{"topic":{"id":9142,"title":"深度学习个人笔记分享","slug":"topic","categoryId":54,"createdAt":"2024-01-18 06:58:23.717815","bumpedAt":"2024-08-22 10:25:23.580504","views":1768,"postsCount":40,"highestPostNumber":43,"likeCount":29,"closed":false,"pinned":false,"pinnedGlobally":false,"pinnedAt":null,"pinnedUntil":null,"isBanner":false,"excerpt":"分享一些个人在学习过程中的笔记。参考资料： https://speech.ee.ntu.edu.tw/~hylee/ml/2022-spring.php https://speech.ee.ntu.edu.tw/~hylee/ml/2023-spring.php https://github.com/mli/paper-reading 《动手学深度学习》","deletedAt":null,"isAnonymous":false,"category":{"id":54,"name":"学术专区","slug":"academic","color":"9EB83B"},"author":{"id":31,"username":"TsuinoSora","name":null,"avatarUrl":"/uploads/image/77/06/7706643ba6d3ee1ec9d0070db0ad8a22bed8651fa8eb829cf44df4d20fa24b3b.avif"},"tags":[]},"posts":[{"id":22087,"postNumber":1,"raw":"分享一些个人在学习过程中的笔记。\n\n参考资料：\n\n- https://speech.ee.ntu.edu.tw/~hylee/ml/2022-spring.php\n- https://speech.ee.ntu.edu.tw/~hylee/ml/2023-spring.php\n- https://github.com/mli/paper-reading\n- [《动手学深度学习》](https://zh.d2l.ai/)","cooked":"

分享一些个人在学习过程中的笔记。

参考资料：

\n","createdAt":"2024-01-18 06:58:23.793579","updatedAt":"2024-03-04 03:37:11.796839","replyToPostNumber":null,"replyCount":0,"likeCount":14,"wordCount":37,"liked":false,"bookmarked":false,"polls":[],"reactions":[{"reaction":"❤️","count":14,"users":[{"id":49,"username":"dokuro","name":null},{"id":172,"username":"HCCH","name":"HCCH"},{"id":307,"username":"Fast","name":null},{"id":346,"username":"hoist","name":null},{"id":876,"username":"秦川楚客","name":null},{"id":1888,"username":"carrortsky","name":"Luo"},{"id":1941,"username":"john","name":"John"},{"id":2276,"username":"2832128540","name":"Schrdinger‘s dog"},{"id":3332,"username":"rijin","name":"\");drop table;--"},{"id":3988,"username":"Edmund123","name":"Edmund123"},{"id":4263,"username":"Murphy","name":null},{"id":5077,"username":"Ayato","name":"桃泽"},{"id":5309,"username":"zhui","name":null},{"id":7444,"username":"周末","name":"😀"}],"hasReacted":false}],"wiki":false,"version":3,"publicVersion":3,"deletedAt":null,"isAnonymous":false,"user":{"id":31,"username":"TsuinoSora","name":null,"avatarUrl":"/uploads/image/77/06/7706643ba6d3ee1ec9d0070db0ad8a22bed8651fa8eb829cf44df4d20fa24b3b.avif","admin":false,"moderator":true,"signatureRaw":null,"signatureCooked":null}},{"id":22090,"postNumber":4,"raw":"# 机器学习基础理论 & 为什么选择深度学习\n\n[basic-thoery-of-ml-why-deep.pdf|attachment](/uploads/pdf/d3/b5/d3b5be76311a31a22467895a2f8a9584aa7db23f8aa48216d4510d2238796d5a.pdf) (483.5 KB)","cooked":"

机器学习基础理论 & 为什么选择深度学习

basic-thoery-of-ml-why-deep.pdf|attachment (483.5 KB)

\n","createdAt":"2024-01-18 07:11:14.119236","updatedAt":"2024-01-18 07:15:46.405109","replyToPostNumber":null,"replyCount":0,"likeCount":0,"wordCount":22,"liked":false,"bookmarked":false,"polls":[],"reactions":[],"wiki":false,"version":1,"publicVersion":1,"deletedAt":null,"isAnonymous":false,"user":{"id":31,"username":"TsuinoSora","name":null,"avatarUrl":"/uploads/image/77/06/7706643ba6d3ee1ec9d0070db0ad8a22bed8651fa8eb829cf44df4d20fa24b3b.avif","admin":false,"moderator":true,"signatureRaw":null,"signatureCooked":null}},{"id":22091,"postNumber":5,"raw":"# 机器学习框架及一些问题的解决\n\n[frame-fix.pdf|attachment](/uploads/pdf/63/ec/63eced41ca6ffc4f2e85b9b578549ee24d88a1fa996bb272f0121b524efe00ba.pdf) (1007.5 KB)","cooked":"

机器学习框架及一些问题的解决

frame-fix.pdf|attachment (1007.5 KB)

\n","createdAt":"2024-01-18 07:11:36.558378","updatedAt":"2024-01-18 07:16:17.115362","replyToPostNumber":null,"replyCount":0,"likeCount":0,"wordCount":18,"liked":false,"bookmarked":false,"polls":[],"reactions":[],"wiki":false,"version":1,"publicVersion":1,"deletedAt":null,"isAnonymous":false,"user":{"id":31,"username":"TsuinoSora","name":null,"avatarUrl":"/uploads/image/77/06/7706643ba6d3ee1ec9d0070db0ad8a22bed8651fa8eb829cf44df4d20fa24b3b.avif","admin":false,"moderator":true,"signatureRaw":null,"signatureCooked":null}},{"id":22092,"postNumber":6,"raw":"# 卷积神经网络\n\n[CNN.pdf|attachment](/uploads/pdf/b2/e1/b2e1363442e6edc83b773ac8f5dfcc52b447d43b53d26ee7a6622aa8077ee050.pdf) (1.8 MB)","cooked":"

卷积神经网络

CNN.pdf|attachment (1.8 MB)

\n","createdAt":"2024-01-18 07:11:58.315944","updatedAt":"2024-01-18 07:15:24.203435","replyToPostNumber":null,"replyCount":0,"likeCount":0,"wordCount":10,"liked":false,"bookmarked":false,"polls":[],"reactions":[],"wiki":false,"version":1,"publicVersion":1,"deletedAt":null,"isAnonymous":false,"user":{"id":31,"username":"TsuinoSora","name":null,"avatarUrl":"/uploads/image/77/06/7706643ba6d3ee1ec9d0070db0ad8a22bed8651fa8eb829cf44df4d20fa24b3b.avif","admin":false,"moderator":true,"signatureRaw":null,"signatureCooked":null}},{"id":22093,"postNumber":7,"raw":"# 自注意力机制与 Transformer\n\n[self-attention-transformer.pdf|attachment](/uploads/pdf/29/e0/29e02f5a4502e6e28dd1e436e5e10cd7ffeec187f027b30f19407fa7d0053d0c.pdf) (2.4 MB)","cooked":"

自注意力机制与 Transformer

self-attention-transformer.pdf|attachment (2.4 MB)

\n","createdAt":"2024-01-18 07:12:32.736761","updatedAt":"2024-01-18 07:14:44.831525","replyToPostNumber":null,"replyCount":0,"likeCount":0,"wordCount":12,"liked":false,"bookmarked":false,"polls":[],"reactions":[],"wiki":false,"version":1,"publicVersion":1,"deletedAt":null,"isAnonymous":false,"user":{"id":31,"username":"TsuinoSora","name":null,"avatarUrl":"/uploads/image/77/06/7706643ba6d3ee1ec9d0070db0ad8a22bed8651fa8eb829cf44df4d20fa24b3b.avif","admin":false,"moderator":true,"signatureRaw":null,"signatureCooked":null}},{"id":22094,"postNumber":8,"raw":"# 对比学习综述\n\n[constrastive-learning.pdf|attachment](/uploads/pdf/f8/83/f883aa1f0ccfb8af3ac69bd9f136ced9ec5d3ebf768e95f6986cd8701791ddfd.pdf) (1.4 MB)","cooked":"

对比学习综述

constrastive-learning.pdf|attachment (1.4 MB)

\n","createdAt":"2024-01-18 07:14:09.277115","updatedAt":"2024-01-18 07:14:09.277115","replyToPostNumber":null,"replyCount":0,"likeCount":0,"wordCount":10,"liked":false,"bookmarked":false,"polls":[],"reactions":[],"wiki":false,"version":1,"publicVersion":1,"deletedAt":null,"isAnonymous":false,"user":{"id":31,"username":"TsuinoSora","name":null,"avatarUrl":"/uploads/image/77/06/7706643ba6d3ee1ec9d0070db0ad8a22bed8651fa8eb829cf44df4d20fa24b3b.avif","admin":false,"moderator":true,"signatureRaw":null,"signatureCooked":null}},{"id":22095,"postNumber":9,"raw":"good","cooked":"

good

\n","createdAt":"2024-01-18 07:19:55.681912","updatedAt":"2024-01-18 07:19:55.681912","replyToPostNumber":null,"replyCount":0,"likeCount":0,"wordCount":1,"liked":false,"bookmarked":false,"polls":[],"reactions":[],"wiki":false,"version":1,"publicVersion":1,"deletedAt":null,"isAnonymous":false,"user":{"id":1941,"username":"john","name":"John","avatarUrl":"/uploads/image/5a/04/5a0499fe826e0ab0e29d97bc3d98bbc1dfebf81309f4b4426ae0c7b62bd908e8.avif","admin":false,"moderator":false,"signatureRaw":"自黄药师黄仁勋来，世人皆爱猫。余独爱老鼠之，可爱之态，活泼之动，实为人所难以抗拒。喵~\n\n:mouse_hole: :mouse: :mouse: :computer_mouse: :mouse_trap: :sob_mouse: :sweat_smile_mouse: :sleeping_mouse: :rolling_on_the_floor_laughing_mouse: :revolving_hearts_mouse: :pleading_face_mouse: :nerd_face_mouse: :mouse_hole: :melting_face_mouse: :mask_mouse: :innocent_mouse: :hugging_face_mouse: :frog_mouse: :face_with_cowboy_hat_mouse: :clown_face_mouse: :cherry_blossom_mouse: :cat_mouse:","signatureCooked":"

自黄药师黄仁勋来，世人皆爱猫。余独爱老鼠之，可爱之态，活泼之动，实为人所难以抗拒。喵~

$\":mouse_hole:\"$ 🐁 🐁 🖱️ 🪤 $\":sob_mouse:\"$ $\":sweat_smile_mouse:\"$ $\":sleeping_mouse:\"$ $\":rolling_on_the_floor_laughing_mouse:\"$ $\":revolving_hearts_mouse:\"$ $\":pleading_face_mouse:\"$ $\":nerd_face_mouse:\"$ $\":mouse_hole:\"$ $\":melting_face_mouse:\"$ $\":mask_mouse:\"$ $\":innocent_mouse:\"$ $\":hugging_face_mouse:\"$ $\":frog_mouse:\"$ $\":face_with_cowboy_hat_mouse:\"$ $\":clown_face_mouse:\"$ $\":cherry_blossom_mouse:\"$ $\":cat_mouse:\"$

\n"}},{"id":22102,"postNumber":10,"raw":"好好好，狠狠支持了 :hooray: 不知道我大二暑假能不能用到 😰","cooked":"

好好好，狠狠支持了 🥳 不知道我大二暑假能不能用到 😰

\n","createdAt":"2024-01-18 08:29:37.099283","updatedAt":"2024-01-18 10:40:20.245164","replyToPostNumber":null,"replyCount":0,"likeCount":2,"wordCount":24,"liked":false,"bookmarked":false,"polls":[],"reactions":[{"reaction":"❤️","count":2,"users":[{"id":3594,"username":"zecrious","name":"勿1小win而不麻"},{"id":4260,"username":"Rick","name":"Rick"}],"hasReacted":false}],"wiki":false,"version":3,"publicVersion":3,"deletedAt":null,"isAnonymous":false,"user":{"id":30,"username":"一只玉米人","name":"玉米🌽","avatarUrl":"/uploads/blob/14/96/14961aeaa46ccbb75a61eec0c27b26e74fb5d1c00ccdcbbe898ca9ea37087dd6.svg","admin":false,"moderator":true,"signatureRaw":"*年月擦身过，暂且问，会更好吗？*","signatureCooked":"

年月擦身过，暂且问，会更好吗？

\n"}},{"id":22109,"postNumber":11,"raw":"好强 :flushed:","cooked":"

好强 😳

\n","createdAt":"2024-01-18 09:31:48.991945","updatedAt":"2024-01-18 09:31:48.991945","replyToPostNumber":null,"replyCount":0,"likeCount":0,"wordCount":3,"liked":false,"bookmarked":false,"polls":[],"reactions":[],"wiki":false,"version":1,"publicVersion":1,"deletedAt":null,"isAnonymous":false,"user":{"id":32,"username":"ScorpioLuc","name":"池鹭","avatarUrl":"/uploads/image/c9/80/c9802bb2950fa972f099d0a7731d8b241263989889c3c325000ad5df64c5d8ff.avif","admin":false,"moderator":false,"signatureRaw":null,"signatureCooked":null}},{"id":22434,"postNumber":12,"raw":"# 杂项\n\n会持续更新\n\n### 正则化\n\nhttps://zh.wikipedia.org/zh-hans/%E6%AD%A3%E5%88%99%E5%8C%96_(%E6%95%B0%E5%AD%A6)\n\n**`Weight Decay` - L2 正则项**\n\nwe multiply the sum of squares with another smaller number. This number is called weight decay or wd.\n\nOur loss function now looks as follows:\n\n`Loss = MSE(y_hat, y) + wd * sum(w^2)`\n\n减小模型复杂度（弹性），降低过拟合的概率\n\n### 残差（residual）和残差网络\n\nhttps://www.cnblogs.com/wuliytTaotao/p/9560205.html\n\n### CNN 输出大小公式\n\n$W$ 为输入大小，$F$ 为卷积核大小，$P$ 为填充大小（Padding），$S$ 为步长（Stride），$N$ 为输出大小，则有：\n\n$$\nN = \\frac {W - F + 2P} {S} + 1\n$$\n\n### 熵（Entropy）；KL 散度（KL Divergence）\n\n#### 熵\n\n若一个离散随机变量 $\\mathbf{X}$ 的可能取值为 $X=\\{x_1,x_2,\\cdots,x_n\\}$，而对应的概率为 $p_i = p(X = x_i)$，则随机变量 $\\mathbf{X}$ 的熵定义为：\n\n$$\n\nH(X) = - \\sum_{i=1}^{n} p(x_i) \\log p(x_{i})\n\n$$\n\n#### 交叉熵\n\n针对上述离散变量的概率分布 $p(x)、q(x)$ 而言，其交叉熵定义为：\n\n$$\nH(p,q) = - \\sum_{x} p(x) \\log q(x)\n$$\n\n> 在信息论中，交叉熵可认为是对预测分布 $q(x)$ 用真实分布 $p(x)$ 来进行编码时所需要的信息量大小。\n\n#### KL 散度（KL Divergence）\n\nKL 散度可以用来衡量两个分布之间的差异：\n\n$$\n\\begin{aligned}\n\nD_{KL}(p||q) = & H(p,q)-H(p) \\\\\n\n=& - \\sum_{x} p(x) \\log q(x) - \\sum_{x}-p(x)\\log p(x)\\\\\n\n=& -\\sum_{x}p(x) \\log \\frac{q(x)}{p(x)}\n\n\\end{aligned}\n$$\n\n> 交叉熵损失函数实际上就是 KL 散度，只不过因为分类问题的特殊性，其式与交叉熵相同：\n\n在分类问题中，我们通过 `softmax` 函数得到一个分布，我们要通过 **KL 散度** 衡量这个分布和目标分布（一个 one-hot 向量）的差异性：\n\n$$\n\nD_{KL}(p||q) = H(p,q)-H(p)\n\n$$\n\n我们易知 $H(p) = 0$；且对于真实分布（一个 one-hot 向量）只有概率为 1 的一项是非 0 的，所以有以下化简：\n\n$$\n\\begin{aligned}\n\nD_{KL}(p||q) = & H(p,q)\\\\\n\n=& - \\sum_{x} p(x) \\log q(x)\\\\\n\n=& -\\log \\text{softmax} \\ q(x_i)\\\\\n\n=& -\\log \\frac{\\exp (x_i)}{\\sum_j \\exp (x_j)}\n\n\\end{aligned}\n$$\n\n**KL 散度的数学性质**\n\n- 正定性\n\n$$\nD_{KL}(p||q) \\ge 0\n$$\n\n- 不对称性\n\nKL 散度并不是一个真正的度量或者距离，因为它不具有对称性：\n\n$$\nD(p || q) \\ne D(q || p)\n$$\n\n#### JS 散度（JS Divergence）\n\nJS 散度度量了两个概率分布的相似度，基于 KL 散度的变体，解决了 KL 散度非对称的问题。一般地，JS 散度是对称的，其取值是 0 到 1 之间。\n\n$$\nJSD(P||Q)=\\frac{1}{2}D_{KL}(P||\\frac{P+Q}{2})+\\frac{1}{2}D_{KL}(Q||\\frac{P+Q}{2})\n$$\n\n> 对于 JS 散度，当两个分布完全不重合时，JS 散度始终为 1（$\\log 2$）","cooked":"

杂项

会持续更新

正则化

正则化 (数学) - 维基百科，自由的百科全书

Weight Decay - L2 正则项

we multiply the sum of squares with another smaller number. This number is called weight decay or wd.

Our loss function now looks as follows:

Loss = MSE(y_hat, y) + wd * sum(w^2)

减小模型复杂度（弹性），降低过拟合的概率

残差（residual）和残差网络

\n\n

CNN 输出大小公式

$W$ 为输入大小， $F$ 为卷积核大小， $P$ 为填充大小（Padding）， $S$ 为步长（Stride）， $N$ 为输出大小，则有：

\nN = \\frac {W - F + 2P} {S} + 1\n

熵（Entropy）；KL 散度（KL Divergence）

熵

若一个离散随机变量 $\\mathbf{X}$ 的可能取值为 $X=\\{x_1,x_2,\\cdots,x_n\\}$ ，而对应的概率为 $p_i = p(X = x_i)$ ，则随机变量 $\\mathbf{X}$ 的熵定义为：

\nH(X) = - \\sum_{i=1}^{n} p(x_i) \\log p(x_{i})\n

交叉熵

针对上述离散变量的概率分布 $p(x)、q(x)$ 而言，其交叉熵定义为：

\nH(p,q) = - \\sum_{x} p(x) \\log q(x)\n

\n
在信息论中，交叉熵可认为是对预测分布 $q(x)$ 用真实分布 $p(x)$ 来进行编码时所需要的信息量大小。
\n

KL 散度（KL Divergence）

KL 散度可以用来衡量两个分布之间的差异：

\n\\begin{aligned}\nD_{KL}(p||q) = & H(p,q)-H(p) \\\\\n=& - \\sum_{x} p(x) \\log q(x) - \\sum_{x}-p(x)\\log p(x)\\\\\n=& -\\sum_{x}p(x) \\log \\frac{q(x)}{p(x)}\n\\end{aligned}\n

\n
交叉熵损失函数实际上就是 KL 散度，只不过因为分类问题的特殊性，其式与交叉熵相同：
\n

在分类问题中，我们通过 softmax 函数得到一个分布，我们要通过 KL 散度 衡量这个分布和目标分布（一个 one-hot 向量）的差异性：

\nD_{KL}(p||q) = H(p,q)-H(p)\n

我们易知 $H(p) = 0$ ；且对于真实分布（一个 one-hot 向量）只有概率为 1 的一项是非 0 的，所以有以下化简：

\n\\begin{aligned}\nD_{KL}(p||q) = & H(p,q)\\\\\n=& - \\sum_{x} p(x) \\log q(x)\\\\\n=& -\\log \\text{softmax} \\ q(x_i)\\\\\n=& -\\log \\frac{\\exp (x_i)}{\\sum_j \\exp (x_j)}\n\\end{aligned}\n

KL 散度的数学性质

正定性

\nD_{KL}(p||q) \\ge 0\n

不对称性

KL 散度并不是一个真正的度量或者距离，因为它不具有对称性：

\nD(p || q) \\ne D(q || p)\n

JS 散度（JS Divergence）

JS 散度度量了两个概率分布的相似度，基于 KL 散度的变体，解决了 KL 散度非对称的问题。一般地，JS 散度是对称的，其取值是 0 到 1 之间。

\nJSD(P||Q)=\\frac{1}{2}D_{KL}(P||\\frac{P+Q}{2})+\\frac{1}{2}D_{KL}(Q||\\frac{P+Q}{2})\n

\n
对于 JS 散度，当两个分布完全不重合时，JS 散度始终为 1（ $\\log 2$ ）
\n

\n","createdAt":"2024-01-22 06:09:52.136693","updatedAt":"2024-01-22 06:09:52.136693","replyToPostNumber":null,"replyCount":0,"likeCount":0,"wordCount":644,"liked":false,"bookmarked":false,"polls":[],"reactions":[],"wiki":false,"version":1,"publicVersion":1,"deletedAt":null,"isAnonymous":false,"user":{"id":31,"username":"TsuinoSora","name":null,"avatarUrl":"/uploads/image/77/06/7706643ba6d3ee1ec9d0070db0ad8a22bed8651fa8eb829cf44df4d20fa24b3b.avif","admin":false,"moderator":true,"signatureRaw":null,"signatureCooked":null}},{"id":22453,"postNumber":13,"raw":"这课看着很不错诶。我也来学一学吧。\n我似乎没寻找到他作业如何发布的，莫非是 private 吗？","cooked":"

这课看着很不错诶。我也来学一学吧。
\n我似乎没寻找到他作业如何发布的，莫非是 private 吗？

\n","createdAt":"2024-01-22 13:06:08.988707","updatedAt":"2024-01-22 13:06:31.376452","replyToPostNumber":null,"replyCount":1,"likeCount":0,"wordCount":39,"liked":false,"bookmarked":false,"polls":[],"reactions":[],"wiki":false,"version":1,"publicVersion":1,"deletedAt":null,"isAnonymous":false,"user":{"id":346,"username":"hoist","name":null,"avatarUrl":"/uploads/image/0d/cd/0dcdf69d213b6798002779ad4b729567cd4d889ee4a56c07a8c4454e8b563219.avif","admin":false,"moderator":false,"signatureRaw":null,"signatureCooked":null}},{"id":22461,"postNumber":14,"raw":"李宏毅的课吗，作业就在课程主页下面。个人认为他的课程坡度还是挺大的，光听不练/不记笔记感觉还是差点。每一年作业基本都是一样的，2022 的做的不好可以试试 2021，2023 的。","cooked":"

李宏毅的课吗，作业就在课程主页下面。个人认为他的课程坡度还是挺大的，光听不练/不记笔记感觉还是差点。每一年作业基本都是一样的，2022 的做的不好可以试试 2021，2023 的。

\n","createdAt":"2024-01-22 15:41:18.130584","updatedAt":"2024-01-22 15:41:18.130584","replyToPostNumber":13,"replyCount":1,"likeCount":1,"wordCount":75,"liked":false,"bookmarked":false,"polls":[],"reactions":[{"reaction":"❤️","count":1,"users":[{"id":3594,"username":"zecrious","name":"勿1小win而不麻"}],"hasReacted":false}],"wiki":false,"version":1,"publicVersion":1,"deletedAt":null,"isAnonymous":false,"user":{"id":31,"username":"TsuinoSora","name":null,"avatarUrl":"/uploads/image/77/06/7706643ba6d3ee1ec9d0070db0ad8a22bed8651fa8eb829cf44df4d20fa24b3b.avif","admin":false,"moderator":true,"signatureRaw":null,"signatureCooked":null}},{"id":22462,"postNumber":15,"raw":"我基本是他的课和他的作业和李沐的带读论文一起进行，中途还有无数次专门去找资料学习某个小知识。","cooked":"

我基本是他的课和他的作业和李沐的带读论文一起进行，中途还有无数次专门去找资料学习某个小知识。

\n","createdAt":"2024-01-22 15:43:02.914965","updatedAt":"2024-01-22 15:43:02.914965","replyToPostNumber":14,"replyCount":0,"likeCount":0,"wordCount":46,"liked":false,"bookmarked":false,"polls":[],"reactions":[],"wiki":false,"version":1,"publicVersion":1,"deletedAt":null,"isAnonymous":false,"user":{"id":31,"username":"TsuinoSora","name":null,"avatarUrl":"/uploads/image/77/06/7706643ba6d3ee1ec9d0070db0ad8a22bed8651fa8eb829cf44df4d20fa24b3b.avif","admin":false,"moderator":true,"signatureRaw":null,"signatureCooked":null}},{"id":22531,"postNumber":16,"raw":"真强","cooked":"

真强

\n","createdAt":"2024-01-23 13:48:59.743638","updatedAt":"2024-01-23 13:48:59.743638","replyToPostNumber":null,"replyCount":0,"likeCount":0,"wordCount":2,"liked":false,"bookmarked":false,"polls":[],"reactions":[],"wiki":false,"version":1,"publicVersion":1,"deletedAt":null,"isAnonymous":false,"user":{"id":53,"username":"Rachel","name":null,"avatarUrl":"/uploads/image/48/ae/48ae0d54cf2d42d31c7886b6d070d7ad42c6c0519f22149410202cb3d41cd731.avif","admin":false,"moderator":false,"signatureRaw":null,"signatureCooked":null}},{"id":25038,"postNumber":17,"raw":"# GAN\n\n[GAN.pdf|attachment](/uploads/pdf/a0/09/a009a2ca080834e283f0006d34484282b0a32027e43fa4e1a3873bc23cc553cb.pdf) (3.7 MB)","cooked":"

GAN

GAN.pdf|attachment (3.7 MB)

\n","createdAt":"2024-02-29 09:29:23.788235","updatedAt":"2024-02-29 09:30:07.493152","replyToPostNumber":null,"replyCount":0,"likeCount":0,"wordCount":5,"liked":false,"bookmarked":false,"polls":[],"reactions":[],"wiki":false,"version":1,"publicVersion":1,"deletedAt":null,"isAnonymous":false,"user":{"id":31,"username":"TsuinoSora","name":null,"avatarUrl":"/uploads/image/77/06/7706643ba6d3ee1ec9d0070db0ad8a22bed8651fa8eb829cf44df4d20fa24b3b.avif","admin":false,"moderator":true,"signatureRaw":null,"signatureCooked":null}},{"id":25039,"postNumber":18,"raw":"Pytorch 要学多少学到哪里才能入门这个课呀","cooked":"

Pytorch 要学多少学到哪里才能入门这个课呀

\n","createdAt":"2024-02-29 09:37:02.474087","updatedAt":"2024-02-29 09:37:02.474087","replyToPostNumber":null,"replyCount":1,"likeCount":0,"wordCount":17,"liked":false,"bookmarked":false,"polls":[],"reactions":[],"wiki":false,"version":1,"publicVersion":1,"deletedAt":null,"isAnonymous":false,"user":{"id":3074,"username":"Trifolium","name":"白车草","avatarUrl":"/uploads/image/e9/21/e92121eac2b612e7ee2ed14db996c0fe51665ec0669ec0acc41dd23c17ae11cc.avif","admin":false,"moderator":false,"signatureRaw":null,"signatureCooked":null}},{"id":25046,"postNumber":19,"raw":"会基础的 Python 语法，遇到啥不会的直接搜索就能入门。\n\n另外不写作业的话一点代码都不写也没学","cooked":"

会基础的 Python 语法，遇到啥不会的直接搜索就能入门。

另外不写作业的话一点代码都不写也没学

\n","createdAt":"2024-02-29 10:01:58.771061","updatedAt":"2024-02-29 10:02:29.228892","replyToPostNumber":18,"replyCount":0,"likeCount":1,"wordCount":41,"liked":false,"bookmarked":false,"polls":[],"reactions":[{"reaction":"❤️","count":1,"users":[{"id":3074,"username":"Trifolium","name":"白车草"}],"hasReacted":false}],"wiki":false,"version":1,"publicVersion":1,"deletedAt":null,"isAnonymous":false,"user":{"id":31,"username":"TsuinoSora","name":null,"avatarUrl":"/uploads/image/77/06/7706643ba6d3ee1ec9d0070db0ad8a22bed8651fa8eb829cf44df4d20fa24b3b.avif","admin":false,"moderator":true,"signatureRaw":null,"signatureCooked":null}},{"id":25086,"postNumber":20,"raw":"我看这课 ML 2023 全部改成了生成式 AI、是觉得会全面迭代技术吗？你对此什么看法。\n另外我对这课程无比佩服，竟然能大幅度调整大纲","cooked":"

我看这课 ML 2023 全部改成了生成式 AI、是觉得会全面迭代技术吗？你对此什么看法。
\n另外我对这课程无比佩服，竟然能大幅度调整大纲

\n","createdAt":"2024-02-29 13:03:10.123187","updatedAt":"2024-02-29 13:03:10.123187","replyToPostNumber":null,"replyCount":1,"likeCount":0,"wordCount":57,"liked":false,"bookmarked":false,"polls":[],"reactions":[],"wiki":false,"version":1,"publicVersion":1,"deletedAt":null,"isAnonymous":false,"user":{"id":346,"username":"hoist","name":null,"avatarUrl":"/uploads/image/0d/cd/0dcdf69d213b6798002779ad4b729567cd4d889ee4a56c07a8c4454e8b563219.avif","admin":false,"moderator":false,"signatureRaw":null,"signatureCooked":null}},{"id":25103,"postNumber":21,"raw":"李宏毅老师的课跟热点跟的很勤。一直都是每年开课根据过去一年的 AI 发展较大幅度的更改大纲，提供比较新的技术。比如 RNN 自从 Transformer 大范围使用后就变成了选学内容，学生想学去看之前的录像就好了。\n\n2023 的版本虽然作业还是沿着 Regression -> CNN -> Attention 来的，但是上课内容基本全是生成式 AI，老师基本默认了不会走“回头路”一般。你们学生不会做作业那就看我以前上课的录像好了。\n\n再看今年 2024 老师开得课，直接没有 ML 课了，改成叫 GENERATIVE AI 了。他可能觉得讲过的不必再讲一遍吧。","cooked":"

李宏毅老师的课跟热点跟的很勤。一直都是每年开课根据过去一年的 AI 发展较大幅度的更改大纲，提供比较新的技术。比如 RNN 自从 Transformer 大范围使用后就变成了选学内容，学生想学去看之前的录像就好了。

2023 的版本虽然作业还是沿着 Regression -> CNN -> Attention 来的，但是上课内容基本全是生成式 AI，老师基本默认了不会走“回头路”一般。你们学生不会做作业那就看我以前上课的录像好了。

再看今年 2024 老师开得课，直接没有 ML 课了，改成叫 GENERATIVE AI 了。他可能觉得讲过的不必再讲一遍吧。

\n","createdAt":"2024-02-29 13:45:09.242772","updatedAt":"2024-02-29 13:45:09.242772","replyToPostNumber":20,"replyCount":0,"likeCount":1,"wordCount":205,"liked":false,"bookmarked":false,"polls":[],"reactions":[{"reaction":"❤️","count":1,"users":[{"id":346,"username":"hoist","name":null}],"hasReacted":false}],"wiki":false,"version":1,"publicVersion":1,"deletedAt":null,"isAnonymous":false,"user":{"id":31,"username":"TsuinoSora","name":null,"avatarUrl":"/uploads/image/77/06/7706643ba6d3ee1ec9d0070db0ad8a22bed8651fa8eb829cf44df4d20fa24b3b.avif","admin":false,"moderator":true,"signatureRaw":null,"signatureCooked":null}},{"id":25310,"postNumber":23,"raw":"# Bert & GPT\n\n[BERT-GPT.pdf|attachment](/uploads/pdf/cd/3c/cd3ce269eadc25332a8fd4104672d2872949f9e8b8e6320cbeb051561817faca.pdf) (1.4 MB)","cooked":"

Bert & GPT

BERT-GPT.pdf|attachment (1.4 MB)

\n","createdAt":"2024-03-02 06:20:07.900956","updatedAt":"2024-03-02 06:20:07.900956","replyToPostNumber":null,"replyCount":0,"likeCount":0,"wordCount":7,"liked":false,"bookmarked":false,"polls":[],"reactions":[],"wiki":false,"version":1,"publicVersion":1,"deletedAt":null,"isAnonymous":false,"user":{"id":31,"username":"TsuinoSora","name":null,"avatarUrl":"/uploads/image/77/06/7706643ba6d3ee1ec9d0070db0ad8a22bed8651fa8eb829cf44df4d20fa24b3b.avif","admin":false,"moderator":true,"signatureRaw":null,"signatureCooked":null}}],"hasMore":true,"lastReadPostNumber":1,"hasUrlPostNumber":true}