Andrew Ng "Machine Learning" Jupyter Notebook version release notes! Illustrations, formulas, exercises have a

Click on the top " AI proper way ", select the "star" Public No.

Heavy dry goods, the first time served

I long ago wrote in the article " How to machine learning portal ", in relation devaluation over Andrew Ng opened on Coursera "Machine Learning" course. This course is the biggest feature is basically no complex mathematical theories and formulas are derived, is ideal for entry!

This course is posted on Coursera, many readers easy to confuse it with another course CS229 Andrew Ng's. In fact, "Machine Learning" on today talking about Coursera easier.

The curriculum addresses:

https://www.coursera.org/learn/machine-learning

The official introduction to this course is: This course will introduce a wide range of machine learning, data mining and statistical pattern recognition. Related Topics include: (i) supervised learning (parametric and non-parametric algorithms, support vector machine, kernel function and neural networks). (Ii) unsupervised learning (clustering, dimension reduction, recommendation system and deep learning). (Iii) examples of machine learning (bias / variance theory; machine learning and innovation in the field of AI). The course will quote many examples and applications, you also need to learn how to apply learning in different areas of algorithms, such as intelligent robots (perception and control), text comprehension (Internet search and spam filtering), computer vision, medical informatics, audio, database mining and other fields.

This course covers the main basic knowledge of machine learning, such as: linear regression, logistic regression, support vector machines, neural networks, K-Means, anomaly detection, and so on. And of course there are no complicated formulas are derived and theoretical analysis. Ng purpose is to allow the machine learning beginners can quickly learn the knowledge of the entire machine as a whole have more understanding for quick entry.

I have put together a mind map of the course:

If readers want to see a clear version of this mind map (including the original .xmind file), the following is the download link:

https://pan.baidu.com/s/16065BpNAP7JEx_PpFHLSOw

提取码：xcmi

好了，回归正题！这门课的笔记网上有很多，但今天推荐一份不错的全面的笔记。该笔记的特点是所有内容都是基于 Jupyter Notebook 写的，集图片、公式、代码、练习题于一体，非常方便使用！

下面来详细看一下！

这份笔记的作者是 Halfrost-Field 冰霜之地，笔记发布地址为：

https://github.com/halfrost/Halfrost-Field

所属于 Machine Learning 一栏。

笔记包含了课程 11 周完整内容，每一周单独对应一个 Jupyter Botebook 文件。

下面举几个代表看一下！

1. 支持向量机（SVM）

打开 Support_Vector_Machines.ipynb文件，逻辑回归损失函数的公式推导：

高斯核函数的可视化理解：

还有作者对 SVM 重要参数的归纳总结：

笔记内容结合了吴恩达的课件，归纳的比较到位！

每一份笔记的最后，还配备相应的习题：

2. 神经网络学习

打开 Neural_Networks_Learning.ipynb 文件，Backpropagation Algorithm 反向传播算法的推导：

反向传播算法实现：

function g = sigmoid(z)
    g = 1.0 ./ (1.0 + exp(-z));
end

function g = sigmoidGradient(z)
    g = sigmoid(z) .* (1 - sigmoid(z));
end

delta3 = a3 - Y;

Theta2_grad = 1 / m * delta3' * [ones(m, 1), a2] + ...
  lambda / m * [zeros(K, 1), Theta2(:, 2:end)]; # 正规化部分

delta2 = (delta3 * Theta2 .* sigmoidGradient([ones(m, 1), z2]));
delta2 = delta2(:, 2:end); # 反向计算多一个偏移参数误差，除去

Theta1_grad = 1 / m * delta2' * [ones(m, 1), a1] + ...
  lambda / m * [zeros(H, 1), Theta1(:, 2:end)]; # 正规化部分