The previous blog shows how to build a neural network manualy from scratch in numpy with matrix/vector multiply and add. Although there are many packages can do this easily and quickly with a few lines of scripts, it is still a good idea to understand the logic behind the packages. This part is from a good blog which use an example predicitng the words in the sentence to explain how to build RNN manually. RNN is a little more complicated than the neural network in the previous blog because the current time status and ourput in RNN will depends on the status in the previous time. So the Backpropagation part will be more complicated. I try to give the details in mathematic formula about how to get the gradients recursively in the partial derivatives.

Use two examples to show how to build the neural network from scratch: define the activation function on each layer, define the lost function, and calculate the partial derivatives using chain rules.

这是以前看卷积神经网络做的笔记.我很喜欢另外一种学习方法:不是从理解什么是CNN开始，也不是去关注后面的数学公式，也不是怎么来用随机梯度下降的办法来最小化cross-entropy函数求解参数。而是直接用tensorflow来build一个简单的卷积神经网络，观察它是怎么工作的，最后的结果是什麽。理解了这些以后，然后在回过头来理解那些理论的部分和数学的部分。这篇blog的绝大部分内容都是来自于stanford的卷积神经网络课程的讲义。用中文整理出来，因为还是看中文快一些。

Recently I read an article from zhihu. It talks about bitcoin arbitrage traiding with risk hedging. It reminds me the 'one upon a time' in full swing.

这是我觉得很好的一篇知乎文章，原作者不仅简明扼要的把一个问题转化为数学问题，把书本上关于Ito积分应用到现实问题中，而且给出了一个Simulation解，做到了让人真正明白是怎么回事。

SQLAlchemy introduction

MNIST is used as example in tensorflow. Here are the examples of how to use logistic regression and KNN(K nearest neighbors) to analyze this data.

summary about http.request, GET and POST, form, and route in flask

It is to provide an online IFRS9/CECL lifetime credit loss calculation engine.