python可以这样学读后感_《Python深度学习》读后感

Background

Last month, I teamed up with two friends to take part in a competition which requires to recognize bank cards basing on deep learning.

However, I did not know about deep learning at all! Actually, I messed my 校创 last year because I did not understand neural network and I could not build one.

Then I found a book in my shelf called Python深度学习 by N.D.Lewis. It’s a very thin book. And I finished reading it this morning.

Now, you may ask,

Do you know how to do the bank card recognition?

My answer is NO.

Have you built a neural network?

My answer is also NO.

Can you?

I don’t know.

What do you know?

Some terms.

That’s why I decide to call this post 读后感 rather than 教程.

What is Deep Learning & what is DNN?深度学习是机器学习与神经网络、人工智能、图形化建模、优化、模式识别和信号处理等技术融合后产生的一个领域。

简单来说，深度神经网络是一种用多层抽象来表示概念或特征的方式。

Neural Networks usually consist of three parts: an input layer, the hidden layer(s) and an output layer. There are neurons in each layer, which take several inputs, produce an output according to the weighted sum and the activation function, then pass the output forward, which work like the real neurons.(Don’t ask me about this. I haven’t studied Biology for 4 years.)(You can also find some pictures to help you understand.)

When the network is learning, the neurons change their weights to make the output closer to the correct output. Sometimes it can even find the inner connections in the data without the correct output.

In my opinion, DNN is just NN with multiple hidden layers.

It is very different from my understanding of programming. Without machine learning and neural network, we can find or invent the rules to make the programs work as we expect.

However, we can also make the programs produce almost correct outputs even if we can not find the rules or patterns with the help of machine leaning and neural network.

Why is it powerful?隐藏层是DNN的秘密所在。

深度学习的威力来自于用适量的并行非线性步骤对非线性数据进行分类或预测的能力。

如果决策规则可以通过复杂函数进行数学建模，那么就可以使用DNN。重要的是，你不需要事先知道函数的形式。

If you have a lot of data, and you want to find out the inner connections to do classfication or prediction, which do not have to be perfectly precise, DNN is a good choice.

Some techniques

After reading the book, you may draw conclusions like these:

The learning rate should not be too high or too low.

Too many or too few neurons will affect the performance.

Dropout may improve the performance.

Perhaps another activation function works better.

……

There is no optimal choice, we have to find the most suitable model for a specific case in experiments. Experimenting is also called 调参 in Chinese.

So I still know nothing about DNN.

深度学习需要耐心和试验。

数据科学是由偶尔的成功点缀的一系列失败。

The author intends to encourage us to embrace failures. However, I feel more than depressed.

The author mainly use sknn.mlp to develop his neural networks, and he uses python2. But I only have python3 on my laptop…

Anyway, it’s definitely a good book for starters. It’s not difficuit to understand and it’s not so thick that you may feel desperate.

See you Next time.

P.S.

I should have done my lab work and reviewed for the final exam of CE162.

But I feel like updating my blog.

This is my second post in English, I will still use English if I am not desperate.

My next post is probably about F1 2016, a game that I have been playing on my iPad.