当前位置：网站首页>Finally, someone explained the supervised learning clearly

Finally, someone explained the supervised learning clearly

2022-07-02 13:04:00 【Big data V】

Reading guide ： Among various methods of machine learning , Supervised learning is by far the most impressive achievement . This article introduces the basic principles of supervised learning in solving problems such as pneumonia diagnosis .

author ： Paul · Perota （Paolo Perrotta）

source ： big data DT（ID：hzdashuju）

01 What is supervised learning

We should carry out supervised learning , We need to start with a set of sample data , Each sample has a label that the computer can learn . for example ：

As you can see , Samples can be many different things ： data 、 Text 、 voice 、 Video etc. . Besides , The label can be The number , It can also be type . The value tag is just a value , Like temperature – Lemonade converter . The type tag represents a category in a predefined collection , For example, in the example of dog breed detector .

Use some imagination , You can think of many other examples to predict things , Predict other things based on values or type tags .

We assume that some labeled samples have been collected . There are now two stages of supervised learning ：

Stage 1： Training phase

We provide labeled samples to an algorithm for discovering patterns . for example , The algorithm may notice , All the scanning images of pneumonia have some common features （ It may be some opaque areas ）, These features are not found in the non pneumonia scan . This stage is called the training stage , Because the algorithm will look at the sample data again and again , And learn to recognize these patterns .

Stage 2： Prediction stage

Now the algorithm knows what pneumonia looks like , Then switch to the prediction stage . We can harvest the results of training work at this stage . Show the trained Algorithm Not marked Of X Light scanning pictures , The algorithm will tell us whether it has the characteristics of pneumonia .

Here is another example of supervised learning —— A system that can recognize animal types . Each input data is a picture of an animal , The label of each sample is the species of animals in the picture . In the training phase , We show the algorithm the tagged image . In the prediction phase , We show the algorithm an unlabeled image , The algorithm is required to guess the label of the image .

I've said that before , Computer programs can be used in the process of machine learning “ Work out ” data . Supervised learning is an example of this process . In the traditional programming process , You can write a program to let the computer calculate the output from the input ; In supervised learning , Just give the sample data of program input and output , The computer can learn how to calculate an output from an input by itself .

Now that you have read a farsighted explanation of supervised learning , Then there may be more problems than when you first learned . We said , The supervised learning procedure is in the sample data “ Pay attention to common features ”, and “ Discovery patterns ”— But how did it do it ？ Let's start at an abstract level , See how this magic is realized .

02 The mathematical principle behind magic

Supervise the use of learning system Function fitting This mathematical concept is used to understand the relationship between sample data and its labels . Next, we will introduce the basic principle of this mathematical concept with specific examples .

Imagine , There is a solar panel on your roof . You are like a supervised learning system , Learn how solar panels generate energy , And predict the amount of energy generated in a certain period of time in the future .

It takes time to predict the energy output of solar panels 、 Weather and other variables . Time should be an important variable , So you decide to focus on the variable of time . For the real supervised learning process , You should start by collecting sample data on the amount of energy produced by solar panels at different times of the day . After several weeks of random sampling , You get the following data list ：

Each row in the above table contains The input variable （ Time ） And labels （ Energy value generated ） Sample data for , Just like the system that recognizes animals , Animal pictures are input , Animal names are labels .

If you draw these sample data into a chart , Then we can visually see the relationship between time and solar panel capacity ：

We know at a glance , Solar panels do not generate energy at night , And the energy value reached its peak at noon . As shown in the figure below , Although the supervised learning system is not as smart as the human brain , But it can approximate the sample data into a function , So as to realize the understanding of data .

It is not easy to find a fitting function that is close to the sample data . however , The subsequent prediction stage is much simpler . The system will forget all the sample information , And use the found fitting function to predict the energy generated by the solar panel at some time in the future , For example, the energy generated at noon is shown in the following figure ：

This is what I call supervised learning, which realizes the algorithm function through function fitting . The actual sample data received by the supervised learning system is usually chaotic and incomplete . In the data training stage , We usually need to construct a relatively simple function to approximate the complex actual data . In the prediction phase , Then the constructed fitting function is used to predict the unknown data .

As a programmer , You are used to thinking about many situations that may go wrong . therefore , You may already be considering ways to complicate the processing of sample data . for example , The energy output of solar panels is not only related to time , It will also be affected by other factors , For example, the influence of clouds or months .

If you collect data on all these variables , Then we will get a multidimensional point cloud , It will not be possible to visualize these point cloud data using a simple chart . Again , For solar panels , What we need to predict is the value tag . You may want to know how to convert this kind of numeric label into non numeric label （ Such as the name of animals ）, That's the category label .

You only need to know a little now ： No matter how many complex things you superimpose on it , The basic idea of supervised learning is the same as what we just described —— Find a pile of sample data , Find a function that can approximate these sample data .

Modern supervised learning systems are very good at this kind of fitting . in fact , This fitting function can be powerful enough to fit extremely complex functional relationships —— for example X The relationship between light scanning pictures and diagnosis conclusion . Of course , The function used to fit these corresponding relationships will be very complex for us humans . However , For computer systems, it's a piece of cake .

This article is excerpted from 《 Machine learning programming ： From coding to deep learning 》, Issued under the authority of the publisher .（ISBN：978-7-111-68091-8）

《 Machine learning programming ： From coding to deep learning 》

Click on the picture above to learn and buy

Please contact wechat for reprint ：DoctorData

Recommended language ： Language humor , Examples are vivid , Suitable for zero basic readers to learn machine learning . Suitable for intelligent science and Technology 、 Data science and big data technology 、 Introduction to machine learning for undergraduate or graduate students of computer science and technology and related majors , It can also be used as a reference for engineers and self-study readers .