[deep learning] 3 minutes introduction

First explain AI And deep learning ： AI is a big circle , Including computer vision 、 Natural person treatment 、 Data mining, etc ; What is called machine learning ？ Draw a small circle . What is deep learning ？ Equivalent to deep learning is also a part of machine learning .

machine learning （ Machine Learning , ML ） Learning from limited observational data （ or “ guess ”） A general law , And using these laws to predict the unknown data ．

Machine learning is an important branch of artificial intelligence , And gradually become the key factor to promote the development of artificial intelligence . Traditional machine learning mainly focuses on how to learn a prediction model ． Generally, you need to first express the data as a set of features （ Feature ）, The representation of a feature can be a continuous number 、 Discrete symbols or other forms . Then input these characteristics into the prediction model , And output the prediction results ． This kind of machine learning can be regarded as shallow learning （ Shallow Learning ）． An important feature of shallow learning is that it does not involve feature learning , Its features are mainly extracted by manual experience or feature transformation methods .

When we use machine learning to solve practical tasks , Will face a variety of data forms , Like sound 、 Images 、 Text etc. ． The feature construction methods of different data vary greatly ． For data such as images , We can naturally express it as a continuous vector ． There are many ways to express image data as vectors , For example, directly set all pixel values of an image （ Grayscale value or RGB value ） Form a continuous vector . For text data , Because it is generally composed of discrete symbols , And each symbol is represented as meaningless coding inside the computer , So it is usually difficult to find a suitable expression ． therefore , Using machine learning model in practical tasks generally includes the following steps ：
（ 1 ） Data preprocessing ： After data preprocessing , Such as removing noise, etc ． For example, in text categorization , Remove stop words, etc ;
（ 2 ） feature extraction ： Extract some effective features from the original data ． For example, in image classification , Edge extraction 、 Scale invariant feature transformation （ Scale Invariant Feature Transform , SIFT ） Characteristics, etc ;
（ 3 ） Feature conversion ： Some processing of features , For example, peacekeeping and upgrading . Many feature transformation methods are also machine learning methods . Dimensionality reduction includes feature extraction （ Feature Extraction ） And feature selection （ Feature Selection ） Two ways ． The commonly used feature transformation methods are principal component analysis （ Principal Components Analysis , PCA ）、 Principal component analysis .

Deep study what it does ？ Deep learning neural network should not be classified as an algorithm , What should I be , It should be regarded as a method of feature extraction . In machine learning and data mining or other tasks , What is the most difficult part of all intelligent tasks ？ The most difficult thing is not your algorithm , It's a data level , Like this person , He is going to cook a good meal , Can he make it without some good ingredients ？ I'm sure I can't do it . Let me tell you a little about this , Characteristics are very, very important , What kind of characteristics , What kind of data can be more suitable for our model and you . At present, the direction of this problem is more important . To summarize , Let's continue the basic process .

First step , Take the data first ;

The second step , Talk about a project ;

The third step , Choose an algorithm to model ;

Step four , Balanced application

What is the difference between deep learning and machine learning in dealing with Feature Engineering ？ I take out the following data , That's how I choose this feature , How to combine that feature ？ How can I extract some information from this feature , How to make the data easier to transform , Can get some more valuable input . We often think about a problem , I think everyone is in the process of considering this problem , Have you ever thought about one thing ？ Is there an algorithm after he gets the data , He himself will choose some features here ？ For example, after some algorithms come , He will learn what characteristics are good by himself , What's wrong , He slowly thought of good features , Which features fit together , What features are appropriate for decomposition and fusion ？ If there is such an algorithm to help us do tasks , Is it easier for us to do it by ourselves ？ Deep learning can .

Why do you say that? , Deep learning is the most integrated with artificial intelligence , Because it solves the problem of intelligence to a certain extent , Why? ？ Because you have to manually select data , Manually select features , Choose an algorithm manually , Get a result manually , Show some special intelligent feelings ？ It makes us feel as if we are doing the task , It's just the realization of this mathematical formula . This is my understanding of machine learning .

But deep learning , After taking the raw data , We can let this network really learn what characteristics , It is more suitable , How to combine is more appropriate . This is a realm that deep learning can achieve . Deep learning is a part of machine learning , It is also the core part of reality . What is the biggest point to solve ？ It solves the problem of Feature Engineering . How to do it ？

Sum up a sentence , What data characteristics determine what , It determines that you are going to make a model at present . So , Data preprocessing and feature extraction are always the first step , It is always the core part, and your choice of Algorithm , For example, I use SVM, Also use logical regression , Also use random forest . How can I adjust this parameter ？ No matter how you choose the algorithm and parameters , It just determines how we approach such an upper limit ？ So you should know a little , Special projects are more important , This thing has to go to the core than algorithms and parameters . Then you may ask a question , How should I choose features, for example, now , I still have a good understanding of some numerical characteristics . If you get some text data , If you get some image data , How can I choose features ？ This number is very, very difficult .

We should think about this parameter as much as possible , It is not this parameter, this feature how to do ？ Then I got one X after , For example, a x1 Is there a git page in the logic meeting , And then there's a x2 It is difficult to reach the characteristics , What characteristics do you want ？ Its influence on decoupling is also relatively difficult . It is difficult to find features , This is the core difficulty . Then how to learn in depth in deep learning? Everyone will regard it as a black box .

Neural network is a black box , How did it do it ？ After he took away the original data , Will do all kinds of complex transformation operations on this raw data , I can paint this thing black first , You can treat this thing as a black box , Because there are many things in the middle , Now I may not know a few details. Let's talk about one of them slowly , Now you need to know a little about behavior .

After having an input data , The big black box of neural network can automatically extract the features of the input data , Take them out of all kinds of features . This feature enables the computer to recognize . For example, how does the computer think it can recognize this 3, He will extract these three numbers , A series of characteristics he knows . So here he has a learning process , Computers really learn , Learning what kind of characteristics is the most suitable for him . This is the core of deep learning. In the future, we will encounter some deep learning problems , In the process of thinking, you will know a little . You may see that when you go to Baidu to check, you will hear various in-depth explanations . I think some of these explanations may be one-sided , I suggest you study in depth according to my form , It is really to learn what characteristics are the most appropriate . What would you like to add after having features ？ If you are willing to add logic and return to illogical, you will do typing , You can even transmit this feature to SM It's OK to be in the middle . So the core of this problem is how to carve . Then he .

The way , Deep study what it can do ？ What do you call it now ？ A driverless car . What does driverless car mean ？ Now this car doesn't need people to drive , He drives by himself , When driving by himself, does he have to see whether there is a car in front of him , Is there anyone in the back ？ The car ahead is running at an hourly rate 80 Drive at km , I spend every hour 100 Drive at km , Tesla , His driverless car had a fatal accident . Why is a car driving on the highway ？ There is a tractor parked in front , This tractor is foreign. It has some treading, spraying, painting and so on , Spray like a blue sky and white clouds , Tesla's driverless system is not recognized , In front is a tractor , Thought it was blue sky and white clouds , Jingle hit it , Is there something wrong ？

therefore , What is the core of deep learning ？ Some traditional data mining tasks may not use so much , Of course, there is no limit to this thing , More is the use of vision . In fact, text data needs to be modeled . For some face recognition with more face detection , Face detected the airport , People used to hold tickets , Hands, hands, hands , See if you stare at you alone for a long time , Now people don't have to look at you much when checking in , Take out your ID card and look at the corresponding camera , The system says it has been recognized , Then you can pass .

Deep learning , It is not particularly well supported on the mobile end . Why? ？ Because deep learning is equivalent to a disadvantage , The amount of calculation is too large , Let's talk about logistic regression or random score or random score , How many parameters can you say? Dozens of parameters can survive, right ？ But in the neural network , Let me start with a number , You may sound scary , A parameter of tens of millions of levels , Why it can go automatically depends on the characteristics ？ It depends on such a huge level of a parameter , Tens of millions of parameters are actually relatively few , Hundreds of millions of parameters are very common .

Let's think , Next, I will use an optimization algorithm to optimize this deep learning network , Optimize this neural network , Will the speed be very, very slow ？ Because you have to adjust the parameters of thousands of levels every time 1000 All the parameters , And every time you tune 50 Parameters , Can the speed be the same ？ Definitely not the same . So the biggest problem of deep learning is that the speed may be too slow . On some mobile terminals , For example, you want to transplant some products to some mobile phones , You use the deep learning algorithm to be sure that it is much stronger than the traditional model algorithm , But it just can't speed up , This may be its biggest problem .

Then I will tell you , Now medicine is particularly popular in social learning , Do some tests, for example, in some cells , It is common to detect some cancer cells . Detect the variation of cells in phase , Testing expectations is frustration , What it's like to do something complicated .

such as , These face changing operations seem to be a complicated process . It's not hard , Neural network to do this thing is relatively easy , What else to do ？ Now you haven't seen some old movies , Then new ones, such as Sunday professional films , Then a long time ago, I couldn't remember when I went to school . When I first entered the University , When I first entered the University, there was the Titanic , Is the Titanic out 3D Version ？ Then you say 3D What is the difference between the version and the old version ？ That just means 3D Version it looks clearer , Is it a clearer version , Talk about whether it is a reconstruction of resolution. Here you let the neural network automatically do a coloring , Sure , It is also possible to do some super-resolution reconstruction automatically .

There are many changes in the neural network , Basically, what you can do is , As long as you can basically think of something , The Internet 80% Can give you to complete . Every year in major conferences , Neural networks are all kinds of deformations , There are too many delayed versions , What you can play is actually covered , It's basically all aspects of life , But the core is still in images and texts . Then this picture can be said to be a point of the rise of deep learning . First of all, let me tell you about this history , Fei Fei Li , Is specialized in computer vision , She wondered if she could create a dataset ？ Six years , She called on how many colleges and universities in the United States , Go to label the image . What is collection and annotation ？ Let's play this deep learning or active learning , Are we usually a supervised problem . Next, we will talk about neural networks , All algorithms are also supervised . So for example, in the above tasks , Do you have to have a mark of what ？ Where is a real location of the face ？ You need to do a critical language test , Are you in the training process , You know where its real location is ？ Don't underestimate such a labeling work .

Completed a data set , It's called a imagine net It also exists now , At that time, you can see that the data inside is a public photograph , Then everyone can come down . The data in Baidu online disk is very, very large , But I don't recommend you to play , Because it's too big , You don't do some research on this data set , Don't do some big movie projects , We may not use it for the time being .
How many do you see 14m One of the imagine 22k A kind of , Basically, its data volume can be said to be the largest Image Classification Library at present . The objects involved in it are basically you see 22 Multiply by 1000 How many kinds? 2 More than ten thousand kinds of tables , It includes all kinds of tables you can think of in life . Then this is when he created such a data set , After creating the dataset , Then he held such a competition , It's called a imagenet The same category competition , It was 09 year , Deep learning is not mentioned at all , No one makes this thing . Why? ？ Because the first point is that it is slow . Second point , This thing has an effect on the pattern , You haven't played before , Because the effect may also be average , until 12 year .

12 year , There is a man named Alex People who , He was in 12 In the image classification competition, I won a champion , He used a neural network in deep learning to complete the game . The first place he won was more than ten percentage points higher than the second place . The second algorithm is what to use ？ It is done with the lifting algorithm of the machine learning algorithm we mentioned before . This time, everyone suddenly found something , Neural network is an algorithm for image classification , It can achieve such a good effect in the field of computer vision . So from this 12 Year begins , Deep learning is the official rise , Everyone is beginning to recognize this thing . For example, in-depth studies , More and more scholars , More and more people are involved in this field .

Our computer vision natural processing is developing rapidly , What is the main reason ？ 12 Such a turning point in history , Some papers of major conferences every year , Many excellent articles will be sent out about some competition situations of major competitions , These weeks' articles become everyone's own to play , Follow up practical projects , The company's business level projects are all done with these papers . This is the starting point .
Then you need to know a little about such a picture , It's about how to do it in depth, and the effect can be better , Here, let's take a look at . Here is a report from Baidu Zhongkan before. Take out a picture , He said such a thing , Baidu has many AI products , When the data scale is small , You can see when the data scale is small , It can be said that there is a difference between deep learning and traditional AI algorithms ？ No difference . The thief of deep learning algorithm is slow , The traditional algorithm is quite fast . Do you say you use deep learning algorithm when the data scale is small ？ Don't use it at all . How much room does the deep learning algorithm have to play ？ The larger the data size, the better , Usually, there are few thousands of data sets , It is generally said that only tens of thousands, even millions of data sets can be done with deep learning algorithms .