Scratch 平台的神经网络（R 语言） #1867

CACppuccino · 2017-07-07T18:18:04Z

请校对 @sqrthree #1856

update20170706

lsvih · 2017-07-11T04:50:15Z

校对认领

linhe0x0 · 2017-07-11T04:50:20Z

@lsvih 好的呢 🍺

lsvih · 2017-07-11T04:56:24Z

TODO/neural-networks-from-scratch-in-r.md


-This post is for those of you with a statistics/econometrics background but not necessarily a machine-learning one and for those of you who want some guidance in building a neural-network from scratch in R to better understand how everything fits (and how it doesn’t).
+这篇文章是针对那些有着统计或经济学背景，却缺少机器学习知识，希望能够借助一些指导，通过 R 语言的 Scratch 平台更好的理解如何正确地使用机器学习。


->

这篇文章是为了那些有着统计或经济学背景，却缺少机器学习知识的人们所写，也是为了那些希望能够借助一些指导，通过 R 语言的 Scratch 平台更好地正确理解如何使用机器学习的人们所写。

这样还是有些不通顺，希望作者再看看怎么翻比较好

lsvih · 2017-07-11T04:56:35Z

TODO/neural-networks-from-scratch-in-r.md


-This post is for those of you with a statistics/econometrics background but not necessarily a machine-learning one and for those of you who want some guidance in building a neural-network from scratch in R to better understand how everything fits (and how it doesn’t).
+这篇文章是针对那些有着统计或经济学背景，却缺少机器学习知识，希望能够借助一些指导，通过 R 语言的 Scratch 平台更好的理解如何正确地使用机器学习。

 Andrej Karpathy [wrote](https://medium.com/@karpathy/yes-you-should-understand-backprop-e2f06eab496b) that when CS231n (Deep Learning at Stanford) was offered:


删除原文

lsvih · 2017-07-11T04:59:51Z

TODO/neural-networks-from-scratch-in-r.md


-Why bother with backpropagation when all frameworks do it for you automatically and there are more interesting deep-learning problems to consider?
+如果所有的反向传播算法已经被框架为你自动计算完毕了，你又何苦折磨自己而不去探寻更多有趣的深度学习问题呢？


->

如果框架已经为你自动计算完了所有的反向传播算法

lsvih · 2017-07-11T05:00:57Z

TODO/neural-networks-from-scratch-in-r.md


-Starting from a linear regression we will work through the maths and the code all the way to a deep-neural-network (DNN) in the accompanying R-notebooks. Hopefully to show that very little is actually new information.
+从线性回归开始，借着 R 语言笔记，通过解决一系列的数学和编程问题直至深度神经网络（DNN）。希望能够显示出来你需要学习的新知识其实只有很少的一部分。


R 语言笔记 -> R-notebook

lsvih · 2017-07-11T05:02:28Z

TODO/neural-networks-from-scratch-in-r.md


 ![](https://cdn-images-1.medium.com/max/800/1*nzwaX3XqlaRGAf0kpN9ShA.png)

-**Notebooks:**
+**笔记**


Notebook：

// 这些都是 ipython notebook，感觉直接翻成笔记不太合适，不如保留原名，您觉得呢？下面的“笔记”同理。

感觉内容还是笔记，或者说 ipython notebook 有什么很特别的地方？这个我不是很清楚

lsvih · 2017-07-11T05:36:05Z

TODO/neural-networks-from-scratch-in-r.md

@@ -376,9 +377,10 @@ We can visualise this map with the following function:

 ![](https://cdn-images-1.medium.com/max/800/1*s-TR-n5n2-4ZwwZ962X3LQ.png)

-Running this function we notice how computationally intensive the process is (compared to a standard fully-connected layer). If these feature maps are not useful ‘features’ (i.e. the loss is difficult to decrease when these are used) then back-propagation will mean we will get different weights which correspond to different feature-maps; which will become more useful to make the classification.
+在运行这个函数的时候我们意识到了整个过程是如何地高密度计算（与标准的全连接神经层相比）。如果这些 feature map 合不是那些那么有用的集合（也就是说，损失在被使用的时候很难被降低）然后反向传播会意味着我们将会得到不同的权重，与不同的 feature map 合相关联，对于构造聚类很有帮助。


，对于构造聚类很有帮助。 -> 。因此这个函数对于计算聚类很有用。

lsvih · 2017-07-11T05:36:15Z

TODO/neural-networks-from-scratch-in-r.md


    filter_map <- lapply(X=c(1:64), FUN=function(x){
        # Random matrix of 0, 1, -1
        conv_rand <- matrix(sample.int(3, size=9, replace = TRUE), ncol=3)-2
        convolution(input_img = r_img, filter = conv_rand, show=FALSE, out=TRUE)
    })

-We can visualise this map with the following function:
+我们可以用以下的函数可视化这个地图：


地图 -> map

lsvih · 2017-07-11T05:37:04Z

TODO/neural-networks-from-scratch-in-r.md


    conv_emboss <- matrix(c(2,0,0,0,-1,0,0,0,-1), nrow = 3)
    convolution(input_img = r_img, filter = conv_emboss)

-You can check the notebook to see the result, however this seems to extract the edges from a picture. Other, convolutions can ‘sharpen’ an image, like this 3x3 filter:
+你可以查看笔记来看结果，然而这看起来是从图片中提取线段。否则，卷积可以‘尖锐化’一张图片，就像一个3*3的过滤器：


尖锐化 -> 锐化

lsvih · 2017-07-11T05:37:17Z

TODO/neural-networks-from-scratch-in-r.md


    conv_emboss <- matrix(c(2,0,0,0,-1,0,0,0,-1), nrow = 3)
    convolution(input_img = r_img, filter = conv_emboss)

-You can check the notebook to see the result, however this seems to extract the edges from a picture. Other, convolutions can ‘sharpen’ an image, like this 3x3 filter:


笔记同上

lsvih · 2017-07-11T05:37:40Z

TODO/neural-networks-from-scratch-in-r.md


-In the previous example we looked at a standard neural-net classifying handwritten text. In that network each neuron from layer i, was connected to each neuron at layer j — our ‘window’ was the whole image. This means if we learn what the digit “2” looks like; we may not recognise it when it is written upside down by mistake, because we have only seen it upright. CNNs have the advantage of looking at small bits of the digit “2” and finding patterns between patterns between patterns. This means that a lot of the features it extracts may be immune to rotation, skew, etc. For more detail, Brandon explains [here ](https://www.youtube.com/watch?v=FmpDIaiMIeA)what a CNN actually is in detail.
+在之前的例子中我们观察的是一个标准的神经网络对手写字体的归类。在神经网络中的 i 层的每个神经元，与 j 层的每个神经元相连-我们所框中的是整个图像（译者注：与 CNN 之前的5*5像素的框不同）。这意味着如果我们学习了数字2的样子，我们可能无法在它被错误地颠倒的时候识别出来，因为我们只见过它正的样子。CNN 在观察数字2的小的比特时并且在比较样式的时候有很大的优势。这意味着很多被提取出的特征对各种旋转，歪斜等是免疫的（译者注：即适用于所有变形）。对于更多的细节，Brandon 在[这里](https://www.youtube.com/watch?v=FmpDIaiMIeA)解释了什么是真正的 CNN。


数字中文空格

CACppuccino

@sqrthree @lsvih 根据校对已经在本地更改，有个别地方不是很同意。
还需要等二校吗，已经一周了

CACppuccino · 2017-07-11T15:37:15Z

TODO/neural-networks-from-scratch-in-r.md


 ![](https://cdn-images-1.medium.com/max/800/1*nzwaX3XqlaRGAf0kpN9ShA.png)

-**Notebooks:**
+**笔记**


感觉内容还是笔记，或者说 ipython notebook 有什么很特别的地方？这个我不是很清楚

CACppuccino · 2017-07-11T15:42:44Z

TODO/neural-networks-from-scratch-in-r.md


-When we have just two mutually-exclusive outcomes we would use a binomial logistic regression. With more than two outcomes (or “classes”), which are mutually-exclusive (e.g. this plane will be delayed by less than 5 minutes, 5–10 minutes, or more than 10 minutes), we would use a multinomial logistic regression (or “softmax”). In the case of many (n)classes that are not mutually-exclusive (e.g. this post references “R” and “neural-networks” and “statistics”), we can fit n-binomial logistic regressions.
+当我们只有两个互斥的结果时我们将使用一个二项逻辑回归。当候选结果（或者分类）多于两个时，即多项互斥（例如：这架飞机延误时间可能在5分钟内、5-10分钟或多于10分钟），我们将使用多项逻辑回归（或者“Softmax 回归”）（译者注：Softmax 函数是逻辑函数的一种推广，更多知识见[知乎](https://www.zhihu.com/question/23765351)）。在这种情况下许多类别不是互斥的（例如：这篇文章中的“R”，“神经网络”和“统计学”），我们可以采用二项式逻辑回归（译者注：不是二项逻辑回归）。


这里我想强调的是二项与二项式的区别，主要防止读者马虎。你觉得怎样解释比较好呢？

CACppuccino · 2017-07-11T15:45:57Z

TODO/neural-networks-from-scratch-in-r.md

@@ -95,24 +94,24 @@ This can be implemented in R, like so:
        beta_hat <- beta_hat - (lr*delta)
      }

-Running this for 200 iterations gets us to same gradient and coefficient as the closed-form solution. Aside from being a stepping stone to a neural-network (where we use GD), this iterative method can be useful in practice when the the closed-form solution cannot be calculated because the matrix is too big to invert (to fit into memory).
+200次的迭代之后我们会得到和闭包方法一样的梯度与参数。除了这代表着我们的进步意外（我们使用了 GD），这个迭代方法在当闭包方法因矩阵过大而无法计算逆的时候也非常有用（因为有内存的限制）。


改成了无法计算矩阵的逆并加入了逗号断句，个人认为把具体原因解释出来会更清楚些

CACppuccino · 2017-07-11T15:46:39Z

TODO/neural-networks-from-scratch-in-r.md


-1. We use an ‘activation’/link function called the logistic-sigmoid to squash the output to a probability bounded by 0 and 1
-2. Instead of minimising the quadratic loss we minimise the negative log-likelihood of the bernoulli distribution
+1. 我们使用一种称为 logistic-sigmoid 的 ‘激活’/链接函数来将输出压缩至0到1的范围内


这个意思就不一样了

CACppuccino · 2017-07-11T16:10:52Z

TODO/neural-networks-from-scratch-in-r.md


-1. We use an ‘activation’/link function called the logistic-sigmoid to squash the output to a probability bounded by 0 and 1
-2. Instead of minimising the quadratic loss we minimise the negative log-likelihood of the bernoulli distribution
+1. 我们使用一种称为 logistic-sigmoid 的 ‘激活’/链接函数来将输出压缩至0到1的范围内


你可以看下 sigmoid 函数的图像，是一条连续的曲线，后面的二项分类还是要规定一条线的，这里的sigmoid只是产生0~1的数值

linhe0x0 · 2017-07-11T16:15:10Z

不等了，目前已有的校对意见有争议的讨论下，明天就 merge。

lsvih · 2017-07-12T01:13:37Z

@CACppuccino sigmoid function 是我的错，看到 sigmoid 就直接相当然了
那个 notebook 坚持意见。。因为我阅读译文的时候，每次看到“笔记”就会迷惑一下 —— 直到看到后面付的链接是 jupyter notebook 才明白。如果读者看到“notebook”，自然而然就会了解指的是文中所附的 notebook 文件。

另外您本地改好之后麻烦 commit 一下哈

可以merge了

CACppuccino · 2017-07-12T01:59:40Z

@sqrthree 根据校对者建议修改完毕可以merge了

linhe0x0 · 2017-07-12T03:04:39Z

已经 merge 啦~ 快快麻溜发布到掘金专栏然后给我发下链接，方便及时添加积分哟。

CACppuccino · 2017-07-12T06:10:26Z

@sqrthree 老哥你在一周前把这篇就给分享了。。我该怎么办

linhe0x0 · 2017-07-12T06:36:26Z

😄 发原创呀

CACppuccino · 2017-07-13T11:59:21Z

@sqrthree 已经发布 https://juejin.im/post/5965cf75f265da6c4741adc4

CACppuccino added 2 commits July 6, 2017 10:23

Merge pull request #3 from xitu/master

3d27a86

update20170706

初稿完毕

2540747

linhe0x0 added the 校对认领 label Jul 7, 2017

linhe0x0 mentioned this pull request Jul 7, 2017

Scratch 平台的神经网络（R 语言） #1856

Closed

CACppuccino changed the title ~~初稿完毕~~ Scratch 平台的神经网络（R 语言） Jul 7, 2017

linhe0x0 added the AI label Jul 8, 2017

linhe0x0 added the 正在校对 label Jul 11, 2017

lsvih requested changes Jul 11, 2017

View reviewed changes

CACppuccino commented Jul 11, 2017

View reviewed changes

linhe0x0 added the merging-soon-if-no-other-reviewer label Jul 11, 2017

lsvih approved these changes Jul 12, 2017

View reviewed changes

修改完毕

cda95f2

可以merge了

linhe0x0 approved these changes Jul 12, 2017

View reviewed changes

linhe0x0 merged commit 248377f into xitu:master Jul 12, 2017

linhe0x0 added 翻译完成 and removed merging-soon-if-no-other-reviewer 校对认领正在校对 labels Jul 12, 2017

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Scratch 平台的神经网络（R 语言） #1867

Scratch 平台的神经网络（R 语言） #1867

CACppuccino commented Jul 7, 2017

lsvih commented Jul 11, 2017

linhe0x0 commented Jul 11, 2017

lsvih Jul 11, 2017

lsvih Jul 11, 2017

lsvih Jul 11, 2017

lsvih Jul 11, 2017

lsvih Jul 11, 2017

CACppuccino Jul 11, 2017

lsvih Jul 11, 2017

lsvih Jul 11, 2017

lsvih Jul 11, 2017

lsvih Jul 11, 2017

lsvih Jul 11, 2017

CACppuccino left a comment

CACppuccino Jul 11, 2017

CACppuccino Jul 11, 2017

CACppuccino Jul 11, 2017

CACppuccino Jul 11, 2017

CACppuccino Jul 11, 2017

linhe0x0 commented Jul 11, 2017

lsvih commented Jul 12, 2017

CACppuccino commented Jul 12, 2017

linhe0x0 commented Jul 12, 2017

CACppuccino commented Jul 12, 2017

linhe0x0 commented Jul 12, 2017

CACppuccino commented Jul 13, 2017


		This post is for those of you with a statistics/econometrics background but not necessarily a machine-learning one and for those of you who want some guidance in building a neural-network from scratch in R to better understand how everything fits (and how it doesn’t).
		这篇文章是针对那些有着统计或经济学背景，却缺少机器学习知识，希望能够借助一些指导，通过 R 语言的 Scratch 平台更好的理解如何正确地使用机器学习。


		Why bother with backpropagation when all frameworks do it for you automatically and there are more interesting deep-learning problems to consider?
		如果所有的反向传播算法已经被框架为你自动计算完毕了，你又何苦折磨自己而不去探寻更多有趣的深度学习问题呢？


		Starting from a linear regression we will work through the maths and the code all the way to a deep-neural-network (DNN) in the accompanying R-notebooks. Hopefully to show that very little is actually new information.
		从线性回归开始，借着 R 语言笔记，通过解决一系列的数学和编程问题直至深度神经网络（DNN）。希望能够显示出来你需要学习的新知识其实只有很少的一部分。


		In the previous example we looked at a standard neural-net classifying handwritten text. In that network each neuron from layer i, was connected to each neuron at layer j — our ‘window’ was the whole image. This means if we learn what the digit “2” looks like; we may not recognise it when it is written upside down by mistake, because we have only seen it upright. CNNs have the advantage of looking at small bits of the digit “2” and finding patterns between patterns between patterns. This means that a lot of the features it extracts may be immune to rotation, skew, etc. For more detail, Brandon explains [here ](https://www.youtube.com/watch?v=FmpDIaiMIeA)what a CNN actually is in detail.
		在之前的例子中我们观察的是一个标准的神经网络对手写字体的归类。在神经网络中的 i 层的每个神经元，与 j 层的每个神经元相连-我们所框中的是整个图像（译者注：与 CNN 之前的5*5像素的框不同）。这意味着如果我们学习了数字2的样子，我们可能无法在它被错误地颠倒的时候识别出来，因为我们只见过它正的样子。CNN 在观察数字2的小的比特时并且在比较样式的时候有很大的优势。这意味着很多被提取出的特征对各种旋转，歪斜等是免疫的（译者注：即适用于所有变形）。对于更多的细节，Brandon 在[这里](https://www.youtube.com/watch?v=FmpDIaiMIeA)解释了什么是真正的 CNN。


		When we have just two mutually-exclusive outcomes we would use a binomial logistic regression. With more than two outcomes (or “classes”), which are mutually-exclusive (e.g. this plane will be delayed by less than 5 minutes, 5–10 minutes, or more than 10 minutes), we would use a multinomial logistic regression (or “softmax”). In the case of many (n)classes that are not mutually-exclusive (e.g. this post references “R” and “neural-networks” and “statistics”), we can fit n-binomial logistic regressions.
		当我们只有两个互斥的结果时我们将使用一个二项逻辑回归。当候选结果（或者分类）多于两个时，即多项互斥（例如：这架飞机延误时间可能在5分钟内、5-10分钟或多于10分钟），我们将使用多项逻辑回归（或者“Softmax 回归”）（译者注：Softmax 函数是逻辑函数的一种推广，更多知识见[知乎](https://www.zhihu.com/question/23765351)）。在这种情况下许多类别不是互斥的（例如：这篇文章中的“R”，“神经网络”和“统计学”），我们可以采用二项式逻辑回归（译者注：不是二项逻辑回归）。

Scratch 平台的神经网络（R 语言） #1867

Scratch 平台的神经网络（R 语言） #1867

Conversation

CACppuccino commented Jul 7, 2017

lsvih commented Jul 11, 2017

linhe0x0 commented Jul 11, 2017

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

CACppuccino left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

linhe0x0 commented Jul 11, 2017

lsvih commented Jul 12, 2017

CACppuccino commented Jul 12, 2017

linhe0x0 commented Jul 12, 2017

CACppuccino commented Jul 12, 2017

linhe0x0 commented Jul 12, 2017

CACppuccino commented Jul 13, 2017