{"version":2,"relationshipURLBase":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p01.html","relationshipResolution":"Resolve relative relationship URLs against relationshipURLBase; complete records and absolute URLs are in blocks.","title":"训练结束后，网络留下了什么？","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p01.html","markdown":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p01.md","pages":[{"id":"h-i-parameters-and-response","title":"训练保存参数，输入才产生表示","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}],"aliases":[{"oldId":"explanation","targetId":"h-i-parameters-and-response","operation":"split-explanation-range"},{"oldId":"b-98557719d66c29f742fd97ae","targetId":"b-98557719d66c29f742fd97ae","page":"h-i-parameters-and-response","operation":"move","beforeSHA256":"d71cf15e707c8a3674b4c90a599efee783cb84da5b8c8236a1d01c3add7b4958","afterSHA256":"d71cf15e707c8a3674b4c90a599efee783cb84da5b8c8236a1d01c3add7b4958","reason":"Same source body in authored semantic page"},{"oldId":"b-0d3ac5570914febfca13703a","targetId":"b-0d3ac5570914febfca13703a","page":"h-i-parameters-and-response","operation":"move","beforeSHA256":"c5e0d1a427d4be3e028ae79de9968a58185de4e8ae0ad9ddea8452875b9f86bd","afterSHA256":"c5e0d1a427d4be3e028ae79de9968a58185de4e8ae0ad9ddea8452875b9f86bd","reason":"Same source body in authored semantic page"},{"oldId":"b-74178ee773393a35124115fc","targetId":"b-74178ee773393a35124115fc","page":"h-i-parameters-and-response","operation":"move","beforeSHA256":"5a4ecd6807364900e549d8da58911f6fefa0d1bdf7274e83e148eb19208b83dc","afterSHA256":"5a4ecd6807364900e549d8da58911f6fefa0d1bdf7274e83e148eb19208b83dc","reason":"Same source body in authored semantic page"},{"oldId":"b-e4fe6d59582b1c840d4ff9a2","targetId":"b-e4fe6d59582b1c840d4ff9a2","page":"h-i-parameters-and-response","operation":"rewrite","beforeSHA256":"402c23f9efa51feea79aa46d26d680f3a48b5804add370c763f4cabc831692eb","afterSHA256":"9b3416b515f8c1db965fb6797f493ef209f4f394b5a309e4e3d04bb30083a257","reason":"Same source identity with figure semantics or reading context enriched"},{"oldId":"b-55476ebb13258f494524a47d","targetId":"b-55476ebb13258f494524a47d","page":"h-i-parameters-and-response","operation":"move","beforeSHA256":"786ce10a6b32a492259a3dccd0a54e61cd41d28a34382fb7b5386c5f03b7ad69","afterSHA256":"786ce10a6b32a492259a3dccd0a54e61cd41d28a34382fb7b5386c5f03b7ad69","reason":"Same source body in authored semantic page"},{"oldId":"b-8562e8e9a534c6c6989cc307","targetId":"b-8562e8e9a534c6c6989cc307","page":"h-i-parameters-and-response","operation":"move","beforeSHA256":"1fe45001cb450b4ebda1020fb108e200834cbfe4f64f1df6b6f1b538cfcda905","afterSHA256":"1fe45001cb450b4ebda1020fb108e200834cbfe4f64f1df6b6f1b538cfcda905","reason":"Same source body in authored semantic page"},{"oldId":"b-ce9fd754022f56966002a96d","targetId":"b-ce9fd754022f56966002a96d","page":"h-i-parameters-and-response","operation":"move","beforeSHA256":"f749d7bbdbe8fc304b346ba2a2c67e965ddab66bcee94cec964566a00d2375fb","afterSHA256":"f749d7bbdbe8fc304b346ba2a2c67e965ddab66bcee94cec964566a00d2375fb","reason":"Same source body in authored semantic page"},{"oldId":"b-dcc0c92d0df5e100f443c08d","targetId":"b-dcc0c92d0df5e100f443c08d","page":"h-i-parameters-and-response","operation":"move","beforeSHA256":"d1fe64ee75a4b06b9419999d1d52b7071408348e3e14b074753cec2b1e237133","afterSHA256":"d1fe64ee75a4b06b9419999d1d52b7071408348e3e14b074753cec2b1e237133","reason":"Same source body in authored semantic page"},{"oldId":"b-ae582dc8c7672f71f23de853","targetId":"b-ae582dc8c7672f71f23de853","page":"h-i-parameters-and-response","operation":"move","beforeSHA256":"8787e6222b812c908f9fb8b9ee10c780670ad4f78049032105346bdc08842346","afterSHA256":"8787e6222b812c908f9fb8b9ee10c780670ad4f78049032105346bdc08842346","reason":"Same source body in authored semantic page"},{"oldId":"b-e4742fddddb5420e5946dfd6","targetId":"b-e4742fddddb5420e5946dfd6","page":"h-i-parameters-and-response","operation":"move","beforeSHA256":"8192daa71ce48b219345b5a9c01e4c9ab2830a269b9191c681a73780bdb9010f","afterSHA256":"8192daa71ce48b219345b5a9c01e4c9ab2830a269b9191c681a73780bdb9010f","reason":"Same source body in authored semantic page"},{"oldId":"b-ec5ace145dc2cd3e8cbd35c2","targetId":"b-ec5ace145dc2cd3e8cbd35c2","page":"h-i-parameters-and-response","operation":"move","beforeSHA256":"cf44fcc62782fb4f6bdc42c16c5187d57556a03cb7ce6065e7bd9c36d8e7d733","afterSHA256":"cf44fcc62782fb4f6bdc42c16c5187d57556a03cb7ce6065e7bd9c36d8e7d733","reason":"Same source body in authored semantic page"},{"oldId":"b-e111a15d1bab792a0952e624","targetId":"b-e111a15d1bab792a0952e624","page":"h-i-parameters-and-response","operation":"move","beforeSHA256":"fba0c350f1a9b4c09b4ed3883411f09fd8fa8256542a1edd0f23a2dfffc43c89","afterSHA256":"fba0c350f1a9b4c09b4ed3883411f09fd8fa8256542a1edd0f23a2dfffc43c89","reason":"Same source body in authored semantic page"},{"oldId":"b-6dc3e7bc0319b506950ee4bd","targetId":"b-6dc3e7bc0319b506950ee4bd","page":"h-i-parameters-and-response","operation":"move","beforeSHA256":"3a8a73c94e1809a04bafabec28bccaa574b8d6d76c45ddbc06eb05becf948b0c","afterSHA256":"3a8a73c94e1809a04bafabec28bccaa574b8d6d76c45ddbc06eb05becf948b0c","reason":"Same source body in authored semantic page"},{"oldId":"b-43aa8584425bd6a28179ea61","targetId":"b-43aa8584425bd6a28179ea61","page":"h-i-parameters-and-response","operation":"move","beforeSHA256":"6854020f9fdc983b5a4d5c73e07730ef5b52fa474476a1992d35e32363c5f183","afterSHA256":"6854020f9fdc983b5a4d5c73e07730ef5b52fa474476a1992d35e32363c5f183","reason":"Same source body in authored semantic page"},{"oldId":"b-2e468168d1803f7069fc3a58","targetId":"b-2e468168d1803f7069fc3a58","page":"h-i-parameters-and-response","operation":"move","beforeSHA256":"6854020f9fdc983b5a4d5c73e07730ef5b52fa474476a1992d35e32363c5f183","afterSHA256":"6854020f9fdc983b5a4d5c73e07730ef5b52fa474476a1992d35e32363c5f183","reason":"Same source body in authored semantic page"},{"oldId":"h-i-parameters-and-response-title","targetId":"b-329f582e57e8a3fdc6e18a54","operation":"retain-source-identity"}],"blocks":[{"id":"b-329f582e57e8a3fdc6e18a54","type":"h2","scope":"h-i-parameters-and-response","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p01.html#b-329f582e57e8a3fdc6e18a54","content":"## 训练保存参数，输入才产生表示","sha256":"d64c47160f252548c2e294ccbb0f1033858eedfd865c8262c8e473913c0c3587","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":"h-i-parameters-and-response","title":"训练保存参数，输入才产生表示","number":1},"sectionTitle":"训练保存参数，输入才产生表示","currentQuestion":"训练留下什么，训练答案又从哪里来？","sourceSlides":["l3-r2-p01"],"knowledgeBlock":null,"formulas":[],"previousBlock":null,"nextBlock":"b-98557719d66c29f742fd97ae","prerequisites":[{"id":"b-74178ee773393a35124115fc","url":"#b-74178ee773393a35124115fc","role":"problem-context"}],"relatedBlocks":[{"id":"b-e4fe6d59582b1c840d4ff9a2","url":"#b-e4fe6d59582b1c840d4ff9a2","relation":"figure-and-explanation"}]},{"id":"b-98557719d66c29f742fd97ae","type":"h2","scope":"explanation","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p01.html#b-98557719d66c29f742fd97ae","content":"## 解释、算例与证据","sha256":"d71cf15e707c8a3674b4c90a599efee783cb84da5b8c8236a1d01c3add7b4958","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":"h-i-parameters-and-response","title":"训练保存参数，输入才产生表示","number":1},"sectionTitle":"训练保存参数，输入才产生表示","currentQuestion":"训练留下什么，训练答案又从哪里来？","sourceSlides":["l3-r2-p01"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-329f582e57e8a3fdc6e18a54","nextBlock":"b-0d3ac5570914febfca13703a","prerequisites":[{"id":"b-329f582e57e8a3fdc6e18a54","url":"#b-329f582e57e8a3fdc6e18a54","role":"section-title"},{"id":"b-74178ee773393a35124115fc","url":"#b-74178ee773393a35124115fc","role":"problem-context"}],"relatedBlocks":[{"id":"b-e4fe6d59582b1c840d4ff9a2","url":"#b-e4fe6d59582b1c840d4ff9a2","relation":"figure-and-explanation"}]},{"id":"b-0d3ac5570914febfca13703a","type":"h1","scope":"document","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p01.html#b-0d3ac5570914febfca13703a","content":"# P01 · 训练结束后，网络留下了什么？","sha256":"c5e0d1a427d4be3e028ae79de9968a58185de4e8ae0ad9ddea8452875b9f86bd","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":"h-i-parameters-and-response","title":"训练保存参数，输入才产生表示","number":1},"sectionTitle":"训练保存参数，输入才产生表示","currentQuestion":"训练留下什么，训练答案又从哪里来？","sourceSlides":["l3-r2-p01"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-98557719d66c29f742fd97ae","nextBlock":"b-74178ee773393a35124115fc","prerequisites":[{"id":"b-98557719d66c29f742fd97ae","url":"#b-98557719d66c29f742fd97ae","role":"section-title"},{"id":"b-74178ee773393a35124115fc","url":"#b-74178ee773393a35124115fc","role":"problem-context"}],"relatedBlocks":[{"id":"b-e4fe6d59582b1c840d4ff9a2","url":"#b-e4fe6d59582b1c840d4ff9a2","relation":"figure-and-explanation"}]},{"id":"b-74178ee773393a35124115fc","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p01.html#b-74178ee773393a35124115fc","content":"先把第二讲的训练闭环收成可复用对象：参数、当前表示、类别头。由此引出本讲为何讨论训练信号和新任务。","sha256":"5a4ecd6807364900e549d8da58911f6fefa0d1bdf7274e83e148eb19208b83dc","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":"h-i-parameters-and-response","title":"训练保存参数，输入才产生表示","number":1},"sectionTitle":"训练保存参数，输入才产生表示","currentQuestion":"训练留下什么，训练答案又从哪里来？","sourceSlides":["l3-r2-p01"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-0d3ac5570914febfca13703a","nextBlock":"b-e4fe6d59582b1c840d4ff9a2","prerequisites":[{"id":"b-0d3ac5570914febfca13703a","url":"#b-0d3ac5570914febfca13703a","role":"section-title"}],"relatedBlocks":[{"id":"b-e4fe6d59582b1c840d4ff9a2","url":"#b-e4fe6d59582b1c840d4ff9a2","relation":"figure-and-explanation"}]},{"id":"b-e4fe6d59582b1c840d4ff9a2","type":"figure","scope":"document","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p01.html#b-e4fe6d59582b1c840d4ff9a2","content":"图内标签与数值：P01 学习表示机制图，状态 2；同一场景；是什么？；杯子 · 书；在哪里？哪一只？；书左边的红杯；杯身与杯柄的边界；学习信号；区域与像素；语言与任务\n\n红蓝两杯与书的同一底图，类别标签逐步增加两个杯框和语言指代。\n\n- 同一场景\n- 是什么？\n- 杯子 · 书\n- 在哪里？哪一只？\n- 书左边的红杯\n- 杯身与杯柄的边界\n- 学习信号\n- 区域与像素\n- 语言与任务红蓝两杯与书的同一底图，类别标签逐步增加两个杯框和语言指代。 [查看对应图解状态](/#/l3-r2-p01/1)[查看完整图解](/course/figures/l3-r2-p01-1.html)","sha256":"9b3416b515f8c1db965fb6797f493ef209f4f394b5a309e4e3d04bb30083a257","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":"h-i-parameters-and-response","title":"训练保存参数，输入才产生表示","number":1},"sectionTitle":"训练保存参数，输入才产生表示","currentQuestion":"训练留下什么，训练答案又从哪里来？","sourceSlides":["l3-r2-p01"],"knowledgeBlock":null,"formulas":[],"figure":{"caption":"红蓝两杯与书的同一底图，类别标签逐步增加两个杯框和语言指代。 查看对应图解状态","labels":["同一场景","是什么？","杯子 · 书","在哪里？哪一只？","书左边的红杯","杯身与杯柄的边界","学习信号","区域与像素","语言与任务"],"description":"红蓝两杯与书的同一底图，类别标签逐步增加两个杯框和语言指代。同一场景是什么？杯子 · 书在哪里？哪一只？书左边的红杯杯身与杯柄的边界学习信号区域与像素语言与任务","evidenceIdentity":"红蓝两杯与书的同一底图，类别标签逐步增加两个杯框和语言指代。 查看对应图解状态","images":[{"src":"/assets/display/584c419ba431799ff41df983.webp","alt":""}]},"previousBlock":"b-74178ee773393a35124115fc","nextBlock":"b-55476ebb13258f494524a47d","prerequisites":[{"id":"b-0d3ac5570914febfca13703a","url":"#b-0d3ac5570914febfca13703a","role":"section-title"},{"id":"b-74178ee773393a35124115fc","url":"#b-74178ee773393a35124115fc","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-55476ebb13258f494524a47d","type":"h2","scope":"narrative","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p01.html#b-55476ebb13258f494524a47d","content":"## 课堂讲述","sha256":"786ce10a6b32a492259a3dccd0a54e61cd41d28a34382fb7b5386c5f03b7ad69","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":"h-i-parameters-and-response","title":"训练保存参数，输入才产生表示","number":1},"sectionTitle":"训练保存参数，输入才产生表示","currentQuestion":"训练留下什么，训练答案又从哪里来？","sourceSlides":["l3-r2-p01"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-e4fe6d59582b1c840d4ff9a2","nextBlock":"b-8562e8e9a534c6c6989cc307","prerequisites":[{"id":"b-0d3ac5570914febfca13703a","url":"#b-0d3ac5570914febfca13703a","role":"section-title"},{"id":"b-74178ee773393a35124115fc","url":"#b-74178ee773393a35124115fc","role":"problem-context"}],"relatedBlocks":[{"id":"b-e4fe6d59582b1c840d4ff9a2","url":"#b-e4fe6d59582b1c840d4ff9a2","relation":"figure-and-explanation"}]},{"id":"b-8562e8e9a534c6c6989cc307","type":"p","scope":"narrative","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p01.html#b-8562e8e9a534c6c6989cc307","content":"上一讲，我们把网络拆开，看它怎样组织计算、怎样根据误差改变参数。今天把训练过程先停下来，留下已经训练好的网络。我要问的是：它到底给我们留下了什么？是一张类别对照表，还是一种可以继续使用的看图方式？\n看这张桌面照片。说出“有杯子”只是一个问题；找到杯子的位置，把杯子的像素分出来，再根据“书左边的杯子”找到目标，是另外几个问题。我们不希望每换一个问题，就把过去学到的东西全部作废。\n今天沿三步往前走：先看这些表示还能怎样学，再看它怎样支持区域和像素，最后看语言怎样加入。每一步都追问同一件事：什么留下来，什么必须重新学习。","sha256":"1fe45001cb450b4ebda1020fb108e200834cbfe4f64f1df6b6f1b538cfcda905","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":"h-i-parameters-and-response","title":"训练保存参数，输入才产生表示","number":1},"sectionTitle":"训练保存参数，输入才产生表示","currentQuestion":"训练留下什么，训练答案又从哪里来？","sourceSlides":["l3-r2-p01"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-55476ebb13258f494524a47d","nextBlock":"b-ce9fd754022f56966002a96d","prerequisites":[{"id":"b-55476ebb13258f494524a47d","url":"#b-55476ebb13258f494524a47d","role":"section-title"},{"id":"b-74178ee773393a35124115fc","url":"#b-74178ee773393a35124115fc","role":"problem-context"}],"relatedBlocks":[{"id":"b-e4fe6d59582b1c840d4ff9a2","url":"#b-e4fe6d59582b1c840d4ff9a2","relation":"figure-and-explanation"}]},{"id":"b-ce9fd754022f56966002a96d","type":"h3","scope":"representation-after-training","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p01.html#b-ce9fd754022f56966002a96d","content":"### 训练留下参数；一次输入才产生表示","sha256":"f749d7bbdbe8fc304b346ba2a2c67e965ddab66bcee94cec964566a00d2375fb","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":"h-i-parameters-and-response","title":"训练保存参数，输入才产生表示","number":1},"sectionTitle":"训练留下参数；一次输入才产生表示","currentQuestion":"训练留下什么，训练答案又从哪里来？","sourceSlides":["l3-r2-p01"],"knowledgeBlock":"representation-after-training","formulas":[],"previousBlock":"b-8562e8e9a534c6c6989cc307","nextBlock":"b-dcc0c92d0df5e100f443c08d","prerequisites":[{"id":"b-55476ebb13258f494524a47d","url":"#b-55476ebb13258f494524a47d","role":"section-title"},{"id":"b-74178ee773393a35124115fc","url":"#b-74178ee773393a35124115fc","role":"problem-context"}],"relatedBlocks":[{"id":"b-e4fe6d59582b1c840d4ff9a2","url":"#b-e4fe6d59582b1c840d4ff9a2","relation":"figure-and-explanation"},{"id":"b-dcc0c92d0df5e100f443c08d","url":"#b-dcc0c92d0df5e100f443c08d","relation":"same-reasoning"},{"id":"b-ae582dc8c7672f71f23de853","url":"#b-ae582dc8c7672f71f23de853","relation":"same-reasoning"},{"id":"b-e4742fddddb5420e5946dfd6","url":"#b-e4742fddddb5420e5946dfd6","relation":"same-reasoning"}]},{"id":"b-dcc0c92d0df5e100f443c08d","type":"p","scope":"representation-after-training","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p01.html#b-dcc0c92d0df5e100f443c08d","content":"训练时，我们不断用图像、目标和损失修改参数 $\\theta$。训练结束后保存的是参数文件；把新图 $x$ 送入它，才得到表示 $h=f_\\theta(x)$。这两样东西不能混叫“特征”：前者是可复用的计算规则，后者是某个输入在这套规则下的输出。","sha256":"d1fe64ee75a4b06b9419999d1d52b7071408348e3e14b074753cec2b1e237133","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":"h-i-parameters-and-response","title":"训练保存参数，输入才产生表示","number":1},"sectionTitle":"训练留下参数；一次输入才产生表示","currentQuestion":"训练留下什么，训练答案又从哪里来？","sourceSlides":["l3-r2-p01"],"knowledgeBlock":"representation-after-training","formulas":[{"latex":"\\theta","definitionContext":"训练时，我们不断用图像、目标和损失修改参数 $\\theta$。训练结束后保存的是参数文件；把新图 $x$ 送入它，才得到表示 $h=f_\\theta(x)$。这两样东西不能混叫“特征”：前者是可复用的计算规则，后者是某个输入在这套规则下的输出。\n分类头把 $h$ 变成类别分数 $a=Wh+b$。换成“哪一只杯子在左边”时，原分类头的类别名单往往无用，但前面识别轮廓、纹理和部件的计算可能仍有用。这就是移除原头、保留编码器的起点。它提出了一个可检验的假设：**旧任务塑造的中间计算，能降低新任务学习的困难。**\n看当前图时，先指向“图像→表示”，再指向“表示→类别”。本讲第一部分改变前一段怎样学习，第二部分改变后一段要交付什么；不要求把所有问题塞进同一个输出向量。"},{"latex":"x","definitionContext":"训练时，我们不断用图像、目标和损失修改参数 $\\theta$。训练结束后保存的是参数文件；把新图 $x$ 送入它，才得到表示 $h=f_\\theta(x)$。这两样东西不能混叫“特征”：前者是可复用的计算规则，后者是某个输入在这套规则下的输出。\n分类头把 $h$ 变成类别分数 $a=Wh+b$。换成“哪一只杯子在左边”时，原分类头的类别名单往往无用，但前面识别轮廓、纹理和部件的计算可能仍有用。这就是移除原头、保留编码器的起点。它提出了一个可检验的假设：**旧任务塑造的中间计算，能降低新任务学习的困难。**\n看当前图时，先指向“图像→表示”，再指向“表示→类别”。本讲第一部分改变前一段怎样学习，第二部分改变后一段要交付什么；不要求把所有问题塞进同一个输出向量。"},{"latex":"h=f_\\theta(x)","definitionContext":"训练时，我们不断用图像、目标和损失修改参数 $\\theta$。训练结束后保存的是参数文件；把新图 $x$ 送入它，才得到表示 $h=f_\\theta(x)$。这两样东西不能混叫“特征”：前者是可复用的计算规则，后者是某个输入在这套规则下的输出。\n分类头把 $h$ 变成类别分数 $a=Wh+b$。换成“哪一只杯子在左边”时，原分类头的类别名单往往无用，但前面识别轮廓、纹理和部件的计算可能仍有用。这就是移除原头、保留编码器的起点。它提出了一个可检验的假设：**旧任务塑造的中间计算，能降低新任务学习的困难。**\n看当前图时，先指向“图像→表示”，再指向“表示→类别”。本讲第一部分改变前一段怎样学习，第二部分改变后一段要交付什么；不要求把所有问题塞进同一个输出向量。"}],"previousBlock":"b-ce9fd754022f56966002a96d","nextBlock":"b-ae582dc8c7672f71f23de853","prerequisites":[{"id":"b-ce9fd754022f56966002a96d","url":"#b-ce9fd754022f56966002a96d","role":"section-title"},{"id":"b-74178ee773393a35124115fc","url":"#b-74178ee773393a35124115fc","role":"problem-context"}],"relatedBlocks":[{"id":"b-e4fe6d59582b1c840d4ff9a2","url":"#b-e4fe6d59582b1c840d4ff9a2","relation":"figure-and-explanation"},{"id":"b-ce9fd754022f56966002a96d","url":"#b-ce9fd754022f56966002a96d","relation":"same-reasoning"},{"id":"b-ae582dc8c7672f71f23de853","url":"#b-ae582dc8c7672f71f23de853","relation":"same-reasoning"},{"id":"b-e4742fddddb5420e5946dfd6","url":"#b-e4742fddddb5420e5946dfd6","relation":"same-reasoning"}]},{"id":"b-ae582dc8c7672f71f23de853","type":"p","scope":"representation-after-training","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p01.html#b-ae582dc8c7672f71f23de853","content":"分类头把 $h$ 变成类别分数 $a=Wh+b$。换成“哪一只杯子在左边”时，原分类头的类别名单往往无用，但前面识别轮廓、纹理和部件的计算可能仍有用。这就是移除原头、保留编码器的起点。它提出了一个可检验的假设：**旧任务塑造的中间计算，能降低新任务学习的困难。**","sha256":"8787e6222b812c908f9fb8b9ee10c780670ad4f78049032105346bdc08842346","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":"h-i-parameters-and-response","title":"训练保存参数，输入才产生表示","number":1},"sectionTitle":"训练留下参数；一次输入才产生表示","currentQuestion":"训练留下什么，训练答案又从哪里来？","sourceSlides":["l3-r2-p01"],"knowledgeBlock":"representation-after-training","formulas":[{"latex":"h","definitionContext":"训练时，我们不断用图像、目标和损失修改参数 $\\theta$。训练结束后保存的是参数文件；把新图 $x$ 送入它，才得到表示 $h=f_\\theta(x)$。这两样东西不能混叫“特征”：前者是可复用的计算规则，后者是某个输入在这套规则下的输出。\n分类头把 $h$ 变成类别分数 $a=Wh+b$。换成“哪一只杯子在左边”时，原分类头的类别名单往往无用，但前面识别轮廓、纹理和部件的计算可能仍有用。这就是移除原头、保留编码器的起点。它提出了一个可检验的假设：**旧任务塑造的中间计算，能降低新任务学习的困难。**\n看当前图时，先指向“图像→表示”，再指向“表示→类别”。本讲第一部分改变前一段怎样学习，第二部分改变后一段要交付什么；不要求把所有问题塞进同一个输出向量。"},{"latex":"a=Wh+b","definitionContext":"训练时，我们不断用图像、目标和损失修改参数 $\\theta$。训练结束后保存的是参数文件；把新图 $x$ 送入它，才得到表示 $h=f_\\theta(x)$。这两样东西不能混叫“特征”：前者是可复用的计算规则，后者是某个输入在这套规则下的输出。\n分类头把 $h$ 变成类别分数 $a=Wh+b$。换成“哪一只杯子在左边”时，原分类头的类别名单往往无用，但前面识别轮廓、纹理和部件的计算可能仍有用。这就是移除原头、保留编码器的起点。它提出了一个可检验的假设：**旧任务塑造的中间计算，能降低新任务学习的困难。**\n看当前图时，先指向“图像→表示”，再指向“表示→类别”。本讲第一部分改变前一段怎样学习，第二部分改变后一段要交付什么；不要求把所有问题塞进同一个输出向量。"}],"previousBlock":"b-dcc0c92d0df5e100f443c08d","nextBlock":"b-e4742fddddb5420e5946dfd6","prerequisites":[{"id":"b-ce9fd754022f56966002a96d","url":"#b-ce9fd754022f56966002a96d","role":"section-title"},{"id":"b-74178ee773393a35124115fc","url":"#b-74178ee773393a35124115fc","role":"problem-context"}],"relatedBlocks":[{"id":"b-e4fe6d59582b1c840d4ff9a2","url":"#b-e4fe6d59582b1c840d4ff9a2","relation":"figure-and-explanation"},{"id":"b-ce9fd754022f56966002a96d","url":"#b-ce9fd754022f56966002a96d","relation":"same-reasoning"},{"id":"b-dcc0c92d0df5e100f443c08d","url":"#b-dcc0c92d0df5e100f443c08d","relation":"same-reasoning"},{"id":"b-e4742fddddb5420e5946dfd6","url":"#b-e4742fddddb5420e5946dfd6","relation":"same-reasoning"}]},{"id":"b-e4742fddddb5420e5946dfd6","type":"p","scope":"representation-after-training","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p01.html#b-e4742fddddb5420e5946dfd6","content":"看当前图时，先指向“图像→表示”，再指向“表示→类别”。本讲第一部分改变前一段怎样学习，第二部分改变后一段要交付什么；不要求把所有问题塞进同一个输出向量。","sha256":"8192daa71ce48b219345b5a9c01e4c9ab2830a269b9191c681a73780bdb9010f","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":"h-i-parameters-and-response","title":"训练保存参数，输入才产生表示","number":1},"sectionTitle":"训练留下参数；一次输入才产生表示","currentQuestion":"训练留下什么，训练答案又从哪里来？","sourceSlides":["l3-r2-p01"],"knowledgeBlock":"representation-after-training","formulas":[],"previousBlock":"b-ae582dc8c7672f71f23de853","nextBlock":"b-ec5ace145dc2cd3e8cbd35c2","prerequisites":[{"id":"b-ce9fd754022f56966002a96d","url":"#b-ce9fd754022f56966002a96d","role":"section-title"},{"id":"b-74178ee773393a35124115fc","url":"#b-74178ee773393a35124115fc","role":"problem-context"}],"relatedBlocks":[{"id":"b-e4fe6d59582b1c840d4ff9a2","url":"#b-e4fe6d59582b1c840d4ff9a2","relation":"figure-and-explanation"},{"id":"b-ce9fd754022f56966002a96d","url":"#b-ce9fd754022f56966002a96d","relation":"same-reasoning"},{"id":"b-dcc0c92d0df5e100f443c08d","url":"#b-dcc0c92d0df5e100f443c08d","relation":"same-reasoning"},{"id":"b-ae582dc8c7672f71f23de853","url":"#b-ae582dc8c7672f71f23de853","relation":"same-reasoning"}]},{"id":"b-ec5ace145dc2cd3e8cbd35c2","type":"h2","scope":"practice","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p01.html#b-ec5ace145dc2cd3e8cbd35c2","content":"## 沿这个问题继续","sha256":"cf44fcc62782fb4f6bdc42c16c5187d57556a03cb7ce6065e7bd9c36d8e7d733","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":"h-i-parameters-and-response","title":"训练保存参数，输入才产生表示","number":1},"sectionTitle":"训练保存参数，输入才产生表示","currentQuestion":"训练留下什么，训练答案又从哪里来？","sourceSlides":["l3-r2-p01"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-e4742fddddb5420e5946dfd6","nextBlock":"b-e111a15d1bab792a0952e624","prerequisites":[{"id":"b-ce9fd754022f56966002a96d","url":"#b-ce9fd754022f56966002a96d","role":"section-title"},{"id":"b-74178ee773393a35124115fc","url":"#b-74178ee773393a35124115fc","role":"problem-context"}],"relatedBlocks":[{"id":"b-e4fe6d59582b1c840d4ff9a2","url":"#b-e4fe6d59582b1c840d4ff9a2","relation":"figure-and-explanation"}]},{"id":"b-e111a15d1bab792a0952e624","type":"p","scope":"practice","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p01.html#b-e111a15d1bab792a0952e624","content":"读“训练留下参数”后，尝试给自己的科研输入、编码器与输出各写一个实例；暂不运行实验。","sha256":"fba0c350f1a9b4c09b4ed3883411f09fd8fa8256542a1edd0f23a2dfffc43c89","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":"h-i-parameters-and-response","title":"训练保存参数，输入才产生表示","number":1},"sectionTitle":"训练保存参数，输入才产生表示","currentQuestion":"训练留下什么，训练答案又从哪里来？","sourceSlides":["l3-r2-p01"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-ec5ace145dc2cd3e8cbd35c2","nextBlock":"b-4f328559c86fbbe412b4ce1d","prerequisites":[{"id":"b-ec5ace145dc2cd3e8cbd35c2","url":"#b-ec5ace145dc2cd3e8cbd35c2","role":"section-title"},{"id":"b-74178ee773393a35124115fc","url":"#b-74178ee773393a35124115fc","role":"problem-context"}],"relatedBlocks":[{"id":"b-e4fe6d59582b1c840d4ff9a2","url":"#b-e4fe6d59582b1c840d4ff9a2","relation":"figure-and-explanation"}]},{"id":"b-4f328559c86fbbe412b4ce1d","type":"h4","scope":"h-i-parameters-and-response","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p01.html#b-4f328559c86fbbe412b4ce1d","content":"#### 图示条件与公式约定","sha256":"ef7f915c1ae63ed67bd23fecf050249273ba0efa8965d2c8d1c1c6b2d3178122","chapter":{"id":"I","title":"学习信号与表示"},"unit":{"id":"i-reuse-and-targets","title":"训练留下什么，训练答案又从哪里来？"},"page":{"id":"h-i-parameters-and-response","title":"训练保存参数，输入才产生表示","number":1},"sectionTitle":"训练保存参数，输入才产生表示","currentQuestion":"训练留下什么，训练答案又从哪里来？","sourceSlides":["l3-r2-p01"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-e111a15d1bab792a0952e624","nextBlock":"b-6dc3e7bc0319b506950ee4bd","prerequisites":[{"id":"b-ec5ace145dc2cd3e8cbd35c2","url":"#b-ec5ace145dc2cd3e8cbd35c2","role":"section-title"},{"id":"b-74178ee773393a35124115fc","url":"#b-74178ee773393a35124115fc","role":"problem-context"}],"relatedBlocks":[{"id":"b-e4fe6d59582b1c840d4ff9a2","url":"#b-e4fe6d59582b1c840d4ff9a2","relation":"figure-and-explanation"}]},{"id":"b-6dc3e7bc0319b506950ee4bd","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p01.html#b-6dc3e7bc0319b506950ee4bd","content":"复用现有两杯一书原创场景；没有在画面中添加不存在的零件。框及文字为教学标注，不是模型输出。","sha256":"3a8a73c94e1809a04bafabec28bccaa574b8d6d76c45ddbc06eb05becf948b0c","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":"h-i-parameters-and-response","title":"训练保存参数，输入才产生表示","number":1},"sectionTitle":"训练保存参数，输入才产生表示","currentQuestion":"训练留下什么，训练答案又从哪里来？","sourceSlides":["l3-r2-p01"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-4f328559c86fbbe412b4ce1d","nextBlock":"b-3146dab65f0e0c3c5a2021ab","prerequisites":[{"id":"b-4f328559c86fbbe412b4ce1d","url":"#b-4f328559c86fbbe412b4ce1d","role":"section-title"},{"id":"b-74178ee773393a35124115fc","url":"#b-74178ee773393a35124115fc","role":"problem-context"}],"relatedBlocks":[{"id":"b-e4fe6d59582b1c840d4ff9a2","url":"#b-e4fe6d59582b1c840d4ff9a2","relation":"figure-and-explanation"}]},{"id":"b-3146dab65f0e0c3c5a2021ab","type":"h4","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p01.html#b-3146dab65f0e0c3c5a2021ab","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-reuse-and-targets","title":"训练留下什么，训练答案又从哪里来？"},"page":{"id":"h-i-parameters-and-response","title":"训练保存参数，输入才产生表示","number":1},"sectionTitle":"训练保存参数，输入才产生表示","currentQuestion":"训练留下什么，训练答案又从哪里来？","sourceSlides":["l3-r2-p01"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-6dc3e7bc0319b506950ee4bd","nextBlock":"b-43aa8584425bd6a28179ea61","prerequisites":[{"id":"b-4f328559c86fbbe412b4ce1d","url":"#b-4f328559c86fbbe412b4ce1d","role":"section-title"},{"id":"b-74178ee773393a35124115fc","url":"#b-74178ee773393a35124115fc","role":"problem-context"}],"relatedBlocks":[{"id":"b-e4fe6d59582b1c840d4ff9a2","url":"#b-e4fe6d59582b1c840d4ff9a2","relation":"figure-and-explanation"}]},{"id":"b-43aa8584425bd6a28179ea61","type":"ul","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p01.html#b-43aa8584425bd6a28179ea61","content":"- [可迁移表示与多模态感知](/supplement.html?return=l3-r2-p01/0#/l3-representation-opening/0) 本页说明","sha256":"6854020f9fdc983b5a4d5c73e07730ef5b52fa474476a1992d35e32363c5f183","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":"h-i-parameters-and-response","title":"训练保存参数，输入才产生表示","number":1},"sectionTitle":"训练保存参数，输入才产生表示","currentQuestion":"训练留下什么，训练答案又从哪里来？","sourceSlides":["l3-r2-p01"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-3146dab65f0e0c3c5a2021ab","nextBlock":"b-2e468168d1803f7069fc3a58","prerequisites":[{"id":"b-3146dab65f0e0c3c5a2021ab","url":"#b-3146dab65f0e0c3c5a2021ab","role":"section-title"},{"id":"b-74178ee773393a35124115fc","url":"#b-74178ee773393a35124115fc","role":"problem-context"}],"relatedBlocks":[{"id":"b-e4fe6d59582b1c840d4ff9a2","url":"#b-e4fe6d59582b1c840d4ff9a2","relation":"figure-and-explanation"}]},{"id":"b-2e468168d1803f7069fc3a58","type":"li","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p01.html#b-2e468168d1803f7069fc3a58","content":"- [可迁移表示与多模态感知](/supplement.html?return=l3-r2-p01/0#/l3-representation-opening/0) 本页说明","sha256":"6854020f9fdc983b5a4d5c73e07730ef5b52fa474476a1992d35e32363c5f183","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":"h-i-parameters-and-response","title":"训练保存参数，输入才产生表示","number":1},"sectionTitle":"训练保存参数，输入才产生表示","currentQuestion":"训练留下什么，训练答案又从哪里来？","sourceSlides":["l3-r2-p01"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-43aa8584425bd6a28179ea61","nextBlock":null,"prerequisites":[{"id":"b-3146dab65f0e0c3c5a2021ab","url":"#b-3146dab65f0e0c3c5a2021ab","role":"section-title"},{"id":"b-74178ee773393a35124115fc","url":"#b-74178ee773393a35124115fc","role":"problem-context"}],"relatedBlocks":[{"id":"b-e4fe6d59582b1c840d4ff9a2","url":"#b-e4fe6d59582b1c840d4ff9a2","relation":"figure-and-explanation"}]}]}