f4ffc","url":"#b-969d668c82c7a1ced11f4ffc","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-969d668c82c7a1ced11f4ffc","type":"p","scope":"h-i-mae-evaluation-interface","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p22.html#b-969d668c82c7a1ced11f4ffc","content":"原论文的端到端微调、最后块微调和线性读出，允许改变的计算不同。","sha256":"c836161bd0888336e9a315b4c08bfe5b94eb6d654ef0c20ca4af1d55df44fea0","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-masked-prediction","title":"限制可见信息，怎样构成可验证的重建任务？"},"page":{"id":"h-i-mae-evaluation-interface","title":"MAE 的正面证据，也要说清使用接口","number":3},"sectionTitle":"MAE 的正面证据，也要说清使用接口","currentQuestion":"限制可见信息，怎样构成可验证的重建任务？","sourceSlides":["l3-r2-p22","l3-r2-p23"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-bd9605313c93f4f5577a074c","nextBlock":"b-0d44129b7af29f106bc8a208","prerequisites":[{"id":"b-bd9605313c93f4f5577a074c","url":"#b-bd9605313c93f4f5577a074c","role":"section-title"}],"relatedBlocks":[]},{"id":"b-0d44129b7af29f106bc8a208","type":"h3","scope":"mae-finetuning-evidence","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p22.html#b-0d44129b7af29f106bc8a208","content":"### MAE 的正面证据也依赖评价接口","sha256":"8ce3eef83eb109d3b7630255be43a3fdb744e2cfc074570b15c8fd7432a1486c","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-masked-prediction","title":"限制可见信息，怎样构成可验证的重建任务？"},"page":{"id":"h-i-mae-evaluation-interface","title":"MAE 的正面证据，也要说清使用接口","number":3},"sectionTitle":"MAE 的正面证据也依赖评价接口","currentQuestion":"限制可见信息，怎样构成可验证的重建任务？","sourceSlides":["l3-r2-p22","l3-r2-p23"],"knowledgeBlock":"mae-finetuning-evidence","formulas":[],"previousBlock":"b-969d668c82c7a1ced11f4ffc","nextBlock":"b-414280c80a3926dbe0cab732","prerequisites":[{"id":"b-bd9605313c93f4f5577a074c","url":"#b-bd9605313c93f4f5577a074c","role":"section-title"},{"id":"b-969d668c82c7a1ced11f4ffc","url":"#b-969d668c82c7a1ced11f4ffc","role":"problem-context"}],"relatedBlocks":[{"id":"b-414280c80a3926dbe0cab732","url":"#b-414280c80a3926dbe0cab732","relation":"same-reasoning"},{"id":"b-6ef915ec7f990490a5f7f770","url":"#b-6ef915ec7f990490a5f7f770","relation":"same-reasoning"},{"id":"b-5d70faa61f04bf4735c98958","url":"#b-5d70faa61f04bf4735c98958","relation":"same-reasoning"}]},{"id":"b-414280c80a3926dbe0cab732","type":"p","scope":"mae-finetuning-evidence","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p22.html#b-414280c80a3926dbe0cab732","content":"MAE论文Table 3使用ImageNet-1K图像作自监督预训练，再端到端微调分类。ViT-L/16在224输入的top-1为85.9%；同表从零训练实现为82.6%。这里的结果支持：遮挡像素目标可以为较大视觉模型提供有效起点。它不是只看重建图好不好看。","sha256":"8e45489df883e86d34df4691779f3cf7e9e27251094df0e7a920779e4922e03c","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-masked-prediction","title":"限制可见信息，怎样构成可验证的重建任务？"},"page":{"id":"h-i-mae-evaluation-interface","title":"MAE 的正面证据，也要说清使用接口","number":3},"sectionTitle":"MAE 的正面证据也依赖评价接口","currentQuestion":"限制可见信息，怎样构成可验证的重建任务？","sourceSlides":["l3-r2-p22","l3-r2-p23"],"knowledgeBlock":"mae-finetuning-evidence","formulas":[],"previousBlock":"b-0d44129b7af29f106bc8a208","nextBlock":"b-6ef915ec7f990490a5f7f770","prerequisites":[{"id":"b-0d44129b7af29f106bc8a208","url":"#b-0d44129b7af29f106bc8a208","role":"section-title"},{"id":"b-969d668c82c7a1ced11f4ffc","url":"#b-969d668c82c7a1ced11f4ffc","role":"problem-context"}],"relatedBlocks":[{"id":"b-0d44129b7af29f106bc8a208","url":"#b-0d44129b7af29f106bc8a208","relation":"same-reasoning"},{"id":"b-6ef915ec7f990490a5f7f770","url":"#b-6ef915ec7f990490a5f7f770","relation":"same-reasoning"},{"id":"b-5d70faa61f04bf4735c98958","url":"#b-5d70faa61f04bf4735c98958","relation":"same-reasoning"}]},{"id":"b-6ef915ec7f990490a5f7f770","type":"p","scope":"mae-finetuning-evidence","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p22.html#b-6ef915ec7f990490a5f7f770","content":"论文Figure 9进一步指出，同一ViT-L表示的线性读出为73.5%，只微调最后一个块即可到81.0%。因此线性读出低，并不自动意味着这份表示不适合适配；评价允许改变多少计算，会影响判断。","sha256":"f2f725904b17bcd54b2673d5a68bc81103ae385a1e35a3653c410868e548ec52","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-masked-prediction","title":"限制可见信息，怎样构成可验证的重建任务？"},"page":{"id":"h-i-mae-evaluation-interface","title":"MAE 的正面证据，也要说清使用接口","number":3},"sectionTitle":"MAE 的正面证据也依赖评价接口","currentQuestion":"限制可见信息，怎样构成可验证的重建任务？","sourceSlides":["l3-r2-p22","l3-r2-p23"],"knowledgeBlock":"mae-finetuning-evidence","formulas":[],"previousBlock":"b-414280c80a3926dbe0cab732","nextBlock":"b-5d70faa61f04bf4735c98958","prerequisites":[{"id":"b-0d44129b7af29f106bc8a208","url":"#b-0d44129b7af29f106bc8a208","role":"section-title"},{"id":"b-969d668c82c7a1ced11f4ffc","url":"#b-969d668c82c7a1ced11f4ffc","role":"problem-context"}],"relatedBlocks":[{"id":"b-0d44129b7af29f106bc8a208","url":"#b-0d44129b7af29f106bc8a208","relation":"same-reasoning"},{"id":"b-414280c80a3926dbe0cab732","url":"#b-414280c80a3926dbe0cab732","relation":"same-reasoning"},{"id":"b-5d70faa61f04bf4735c98958","url":"#b-5d70faa61f04bf4735c98958","relation":"same-reasoning"}]},{"id":"b-5d70faa61f04bf4735c98958","type":"p","scope":"mae-finetuning-evidence","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p22.html#b-5d70faa61f04bf4735c98958","content":"这些是论文在明确训练与评价条件下的结果，不是本课程LAB 08的数字。阅读时区分三层：MAE原论文、当前教学卷积模型、页面里的16块索引图。它们分别提供研究实证、可运行学习链和透明机制。","sha256":"857539d92a95c906eb9fcd172edca91e7dce5831d1fd66022c510cf90947bf93","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-masked-prediction","title":"限制可见信息，怎样构成可验证的重建任务？"},"page":{"id":"h-i-mae-evaluation-interface","title":"MAE 的正面证据，也要说清使用接口","number":3},"sectionTitle":"MAE 的正面证据也依赖评价接口","currentQuestion":"限制可见信息，怎样构成可验证的重建任务？","sourceSlides":["l3-r2-p22","l3-r2-p23"],"knowledgeBlock":"mae-finetuning-evidence","formulas":[],"previousBlock":"b-6ef915ec7f990490a5f7f770","nextBlock":null,"prerequisites":[{"id":"b-0d44129b7af29f106bc8a208","url":"#b-0d44129b7af29f106bc8a208","role":"section-title"},{"id":"b-969d668c82c7a1ced11f4ffc","url":"#b-969d668c82c7a1ced11f4ffc","role":"problem-context"}],"relatedBlocks":[{"id":"b-0d44129b7af29f106bc8a208","url":"#b-0d44129b7af29f106bc8a208","relation":"same-reasoning"},{"id":"b-414280c80a3926dbe0cab732","url":"#b-414280c80a3926dbe0cab732","relation":"same-reasoning"},{"id":"b-6ef915ec7f990490a5f7f770","url":"#b-6ef915ec7f990490a5f7f770","relation":"same-reasoning"}]},{"id":"b-01394caa0965f81884dda6bb","type":"h2","scope":"h-i-architecture-and-target","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p22.html#b-01394caa0965f81884dda6bb","content":"## 选读：结构和重建目标需要共同检验","sha256":"ef9750b36830c682223b58dc3efc9b336446956d2392a11c09b4158baf6b30b1","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-architecture-and-target","title":"选读：结构和重建目标需要共同检验","number":4},"sectionTitle":"选读：结构和重建目标需要共同检验","currentQuestion":"哪些细节需要进一步对照？","sourceSlides":["l3-r2-p19","l3-r2-p22"],"knowledgeBlock":null,"formulas":[],"previousBlock":null,"nextBlock":"b-22a7a6fbc5e1014881f46408","prerequisites":[{"id":"b-22a7a6fbc5e1014881f46408","url":"#b-22a7a6fbc5e1014881f46408","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-22a7a6fbc5e1014881f46408","type":"p","scope":"h-i-architecture-and-target","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p22.html#b-22a7a6fbc5e1014881f46408","content":"比较ConvNeXt V1/V2与监督/FCMAE两列，先保留训练预算和证据范围。","sha256":"3e9b9603a3bd310b5876bc7cd18d5d692355daac27e3aaa41eea5afc9419672a","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-architecture-and-target","title":"选读：结构和重建目标需要共同检验","number":4},"sectionTitle":"选读：结构和重建目标需要共同检验","currentQuestion":"哪些细节需要进一步对照？","sourceSlides":["l3-r2-p19","l3-r2-p22"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-01394caa0965f81884dda6bb","nextBlock":"b-4b5bb5061034104a36307b53","prerequisites":[{"id":"b-01394caa0965f81884dda6bb","url":"#b-01394caa0965f81884dda6bb","role":"section-title"}],"relatedBlocks":[]},{"id":"b-4b5bb5061034104a36307b53","type":"h3","scope":"architecture-objective-co-design","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p22.html#b-4b5bb5061034104a36307b53","content":"### 选读：结构与遮挡目标为什么要共同设计","sha256":"60344dd625f5f286022e81a326bb4283bca502621019d629ef7897d2863f6fe7","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-architecture-and-target","title":"选读：结构和重建目标需要共同检验","number":4},"sectionTitle":"选读：结构与遮挡目标为什么要共同设计","currentQuestion":"哪些细节需要进一步对照？","sourceSlides":["l3-r2-p19","l3-r2-p22"],"knowledgeBlock":"architecture-objective-co-design","formulas":[],"previousBlock":"b-22a7a6fbc5e1014881f46408","nextBlock":"b-83fd46bf8aef1bcacf079a65","prerequisites":[{"id":"b-01394caa0965f81884dda6bb","url":"#b-01394caa0965f81884dda6bb","role":"section-title"},{"id":"b-22a7a6fbc5e1014881f46408","url":"#b-22a7a6fbc5e1014881f46408","role":"problem-context"}],"relatedBlocks":[{"id":"b-83fd46bf8aef1bcacf079a65","url":"#b-83fd46bf8aef1bcacf079a65","relation":"same-reasoning"},{"id":"b-c35729d284ad52e8061a7abb","url":"#b-c35729d284ad52e8061a7abb","relation":"same-reasoning"},{"id":"b-7f98852cbc75ebe13c692c9b","url":"#b-7f98852cbc75ebe13c692c9b","relation":"same-reasoning"},{"id":"b-7a68f9594175c2e15c84b528","url":"#b-7a68f9594175c2e15c84b528","relation":"same-reasoning"}]},{"id":"b-83fd46bf8aef1bcacf079a65","type":"p","scope":"architecture-objective-co-design","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p22.html#b-83fd46bf8aef1bcacf079a65","content":"把原MAE目标直接搬到卷积网络，并不意味着所有结构假设仍成立。ConvNeXt V2将遮挡图看成稀疏空间输入，用FCMAE组织学习，并引入GRN改善通道响应的多样性。这提供了一个正面研究例子：目标与网络结构可以一起设计。","sha256":"f90a46d599e1e60c5744399504a4b565df7745bd6e5f034a33ed662df7be5117","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-architecture-and-target","title":"选读：结构和重建目标需要共同检验","number":4},"sectionTitle":"选读：结构与遮挡目标为什么要共同设计","currentQuestion":"哪些细节需要进一步对照？","sourceSlides":["l3-r2-p19","l3-r2-p22"],"knowledgeBlock":"architecture-objective-co-design","formulas":[],"previousBlock":"b-4b5bb5061034104a36307b53","nextBlock":"b-c35729d284ad52e8061a7abb","prerequisites":[{"id":"b-4b5bb5061034104a36307b53","url":"#b-4b5bb5061034104a36307b53","role":"section-title"},{"id":"b-22a7a6fbc5e1014881f46408","url":"#b-22a7a6fbc5e1014881f46408","role":"problem-context"}],"relatedBlocks":[{"id":"b-4b5bb5061034104a36307b53","url":"#b-4b5bb5061034104a36307b53","relation":"same-reasoning"},{"id":"b-c35729d284ad52e8061a7abb","url":"#b-c35729d284ad52e8061a7abb","relation":"same-reasoning"},{"id":"b-7f98852cbc75ebe13c692c9b","url":"#b-7f98852cbc75ebe13c692c9b","relation":"same-reasoning"},{"id":"b-7a68f9594175c2e15c84b528","url":"#b-7a68f9594175c2e15c84b528","relation":"same-reasoning"}]},{"id":"b-c35729d284ad52e8061a7abb","type":"p","scope":"architecture-objective-co-design","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p22.html#b-c35729d284ad52e8061a7abb","content":"原论文Table 3的Base比较（ImageNet-1K top-1，%）：","sha256":"5f412f089df24a97ca57f905320ac07758cc419ed15af144a02d763a3d5b011d","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-architecture-and-target","title":"选读：结构和重建目标需要共同检验","number":4},"sectionTitle":"选读：结构与遮挡目标为什么要共同设计","currentQuestion":"哪些细节需要进一步对照？","sourceSlides":["l3-r2-p19","l3-r2-p22"],"knowledgeBlock":"architecture-objective-co-design","formulas":[],"previousBlock":"b-83fd46bf8aef1bcacf079a65","nextBlock":"b-7f98852cbc75ebe13c692c9b","prerequisites":[{"id":"b-4b5bb5061034104a36307b53","url":"#b-4b5bb5061034104a36307b53","role":"section-title"},{"id":"b-22a7a6fbc5e1014881f46408","url":"#b-22a7a6fbc5e1014881f46408","role":"problem-context"}],"relatedBlocks":[{"id":"b-4b5bb5061034104a36307b53","url":"#b-4b5bb5061034104a36307b53","relation":"same-reasoning"},{"id":"b-83fd46bf8aef1bcacf079a65","url":"#b-83fd46bf8aef1bcacf079a65","relation":"same-reasoning"},{"id":"b-7f98852cbc75ebe13c692c9b","url":"#b-7f98852cbc75ebe13c692c9b","relation":"same-reasoning"},{"id":"b-7a68f9594175c2e15c84b528","url":"#b-7a68f9594175c2e15c84b528","relation":"same-reasoning"}]},{"id":"b-7f98852cbc75ebe13c692c9b","type":"table","scope":"architecture-objective-co-design","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p22.html#b-7f98852cbc75ebe13c692c9b","content":"| 架构 | 监督训练 | FCMAE预训练后微调 |\n| --- | --- | --- |\n| V1-Base | 83.8 | 83.7 |\n| V2-Base | 84.3 | 84.6 |","sha256":"34076280b7dc70991b3bc8eef7d2adec72f68cfb32c548503d53852f2ee695d6","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-architecture-and-target","title":"选读：结构和重建目标需要共同检验","number":4},"sectionTitle":"选读：结构与遮挡目标为什么要共同设计","currentQuestion":"哪些细节需要进一步对照？","sourceSlides":["l3-r2-p19","l3-r2-p22"],"knowledgeBlock":"architecture-objective-co-design","formulas":[],"table":{"headers":["架构","监督训练","FCMAE预训练后微调"],"rows":[["架构","监督训练","FCMAE预训练后微调"],["V1-Base","83.8","83.7"],["V2-Base","84.3","84.6"]]},"previousBlock":"b-c35729d284ad52e8061a7abb","nextBlock":"b-7a68f9594175c2e15c84b528","prerequisites":[{"id":"b-4b5bb5061034104a36307b53","url":"#b-4b5bb5061034104a36307b53","role":"section-title"},{"id":"b-22a7a6fbc5e1014881f46408","url":"#b-22a7a6fbc5e1014881f46408","role":"problem-context"}],"relatedBlocks":[{"id":"b-4b5bb5061034104a36307b53","url":"#b-4b5bb5061034104a36307b53","relation":"same-reasoning"},{"id":"b-83fd46bf8aef1bcacf079a65","url":"#b-83fd46bf8aef1bcacf079a65","relation":"same-reasoning"},{"id":"b-c35729d284ad52e8061a7abb","url":"#b-c35729d284ad52e8061a7abb","relation":"same-reasoning"},{"id":"b-7a68f9594175c2e15c84b528","url":"#b-7a68f9594175c2e15c84b528","relation":"same-reasoning"}]},{"id":"b-7a68f9594175c2e15c84b528","type":"p","scope":"architecture-objective-co-design","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p22.html#b-7a68f9594175c2e15c84b528","content":"FCMAE列先预训练800轮再微调，不能称等总预算。V1中换目标为−0.1，V2中为+0.3；这提示组合值得研究，却不能仅凭四个数识别一个严格因果交互。这里是论文证据，和LAB05的patch词典拟合不是同一实验。读完回到P22的用途评价，不再另开主讲章节。","sha256":"bc8d8b41f24cb8500fbdfecfb6289b5ef4dc3c5bcc086c466aaaa7caad8b4b15","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-architecture-and-target","title":"选读：结构和重建目标需要共同检验","number":4},"sectionTitle":"选读：结构与遮挡目标为什么要共同设计","currentQuestion":"哪些细节需要进一步对照？","sourceSlides":["l3-r2-p19","l3-r2-p22"],"knowledgeBlock":"architecture-objective-co-design","formulas":[],"previousBlock":"b-7f98852cbc75ebe13c692c9b","nextBlock":"b-1bca931ab02760a029960de2","prerequisites":[{"id":"b-4b5bb5061034104a36307b53","url":"#b-4b5bb5061034104a36307b53","role":"section-title"},{"id":"b-22a7a6fbc5e1014881f46408","url":"#b-22a7a6fbc5e1014881f46408","role":"problem-context"}],"relatedBlocks":[{"id":"b-4b5bb5061034104a36307b53","url":"#b-4b5bb5061034104a36307b53","relation":"same-reasoning"},{"id":"b-83fd46bf8aef1bcacf079a65","url":"#b-83fd46bf8aef1bcacf079a65","relation":"same-reasoning"},{"id":"b-c35729d284ad52e8061a7abb","url":"#b-c35729d284ad52e8061a7abb","relation":"same-reasoning"},{"id":"b-7f98852cbc75ebe13c692c9b","url":"#b-7f98852cbc75ebe13c692c9b","relation":"same-reasoning"}]},{"id":"b-1bca931ab02760a029960de2","type":"h2","scope":"practice","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p22.html#b-1bca931ab02760a029960de2","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-architecture-and-target","title":"选读：结构和重建目标需要共同检验","number":4},"sectionTitle":"选读：结构和重建目标需要共同检验","currentQuestion":"哪些细节需要进一步对照？","sourceSlides":["l3-r2-p19","l3-r2-p22"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-7a68f9594175c2e15c84b528","nextBlock":"b-c01907a791608f1f0f1f9bb0","prerequisites":[{"id":"b-4b5bb5061034104a36307b53","url":"#b-4b5bb5061034104a36307b53","role":"section-title"},{"id":"b-22a7a6fbc5e1014881f46408","url":"#b-22a7a6fbc5e1014881f46408","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-c01907a791608f1f0f1f9bb0","type":"p","scope":"practice","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p22.html#b-c01907a791608f1f0f1f9bb0","content":"在LAB 08对照清洁、破坏、预测、误差四列，再读冻结5NN表；不把两种指标相互替代。","sha256":"5bf6b9348a61224c9e47f1af545ae9404fd41534e1648bc6ed1b80d86160b455","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-architecture-and-target","title":"选读：结构和重建目标需要共同检验","number":4},"sectionTitle":"选读：结构和重建目标需要共同检验","currentQuestion":"哪些细节需要进一步对照？","sourceSlides":["l3-r2-p19","l3-r2-p22"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-1bca931ab02760a029960de2","nextBlock":"b-d6417fe286de521b098b1449","prerequisites":[{"id":"b-1bca931ab02760a029960de2","url":"#b-1bca931ab02760a029960de2","role":"section-title"},{"id":"b-22a7a6fbc5e1014881f46408","url":"#b-22a7a6fbc5e1014881f46408","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-d6417fe286de521b098b1449","type":"h4","scope":"practice","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p22.html#b-d6417fe286de521b098b1449","content":"#### 按问题回查的讲解与短练习（1）","sha256":"0480869f0a66009b015a757d67aed6061011e378987f98a024f4e8bfc5386152","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-architecture-and-target","title":"选读：结构和重建目标需要共同检验","number":4},"sectionTitle":"选读：结构和重建目标需要共同检验","currentQuestion":"哪些细节需要进一步对照？","sourceSlides":["l3-r2-p19","l3-r2-p22"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-c01907a791608f1f0f1f9bb0","nextBlock":null,"prerequisites":[{"id":"b-1bca931ab02760a029960de2","url":"#b-1bca931ab02760a029960de2","role":"section-title"},{"id":"b-22a7a6fbc5e1014881f46408","url":"#b-22a7a6fbc5e1014881f46408","role":"problem-context"}],"relatedBlocks":[]}]}