{"version":2,"relationshipURLBase":"https://codingai-lec03.pages.dev/course/notebooks/b8320404134f4a2e858ccfe7d50a71d3.html","relationshipResolution":"Resolve relative relationship URLs against relationshipURLBase; complete records and absolute URLs are in blocks.","title":"CLS 05 · 对比学习：视图、投影与历史字典","url":"https://codingai-lec03.pages.dev/course/notebooks/b8320404134f4a2e858ccfe7d50a71d3.html","markdown":"https://codingai-lec03.pages.dev/course/notebooks/b8320404134f4a2e858ccfe7d50a71d3.md","pages":[{"id":"reading-opening","title":"CLS 05 · 对比学习：视图、投影与历史字典","unit":"训练留下什么，训练答案又从哪里来？","unitId":"i-reuse-and-targets","chapter":"I","number":1},{"id":"reading-block-a7a8f9c43bc64c6e8609e87186f7cb77","title":"从两张视图到一个可微分的选择题","unit":"训练留下什么，训练答案又从哪里来？","unitId":"i-reuse-and-targets","chapter":"I","number":2},{"id":"reading-block-38813722014249e6a8ee5e14b736dca2","title":"队列解决候选数量，EMA 缓和特征坐标系漂移","unit":"训练留下什么，训练答案又从哪里来？","unitId":"i-reuse-and-targets","chapter":"I","number":3},{"id":"reading-block-aa9100c6e89d426aa1200f6c622cd2fb","title":"研究者为什么有理由继续沿这条路线走？","unit":"训练留下什么，训练答案又从哪里来？","unitId":"i-reuse-and-targets","chapter":"I","number":4}],"aliases":[{"oldId":"b-1b967d59b35065ffbdb4c507","targetId":"b-1b967d59b35065ffbdb4c507","page":"reading-opening","operation":"move","beforeSHA256":"e1aab7fb0d675cfd630f7a549a0fdac43e2b691fe3babce3f44effa9ea3097f3","afterSHA256":"e1aab7fb0d675cfd630f7a549a0fdac43e2b691fe3babce3f44effa9ea3097f3","reason":"Same source body in authored semantic page"},{"oldId":"b-f1d9c434dea1b58549058a1c","targetId":"b-f1d9c434dea1b58549058a1c","page":"reading-opening","operation":"move","beforeSHA256":"e91f24e4e4ca40e2099a0d81d30db6568f4bf2074b499eeee6e3959617bd97ac","afterSHA256":"e91f24e4e4ca40e2099a0d81d30db6568f4bf2074b499eeee6e3959617bd97ac","reason":"Same source body in authored semantic page"},{"oldId":"b-5c2e8c0078648b732439cfff","targetId":"b-5c2e8c0078648b732439cfff","page":"reading-opening","operation":"move","beforeSHA256":"c023933b2170d7e6bbe13f73a3029a2b1c1858f66f78f1255a4634227653e856","afterSHA256":"c023933b2170d7e6bbe13f73a3029a2b1c1858f66f78f1255a4634227653e856","reason":"Same source body in authored semantic page"},{"oldId":"b-7af6e080f43ced9a2c2ec010","targetId":"b-7af6e080f43ced9a2c2ec010","page":"reading-opening","operation":"move","beforeSHA256":"f22152fd55dcdb0090b5653d37145694fd19388d31e10226549464fe8b483578","afterSHA256":"f22152fd55dcdb0090b5653d37145694fd19388d31e10226549464fe8b483578","reason":"Same source body in authored semantic page"},{"oldId":"b-53d7b18ac411d732d5c4ef84","targetId":"b-53d7b18ac411d732d5c4ef84","page":"reading-block-a7a8f9c43bc64c6e8609e87186f7cb77","operation":"move","beforeSHA256":"9412c334ba69cecafc029ab0c22814d9df928a6c4d41cd69390e96b2b55ccfe2","afterSHA256":"9412c334ba69cecafc029ab0c22814d9df928a6c4d41cd69390e96b2b55ccfe2","reason":"Same source body in authored semantic page"},{"oldId":"b-24565fded051a931ab4fd954","targetId":"b-24565fded051a931ab4fd954","page":"reading-block-a7a8f9c43bc64c6e8609e87186f7cb77","operation":"move","beforeSHA256":"9cb323f152c42ea8e4399cbe3e89d1da1eb963999f394dfd495a7f81ec530058","afterSHA256":"9cb323f152c42ea8e4399cbe3e89d1da1eb963999f394dfd495a7f81ec530058","reason":"Same source body in authored semantic page"},{"oldId":"b-105508ee09d1cbe501a14dc0","targetId":"b-105508ee09d1cbe501a14dc0","page":"reading-block-a7a8f9c43bc64c6e8609e87186f7cb77","operation":"move","beforeSHA256":"f935511521291987e8d32a9d2f118aead493bf7ee5d87ccef6005b301011abf6","afterSHA256":"f935511521291987e8d32a9d2f118aead493bf7ee5d87ccef6005b301011abf6","reason":"Same source body in authored semantic page"},{"oldId":"b-f7d9415c86d35c923d8a0ec2","targetId":"b-f7d9415c86d35c923d8a0ec2","page":"reading-block-a7a8f9c43bc64c6e8609e87186f7cb77","operation":"move","beforeSHA256":"a14cb3bb25ac2e163235dba6292f6c6dc48dcfcf9dc59e99ae843a1fa8e2f1ed","afterSHA256":"a14cb3bb25ac2e163235dba6292f6c6dc48dcfcf9dc59e99ae843a1fa8e2f1ed","reason":"Same source body in authored semantic page"},{"oldId":"b-7bbeab837f42b319a6b3c014","targetId":"b-7bbeab837f42b319a6b3c014","page":"reading-block-a7a8f9c43bc64c6e8609e87186f7cb77","operation":"move","beforeSHA256":"9eaa2737f13a8bb403958a4b5ea1f8413c4fe21102fc1568aa8462c83c674588","afterSHA256":"9eaa2737f13a8bb403958a4b5ea1f8413c4fe21102fc1568aa8462c83c674588","reason":"Same source body in authored semantic page"},{"oldId":"b-2d3ae6bb9490098f72bcd1ff","targetId":"b-2d3ae6bb9490098f72bcd1ff","page":"reading-block-a7a8f9c43bc64c6e8609e87186f7cb77","operation":"move","beforeSHA256":"f28b0a52119137411dd8ab12c666dd9da5e57c4838b284b24b74ebffb6976cfe","afterSHA256":"f28b0a52119137411dd8ab12c666dd9da5e57c4838b284b24b74ebffb6976cfe","reason":"Same source body in authored semantic page"},{"oldId":"b-a8ec36515d73cfdacf146b55","targetId":"b-a8ec36515d73cfdacf146b55","page":"reading-block-a7a8f9c43bc64c6e8609e87186f7cb77","operation":"move","beforeSHA256":"e8ad5b16d365d3ed5a5c7ce3fde725c74331832f6e62b2f7bf33fd1909200cfd","afterSHA256":"e8ad5b16d365d3ed5a5c7ce3fde725c74331832f6e62b2f7bf33fd1909200cfd","reason":"Same source body in authored semantic page"},{"oldId":"b-32bc13f97b9ff328c8c725f5","targetId":"b-32bc13f97b9ff328c8c725f5","page":"reading-block-a7a8f9c43bc64c6e8609e87186f7cb77","operation":"move","beforeSHA256":"9199b465c0c665ebce4759b252fe603a1a408fbfae1bb3ab94339ee2add2c802","afterSHA256":"9199b465c0c665ebce4759b252fe603a1a408fbfae1bb3ab94339ee2add2c802","reason":"Same source body in authored semantic page"},{"oldId":"b-737b07ec1bcb0daaadc1cc4c","targetId":"b-737b07ec1bcb0daaadc1cc4c","page":"reading-block-a7a8f9c43bc64c6e8609e87186f7cb77","operation":"move","beforeSHA256":"adc84daeb588357c3c31e6da7d6c09ebecd6078e74e8e6636cf48f7a8e775f77","afterSHA256":"adc84daeb588357c3c31e6da7d6c09ebecd6078e74e8e6636cf48f7a8e775f77","reason":"Same source body in authored semantic page"},{"oldId":"b-d077fda68d5d1fea5b0e4bd3","targetId":"b-d077fda68d5d1fea5b0e4bd3","page":"reading-block-a7a8f9c43bc64c6e8609e87186f7cb77","operation":"move","beforeSHA256":"1571dbae1ada9dbf34cdbbd00400cd017e7a05cc5fc496dcd5ffe97ba7a8a7b7","afterSHA256":"1571dbae1ada9dbf34cdbbd00400cd017e7a05cc5fc496dcd5ffe97ba7a8a7b7","reason":"Same source body in authored semantic page"},{"oldId":"b-013ce3b881259d2ea8592c33","targetId":"b-013ce3b881259d2ea8592c33","page":"reading-block-a7a8f9c43bc64c6e8609e87186f7cb77","operation":"move","beforeSHA256":"fa7a35afd1de0c0e4d821a10fd183adf7b8e20a7e2bc1c49bda8198352aefe7b","afterSHA256":"fa7a35afd1de0c0e4d821a10fd183adf7b8e20a7e2bc1c49bda8198352aefe7b","reason":"Same source body in authored semantic page"},{"oldId":"b-2e56bc3cb2d67ff0a6a2fd5f","targetId":"b-2e56bc3cb2d67ff0a6a2fd5f","page":"reading-block-a7a8f9c43bc64c6e8609e87186f7cb77","operation":"move","beforeSHA256":"f56c232bd52e5020fa340c9c136070bd13e9ceee464f986929f7ac9606efd5c1","afterSHA256":"f56c232bd52e5020fa340c9c136070bd13e9ceee464f986929f7ac9606efd5c1","reason":"Same source body in authored semantic page"},{"oldId":"b-23be33fa71156ed44f6bddcf","targetId":"b-23be33fa71156ed44f6bddcf","page":"reading-block-a7a8f9c43bc64c6e8609e87186f7cb77","operation":"move","beforeSHA256":"4c3e6341c20928a51a160dedbf3f400e6361ad7a8d7eb431ea5a89a28ce1eaea","afterSHA256":"4c3e6341c20928a51a160dedbf3f400e6361ad7a8d7eb431ea5a89a28ce1eaea","reason":"Same source body in authored semantic page"},{"oldId":"b-43db1c1f974f6e92ae8e5b82","targetId":"b-43db1c1f974f6e92ae8e5b82","page":"reading-block-a7a8f9c43bc64c6e8609e87186f7cb77","operation":"rewrite","beforeSHA256":"51f3e5310da18b988e306f65684313e7e71e01860bb54a6a00494d8bffce1362","afterSHA256":"aee97eabc966fb919064f5873a200fd72b57e5a97fdd59c88abdc230c48e6942","reason":"Same source identity with figure semantics or reading context enriched"},{"oldId":"b-50664fe575dc42a19ab365aa","targetId":"b-50664fe575dc42a19ab365aa","page":"reading-block-a7a8f9c43bc64c6e8609e87186f7cb77","operation":"move","beforeSHA256":"7ba3018ebf92c10b1fc78f3cedab36def010e2813e96cfe56b3fff05bd46a085","afterSHA256":"7ba3018ebf92c10b1fc78f3cedab36def010e2813e96cfe56b3fff05bd46a085","reason":"Same source body in authored semantic page"},{"oldId":"b-c1cc00506eb9b390c5374f76","targetId":"b-c1cc00506eb9b390c5374f76","page":"reading-block-a7a8f9c43bc64c6e8609e87186f7cb77","operation":"move","beforeSHA256":"9ae01a37f9c65f1d1a18d40eca98d9792d0f52c55c9e1d68ba854b1f528c4cd6","afterSHA256":"9ae01a37f9c65f1d1a18d40eca98d9792d0f52c55c9e1d68ba854b1f528c4cd6","reason":"Same source body in authored semantic page"},{"oldId":"b-2fcb4448e42b61cb00592633","targetId":"b-2fcb4448e42b61cb00592633","page":"reading-block-a7a8f9c43bc64c6e8609e87186f7cb77","operation":"move","beforeSHA256":"60c6c757dc9996e3547d297bf1471b72cf1cb47de89e7adb3fdf2074e6660e62","afterSHA256":"60c6c757dc9996e3547d297bf1471b72cf1cb47de89e7adb3fdf2074e6660e62","reason":"Same source body in authored semantic page"},{"oldId":"b-333c239a6edc21aaec57aa4b","targetId":"b-333c239a6edc21aaec57aa4b","page":"reading-block-a7a8f9c43bc64c6e8609e87186f7cb77","operation":"move","beforeSHA256":"162818d06b0c1920ea801f7a914ddee919bcf66303205da1d7ecfa2c002f0e0d","afterSHA256":"162818d06b0c1920ea801f7a914ddee919bcf66303205da1d7ecfa2c002f0e0d","reason":"Same source body in authored semantic page"},{"oldId":"b-09a1b6414436ed7e492d94e0","targetId":"b-09a1b6414436ed7e492d94e0","page":"reading-block-a7a8f9c43bc64c6e8609e87186f7cb77","operation":"move","beforeSHA256":"1e9d049048183aa6e6e76b4872e6d46739728419873e176868c84cb1e0419b25","afterSHA256":"1e9d049048183aa6e6e76b4872e6d46739728419873e176868c84cb1e0419b25","reason":"Same source body in authored semantic page"},{"oldId":"b-4813bc4f52ee0745be8fe68c","targetId":"b-4813bc4f52ee0745be8fe68c","page":"reading-block-a7a8f9c43bc64c6e8609e87186f7cb77","operation":"move","beforeSHA256":"5d23cd73ca3e47dcbef8d61f23e2fa0ab3f423d612b6dd4408773577fc82375f","afterSHA256":"5d23cd73ca3e47dcbef8d61f23e2fa0ab3f423d612b6dd4408773577fc82375f","reason":"Same source body in authored semantic page"},{"oldId":"b-d805dbac107993e22b0df6a4","targetId":"b-d805dbac107993e22b0df6a4","page":"reading-block-a7a8f9c43bc64c6e8609e87186f7cb77","operation":"move","beforeSHA256":"0c4c6ea20d2153b2e1535a2be1939033ad5c45c42167fba81ef0c782b77d60f1","afterSHA256":"0c4c6ea20d2153b2e1535a2be1939033ad5c45c42167fba81ef0c782b77d60f1","reason":"Same source body in authored semantic page"},{"oldId":"b-5bbe36e223e61eedbbfa9ab2","targetId":"b-5bbe36e223e61eedbbfa9ab2","page":"reading-block-a7a8f9c43bc64c6e8609e87186f7cb77","operation":"move","beforeSHA256":"a659d64a58c06d563b62e4fe28734e91d077998a3700743c08340c069019df20","afterSHA256":"a659d64a58c06d563b62e4fe28734e91d077998a3700743c08340c069019df20","reason":"Same source body in authored semantic page"},{"oldId":"b-7018ef621ca36ddbedc5bbba","targetId":"b-7018ef621ca36ddbedc5bbba","page":"reading-block-a7a8f9c43bc64c6e8609e87186f7cb77","operation":"move","beforeSHA256":"542bbbbbf07478a29b46b11c0e054862781ed72fdbffa19e42b223caf08168a2","afterSHA256":"542bbbbbf07478a29b46b11c0e054862781ed72fdbffa19e42b223caf08168a2","reason":"Same source body in authored semantic page"},{"oldId":"b-55c33e4ba51da7ab55443c01","targetId":"b-55c33e4ba51da7ab55443c01","page":"reading-block-a7a8f9c43bc64c6e8609e87186f7cb77","operation":"move","beforeSHA256":"cdfb9cdcad42224287fc11f2cf2313e29b96100a33841b6a10511d0ba3b8adea","afterSHA256":"cdfb9cdcad42224287fc11f2cf2313e29b96100a33841b6a10511d0ba3b8adea","reason":"Same source body in authored semantic page"},{"oldId":"b-403c6699eab8d0b2e90ca40a","targetId":"b-403c6699eab8d0b2e90ca40a","page":"reading-block-38813722014249e6a8ee5e14b736dca2","operation":"move","beforeSHA256":"13ad202bf0f7dee20c986d719ee6941a3db7967c52a2a292c134601c0ef77683","afterSHA256":"13ad202bf0f7dee20c986d719ee6941a3db7967c52a2a292c134601c0ef77683","reason":"Same source body in authored semantic page"},{"oldId":"b-b68fdd134718b535c23f50b8","targetId":"b-b68fdd134718b535c23f50b8","page":"reading-block-38813722014249e6a8ee5e14b736dca2","operation":"move","beforeSHA256":"19896d78aeb907e9684aabde5af58ecc2db736f45f0fed7e1cd0a4b8f2daf80f","afterSHA256":"19896d78aeb907e9684aabde5af58ecc2db736f45f0fed7e1cd0a4b8f2daf80f","reason":"Same source body in authored semantic page"},{"oldId":"b-86464cca8a354b7ebc841835","targetId":"b-86464cca8a354b7ebc841835","page":"reading-block-38813722014249e6a8ee5e14b736dca2","operation":"move","beforeSHA256":"280028e46b956fac640ebb67c8bb7e918caba37ffff27f805b05c236d69a7bea","afterSHA256":"280028e46b956fac640ebb67c8bb7e918caba37ffff27f805b05c236d69a7bea","reason":"Same source body in authored semantic page"},{"oldId":"b-9ec9f6a6b54160158049a751","targetId":"b-9ec9f6a6b54160158049a751","page":"reading-block-38813722014249e6a8ee5e14b736dca2","operation":"move","beforeSHA256":"938a7eebf48260181920170d01afbceba7b0c1a0530743e2730361f2751f49a1","afterSHA256":"938a7eebf48260181920170d01afbceba7b0c1a0530743e2730361f2751f49a1","reason":"Same source body in authored semantic page"},{"oldId":"b-2cbdb46c591ac509c975e2c0","targetId":"b-2cbdb46c591ac509c975e2c0","page":"reading-block-38813722014249e6a8ee5e14b736dca2","operation":"move","beforeSHA256":"cab2d398a639b456c7177b2b3a95d5f2ae1fe4a58c6ac53373a1b26511498cc0","afterSHA256":"cab2d398a639b456c7177b2b3a95d5f2ae1fe4a58c6ac53373a1b26511498cc0","reason":"Same source body in authored semantic page"},{"oldId":"b-d1a33a857de9ee581a2a75fd","targetId":"b-d1a33a857de9ee581a2a75fd","page":"reading-block-38813722014249e6a8ee5e14b736dca2","operation":"rewrite","beforeSHA256":"96e727a6c9211ec5416b12a6020c16b9356bdd2fb798474197a03dad59eb9117","afterSHA256":"62fdb0da604d426cd683fd27eb2e857aad9dcbe43b3ccb9854372ec221ae6d6f","reason":"Same source identity with figure semantics or reading context enriched"},{"oldId":"b-c4033a2d5ae56a273bb47e48","targetId":"b-c4033a2d5ae56a273bb47e48","page":"reading-block-38813722014249e6a8ee5e14b736dca2","operation":"move","beforeSHA256":"2ba3b260bb0604853faa2a1daad1083b7ec067c1949a2c935ab2a6c3b2be1839","afterSHA256":"2ba3b260bb0604853faa2a1daad1083b7ec067c1949a2c935ab2a6c3b2be1839","reason":"Same source body in authored semantic page"},{"oldId":"b-0aa71c35431a5068d5c9120a","targetId":"b-0aa71c35431a5068d5c9120a","page":"reading-block-38813722014249e6a8ee5e14b736dca2","operation":"move","beforeSHA256":"b69c0c9f503f70eec34c1c4211e649acec947654b830cc37772f274ee7e290a0","afterSHA256":"b69c0c9f503f70eec34c1c4211e649acec947654b830cc37772f274ee7e290a0","reason":"Same source body in authored semantic page"},{"oldId":"b-9e80eff8ee0f98dbbcfd3b7d","targetId":"b-9e80eff8ee0f98dbbcfd3b7d","page":"reading-block-38813722014249e6a8ee5e14b736dca2","operation":"move","beforeSHA256":"912e990e14a7ac966b4697be033887a98e04c708b08a85be668341b6e3b43a3e","afterSHA256":"912e990e14a7ac966b4697be033887a98e04c708b08a85be668341b6e3b43a3e","reason":"Same source body in authored semantic page"},{"oldId":"b-ae0fbac703919117e0856417","targetId":"b-ae0fbac703919117e0856417","page":"reading-block-38813722014249e6a8ee5e14b736dca2","operation":"move","beforeSHA256":"da4b9f90252f4b39b9e979ccc07bb322ee9d5be1cc08040b3f9e718266e5d680","afterSHA256":"da4b9f90252f4b39b9e979ccc07bb322ee9d5be1cc08040b3f9e718266e5d680","reason":"Same source body in authored semantic page"},{"oldId":"b-a1e0c702fbbf81588a8bb5d6","targetId":"b-a1e0c702fbbf81588a8bb5d6","page":"reading-block-38813722014249e6a8ee5e14b736dca2","operation":"move","beforeSHA256":"ca5d95bc4159403ab2f762e0281e0868f2600b5e08854fece5f940b1569dff0a","afterSHA256":"ca5d95bc4159403ab2f762e0281e0868f2600b5e08854fece5f940b1569dff0a","reason":"Same source body in authored semantic page"},{"oldId":"b-06f246687dc35899e829575c","targetId":"b-06f246687dc35899e829575c","page":"reading-block-38813722014249e6a8ee5e14b736dca2","operation":"move","beforeSHA256":"68033e34c77ec9b226141fdb781f4d1793b69e146bd6834e9c7a1c665624bdfc","afterSHA256":"68033e34c77ec9b226141fdb781f4d1793b69e146bd6834e9c7a1c665624bdfc","reason":"Same source body in authored semantic page"},{"oldId":"b-1d599c93d7dcd4a8696e65b4","targetId":"b-1d599c93d7dcd4a8696e65b4","page":"reading-block-38813722014249e6a8ee5e14b736dca2","operation":"rewrite","beforeSHA256":"283c0363f8a37e0f0c189f4e8c0e86abd905b52d25935e162337232801d6ced9","afterSHA256":"7f4f718dcaa14fddb0c1dad9832eb9368cf758d6881a4f1887b3ca4f102fe033","reason":"Same source identity with figure semantics or reading context enriched"},{"oldId":"b-ac6146de7b07c5d432d26167","targetId":"b-ac6146de7b07c5d432d26167","page":"reading-block-38813722014249e6a8ee5e14b736dca2","operation":"move","beforeSHA256":"43da66ee00d050df626f08c02b5c642ae913fe0106ad568f2d377efe68a1386a","afterSHA256":"5e7178080f9b3e931941abe1b04337d20f4b1160cbc39c4f37f0038310ef3636","reason":"Same source body in authored semantic page"},{"oldId":"b-1eda3089913c5ddfaa70d616","targetId":"b-1eda3089913c5ddfaa70d616","page":"reading-block-38813722014249e6a8ee5e14b736dca2","operation":"move","beforeSHA256":"f61a601120cd0a20dd289ba719d67eb1b818f33fab4cf87fd90b2bd3092c4e6d","afterSHA256":"f61a601120cd0a20dd289ba719d67eb1b818f33fab4cf87fd90b2bd3092c4e6d","reason":"Same source body in authored semantic page"},{"oldId":"b-27ed21128af52f1ddbb663e0","targetId":"b-27ed21128af52f1ddbb663e0","page":"reading-block-38813722014249e6a8ee5e14b736dca2","operation":"move","beforeSHA256":"5c898a84901cad07894ad357a6d3e79e5f976c45f22fc65e28c9298a525616f8","afterSHA256":"5c898a84901cad07894ad357a6d3e79e5f976c45f22fc65e28c9298a525616f8","reason":"Same source body in authored semantic page"},{"oldId":"b-5f66df60ad49c4f23d9bc64c","targetId":"b-5f66df60ad49c4f23d9bc64c","page":"reading-block-38813722014249e6a8ee5e14b736dca2","operation":"move","beforeSHA256":"256c571e4cd5bafbfd1e4cccd709b0a6104cfe82b582cc656501b4e5eb015b48","afterSHA256":"256c571e4cd5bafbfd1e4cccd709b0a6104cfe82b582cc656501b4e5eb015b48","reason":"Same source body in authored semantic page"},{"oldId":"b-f2f3b022038383fa2598557f","targetId":"b-f2f3b022038383fa2598557f","page":"reading-block-38813722014249e6a8ee5e14b736dca2","operation":"move","beforeSHA256":"c7be9efa1d234e046121cf806e9454466d3fa4a53b8ba61bd6876e2ae7fedb44","afterSHA256":"c7be9efa1d234e046121cf806e9454466d3fa4a53b8ba61bd6876e2ae7fedb44","reason":"Same source body in authored semantic page"},{"oldId":"b-1766f39023d331d0dd08d954","targetId":"b-1766f39023d331d0dd08d954","page":"reading-block-aa9100c6e89d426aa1200f6c622cd2fb","operation":"move","beforeSHA256":"d39149b53db2953a6ef0c1687755054b5e8519db491258e72ff8994161416a28","afterSHA256":"d39149b53db2953a6ef0c1687755054b5e8519db491258e72ff8994161416a28","reason":"Same source body in authored semantic page"},{"oldId":"b-8d6f2eb248eb5c683fd6d926","targetId":"b-8d6f2eb248eb5c683fd6d926","page":"reading-block-aa9100c6e89d426aa1200f6c622cd2fb","operation":"move","beforeSHA256":"38a51d52dbda974eaf4fab05b73bdc5677b90f775950b000d2ec7cec0518fdc2","afterSHA256":"38a51d52dbda974eaf4fab05b73bdc5677b90f775950b000d2ec7cec0518fdc2","reason":"Same source body in authored semantic page"},{"oldId":"b-702cd32003c0d46a52f0b525","targetId":"b-702cd32003c0d46a52f0b525","page":"reading-block-aa9100c6e89d426aa1200f6c622cd2fb","operation":"rewrite","beforeSHA256":"8c3082b7ba11acc17d24679b827e9d7af44b0d940b49637eb47b882e1a6bde20","afterSHA256":"05be0d423c5b39f2132c94108a586c9d62c94be605786e964a2ae2455d16feca","reason":"Same source identity with figure semantics or reading context enriched"},{"oldId":"b-7013143d12306fa3c1d69dea","targetId":"b-7013143d12306fa3c1d69dea","page":"reading-block-aa9100c6e89d426aa1200f6c622cd2fb","operation":"move","beforeSHA256":"e7b926bf7cbc1fa25056b061f25eefaf0bba37b86c7983c3d39257ed11e0a513","afterSHA256":"e7b926bf7cbc1fa25056b061f25eefaf0bba37b86c7983c3d39257ed11e0a513","reason":"Same source body in authored semantic page"},{"oldId":"b-b41b5f520fb227e149782ef4","targetId":"b-b41b5f520fb227e149782ef4","page":"reading-block-aa9100c6e89d426aa1200f6c622cd2fb","operation":"move","beforeSHA256":"53fd3e352a51589d7970016133506a18b796dbfa05c222caa0740be98bbf51d8","afterSHA256":"53fd3e352a51589d7970016133506a18b796dbfa05c222caa0740be98bbf51d8","reason":"Same source body in authored semantic page"},{"oldId":"b-1a60883338f330ccbf47da8e","targetId":"b-1a60883338f330ccbf47da8e","page":"reading-block-aa9100c6e89d426aa1200f6c622cd2fb","operation":"move","beforeSHA256":"421f861c3e0e6677e1ffead65526d273d60d6a0af426f690c55cf06f485ee332","afterSHA256":"421f861c3e0e6677e1ffead65526d273d60d6a0af426f690c55cf06f485ee332","reason":"Same source body in authored semantic page"},{"oldId":"b-2c43a24dcbff366e21ce289d","targetId":"b-2c43a24dcbff366e21ce289d","page":"reading-block-aa9100c6e89d426aa1200f6c622cd2fb","operation":"move","beforeSHA256":"c3e6c2e2fbe739dd93a965244fb0eff35c8e7734942874e27a0dff60f535eceb","afterSHA256":"c3e6c2e2fbe739dd93a965244fb0eff35c8e7734942874e27a0dff60f535eceb","reason":"Same source body in authored semantic page"},{"oldId":"block-a7a8f9c43bc64c6e8609e87186f7cb77--从两张视图到一个可微分的选择题","targetId":"b-2d3ae6bb9490098f72bcd1ff","operation":"retain-source-identity"},{"oldId":"block-a7a8f9c43bc64c6e8609e87186f7cb77--温度改变竞争强度不改变这组相似度的排序","targetId":"b-2e56bc3cb2d67ff0a6a2fd5f","operation":"retain-source-identity"},{"oldId":"block-a7a8f9c43bc64c6e8609e87186f7cb77--为什么损失作用在-z迁移却常读-h","targetId":"b-d805dbac107993e22b0df6a4","operation":"retain-source-identity"},{"oldId":"block-38813722014249e6a8ee5e14b736dca2--队列解决候选数量ema-缓和特征坐标系漂移","targetId":"b-403c6699eab8d0b2e90ca40a","operation":"retain-source-identity"},{"oldId":"block-38813722014249e6a8ee5e14b736dca2--训练学会区分视图以后还要回答哪个下游问题","targetId":"b-0aa71c35431a5068d5c9120a","operation":"retain-source-identity"},{"oldId":"block-38813722014249e6a8ee5e14b736dca2--同一个颜色干预真的改变了学到的表示","targetId":"b-a1e0c702fbbf81588a8bb5d6","operation":"retain-source-identity"},{"oldId":"block-aa9100c6e89d426aa1200f6c622cd2fb--研究者为什么有理由继续沿这条路线走","targetId":"b-1766f39023d331d0dd08d954","operation":"retain-source-identity"}],"blocks":[{"id":"b-1b967d59b35065ffbdb4c507","type":"h1","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebooks/b8320404134f4a2e858ccfe7d50a71d3.html#b-1b967d59b35065ffbdb4c507","content":"# CLS 05 · 对比学习：视图、投影与历史字典","sha256":"e1aab7fb0d675cfd630f7a549a0fdac43e2b691fe3babce3f44effa9ea3097f3","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-reuse-and-targets","title":"训练留下什么，训练答案又从哪里来？"},"page":{"id":"reading-opening","title":"CLS 05 · 对比学习：视图、投影与历史字典","number":1},"sectionTitle":"CLS 05 · 对比学习：视图、投影与历史字典","currentQuestion":"CLS 05 · 对比学习：视图、投影与历史字典","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":null,"nextBlock":"b-f1d9c434dea1b58549058a1c","prerequisites":[{"id":"b-f1d9c434dea1b58549058a1c","url":"#b-f1d9c434dea1b58549058a1c","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-f1d9c434dea1b58549058a1c","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebooks/b8320404134f4a2e858ccfe7d50a71d3.html#b-f1d9c434dea1b58549058a1c","content":"按问题继续","sha256":"e91f24e4e4ca40e2099a0d81d30db6568f4bf2074b499eeee6e3959617bd97ac","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-reuse-and-targets","title":"训练留下什么，训练答案又从哪里来？"},"page":{"id":"reading-opening","title":"CLS 05 · 对比学习：视图、投影与历史字典","number":1},"sectionTitle":"CLS 05 · 对比学习：视图、投影与历史字典","currentQuestion":"CLS 05 · 对比学习：视图、投影与历史字典","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-1b967d59b35065ffbdb4c507","nextBlock":"b-5c2e8c0078648b732439cfff","prerequisites":[{"id":"b-1b967d59b35065ffbdb4c507","url":"#b-1b967d59b35065ffbdb4c507","role":"section-title"}],"relatedBlocks":[]},{"id":"b-5c2e8c0078648b732439cfff","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebooks/b8320404134f4a2e858ccfe7d50a71d3.html#b-5c2e8c0078648b732439cfff","content":"InfoNCE概率与梯度、h/z、温度、队列陈旧性、实测与评价协议","sha256":"c023933b2170d7e6bbe13f73a3029a2b1c1858f66f78f1255a4634227653e856","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-reuse-and-targets","title":"训练留下什么，训练答案又从哪里来？"},"page":{"id":"reading-opening","title":"CLS 05 · 对比学习：视图、投影与历史字典","number":1},"sectionTitle":"CLS 05 · 对比学习：视图、投影与历史字典","currentQuestion":"CLS 05 · 对比学习：视图、投影与历史字典","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-f1d9c434dea1b58549058a1c","nextBlock":"b-7af6e080f43ced9a2c2ec010","prerequisites":[{"id":"b-1b967d59b35065ffbdb4c507","url":"#b-1b967d59b35065ffbdb4c507","role":"section-title"},{"id":"b-f1d9c434dea1b58549058a1c","url":"#b-f1d9c434dea1b58549058a1c","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-7af6e080f43ced9a2c2ec010","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebooks/b8320404134f4a2e858ccfe7d50a71d3.html#b-7af6e080f43ced9a2c2ec010","content":"可先阅读已有结果，再决定是否运行。重算时请核对各节记录的数据、参数、环境与日期。","sha256":"f22152fd55dcdb0090b5653d37145694fd19388d31e10226549464fe8b483578","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-reuse-and-targets","title":"训练留下什么，训练答案又从哪里来？"},"page":{"id":"reading-opening","title":"CLS 05 · 对比学习：视图、投影与历史字典","number":1},"sectionTitle":"CLS 05 · 对比学习：视图、投影与历史字典","currentQuestion":"CLS 05 · 对比学习：视图、投影与历史字典","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-5c2e8c0078648b732439cfff","nextBlock":null,"prerequisites":[{"id":"b-1b967d59b35065ffbdb4c507","url":"#b-1b967d59b35065ffbdb4c507","role":"section-title"},{"id":"b-f1d9c434dea1b58549058a1c","url":"#b-f1d9c434dea1b58549058a1c","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-53d7b18ac411d732d5c4ef84","type":"p","scope":"block-a7a8f9c43bc64c6e8609e87186f7cb77","url":"https://codingai-lec03.pages.dev/course/notebooks/b8320404134f4a2e858ccfe7d50a71d3.html#b-53d7b18ac411d732d5c4ef84","content":"**学习路径：** [视图、投影头与历史字典](/course/study/contrastive-views.html)","sha256":"9412c334ba69cecafc029ab0c22814d9df928a6c4d41cd69390e96b2b55ccfe2","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-reuse-and-targets","title":"训练留下什么，训练答案又从哪里来？"},"page":{"id":"reading-block-a7a8f9c43bc64c6e8609e87186f7cb77","title":"从两张视图到一个可微分的选择题","number":2},"sectionTitle":"从两张视图到一个可微分的选择题","currentQuestion":"从两张视图到一个可微分的选择题","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":null,"nextBlock":"b-24565fded051a931ab4fd954","prerequisites":[],"relatedBlocks":[]},{"id":"b-24565fded051a931ab4fd954","type":"p","scope":"block-a7a8f9c43bc64c6e8609e87186f7cb77","url":"https://codingai-lec03.pages.dev/course/notebooks/b8320404134f4a2e858ccfe7d50a71d3.html#b-24565fded051a931ab4fd954","content":"[P08同源视图](/course/pages/l3-r2-p08.html) → [P12投影头](/course/pages/l3-r2-p12.html) → [P17一次队列更新](/course/pages/l3-r2-p17.html)","sha256":"9cb323f152c42ea8e4399cbe3e89d1da1eb963999f394dfd495a7f81ec530058","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-reuse-and-targets","title":"训练留下什么，训练答案又从哪里来？"},"page":{"id":"reading-block-a7a8f9c43bc64c6e8609e87186f7cb77","title":"从两张视图到一个可微分的选择题","number":2},"sectionTitle":"从两张视图到一个可微分的选择题","currentQuestion":"从两张视图到一个可微分的选择题","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-53d7b18ac411d732d5c4ef84","nextBlock":"b-105508ee09d1cbe501a14dc0","prerequisites":[{"id":"b-53d7b18ac411d732d5c4ef84","url":"#b-53d7b18ac411d732d5c4ef84","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-105508ee09d1cbe501a14dc0","type":"p","scope":"block-a7a8f9c43bc64c6e8609e87186f7cb77","url":"https://codingai-lec03.pages.dev/course/notebooks/b8320404134f4a2e858ccfe7d50a71d3.html#b-105508ee09d1cbe501a14dc0","content":"先把一张图的两种观察放入同一候选表，算出正例概率；再追踪梯度怎样经过z回到h，以及旧key为什么需要缓慢变化的编码器。[温度的概率与梯度算例](/course/blocks/temperature-ranking-probability-gradient.html)、[队列身份](/course/blocks/moco-key-queue-identity.html)都可以独立引用。","sha256":"f935511521291987e8d32a9d2f118aead493bf7ee5d87ccef6005b301011abf6","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-reuse-and-targets","title":"训练留下什么，训练答案又从哪里来？"},"page":{"id":"reading-block-a7a8f9c43bc64c6e8609e87186f7cb77","title":"从两张视图到一个可微分的选择题","number":2},"sectionTitle":"从两张视图到一个可微分的选择题","currentQuestion":"从两张视图到一个可微分的选择题","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-24565fded051a931ab4fd954","nextBlock":"b-f7d9415c86d35c923d8a0ec2","prerequisites":[{"id":"b-53d7b18ac411d732d5c4ef84","url":"#b-53d7b18ac411d732d5c4ef84","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-f7d9415c86d35c923d8a0ec2","type":"p","scope":"block-a7a8f9c43bc64c6e8609e87186f7cb77","url":"https://codingai-lec03.pages.dev/course/notebooks/b8320404134f4a2e858ccfe7d50a71d3.html#b-f7d9415c86d35c923d8a0ec2","content":"动手只选一个入口：想检查增强留下什么，选[LAB 06](https://deepnote.com/project/38755e4e-23f6-4f5f-8c68-8a2e4114b1c1/notebook/7c4f531113d7473abc174e2e014e1340)；想追踪真实key、FIFO和EMA，选[LAB 07](https://deepnote.com/project/38755e4e-23f6-4f5f-8c68-8a2e4114b1c1/notebook/0d1b8f6fbf6a476c93ffcc169f2ec1f4)。读完用一张图或一段trace指出输入、候选和两种参数更新，再用独立读出评价它是否对任务有用。后文同时保留论文迁移进展和小型实验的失败结果，二者的数据与协议分别写明。","sha256":"a14cb3bb25ac2e163235dba6292f6c6dc48dcfcf9dc59e99ae843a1fa8e2f1ed","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-reuse-and-targets","title":"训练留下什么，训练答案又从哪里来？"},"page":{"id":"reading-block-a7a8f9c43bc64c6e8609e87186f7cb77","title":"从两张视图到一个可微分的选择题","number":2},"sectionTitle":"从两张视图到一个可微分的选择题","currentQuestion":"从两张视图到一个可微分的选择题","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-105508ee09d1cbe501a14dc0","nextBlock":"b-7bbeab837f42b319a6b3c014","prerequisites":[{"id":"b-53d7b18ac411d732d5c4ef84","url":"#b-53d7b18ac411d732d5c4ef84","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-7bbeab837f42b319a6b3c014","type":"p","scope":"block-a7a8f9c43bc64c6e8609e87186f7cb77","url":"https://codingai-lec03.pages.dev/course/notebooks/b8320404134f4a2e858ccfe7d50a71d3.html#b-7bbeab837f42b319a6b3c014","content":"拍同一只杯子，改变光照、裁剪和背景，类别仍然是杯子。若能把这些变化视为同一对象的不同观察，训练目标就可以来自图像自身：两次观察的表示应该接近，其他图像的表示提供区分参照。但如果下一项任务是分辨红杯和蓝杯，“改变颜色仍然算同一个”就不再无害。**对比学习的正例关系，是设计者通过数据和增强作出的任务假设。** 它不是无标签数据自动告诉我们的普遍相同性。","sha256":"9eaa2737f13a8bb403958a4b5ea1f8413c4fe21102fc1568aa8462c83c674588","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-reuse-and-targets","title":"训练留下什么，训练答案又从哪里来？"},"page":{"id":"reading-block-a7a8f9c43bc64c6e8609e87186f7cb77","title":"从两张视图到一个可微分的选择题","number":2},"sectionTitle":"从两张视图到一个可微分的选择题","currentQuestion":"从两张视图到一个可微分的选择题","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-f7d9415c86d35c923d8a0ec2","nextBlock":"b-2d3ae6bb9490098f72bcd1ff","prerequisites":[{"id":"b-53d7b18ac411d732d5c4ef84","url":"#b-53d7b18ac411d732d5c4ef84","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-2d3ae6bb9490098f72bcd1ff","type":"h2","scope":"block-a7a8f9c43bc64c6e8609e87186f7cb77","url":"https://codingai-lec03.pages.dev/course/notebooks/b8320404134f4a2e858ccfe7d50a71d3.html#b-2d3ae6bb9490098f72bcd1ff","content":"## 从两张视图到一个可微分的选择题","sha256":"f28b0a52119137411dd8ab12c666dd9da5e57c4838b284b24b74ebffb6976cfe","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-reuse-and-targets","title":"训练留下什么，训练答案又从哪里来？"},"page":{"id":"reading-block-a7a8f9c43bc64c6e8609e87186f7cb77","title":"从两张视图到一个可微分的选择题","number":2},"sectionTitle":"从两张视图到一个可微分的选择题","currentQuestion":"从两张视图到一个可微分的选择题","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-7bbeab837f42b319a6b3c014","nextBlock":"b-a8ec36515d73cfdacf146b55","prerequisites":[{"id":"b-53d7b18ac411d732d5c4ef84","url":"#b-53d7b18ac411d732d5c4ef84","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-a8ec36515d73cfdacf146b55","type":"p","scope":"block-a7a8f9c43bc64c6e8609e87186f7cb77","url":"https://codingai-lec03.pages.dev/course/notebooks/b8320404134f4a2e858ccfe7d50a71d3.html#b-a8ec36515d73cfdacf146b55","content":"给原图 $x$ 独立采样增强 $t,t'$，得到 $v=t(x),v^+=t'(x)$。编码器 $f_\\theta$ 产生表示 $h=f_\\theta(v)$；投影头 $g_\\phi$ 产生 $z=g_\\phi(h)$，通常再做 $\\ell_2$ 归一化。余弦相似度便是单位向量点积 $s(q,k)=q^Tk$。正例 key 来自同一原图另一视图，候选负例来自其他原图。","sha256":"e8ad5b16d365d3ed5a5c7ce3fde725c74331832f6e62b2f7bf33fd1909200cfd","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-reuse-and-targets","title":"训练留下什么，训练答案又从哪里来？"},"page":{"id":"reading-block-a7a8f9c43bc64c6e8609e87186f7cb77","title":"从两张视图到一个可微分的选择题","number":2},"sectionTitle":"从两张视图到一个可微分的选择题","currentQuestion":"从两张视图到一个可微分的选择题","sourceSlides":[],"knowledgeBlock":null,"formulas":[{"latex":"x","definitionContext":"从两张视图到一个可微分的选择题"},{"latex":"t,t'","definitionContext":"从两张视图到一个可微分的选择题"},{"latex":"v=t(x),v^+=t'(x)","definitionContext":"从两张视图到一个可微分的选择题"},{"latex":"f_\\theta","definitionContext":"从两张视图到一个可微分的选择题"},{"latex":"h=f_\\theta(v)","definitionContext":"从两张视图到一个可微分的选择题"},{"latex":"g_\\phi","definitionContext":"从两张视图到一个可微分的选择题"},{"latex":"z=g_\\phi(h)","definitionContext":"从两张视图到一个可微分的选择题"},{"latex":"\\ell_2","definitionContext":"从两张视图到一个可微分的选择题"},{"latex":"s(q,k)=q^Tk","definitionContext":"从两张视图到一个可微分的选择题"}],"previousBlock":"b-2d3ae6bb9490098f72bcd1ff","nextBlock":"b-32bc13f97b9ff328c8c725f5","prerequisites":[{"id":"b-2d3ae6bb9490098f72bcd1ff","url":"#b-2d3ae6bb9490098f72bcd1ff","role":"section-title"},{"id":"b-53d7b18ac411d732d5c4ef84","url":"#b-53d7b18ac411d732d5c4ef84","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-32bc13f97b9ff328c8c725f5","type":"p","scope":"block-a7a8f9c43bc64c6e8609e87186f7cb77","url":"https://codingai-lec03.pages.dev/course/notebooks/b8320404134f4a2e858ccfe7d50a71d3.html#b-32bc13f97b9ff328c8c725f5","content":"对一个 query $q$、一个正例 $k^+$ 和若干负例 $k_j^-$，写出 InfoNCE 形式：","sha256":"9199b465c0c665ebce4759b252fe603a1a408fbfae1bb3ab94339ee2add2c802","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-reuse-and-targets","title":"训练留下什么，训练答案又从哪里来？"},"page":{"id":"reading-block-a7a8f9c43bc64c6e8609e87186f7cb77","title":"从两张视图到一个可微分的选择题","number":2},"sectionTitle":"从两张视图到一个可微分的选择题","currentQuestion":"从两张视图到一个可微分的选择题","sourceSlides":[],"knowledgeBlock":null,"formulas":[{"latex":"q","definitionContext":"从两张视图到一个可微分的选择题"},{"latex":"k^+","definitionContext":"从两张视图到一个可微分的选择题"},{"latex":"k_j^-","definitionContext":"从两张视图到一个可微分的选择题"}],"previousBlock":"b-a8ec36515d73cfdacf146b55","nextBlock":"b-737b07ec1bcb0daaadc1cc4c","prerequisites":[{"id":"b-2d3ae6bb9490098f72bcd1ff","url":"#b-2d3ae6bb9490098f72bcd1ff","role":"section-title"},{"id":"b-53d7b18ac411d732d5c4ef84","url":"#b-53d7b18ac411d732d5c4ef84","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-737b07ec1bcb0daaadc1cc4c","type":"p","scope":"block-a7a8f9c43bc64c6e8609e87186f7cb77","url":"https://codingai-lec03.pages.dev/course/notebooks/b8320404134f4a2e858ccfe7d50a71d3.html#b-737b07ec1bcb0daaadc1cc4c","content":"$L_q=-\\log\\frac{\\exp(s(q,k^+)/\\tau)}{\\exp(s(q,k^+)/\\tau)+\\sum_j\\exp(s(q,k_j^-)/\\tau)}.$","sha256":"adc84daeb588357c3c31e6da7d6c09ebecd6078e74e8e6636cf48f7a8e775f77","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-reuse-and-targets","title":"训练留下什么，训练答案又从哪里来？"},"page":{"id":"reading-block-a7a8f9c43bc64c6e8609e87186f7cb77","title":"从两张视图到一个可微分的选择题","number":2},"sectionTitle":"从两张视图到一个可微分的选择题","currentQuestion":"从两张视图到一个可微分的选择题","sourceSlides":[],"knowledgeBlock":null,"formulas":[{"latex":"L_q=-\\log\\frac{\\exp(s(q,k^+)/\\tau)}{\\exp(s(q,k^+)/\\tau)+\\sum_j\\exp(s(q,k_j^-)/\\tau)}.","definitionContext":"从两张视图到一个可微分的选择题"}],"previousBlock":"b-32bc13f97b9ff328c8c725f5","nextBlock":"b-d077fda68d5d1fea5b0e4bd3","prerequisites":[{"id":"b-2d3ae6bb9490098f72bcd1ff","url":"#b-2d3ae6bb9490098f72bcd1ff","role":"section-title"},{"id":"b-53d7b18ac411d732d5c4ef84","url":"#b-53d7b18ac411d732d5c4ef84","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-d077fda68d5d1fea5b0e4bd3","type":"p","scope":"block-a7a8f9c43bc64c6e8609e87186f7cb77","url":"https://codingai-lec03.pages.dev/course/notebooks/b8320404134f4a2e858ccfe7d50a71d3.html#b-d077fda68d5d1fea5b0e4bd3","content":"这里 $\\tau>0$ 是温度，分母包含正例。它把“拉近正例、推开负例”写成在当前候选集合里选中正例的交叉熵。候选身份来自采样过程，不是人工类别；两张不同原图可能都是杯子，所以负例并不保证语义类别不同。实例区分与类别聚合之间存在真实张力。","sha256":"1571dbae1ada9dbf34cdbbd00400cd017e7a05cc5fc496dcd5ffe97ba7a8a7b7","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-reuse-and-targets","title":"训练留下什么，训练答案又从哪里来？"},"page":{"id":"reading-block-a7a8f9c43bc64c6e8609e87186f7cb77","title":"从两张视图到一个可微分的选择题","number":2},"sectionTitle":"从两张视图到一个可微分的选择题","currentQuestion":"从两张视图到一个可微分的选择题","sourceSlides":[],"knowledgeBlock":null,"formulas":[{"latex":"\\tau>0","definitionContext":"从两张视图到一个可微分的选择题"}],"previousBlock":"b-737b07ec1bcb0daaadc1cc4c","nextBlock":"b-013ce3b881259d2ea8592c33","prerequisites":[{"id":"b-2d3ae6bb9490098f72bcd1ff","url":"#b-2d3ae6bb9490098f72bcd1ff","role":"section-title"},{"id":"b-53d7b18ac411d732d5c4ef84","url":"#b-53d7b18ac411d732d5c4ef84","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-013ce3b881259d2ea8592c33","type":"p","scope":"block-a7a8f9c43bc64c6e8609e87186f7cb77","url":"https://codingai-lec03.pages.dev/course/notebooks/b8320404134f4a2e858ccfe7d50a71d3.html#b-013ce3b881259d2ea8592c33","content":"SimCLR 的一个 batch 有 $N$ 张原图，生成 $2N$ 个视图。对其中任一视图，排除自身，余下 $2N-1$ 项中有一个正例和 $2N-2$ 个负例。自身相似度通常接近 1，若误把它留进分母，就额外加入一个不可避免的强竞争者。损失平均两个方向：视图 A 找 B，B 也找 A。可核对 [SimCLR §2、式 1 与 Algorithm 1](https://proceedings.mlr.press/v119/chen20j/chen20j.pdf)。","sha256":"fa7a35afd1de0c0e4d821a10fd183adf7b8e20a7e2bc1c49bda8198352aefe7b","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-reuse-and-targets","title":"训练留下什么，训练答案又从哪里来？"},"page":{"id":"reading-block-a7a8f9c43bc64c6e8609e87186f7cb77","title":"从两张视图到一个可微分的选择题","number":2},"sectionTitle":"从两张视图到一个可微分的选择题","currentQuestion":"从两张视图到一个可微分的选择题","sourceSlides":[],"knowledgeBlock":null,"formulas":[{"latex":"N","definitionContext":"从两张视图到一个可微分的选择题"},{"latex":"2N","definitionContext":"从两张视图到一个可微分的选择题"},{"latex":"2N-1","definitionContext":"从两张视图到一个可微分的选择题"},{"latex":"2N-2","definitionContext":"从两张视图到一个可微分的选择题"}],"previousBlock":"b-d077fda68d5d1fea5b0e4bd3","nextBlock":"b-2e56bc3cb2d67ff0a6a2fd5f","prerequisites":[{"id":"b-2d3ae6bb9490098f72bcd1ff","url":"#b-2d3ae6bb9490098f72bcd1ff","role":"section-title"},{"id":"b-53d7b18ac411d732d5c4ef84","url":"#b-53d7b18ac411d732d5c4ef84","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-2e56bc3cb2d67ff0a6a2fd5f","type":"h2","scope":"block-a7a8f9c43bc64c6e8609e87186f7cb77","url":"https://codingai-lec03.pages.dev/course/notebooks/b8320404134f4a2e858ccfe7d50a71d3.html#b-2e56bc3cb2d67ff0a6a2fd5f","content":"## 温度改变竞争强度，不改变这组相似度的排序","sha256":"f56c232bd52e5020fa340c9c136070bd13e9ceee464f986929f7ac9606efd5c1","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-reuse-and-targets","title":"训练留下什么，训练答案又从哪里来？"},"page":{"id":"reading-block-a7a8f9c43bc64c6e8609e87186f7cb77","title":"从两张视图到一个可微分的选择题","number":2},"sectionTitle":"从两张视图到一个可微分的选择题","currentQuestion":"从两张视图到一个可微分的选择题","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-013ce3b881259d2ea8592c33","nextBlock":"b-23be33fa71156ed44f6bddcf","prerequisites":[{"id":"b-2d3ae6bb9490098f72bcd1ff","url":"#b-2d3ae6bb9490098f72bcd1ff","role":"section-title"},{"id":"b-53d7b18ac411d732d5c4ef84","url":"#b-53d7b18ac411d732d5c4ef84","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-23be33fa71156ed44f6bddcf","type":"p","scope":"block-a7a8f9c43bc64c6e8609e87186f7cb77","url":"https://codingai-lec03.pages.dev/course/notebooks/b8320404134f4a2e858ccfe7d50a71d3.html#b-23be33fa71156ed44f6bddcf","content":"固定三项相似度：正例 0.8，两个负例 0.4、0.2。直接代入上式得到：","sha256":"4c3e6341c20928a51a160dedbf3f400e6361ad7a8d7eb431ea5a89a28ce1eaea","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-reuse-and-targets","title":"训练留下什么，训练答案又从哪里来？"},"page":{"id":"reading-block-a7a8f9c43bc64c6e8609e87186f7cb77","title":"从两张视图到一个可微分的选择题","number":2},"sectionTitle":"从两张视图到一个可微分的选择题","currentQuestion":"从两张视图到一个可微分的选择题","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-2e56bc3cb2d67ff0a6a2fd5f","nextBlock":"b-43db1c1f974f6e92ae8e5b82","prerequisites":[{"id":"b-2e56bc3cb2d67ff0a6a2fd5f","url":"#b-2e56bc3cb2d67ff0a6a2fd5f","role":"section-title"},{"id":"b-53d7b18ac411d732d5c4ef84","url":"#b-53d7b18ac411d732d5c4ef84","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-43db1c1f974f6e92ae8e5b82","type":"table","scope":"block-a7a8f9c43bc64c6e8609e87186f7cb77","url":"https://codingai-lec03.pages.dev/course/notebooks/b8320404134f4a2e858ccfe7d50a71d3.html#b-43db1c1f974f6e92ae8e5b82","content":"| 温度 $\\tau$ | 正例概率 | 负例 0.4 的概率 | 负例 0.2 的概率 | 损失 $-\\log p_+$ |\n| --- | --- | --- | --- | --- |\n| 1.0 | 0.4506 | 0.3021 | 0.2473 | 0.7971 |\n| 0.5 | 0.5713 | 0.2567 | 0.1721 | 0.5599 |\n| 0.1 | 0.9796 | 0.0179 | 0.0024 | 0.0206 |","sha256":"aee97eabc966fb919064f5873a200fd72b57e5a97fdd59c88abdc230c48e6942","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-reuse-and-targets","title":"训练留下什么，训练答案又从哪里来？"},"page":{"id":"reading-block-a7a8f9c43bc64c6e8609e87186f7cb77","title":"从两张视图到一个可微分的选择题","number":2},"sectionTitle":"从两张视图到一个可微分的选择题","currentQuestion":"从两张视图到一个可微分的选择题","sourceSlides":[],"knowledgeBlock":null,"formulas":[{"latex":"\\tau","definitionContext":"从两张视图到一个可微分的选择题"},{"latex":"-\\log p_+","definitionContext":"从两张视图到一个可微分的选择题"}],"table":{"headers":["温度 τ\\tauτ","正例概率","负例 0.4 的概率","负例 0.2 的概率","损失 −log⁡p+-\\log p_+−logp+​"],"rows":[["温度 τ\\tauτ","正例概率","负例 0.4 的概率","负例 0.2 的概率","损失 −log⁡p+-\\log p_+−logp+​"],["1.0","0.4506","0.3021","0.2473","0.7971"],["0.5","0.5713","0.2567","0.1721","0.5599"],["0.1","0.9796","0.0179","0.0024","0.0206"]]},"previousBlock":"b-23be33fa71156ed44f6bddcf","nextBlock":"b-50664fe575dc42a19ab365aa","prerequisites":[{"id":"b-2e56bc3cb2d67ff0a6a2fd5f","url":"#b-2e56bc3cb2d67ff0a6a2fd5f","role":"section-title"},{"id":"b-53d7b18ac411d732d5c4ef84","url":"#b-53d7b18ac411d732d5c4ef84","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-50664fe575dc42a19ab365aa","type":"p","scope":"block-a7a8f9c43bc64c6e8609e87186f7cb77","url":"https://codingai-lec03.pages.dev/course/notebooks/b8320404134f4a2e858ccfe7d50a71d3.html#b-50664fe575dc42a19ab365aa","content":"因为正例已经领先，低温度放大它的优势。若把正例改成 0.3、最强负例仍为 0.4，低温度则会更强地惩罚错误排序。不能仅因降低温度使这张表的损失变小，就宣称表示更好了。","sha256":"7ba3018ebf92c10b1fc78f3cedab36def010e2813e96cfe56b3fff05bd46a085","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-reuse-and-targets","title":"训练留下什么，训练答案又从哪里来？"},"page":{"id":"reading-block-a7a8f9c43bc64c6e8609e87186f7cb77","title":"从两张视图到一个可微分的选择题","number":2},"sectionTitle":"从两张视图到一个可微分的选择题","currentQuestion":"从两张视图到一个可微分的选择题","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-43db1c1f974f6e92ae8e5b82","nextBlock":"b-c1cc00506eb9b390c5374f76","prerequisites":[{"id":"b-2e56bc3cb2d67ff0a6a2fd5f","url":"#b-2e56bc3cb2d67ff0a6a2fd5f","role":"section-title"},{"id":"b-53d7b18ac411d732d5c4ef84","url":"#b-53d7b18ac411d732d5c4ef84","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-c1cc00506eb9b390c5374f76","type":"p","scope":"block-a7a8f9c43bc64c6e8609e87186f7cb77","url":"https://codingai-lec03.pages.dev/course/notebooks/b8320404134f4a2e858ccfe7d50a71d3.html#b-c1cc00506eb9b390c5374f76","content":"梯度把这个现象写得更精确。令 $p_j$ 为候选 softmax 概率，则","sha256":"9ae01a37f9c65f1d1a18d40eca98d9792d0f52c55c9e1d68ba854b1f528c4cd6","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-reuse-and-targets","title":"训练留下什么，训练答案又从哪里来？"},"page":{"id":"reading-block-a7a8f9c43bc64c6e8609e87186f7cb77","title":"从两张视图到一个可微分的选择题","number":2},"sectionTitle":"从两张视图到一个可微分的选择题","currentQuestion":"从两张视图到一个可微分的选择题","sourceSlides":[],"knowledgeBlock":null,"formulas":[{"latex":"p_j","definitionContext":"从两张视图到一个可微分的选择题"}],"previousBlock":"b-50664fe575dc42a19ab365aa","nextBlock":"b-2fcb4448e42b61cb00592633","prerequisites":[{"id":"b-2e56bc3cb2d67ff0a6a2fd5f","url":"#b-2e56bc3cb2d67ff0a6a2fd5f","role":"section-title"},{"id":"b-53d7b18ac411d732d5c4ef84","url":"#b-53d7b18ac411d732d5c4ef84","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-2fcb4448e42b61cb00592633","type":"p","scope":"block-a7a8f9c43bc64c6e8609e87186f7cb77","url":"https://codingai-lec03.pages.dev/course/notebooks/b8320404134f4a2e858ccfe7d50a71d3.html#b-2fcb4448e42b61cb00592633","content":"$\\frac{\\partial L}{\\partial s_+}=\\frac{p_+-1}{\\tau},\\qquad\n\\frac{\\partial L}{\\partial s_j^-}=\\frac{p_j}{\\tau}.$","sha256":"60c6c757dc9996e3547d297bf1471b72cf1cb47de89e7adb3fdf2074e6660e62","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-reuse-and-targets","title":"训练留下什么，训练答案又从哪里来？"},"page":{"id":"reading-block-a7a8f9c43bc64c6e8609e87186f7cb77","title":"从两张视图到一个可微分的选择题","number":2},"sectionTitle":"从两张视图到一个可微分的选择题","currentQuestion":"从两张视图到一个可微分的选择题","sourceSlides":[],"knowledgeBlock":null,"formulas":[{"latex":"\\frac{\\partial L}{\\partial s_+}=\\frac{p_+-1}{\\tau},\\qquad\n\\frac{\\partial L}{\\partial s_j^-}=\\frac{p_j}{\\tau}.","definitionContext":"从两张视图到一个可微分的选择题"}],"previousBlock":"b-c1cc00506eb9b390c5374f76","nextBlock":"b-333c239a6edc21aaec57aa4b","prerequisites":[{"id":"b-2e56bc3cb2d67ff0a6a2fd5f","url":"#b-2e56bc3cb2d67ff0a6a2fd5f","role":"section-title"},{"id":"b-53d7b18ac411d732d5c4ef84","url":"#b-53d7b18ac411d732d5c4ef84","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-333c239a6edc21aaec57aa4b","type":"p","scope":"block-a7a8f9c43bc64c6e8609e87186f7cb77","url":"https://codingai-lec03.pages.dev/course/notebooks/b8320404134f4a2e858ccfe7d50a71d3.html#b-333c239a6edc21aaec57aa4b","content":"温度既出现在 $1/\\tau$ 中，也改变 $p$。所以“温度越低，所有梯度越大”不成立：本例正例梯度的绝对值在 $\\tau=0.5$ 时约 0.8575，在 0.1 时反而约 0.2037，因为预测已饱和。最相似负例会获得更多概率和更强排斥，但过度强调困难负例也可能放大假负例或噪声。","sha256":"162818d06b0c1920ea801f7a914ddee919bcf66303205da1d7ecfa2c002f0e0d","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-reuse-and-targets","title":"训练留下什么，训练答案又从哪里来？"},"page":{"id":"reading-block-a7a8f9c43bc64c6e8609e87186f7cb77","title":"从两张视图到一个可微分的选择题","number":2},"sectionTitle":"从两张视图到一个可微分的选择题","currentQuestion":"从两张视图到一个可微分的选择题","sourceSlides":[],"knowledgeBlock":null,"formulas":[{"latex":"1/\\tau","definitionContext":"从两张视图到一个可微分的选择题"},{"latex":"p","definitionContext":"从两张视图到一个可微分的选择题"},{"latex":"\\tau=0.5","definitionContext":"从两张视图到一个可微分的选择题"}],"previousBlock":"b-2fcb4448e42b61cb00592633","nextBlock":"b-09a1b6414436ed7e492d94e0","prerequisites":[{"id":"b-2e56bc3cb2d67ff0a6a2fd5f","url":"#b-2e56bc3cb2d67ff0a6a2fd5f","role":"section-title"},{"id":"b-53d7b18ac411d732d5c4ef84","url":"#b-53d7b18ac411d732d5c4ef84","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-09a1b6414436ed7e492d94e0","type":"p","scope":"block-a7a8f9c43bc64c6e8609e87186f7cb77","url":"https://codingai-lec03.pages.dev/course/notebooks/b8320404134f4a2e858ccfe7d50a71d3.html#b-09a1b6414436ed7e492d94e0","content":"![同一相似度的温度变化与 MoCo 队列更新](/course/repair/assets/cls06-temperature-queue.svg)","sha256":"1e9d049048183aa6e6e76b4872e6d46739728419873e176868c84cb1e0419b25","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-reuse-and-targets","title":"训练留下什么，训练答案又从哪里来？"},"page":{"id":"reading-block-a7a8f9c43bc64c6e8609e87186f7cb77","title":"从两张视图到一个可微分的选择题","number":2},"sectionTitle":"从两张视图到一个可微分的选择题","currentQuestion":"从两张视图到一个可微分的选择题","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-333c239a6edc21aaec57aa4b","nextBlock":"b-4813bc4f52ee0745be8fe68c","prerequisites":[{"id":"b-2e56bc3cb2d67ff0a6a2fd5f","url":"#b-2e56bc3cb2d67ff0a6a2fd5f","role":"section-title"},{"id":"b-53d7b18ac411d732d5c4ef84","url":"#b-53d7b18ac411d732d5c4ef84","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-4813bc4f52ee0745be8fe68c","type":"p","scope":"block-a7a8f9c43bc64c6e8609e87186f7cb77","url":"https://codingai-lec03.pages.dev/course/notebooks/b8320404134f4a2e858ccfe7d50a71d3.html#b-4813bc4f52ee0745be8fe68c","content":"上半图按表中真实计算值画出候选概率；下半图是长度 6 的 FIFO 队列，加入两个新 key 后移出两个最老 key。温度控制一次比较，队列控制比较对象从哪里来。","sha256":"5d23cd73ca3e47dcbef8d61f23e2fa0ab3f423d612b6dd4408773577fc82375f","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-reuse-and-targets","title":"训练留下什么，训练答案又从哪里来？"},"page":{"id":"reading-block-a7a8f9c43bc64c6e8609e87186f7cb77","title":"从两张视图到一个可微分的选择题","number":2},"sectionTitle":"从两张视图到一个可微分的选择题","currentQuestion":"从两张视图到一个可微分的选择题","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-09a1b6414436ed7e492d94e0","nextBlock":"b-d805dbac107993e22b0df6a4","prerequisites":[{"id":"b-2e56bc3cb2d67ff0a6a2fd5f","url":"#b-2e56bc3cb2d67ff0a6a2fd5f","role":"section-title"},{"id":"b-53d7b18ac411d732d5c4ef84","url":"#b-53d7b18ac411d732d5c4ef84","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-d805dbac107993e22b0df6a4","type":"h2","scope":"block-a7a8f9c43bc64c6e8609e87186f7cb77","url":"https://codingai-lec03.pages.dev/course/notebooks/b8320404134f4a2e858ccfe7d50a71d3.html#b-d805dbac107993e22b0df6a4","content":"## 为什么损失作用在 z，迁移却常读 h","sha256":"0c4c6ea20d2153b2e1535a2be1939033ad5c45c42167fba81ef0c782b77d60f1","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-reuse-and-targets","title":"训练留下什么，训练答案又从哪里来？"},"page":{"id":"reading-block-a7a8f9c43bc64c6e8609e87186f7cb77","title":"从两张视图到一个可微分的选择题","number":2},"sectionTitle":"从两张视图到一个可微分的选择题","currentQuestion":"从两张视图到一个可微分的选择题","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-4813bc4f52ee0745be8fe68c","nextBlock":"b-5bbe36e223e61eedbbfa9ab2","prerequisites":[{"id":"b-2e56bc3cb2d67ff0a6a2fd5f","url":"#b-2e56bc3cb2d67ff0a6a2fd5f","role":"section-title"},{"id":"b-53d7b18ac411d732d5c4ef84","url":"#b-53d7b18ac411d732d5c4ef84","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-5bbe36e223e61eedbbfa9ab2","type":"p","scope":"block-a7a8f9c43bc64c6e8609e87186f7cb77","url":"https://codingai-lec03.pages.dev/course/notebooks/b8320404134f4a2e858ccfe7d50a71d3.html#b-5bbe36e223e61eedbbfa9ab2","content":"假设 $h=(u,c)$，$u$ 编码形状，$c$ 编码颜色。若增强经常改变颜色，两个视图可能有相同 $u$、不同 $c$。直接要求整个 $h$ 相同，会压制颜色差异；增加投影头以后，$g(h)=u$ 这样的映射可以在 $z$ 中满足不变性，而 $h$ 仍可能留下颜色。这是一个**可行性构造**，不是声称真实网络一定学成两根独立坐标。","sha256":"a659d64a58c06d563b62e4fe28734e91d077998a3700743c08340c069019df20","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-reuse-and-targets","title":"训练留下什么，训练答案又从哪里来？"},"page":{"id":"reading-block-a7a8f9c43bc64c6e8609e87186f7cb77","title":"从两张视图到一个可微分的选择题","number":2},"sectionTitle":"从两张视图到一个可微分的选择题","currentQuestion":"从两张视图到一个可微分的选择题","sourceSlides":[],"knowledgeBlock":null,"formulas":[{"latex":"h=(u,c)","definitionContext":"从两张视图到一个可微分的选择题"},{"latex":"u","definitionContext":"从两张视图到一个可微分的选择题"},{"latex":"c","definitionContext":"从两张视图到一个可微分的选择题"},{"latex":"u","definitionContext":"从两张视图到一个可微分的选择题"},{"latex":"c","definitionContext":"从两张视图到一个可微分的选择题"},{"latex":"h","definitionContext":"从两张视图到一个可微分的选择题"},{"latex":"g(h)=u","definitionContext":"从两张视图到一个可微分的选择题"},{"latex":"z","definitionContext":"从两张视图到一个可微分的选择题"},{"latex":"h","definitionContext":"从两张视图到一个可微分的选择题"}],"previousBlock":"b-d805dbac107993e22b0df6a4","nextBlock":"b-7018ef621ca36ddbedc5bbba","prerequisites":[{"id":"b-d805dbac107993e22b0df6a4","url":"#b-d805dbac107993e22b0df6a4","role":"section-title"},{"id":"b-53d7b18ac411d732d5c4ef84","url":"#b-53d7b18ac411d732d5c4ef84","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-7018ef621ca36ddbedc5bbba","type":"p","scope":"block-a7a8f9c43bc64c6e8609e87186f7cb77","url":"https://codingai-lec03.pages.dev/course/notebooks/b8320404134f4a2e858ccfe7d50a71d3.html#b-7018ef621ca36ddbedc5bbba","content":"损失计算在 $z$ 不代表编码器没学：链式法则给出 $\\partial L/\\partial\\theta=(\\partial L/\\partial z)(\\partial z/\\partial h)(\\partial h/\\partial\\theta)$。梯度穿过投影头更新编码器。SimCLR 通过实验发现非线性投影头有益，并把投影前表示用于下游；这支持区分目标空间与读出空间，不能推广成“h 永远保留全部被增强抹掉的信息”。图像已经被转成灰度时，该视图根本不含原颜色；若所有训练视图都删掉颜色，表示也没有从这些输入恢复真实颜色的依据。","sha256":"542bbbbbf07478a29b46b11c0e054862781ed72fdbffa19e42b223caf08168a2","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-reuse-and-targets","title":"训练留下什么，训练答案又从哪里来？"},"page":{"id":"reading-block-a7a8f9c43bc64c6e8609e87186f7cb77","title":"从两张视图到一个可微分的选择题","number":2},"sectionTitle":"从两张视图到一个可微分的选择题","currentQuestion":"从两张视图到一个可微分的选择题","sourceSlides":[],"knowledgeBlock":null,"formulas":[{"latex":"z","definitionContext":"从两张视图到一个可微分的选择题"},{"latex":"\\partial L/\\partial\\theta=(\\partial L/\\partial z)(\\partial z/\\partial h)(\\partial h/\\partial\\theta)","definitionContext":"从两张视图到一个可微分的选择题"}],"previousBlock":"b-5bbe36e223e61eedbbfa9ab2","nextBlock":"b-55c33e4ba51da7ab55443c01","prerequisites":[{"id":"b-d805dbac107993e22b0df6a4","url":"#b-d805dbac107993e22b0df6a4","role":"section-title"},{"id":"b-53d7b18ac411d732d5c4ef84","url":"#b-53d7b18ac411d732d5c4ef84","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-55c33e4ba51da7ab55443c01","type":"p","scope":"block-a7a8f9c43bc64c6e8609e87186f7cb77","url":"https://codingai-lec03.pages.dev/course/notebooks/b8320404134f4a2e858ccfe7d50a71d3.html#b-55c33e4ba51da7ab55443c01","content":"在 [LAB 06](https://deepnote.com/workspace/ise3309jingao-5ba66b0a-4283-4f5a-ae06-3f68d3201e99/project/38755e4e-23f6-4f5f-8c68-8a2e4114b1c1/notebook/7c4f531113d7473abc174e2e014e1340?utm_source=openai&utm_medium=mcp&utm_campaign=openaimcp&utm_content=7c4f531113d7473abc174e2e014e1340&utm_term=get_notebook) 中，应同时检查颜色/形状任务、增强设置和 $h/z$ 读出。最有解释力的比较是：同一原图划分、相同模型容量与读出预算，只改变颜色增强；再看形状读出和颜色读出是否朝不同方向变化。不能用颜色损失下降代替形状泛化，也不能把随机二维向量当作实际训练出的特征。","sha256":"cdfb9cdcad42224287fc11f2cf2313e29b96100a33841b6a10511d0ba3b8adea","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-reuse-and-targets","title":"训练留下什么，训练答案又从哪里来？"},"page":{"id":"reading-block-a7a8f9c43bc64c6e8609e87186f7cb77","title":"从两张视图到一个可微分的选择题","number":2},"sectionTitle":"从两张视图到一个可微分的选择题","currentQuestion":"从两张视图到一个可微分的选择题","sourceSlides":[],"knowledgeBlock":null,"formulas":[{"latex":"h/z","definitionContext":"从两张视图到一个可微分的选择题"}],"previousBlock":"b-7018ef621ca36ddbedc5bbba","nextBlock":null,"prerequisites":[{"id":"b-d805dbac107993e22b0df6a4","url":"#b-d805dbac107993e22b0df6a4","role":"section-title"},{"id":"b-53d7b18ac411d732d5c4ef84","url":"#b-53d7b18ac411d732d5c4ef84","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-403c6699eab8d0b2e90ca40a","type":"h2","scope":"block-38813722014249e6a8ee5e14b736dca2","url":"https://codingai-lec03.pages.dev/course/notebooks/b8320404134f4a2e858ccfe7d50a71d3.html#b-403c6699eab8d0b2e90ca40a","content":"## 队列解决候选数量，EMA 缓和特征坐标系漂移","sha256":"13ad202bf0f7dee20c986d719ee6941a3db7967c52a2a292c134601c0ef77683","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-reuse-and-targets","title":"训练留下什么，训练答案又从哪里来？"},"page":{"id":"reading-block-38813722014249e6a8ee5e14b736dca2","title":"队列解决候选数量，EMA 缓和特征坐标系漂移","number":3},"sectionTitle":"队列解决候选数量，EMA 缓和特征坐标系漂移","currentQuestion":"队列解决候选数量，EMA 缓和特征坐标系漂移","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":null,"nextBlock":"b-b68fdd134718b535c23f50b8","prerequisites":[{"id":"b-b68fdd134718b535c23f50b8","url":"#b-b68fdd134718b535c23f50b8","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-b68fdd134718b535c23f50b8","type":"p","scope":"block-38813722014249e6a8ee5e14b736dca2","url":"https://codingai-lec03.pages.dev/course/notebooks/b8320404134f4a2e858ccfe7d50a71d3.html#b-b68fdd134718b535c23f50b8","content":"当前 batch 较小时，每个 query 可比较的负例有限。MoCo 把近期 batch 的 key 特征保存为队列，下一批仍可使用，因而字典大小不完全受当前 batch 限制。长度 6 的队列从 `[a,b,c,d,e,f]` 加入 `[g,h]` 后成为 `[c,d,e,f,g,h]`。里面存的是先前编码好的向量；它不保存那些图片的完整计算图，也不在每一步重算所有旧向量。","sha256":"19896d78aeb907e9684aabde5af58ecc2db736f45f0fed7e1cd0a4b8f2daf80f","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-reuse-and-targets","title":"训练留下什么，训练答案又从哪里来？"},"page":{"id":"reading-block-38813722014249e6a8ee5e14b736dca2","title":"队列解决候选数量，EMA 缓和特征坐标系漂移","number":3},"sectionTitle":"队列解决候选数量，EMA 缓和特征坐标系漂移","currentQuestion":"队列解决候选数量，EMA 缓和特征坐标系漂移","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-403c6699eab8d0b2e90ca40a","nextBlock":"b-86464cca8a354b7ebc841835","prerequisites":[{"id":"b-403c6699eab8d0b2e90ca40a","url":"#b-403c6699eab8d0b2e90ca40a","role":"section-title"}],"relatedBlocks":[]},{"id":"b-86464cca8a354b7ebc841835","type":"p","scope":"block-38813722014249e6a8ee5e14b736dca2","url":"https://codingai-lec03.pages.dev/course/notebooks/b8320404134f4a2e858ccfe7d50a71d3.html#b-86464cca8a354b7ebc841835","content":"这节省了候选编码计算，却产生一致性问题。若编码器每步剧烈变化，昨天向量的某个方向和今天向量的同一方向可能不再表达类似内容。MoCo 使用 query 编码器参数 $\\theta_q$ 做正常梯度更新，key 编码器则采用参数指数移动平均：","sha256":"280028e46b956fac640ebb67c8bb7e918caba37ffff27f805b05c236d69a7bea","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-reuse-and-targets","title":"训练留下什么，训练答案又从哪里来？"},"page":{"id":"reading-block-38813722014249e6a8ee5e14b736dca2","title":"队列解决候选数量，EMA 缓和特征坐标系漂移","number":3},"sectionTitle":"队列解决候选数量，EMA 缓和特征坐标系漂移","currentQuestion":"队列解决候选数量，EMA 缓和特征坐标系漂移","sourceSlides":[],"knowledgeBlock":null,"formulas":[{"latex":"\\theta_q","definitionContext":"队列解决候选数量，EMA 缓和特征坐标系漂移"}],"previousBlock":"b-b68fdd134718b535c23f50b8","nextBlock":"b-9ec9f6a6b54160158049a751","prerequisites":[{"id":"b-403c6699eab8d0b2e90ca40a","url":"#b-403c6699eab8d0b2e90ca40a","role":"section-title"},{"id":"b-b68fdd134718b535c23f50b8","url":"#b-b68fdd134718b535c23f50b8","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-9ec9f6a6b54160158049a751","type":"p","scope":"block-38813722014249e6a8ee5e14b736dca2","url":"https://codingai-lec03.pages.dev/course/notebooks/b8320404134f4a2e858ccfe7d50a71d3.html#b-9ec9f6a6b54160158049a751","content":"$\\theta_k\\leftarrow m\\theta_k+(1-m)\\theta_q.$","sha256":"938a7eebf48260181920170d01afbceba7b0c1a0530743e2730361f2751f49a1","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-reuse-and-targets","title":"训练留下什么，训练答案又从哪里来？"},"page":{"id":"reading-block-38813722014249e6a8ee5e14b736dca2","title":"队列解决候选数量，EMA 缓和特征坐标系漂移","number":3},"sectionTitle":"队列解决候选数量，EMA 缓和特征坐标系漂移","currentQuestion":"队列解决候选数量，EMA 缓和特征坐标系漂移","sourceSlides":[],"knowledgeBlock":null,"formulas":[{"latex":"\\theta_k\\leftarrow m\\theta_k+(1-m)\\theta_q.","definitionContext":"队列解决候选数量，EMA 缓和特征坐标系漂移"}],"previousBlock":"b-86464cca8a354b7ebc841835","nextBlock":"b-2cbdb46c591ac509c975e2c0","prerequisites":[{"id":"b-403c6699eab8d0b2e90ca40a","url":"#b-403c6699eab8d0b2e90ca40a","role":"section-title"},{"id":"b-b68fdd134718b535c23f50b8","url":"#b-b68fdd134718b535c23f50b8","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-2cbdb46c591ac509c975e2c0","type":"p","scope":"block-38813722014249e6a8ee5e14b736dca2","url":"https://codingai-lec03.pages.dev/course/notebooks/b8320404134f4a2e858ccfe7d50a71d3.html#b-2cbdb46c591ac509c975e2c0","content":"取 $m=0.9$，旧 $\\theta_k=(0,2)$，当前 $\\theta_q=(2,4)$，新 key 参数是 $(0.2,2.2)$。若下一步 query 仍相同，新 key 为 $(0.38,2.38)$。平均的是**网络参数**，不是队列向量，也不是 Adam 的梯度动量；旧 key 不会因这一步 EMA 自动刷新。$m=0$ 令 key 立即跟随 query，$m=1$ 则把 key 冻结。较慢变化让历史 key 更可比，但不能消除全部陈旧性。[MoCo §3 与 Algorithm 1](https://arxiv.org/html/1911.05722v3) 给出队列、动量编码器和停止 key 分支梯度的配合。","sha256":"cab2d398a639b456c7177b2b3a95d5f2ae1fe4a58c6ac53373a1b26511498cc0","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-reuse-and-targets","title":"训练留下什么，训练答案又从哪里来？"},"page":{"id":"reading-block-38813722014249e6a8ee5e14b736dca2","title":"队列解决候选数量，EMA 缓和特征坐标系漂移","number":3},"sectionTitle":"队列解决候选数量，EMA 缓和特征坐标系漂移","currentQuestion":"队列解决候选数量，EMA 缓和特征坐标系漂移","sourceSlides":[],"knowledgeBlock":null,"formulas":[{"latex":"m=0.9","definitionContext":"队列解决候选数量，EMA 缓和特征坐标系漂移"},{"latex":"\\theta_k=(0,2)","definitionContext":"队列解决候选数量，EMA 缓和特征坐标系漂移"},{"latex":"\\theta_q=(2,4)","definitionContext":"队列解决候选数量，EMA 缓和特征坐标系漂移"},{"latex":"(0.2,2.2)","definitionContext":"队列解决候选数量，EMA 缓和特征坐标系漂移"},{"latex":"(0.38,2.38)","definitionContext":"队列解决候选数量，EMA 缓和特征坐标系漂移"},{"latex":"m=0","definitionContext":"队列解决候选数量，EMA 缓和特征坐标系漂移"},{"latex":"m=1","definitionContext":"队列解决候选数量，EMA 缓和特征坐标系漂移"}],"previousBlock":"b-9ec9f6a6b54160158049a751","nextBlock":"b-d1a33a857de9ee581a2a75fd","prerequisites":[{"id":"b-403c6699eab8d0b2e90ca40a","url":"#b-403c6699eab8d0b2e90ca40a","role":"section-title"},{"id":"b-b68fdd134718b535c23f50b8","url":"#b-b68fdd134718b535c23f50b8","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-d1a33a857de9ee581a2a75fd","type":"table","scope":"block-38813722014249e6a8ee5e14b736dca2","url":"https://codingai-lec03.pages.dev/course/notebooks/b8320404134f4a2e858ccfe7d50a71d3.html#b-d1a33a857de9ee581a2a75fd","content":"| 组织方式 | 候选从哪里来 | 是否保留旧特征 | 主要计算取舍 |\n| --- | --- | --- | --- |\n| SimCLR 批内比较 | 当前 batch 的两视图 | 不需要队列 | 更多候选通常要求更大 batch 及其通信/激活 |\n| MoCo 字典 | 当前正例与历史 key 队列 | 是 | 候选数与 batch 部分解耦，需控制 key 漂移 |","sha256":"62fdb0da604d426cd683fd27eb2e857aad9dcbe43b3ccb9854372ec221ae6d6f","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-reuse-and-targets","title":"训练留下什么，训练答案又从哪里来？"},"page":{"id":"reading-block-38813722014249e6a8ee5e14b736dca2","title":"队列解决候选数量，EMA 缓和特征坐标系漂移","number":3},"sectionTitle":"队列解决候选数量，EMA 缓和特征坐标系漂移","currentQuestion":"队列解决候选数量，EMA 缓和特征坐标系漂移","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"table":{"headers":["组织方式","候选从哪里来","是否保留旧特征","主要计算取舍"],"rows":[["组织方式","候选从哪里来","是否保留旧特征","主要计算取舍"],["SimCLR 批内比较","当前 batch 的两视图","不需要队列","更多候选通常要求更大 batch 及其通信/激活"],["MoCo 字典","当前正例与历史 key 队列","是","候选数与 batch 部分解耦，需控制 key 漂移"]]},"previousBlock":"b-2cbdb46c591ac509c975e2c0","nextBlock":"b-c4033a2d5ae56a273bb47e48","prerequisites":[{"id":"b-403c6699eab8d0b2e90ca40a","url":"#b-403c6699eab8d0b2e90ca40a","role":"section-title"},{"id":"b-b68fdd134718b535c23f50b8","url":"#b-b68fdd134718b535c23f50b8","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-c4033a2d5ae56a273bb47e48","type":"p","scope":"block-38813722014249e6a8ee5e14b736dca2","url":"https://codingai-lec03.pages.dev/course/notebooks/b8320404134f4a2e858ccfe7d50a71d3.html#b-c4033a2d5ae56a273bb47e48","content":"这是计算组织的不同答案，不是一个方法必然取代另一个。扩大队列会让最旧特征更老；增大 batch 会改变优化与负例分布；调整温度则改变同一批候选如何竞争。做消融时不能三项同时改，再把收益全部归给 EMA。","sha256":"2ba3b260bb0604853faa2a1daad1083b7ec067c1949a2c935ab2a6c3b2be1839","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-reuse-and-targets","title":"训练留下什么，训练答案又从哪里来？"},"page":{"id":"reading-block-38813722014249e6a8ee5e14b736dca2","title":"队列解决候选数量，EMA 缓和特征坐标系漂移","number":3},"sectionTitle":"队列解决候选数量，EMA 缓和特征坐标系漂移","currentQuestion":"队列解决候选数量，EMA 缓和特征坐标系漂移","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-d1a33a857de9ee581a2a75fd","nextBlock":"b-0aa71c35431a5068d5c9120a","prerequisites":[{"id":"b-403c6699eab8d0b2e90ca40a","url":"#b-403c6699eab8d0b2e90ca40a","role":"section-title"},{"id":"b-b68fdd134718b535c23f50b8","url":"#b-b68fdd134718b535c23f50b8","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-0aa71c35431a5068d5c9120a","type":"h2","scope":"block-38813722014249e6a8ee5e14b736dca2","url":"https://codingai-lec03.pages.dev/course/notebooks/b8320404134f4a2e858ccfe7d50a71d3.html#b-0aa71c35431a5068d5c9120a","content":"## 训练学会区分视图以后，还要回答哪个下游问题","sha256":"b69c0c9f503f70eec34c1c4211e649acec947654b830cc37772f274ee7e290a0","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-reuse-and-targets","title":"训练留下什么，训练答案又从哪里来？"},"page":{"id":"reading-block-38813722014249e6a8ee5e14b736dca2","title":"队列解决候选数量，EMA 缓和特征坐标系漂移","number":3},"sectionTitle":"队列解决候选数量，EMA 缓和特征坐标系漂移","currentQuestion":"队列解决候选数量，EMA 缓和特征坐标系漂移","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-c4033a2d5ae56a273bb47e48","nextBlock":"b-9e80eff8ee0f98dbbcfd3b7d","prerequisites":[{"id":"b-403c6699eab8d0b2e90ca40a","url":"#b-403c6699eab8d0b2e90ca40a","role":"section-title"},{"id":"b-b68fdd134718b535c23f50b8","url":"#b-b68fdd134718b535c23f50b8","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-9e80eff8ee0f98dbbcfd3b7d","type":"p","scope":"block-38813722014249e6a8ee5e14b736dca2","url":"https://codingai-lec03.pages.dev/course/notebooks/b8320404134f4a2e858ccfe7d50a71d3.html#b-9e80eff8ee0f98dbbcfd3b7d","content":"冻结线性评价（linear probe）先固定编码器 $f$，仅训练 $\\hat y=Wh+b$；它检验标签能否用指定线性读出从表示取得。kNN 不更新编码器，用训练集表示及标签给测试点投票，结果仍依赖距离、归一化、$k$ 与支持集。微调允许 $f$ 更新，检验的是初始化加适配方案。检测或分割则需要空间特征和任务头，整图线性分类成绩不能自动给出边界能力。","sha256":"912e990e14a7ac966b4697be033887a98e04c708b08a85be668341b6e3b43a3e","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-reuse-and-targets","title":"训练留下什么，训练答案又从哪里来？"},"page":{"id":"reading-block-38813722014249e6a8ee5e14b736dca2","title":"队列解决候选数量，EMA 缓和特征坐标系漂移","number":3},"sectionTitle":"队列解决候选数量，EMA 缓和特征坐标系漂移","currentQuestion":"队列解决候选数量，EMA 缓和特征坐标系漂移","sourceSlides":[],"knowledgeBlock":null,"formulas":[{"latex":"f","definitionContext":"队列解决候选数量，EMA 缓和特征坐标系漂移"},{"latex":"\\hat y=Wh+b","definitionContext":"队列解决候选数量，EMA 缓和特征坐标系漂移"},{"latex":"k","definitionContext":"队列解决候选数量，EMA 缓和特征坐标系漂移"},{"latex":"f","definitionContext":"队列解决候选数量，EMA 缓和特征坐标系漂移"}],"previousBlock":"b-0aa71c35431a5068d5c9120a","nextBlock":"b-ae0fbac703919117e0856417","prerequisites":[{"id":"b-0aa71c35431a5068d5c9120a","url":"#b-0aa71c35431a5068d5c9120a","role":"section-title"},{"id":"b-b68fdd134718b535c23f50b8","url":"#b-b68fdd134718b535c23f50b8","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-ae0fbac703919117e0856417","type":"p","scope":"block-38813722014249e6a8ee5e14b736dca2","url":"https://codingai-lec03.pages.dev/course/notebooks/b8320404134f4a2e858ccfe7d50a71d3.html#b-ae0fbac703919117e0856417","content":"完整协议至少明确：原始样本的训练/验证/测试划分，预训练是否接触无标签测试图，冻结时的 `eval`/BatchNorm 状态，读 $h$ 还是 $z$，标签预算，验证集选超参数的方法。两视图必须在原图划分之后生成。同一原图一张裁剪放训练、另一张放测试，会把近重复泄漏包装成泛化。","sha256":"da4b9f90252f4b39b9e979ccc07bb322ee9d5be1cc08040b3f9e718266e5d680","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-reuse-and-targets","title":"训练留下什么，训练答案又从哪里来？"},"page":{"id":"reading-block-38813722014249e6a8ee5e14b736dca2","title":"队列解决候选数量，EMA 缓和特征坐标系漂移","number":3},"sectionTitle":"队列解决候选数量，EMA 缓和特征坐标系漂移","currentQuestion":"队列解决候选数量，EMA 缓和特征坐标系漂移","sourceSlides":[],"knowledgeBlock":null,"formulas":[{"latex":"h","definitionContext":"队列解决候选数量，EMA 缓和特征坐标系漂移"},{"latex":"z","definitionContext":"队列解决候选数量，EMA 缓和特征坐标系漂移"}],"previousBlock":"b-9e80eff8ee0f98dbbcfd3b7d","nextBlock":"b-a1e0c702fbbf81588a8bb5d6","prerequisites":[{"id":"b-0aa71c35431a5068d5c9120a","url":"#b-0aa71c35431a5068d5c9120a","role":"section-title"},{"id":"b-b68fdd134718b535c23f50b8","url":"#b-b68fdd134718b535c23f50b8","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-a1e0c702fbbf81588a8bb5d6","type":"h3","scope":"block-38813722014249e6a8ee5e14b736dca2","url":"https://codingai-lec03.pages.dev/course/notebooks/b8320404134f4a2e858ccfe7d50a71d3.html#b-a1e0c702fbbf81588a8bb5d6","content":"### 同一个颜色干预，真的改变了学到的表示","sha256":"ca5d95bc4159403ab2f762e0281e0868f2600b5e08854fece5f940b1569dff0a","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-reuse-and-targets","title":"训练留下什么，训练答案又从哪里来？"},"page":{"id":"reading-block-38813722014249e6a8ee5e14b736dca2","title":"队列解决候选数量，EMA 缓和特征坐标系漂移","number":3},"sectionTitle":"队列解决候选数量，EMA 缓和特征坐标系漂移","currentQuestion":"队列解决候选数量，EMA 缓和特征坐标系漂移","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-ae0fbac703919117e0856417","nextBlock":"b-06f246687dc35899e829575c","prerequisites":[{"id":"b-0aa71c35431a5068d5c9120a","url":"#b-0aa71c35431a5068d5c9120a","role":"section-title"},{"id":"b-b68fdd134718b535c23f50b8","url":"#b-b68fdd134718b535c23f50b8","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-06f246687dc35899e829575c","type":"p","scope":"block-38813722014249e6a8ee5e14b736dca2","url":"https://codingai-lec03.pages.dev/course/notebooks/b8320404134f4a2e858ccfe7d50a71d3.html#b-06f246687dc35899e829575c","content":"LAB 06 已在 2026-09-22 本地 CPU 完成真实两视图训练：600 张训练图、240 张保留图，RGB 16×16，卷积编码器输出 48 维 $h$，投影头输出 24 维 $z$，InfoNCE 温度 0.2，batch 60，Adam 学习率 0.002，固定种子 20260922、40 epoch。预训练不使用类别标签；标签只供冻结表示的 5NN 与线性读出使用。对照唯一指定变化是每个视图是否以 50% 概率交换红/蓝通道。","sha256":"68033e34c77ec9b226141fdb781f4d1793b69e146bd6834e9c7a1c665624bdfc","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-reuse-and-targets","title":"训练留下什么，训练答案又从哪里来？"},"page":{"id":"reading-block-38813722014249e6a8ee5e14b736dca2","title":"队列解决候选数量，EMA 缓和特征坐标系漂移","number":3},"sectionTitle":"队列解决候选数量，EMA 缓和特征坐标系漂移","currentQuestion":"队列解决候选数量，EMA 缓和特征坐标系漂移","sourceSlides":[],"knowledgeBlock":null,"formulas":[{"latex":"h","definitionContext":"队列解决候选数量，EMA 缓和特征坐标系漂移"},{"latex":"z","definitionContext":"队列解决候选数量，EMA 缓和特征坐标系漂移"}],"previousBlock":"b-a1e0c702fbbf81588a8bb5d6","nextBlock":"b-1d599c93d7dcd4a8696e65b4","prerequisites":[{"id":"b-a1e0c702fbbf81588a8bb5d6","url":"#b-a1e0c702fbbf81588a8bb5d6","role":"section-title"},{"id":"b-b68fdd134718b535c23f50b8","url":"#b-b68fdd134718b535c23f50b8","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-1d599c93d7dcd4a8696e65b4","type":"table","scope":"block-38813722014249e6a8ee5e14b736dca2","url":"https://codingai-lec03.pages.dev/course/notebooks/b8320404134f4a2e858ccfe7d50a71d3.html#b-1d599c93d7dcd4a8696e65b4","content":"| 表示与训练 | 形状 5NN | 颜色 5NN |\n| --- | --- | --- |\n| 随机编码器 $h$ | 72.50% | 100.00% |\n| 保留颜色增强，训练后 $h$ | 47.50% | 100.00% |\n| 随机换色增强，训练后 $h$ | 85.83% | 49.17% |\n| 随机换色增强，训练后 $z$ | 95.42% | 52.92% |","sha256":"7f4f718dcaa14fddb0c1dad9832eb9368cf758d6881a4f1887b3ca4f102fe033","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-reuse-and-targets","title":"训练留下什么，训练答案又从哪里来？"},"page":{"id":"reading-block-38813722014249e6a8ee5e14b736dca2","title":"队列解决候选数量，EMA 缓和特征坐标系漂移","number":3},"sectionTitle":"队列解决候选数量，EMA 缓和特征坐标系漂移","currentQuestion":"队列解决候选数量，EMA 缓和特征坐标系漂移","sourceSlides":[],"knowledgeBlock":null,"formulas":[{"latex":"h","definitionContext":"队列解决候选数量，EMA 缓和特征坐标系漂移"},{"latex":"h","definitionContext":"队列解决候选数量，EMA 缓和特征坐标系漂移"},{"latex":"h","definitionContext":"队列解决候选数量，EMA 缓和特征坐标系漂移"},{"latex":"z","definitionContext":"队列解决候选数量，EMA 缓和特征坐标系漂移"}],"table":{"headers":["表示与训练","形状 5NN","颜色 5NN"],"rows":[["表示与训练","形状 5NN","颜色 5NN"],["随机编码器 hhh","72.50%","100.00%"],["保留颜色增强，训练后 hhh","47.50%","100.00%"],["随机换色增强，训练后 hhh","85.83%","49.17%"],["随机换色增强，训练后 zzz","95.42%","52.92%"]]},"previousBlock":"b-06f246687dc35899e829575c","nextBlock":"b-ac6146de7b07c5d432d26167","prerequisites":[{"id":"b-a1e0c702fbbf81588a8bb5d6","url":"#b-a1e0c702fbbf81588a8bb5d6","role":"section-title"},{"id":"b-b68fdd134718b535c23f50b8","url":"#b-b68fdd134718b535c23f50b8","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-ac6146de7b07c5d432d26167","type":"p","scope":"block-38813722014249e6a8ee5e14b736dca2","url":"https://codingai-lec03.pages.dev/course/notebooks/b8320404134f4a2e858ccfe7d50a71d3.html#b-ac6146de7b07c5d432d26167","content":"![真实两视图训练与冻结读出](/assets/display/f482e1157acae587ec308bdd.webp)","sha256":"5e7178080f9b3e931941abe1b04337d20f4b1160cbc39c4f37f0038310ef3636","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-reuse-and-targets","title":"训练留下什么，训练答案又从哪里来？"},"page":{"id":"reading-block-38813722014249e6a8ee5e14b736dca2","title":"队列解决候选数量，EMA 缓和特征坐标系漂移","number":3},"sectionTitle":"队列解决候选数量，EMA 缓和特征坐标系漂移","currentQuestion":"队列解决候选数量，EMA 缓和特征坐标系漂移","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-1d599c93d7dcd4a8696e65b4","nextBlock":"b-1eda3089913c5ddfaa70d616","prerequisites":[{"id":"b-a1e0c702fbbf81588a8bb5d6","url":"#b-a1e0c702fbbf81588a8bb5d6","role":"section-title"},{"id":"b-b68fdd134718b535c23f50b8","url":"#b-b68fdd134718b535c23f50b8","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-1eda3089913c5ddfaa70d616","type":"p","scope":"block-38813722014249e6a8ee5e14b736dca2","url":"https://codingai-lec03.pages.dev/course/notebooks/b8320404134f4a2e858ccfe7d50a71d3.html#b-1eda3089913c5ddfaa70d616","content":"保留颜色时模型可利用容易的颜色区别，形状 5NN 甚至低于随机表示；换色使这种线索不可靠，形状读出提高、颜色读出接近二类机会水平。**$h$ 也会丢失颜色，$h$ 优于 $z$ 不是保证**：本次 $z$ 的形状 5NN 反而更高。同一换色 $h$ 的线性形状读出是 98.75%，再次说明“表示质量”不能脱离读出协议。这里报告的是一个固定种子的小型受控实验，不是 SimCLR 在自然图像上的通用排名。","sha256":"f61a601120cd0a20dd289ba719d67eb1b818f33fab4cf87fd90b2bd3092c4e6d","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-reuse-and-targets","title":"训练留下什么，训练答案又从哪里来？"},"page":{"id":"reading-block-38813722014249e6a8ee5e14b736dca2","title":"队列解决候选数量，EMA 缓和特征坐标系漂移","number":3},"sectionTitle":"队列解决候选数量，EMA 缓和特征坐标系漂移","currentQuestion":"队列解决候选数量，EMA 缓和特征坐标系漂移","sourceSlides":[],"knowledgeBlock":null,"formulas":[{"latex":"h","definitionContext":"队列解决候选数量，EMA 缓和特征坐标系漂移"},{"latex":"h","definitionContext":"队列解决候选数量，EMA 缓和特征坐标系漂移"},{"latex":"z","definitionContext":"队列解决候选数量，EMA 缓和特征坐标系漂移"},{"latex":"z","definitionContext":"队列解决候选数量，EMA 缓和特征坐标系漂移"},{"latex":"h","definitionContext":"队列解决候选数量，EMA 缓和特征坐标系漂移"}],"previousBlock":"b-ac6146de7b07c5d432d26167","nextBlock":"b-27ed21128af52f1ddbb663e0","prerequisites":[{"id":"b-a1e0c702fbbf81588a8bb5d6","url":"#b-a1e0c702fbbf81588a8bb5d6","role":"section-title"},{"id":"b-b68fdd134718b535c23f50b8","url":"#b-b68fdd134718b535c23f50b8","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-27ed21128af52f1ddbb663e0","type":"p","scope":"block-38813722014249e6a8ee5e14b736dca2","url":"https://codingai-lec03.pages.dev/course/notebooks/b8320404134f4a2e858ccfe7d50a71d3.html#b-27ed21128af52f1ddbb663e0","content":"LAB 07 的小型 MoCo 使用同一图像来源、真实 FIFO 特征队列与 EMA key 编码器，训练 15 epoch。其 $h$ 形状 5NN 为 38.33%，低于随机基线 72.50%；损失由约 4.79 变到 4.27，也没有保证目标语义变得好读。它完成了训练机制，但给出了一个失败的表示结果。MoCo 与 LAB 06 的训练轮数和组织不同，不能据这些数作方法优劣排名；可以据此检查为什么视图目标、候选和下游形状任务不一致。","sha256":"5c898a84901cad07894ad357a6d3e79e5f976c45f22fc65e28c9298a525616f8","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-reuse-and-targets","title":"训练留下什么，训练答案又从哪里来？"},"page":{"id":"reading-block-38813722014249e6a8ee5e14b736dca2","title":"队列解决候选数量，EMA 缓和特征坐标系漂移","number":3},"sectionTitle":"队列解决候选数量，EMA 缓和特征坐标系漂移","currentQuestion":"队列解决候选数量，EMA 缓和特征坐标系漂移","sourceSlides":[],"knowledgeBlock":null,"formulas":[{"latex":"h","definitionContext":"队列解决候选数量，EMA 缓和特征坐标系漂移"}],"previousBlock":"b-1eda3089913c5ddfaa70d616","nextBlock":"b-5f66df60ad49c4f23d9bc64c","prerequisites":[{"id":"b-a1e0c702fbbf81588a8bb5d6","url":"#b-a1e0c702fbbf81588a8bb5d6","role":"section-title"},{"id":"b-b68fdd134718b535c23f50b8","url":"#b-b68fdd134718b535c23f50b8","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-5f66df60ad49c4f23d9bc64c","type":"p","scope":"block-38813722014249e6a8ee5e14b736dca2","url":"https://codingai-lec03.pages.dev/course/notebooks/b8320404134f4a2e858ccfe7d50a71d3.html#b-5f66df60ad49c4f23d9bc64c","content":"[LAB 07](https://deepnote.com/project/38755e4e-23f6-4f5f-8c68-8a2e4114b1c1/notebook/0d1b8f6fbf6a476c93ffcc169f2ec1f4) 把队列/EMA 精确算例和图像两视图训练放在同一条学习链中。阅读训练输出时，先确认参数确实由损失更新，再比较相同协议下随机初始化与训练后表示；若小数据上的自监督结果没有领先，就讨论增强、数据量和读出限制，而不改变测试口径。","sha256":"256c571e4cd5bafbfd1e4cccd709b0a6104cfe82b582cc656501b4e5eb015b48","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-reuse-and-targets","title":"训练留下什么，训练答案又从哪里来？"},"page":{"id":"reading-block-38813722014249e6a8ee5e14b736dca2","title":"队列解决候选数量，EMA 缓和特征坐标系漂移","number":3},"sectionTitle":"队列解决候选数量，EMA 缓和特征坐标系漂移","currentQuestion":"队列解决候选数量，EMA 缓和特征坐标系漂移","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-27ed21128af52f1ddbb663e0","nextBlock":"b-f2f3b022038383fa2598557f","prerequisites":[{"id":"b-a1e0c702fbbf81588a8bb5d6","url":"#b-a1e0c702fbbf81588a8bb5d6","role":"section-title"},{"id":"b-b68fdd134718b535c23f50b8","url":"#b-b68fdd134718b535c23f50b8","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-f2f3b022038383fa2598557f","type":"p","scope":"block-38813722014249e6a8ee5e14b736dca2","url":"https://codingai-lec03.pages.dev/course/notebooks/b8320404134f4a2e858ccfe7d50a71d3.html#b-f2f3b022038383fa2598557f","content":"从 [CLS 04 的重建](https://deepnote.com/project/38755e4e-23f6-4f5f-8c68-8a2e4114b1c1/notebook/23b0affc4b824723927bd0938fd2bbda) 到本册，目标从恢复像素变为识别视图关系；下一册 [CLS 06](https://deepnote.com/project/38755e4e-23f6-4f5f-8c68-8a2e4114b1c1/notebook/c0645a4db6e04188ab5b119e87c53bed) 再问：没有显式负例时，为什么表示没有全部变成同一个常量？这些路线共同要求我们区分训练信号、实际优化以及表示的使用证据。","sha256":"c7be9efa1d234e046121cf806e9454466d3fa4a53b8ba61bd6876e2ae7fedb44","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-reuse-and-targets","title":"训练留下什么，训练答案又从哪里来？"},"page":{"id":"reading-block-38813722014249e6a8ee5e14b736dca2","title":"队列解决候选数量，EMA 缓和特征坐标系漂移","number":3},"sectionTitle":"队列解决候选数量，EMA 缓和特征坐标系漂移","currentQuestion":"队列解决候选数量，EMA 缓和特征坐标系漂移","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-5f66df60ad49c4f23d9bc64c","nextBlock":null,"prerequisites":[{"id":"b-a1e0c702fbbf81588a8bb5d6","url":"#b-a1e0c702fbbf81588a8bb5d6","role":"section-title"},{"id":"b-b68fdd134718b535c23f50b8","url":"#b-b68fdd134718b535c23f50b8","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-1766f39023d331d0dd08d954","type":"h2","scope":"block-aa9100c6e89d426aa1200f6c622cd2fb","url":"https://codingai-lec03.pages.dev/course/notebooks/b8320404134f4a2e858ccfe7d50a71d3.html#b-1766f39023d331d0dd08d954","content":"## 研究者为什么有理由继续沿这条路线走？","sha256":"d39149b53db2953a6ef0c1687755054b5e8519db491258e72ff8994161416a28","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-reuse-and-targets","title":"训练留下什么，训练答案又从哪里来？"},"page":{"id":"reading-block-aa9100c6e89d426aa1200f6c622cd2fb","title":"研究者为什么有理由继续沿这条路线走？","number":4},"sectionTitle":"研究者为什么有理由继续沿这条路线走？","currentQuestion":"研究者为什么有理由继续沿这条路线走？","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":null,"nextBlock":"b-8d6f2eb248eb5c683fd6d926","prerequisites":[{"id":"b-8d6f2eb248eb5c683fd6d926","url":"#b-8d6f2eb248eb5c683fd6d926","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-8d6f2eb248eb5c683fd6d926","type":"p","scope":"block-aa9100c6e89d426aa1200f6c622cd2fb","url":"https://codingai-lec03.pages.dev/course/notebooks/b8320404134f4a2e858ccfe7d50a71d3.html#b-8d6f2eb248eb5c683fd6d926","content":"MoCo原论文Table 2(b)在同一Faster R-CNN、ResNet-50-C4设置下比较预训练起点：","sha256":"38a51d52dbda974eaf4fab05b73bdc5677b90f775950b000d2ec7cec0518fdc2","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-reuse-and-targets","title":"训练留下什么，训练答案又从哪里来？"},"page":{"id":"reading-block-aa9100c6e89d426aa1200f6c622cd2fb","title":"研究者为什么有理由继续沿这条路线走？","number":4},"sectionTitle":"研究者为什么有理由继续沿这条路线走？","currentQuestion":"研究者为什么有理由继续沿这条路线走？","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-1766f39023d331d0dd08d954","nextBlock":"b-702cd32003c0d46a52f0b525","prerequisites":[{"id":"b-1766f39023d331d0dd08d954","url":"#b-1766f39023d331d0dd08d954","role":"section-title"}],"relatedBlocks":[]},{"id":"b-702cd32003c0d46a52f0b525","type":"table","scope":"block-aa9100c6e89d426aa1200f6c622cd2fb","url":"https://codingai-lec03.pages.dev/course/notebooks/b8320404134f4a2e858ccfe7d50a71d3.html#b-702cd32003c0d46a52f0b525","content":"| 预训练 | box AP | AP50 | AP75 |\n| --- | --- | --- | --- |\n| ImageNet-1M 人工类别监督 | 53.5 | 81.3 | 58.8 |\n| ImageNet-1M MoCo | 55.9 | 81.5 | 62.6 |","sha256":"05be0d423c5b39f2132c94108a586c9d62c94be605786e964a2ae2455d16feca","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-reuse-and-targets","title":"训练留下什么，训练答案又从哪里来？"},"page":{"id":"reading-block-aa9100c6e89d426aa1200f6c622cd2fb","title":"研究者为什么有理由继续沿这条路线走？","number":4},"sectionTitle":"研究者为什么有理由继续沿这条路线走？","currentQuestion":"研究者为什么有理由继续沿这条路线走？","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"table":{"headers":["预训练","box AP","AP50","AP75"],"rows":[["预训练","box AP","AP50","AP75"],["ImageNet-1M 人工类别监督","53.5","81.3","58.8"],["ImageNet-1M MoCo","55.9","81.5","62.6"]]},"previousBlock":"b-8d6f2eb248eb5c683fd6d926","nextBlock":"b-7013143d12306fa3c1d69dea","prerequisites":[{"id":"b-1766f39023d331d0dd08d954","url":"#b-1766f39023d331d0dd08d954","role":"section-title"},{"id":"b-8d6f2eb248eb5c683fd6d926","url":"#b-8d6f2eb248eb5c683fd6d926","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-7013143d12306fa3c1d69dea","type":"p","scope":"block-aa9100c6e89d426aa1200f6c622cd2fb","url":"https://codingai-lec03.pages.dev/course/notebooks/b8320404134f4a2e858ccfe7d50a71d3.html#b-7013143d12306fa3c1d69dea","content":"两者均在VOC trainval07+12检测标注上全层微调24k次，评价为VOC test2007，取5次实验平均；AP是COCO式多IoU阈值指标，AP50/AP75是指定阈值。MoCo相对提高2.4 AP，说明视图对比学出的参数在这套下游协议下提供了有效起点。","sha256":"e7b926bf7cbc1fa25056b061f25eefaf0bba37b86c7983c3d39257ed11e0a513","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-reuse-and-targets","title":"训练留下什么，训练答案又从哪里来？"},"page":{"id":"reading-block-aa9100c6e89d426aa1200f6c622cd2fb","title":"研究者为什么有理由继续沿这条路线走？","number":4},"sectionTitle":"研究者为什么有理由继续沿这条路线走？","currentQuestion":"研究者为什么有理由继续沿这条路线走？","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-702cd32003c0d46a52f0b525","nextBlock":"b-b41b5f520fb227e149782ef4","prerequisites":[{"id":"b-1766f39023d331d0dd08d954","url":"#b-1766f39023d331d0dd08d954","role":"section-title"},{"id":"b-8d6f2eb248eb5c683fd6d926","url":"#b-8d6f2eb248eb5c683fd6d926","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-b41b5f520fb227e149782ef4","type":"p","scope":"block-aa9100c6e89d426aa1200f6c622cd2fb","url":"https://codingai-lec03.pages.dev/course/notebooks/b8320404134f4a2e858ccfe7d50a71d3.html#b-b41b5f520fb227e149782ef4","content":"这不是冻结模型直接会检测：检测头与主干都用检测监督适配。它也不把所有预训练和下游总算力配平。相同论文的另一种主干结构收益不同，因此结论应保留具体模型与协议。","sha256":"53fd3e352a51589d7970016133506a18b796dbfa05c222caa0740be98bbf51d8","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-reuse-and-targets","title":"训练留下什么，训练答案又从哪里来？"},"page":{"id":"reading-block-aa9100c6e89d426aa1200f6c622cd2fb","title":"研究者为什么有理由继续沿这条路线走？","number":4},"sectionTitle":"研究者为什么有理由继续沿这条路线走？","currentQuestion":"研究者为什么有理由继续沿这条路线走？","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-7013143d12306fa3c1d69dea","nextBlock":"b-1a60883338f330ccbf47da8e","prerequisites":[{"id":"b-1766f39023d331d0dd08d954","url":"#b-1766f39023d331d0dd08d954","role":"section-title"},{"id":"b-8d6f2eb248eb5c683fd6d926","url":"#b-8d6f2eb248eb5c683fd6d926","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-1a60883338f330ccbf47da8e","type":"p","scope":"block-aa9100c6e89d426aa1200f6c622cd2fb","url":"https://codingai-lec03.pages.dev/course/notebooks/b8320404134f4a2e858ccfe7d50a71d3.html#b-1a60883338f330ccbf47da8e","content":"这份研究证据和上面的LAB07失败结果并不冲突：前者是论文规模、数据与下游协议下的能力进展；后者是小型教学实现的某次表示结果。机制跑通和用途改善应分别判断。","sha256":"421f861c3e0e6677e1ffead65526d273d60d6a0af426f690c55cf06f485ee332","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-reuse-and-targets","title":"训练留下什么，训练答案又从哪里来？"},"page":{"id":"reading-block-aa9100c6e89d426aa1200f6c622cd2fb","title":"研究者为什么有理由继续沿这条路线走？","number":4},"sectionTitle":"研究者为什么有理由继续沿这条路线走？","currentQuestion":"研究者为什么有理由继续沿这条路线走？","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-b41b5f520fb227e149782ef4","nextBlock":"b-2c43a24dcbff366e21ce289d","prerequisites":[{"id":"b-1766f39023d331d0dd08d954","url":"#b-1766f39023d331d0dd08d954","role":"section-title"},{"id":"b-8d6f2eb248eb5c683fd6d926","url":"#b-8d6f2eb248eb5c683fd6d926","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-2c43a24dcbff366e21ce289d","type":"p","scope":"block-aa9100c6e89d426aa1200f6c622cd2fb","url":"https://codingai-lec03.pages.dev/course/notebooks/b8320404134f4a2e858ccfe7d50a71d3.html#b-2c43a24dcbff366e21ce289d","content":"[把一次学习走通，也读清实现顺序](/course/blocks/moco-one-step-paper-and-code.html) 对照本页图解与LAB07的步内时序；[模拟问答：旧 key 还能比较吗？](/course/blocks/moco-point-at-the-figure-dialogue.html) 提供可指认key身份的模拟问答。第一次动手只选 [LAB 06 · 颜色干预：视图、投影与读出](https://deepnote.com/workspace/ise3309jingao-5ba66b0a-4283-4f5a-ae06-3f68d3201e99/project/38755e4e-23f6-4f5f-8c68-8a2e4114b1c1/notebook/7c4f531113d7473abc174e2e014e1340?utm_source=openai&utm_medium=mcp&utm_campaign=openaimcp&utm_content=7c4f531113d7473abc174e2e014e1340&utm_term=get_notebook) 的颜色干预或 [LAB 07 · MoCo：历史 key、队列与两种更新](https://deepnote.com/workspace/ise3309jingao-5ba66b0a-4283-4f5a-ae06-3f68d3201e99/project/38755e4e-23f6-4f5f-8c68-8a2e4114b1c1/notebook/0d1b8f6fbf6a476c93ffcc169f2ec1f4?utm_source=openai&utm_medium=mcp&utm_campaign=openaimcp&utm_content=0d1b8f6fbf6a476c93ffcc169f2ec1f4&utm_term=get_notebook) 的队列路径之一，不要求两本重复训练。","sha256":"c3e6c2e2fbe739dd93a965244fb0eff35c8e7734942874e27a0dff60f535eceb","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-reuse-and-targets","title":"训练留下什么，训练答案又从哪里来？"},"page":{"id":"reading-block-aa9100c6e89d426aa1200f6c622cd2fb","title":"研究者为什么有理由继续沿这条路线走？","number":4},"sectionTitle":"研究者为什么有理由继续沿这条路线走？","currentQuestion":"研究者为什么有理由继续沿这条路线走？","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-1a60883338f330ccbf47da8e","nextBlock":null,"prerequisites":[{"id":"b-1766f39023d331d0dd08d954","url":"#b-1766f39023d331d0dd08d954","role":"section-title"},{"id":"b-8d6f2eb248eb5c683fd6d926","url":"#b-8d6f2eb248eb5c683fd6d926","role":"problem-context"}],"relatedBlocks":[]}]}