Perhaps the first question you want to understand is the difference and connection between artificial intelligence (AI), machine learning (ML), and deep learning (DL), The picture below tells you clearly,
1 What is machine learning
Since elementary school, we have been dealing with "functions", such as a straight line. If the input is, it is easy to get.
If the input is a piece of audio , you need to simulate something similar to a function, and its output is the audio content
If the input is a picture , you want to simulate a complex function so that its output is what this picture is
If the input is a Go board , you want to simulate a complex function and let it tell you where to go next
Yes, you can think that machine learning is looking for this complex "function", because it is complex, it is uncertain, and it is nonlinear, so you need to design some algorithms for the machine to learn this complex "function" What exactly is it.
Yes, this is machine learning. People will make some decisions, some judgments, and you hope to use some existing data to train the machine so that the machine can also learn how to make decisions, or even do better
2 what is deep learning
Deep learning is a subcategory of machine learning. In other words, it is actually a method of realizing machine learning. With the rapid development of computer software and hardware, people realize that the human brain can be simulated through the neural network (Neural Network), and the word depth means neural The network has many layers.
Looking back, how does your brain think? How does your biology teacher tell you how signals are transmitted in the brain.
At present, deep learning mainly includes
Convolutional Neural Network
Recurrent Neural Network
Generative Adversarial Networks
Deep Reinforcement Learning
These major research directions.
How to start
How to start learning deep learning?
What to learn? No matter which direction you belong to, I think the first thing you should learn is CNN. This series of articles also formally introduces the various architectures of convolutional neural networks to achieve the purpose of getting you started.
How to learn?
books? Learning through books, this is a way that has not been too wrong since ancient times,
First of all, I would like to recommend the first authoritative deep learning textbook "Deep Learning》
In addition, the Chinese translation of this book also went on sale a few weeks ago.
Fortunately, you can download from deeplearningbook-chineseGet the PDF of this book .Of course, if you think your machine learning knowledge needs to be improved, you can try Zhou ZhihuaTeacher's "Machine Learning", or through Li Hang’s "Statistical Learning Method"Make up for it?
In addition, please support the genuine, please do not spread illegal PDF documents.
course? Since ancient times, people have been accustomed to learning by face-to-face teaching. Nowadays, the developed network has brought a lot of video resources for learning.
My most recommended is Fei-Fei Li The CS231n, the course is complete, complete information, teaching content is always at the forefront. The video of the course can also be easily found on YouTube.
CS231n: Convolutional Neural Networks for Visual Recognition
In addition, Zhihu also has a translation of the course notes.
Congratulations! CS231n Official Notes Authorized Translation Collection Published
另一个强力推荐的是台大 李宏毅 老师的课程
Machine Learning (2017,Spring)
Machine Learning (2016,Fall)
Machine Learning and having it deep and structured (2017,Spring)
这门课叫做MLDS,其实就是在讲Deep Learning，而且homework非常有意思，有GAN也有RL，值得推荐此外，李宏毅老师的《一日搞懂深度学习》，其 链接 也在老师的个人网站上，8月13号在台北还有他的talk《一日搞懂生成对抗网络》，之后应该也能在它的主页上找到
同样，还有 林軒田 老师的公开课《机器学习基石》和《机器学习技法》
也是中文授课。
讲得很好，PDF做得也很棒，毕竟老师曾是多届 KDD Cup 的冠军Machine Learning Techniques, Spring 2017
Machine Learning Foundations, Fall 2016
就在前段时间，Andrew Ng教授似乎又有了新动作，
如果你愿意，也可以尝试一下
实践？只有理论没有实践，俗称花把式。
个人认为比较好的方式就是写作业，比如上面课程的作业，比如重新实现github上的一些优秀代码，一定要花时间去实践，这和你要刷题才能领悟真谛是一个道理！
寻求帮助？上知乎，上github，上google，上stackoverflow，上jizhi
上CrossValidated，你需要专业的解答。机器？工欲善其事，必先利其器，请务必准备一张GPU，这样你才能愉快地玩耍！
你应该走过的战场
好，接下来说一说，你需要走过的战场。
你开始了解什么是图片的分类(Image Classification)
你开始用一些简单的线性模型来进行预测(SVM,KNN)
你开始研究什么是最优化，什么是损失函数(Loss function)和梯度下降
你开始研究什么是反向传播(Backpropagation)
你开始研究什么是神经网络(NN)
终于，你踏入了卷积神经网络的战场
你开始学习卷积层，池化，归一化等等
你开始研究 Tensorflow，Keras，Pytorch，Caffe
你开始读别人的代码，你开始自己写代码
你开始知道cifar10,知道ImageNet
你开始研究第一个CNN model: LeNet-5 - Yann LeCun
最开始你有必要研究一下CNN之父LeCun发明的LeNet，虽然现在已经没有用武之地，但是CNN便是从这里开始的。紧接着是一个突破性的架构 「AlexNet」：ImageNet Classification with Deep Convolutional Neural Networks
2012 ImageNet的冠军，真正让CNN开始流行起来然后是它的改进 「ZF Net」： Visualizing and Understanding Convolutional Networks
以及「Network in Network」: Network In Network
你开始逐渐知道一些CNN 训练常用的 「trick」
数据预处理(Data Augmentation)
权重初始化(Xavier/He Weight initial)
批量归一化(Batch Normalization)
正则化(L2/L1/Maxnorm/Dropout)
以及震惊一时的「Vgg Network」: Very Deep Convolutional Networks for Large-Scale Image Recognition
然后进入「GoogleNet」家族：
[v1] Going Deeper with Convolutions
[v2] Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift
[v3] Rethinking the Inception Architecture for Computer Vision
[v4] Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning
你开始知道微调(Fine-tune)，你开始不断改善你的代码
再入残差家族 「Residual Network」:
Deep Residual Learning for Image Recognition
Identity Mappings in Deep Residual Networks
「Wide Residual Network」: Wide Residual Networks
「ResNeXt」: Aggregated Residual Transformations for Deep Neural Networks
最后是 「DenseNet」 Densely Connected Convolutional Networks
至此，你的CNN学得就还算过得去了。
继续前行吧，世界是广阔的，你会达到远方美丽的绿洲！一定会的！
推荐阅读：
精选干货|近半年干货目录汇总----全是通俗易懂的硬货！欢迎阅读！
干货|台湾大学林轩田机器学习基石课程学习笔记2 -- Learning to Answer Yes/No
干货|台湾大学林轩田机器学习基石课程学习笔记3 -- Types of Learning
Welcome to follow the public account to learn and exchange~
