Machine learning
can be thought of as looking for a function, which requires the machine to have the ability to find the function we want based on the information you provide it.
The complete framework structure can be summarized as follows: the first step is to pick a function from a bunch of functions, the second step is to combine the training data so that the machine can measure whether the function you selected is good or bad, and the third step is to let The machine has an automatic method and a good algorithm to pick out the best features. After getting the best function, enter the test data and you can test it.
The following is a brief introduction to related technologies in machine learning, including supervised learning, semi-supervised learning, transfer learning, unsupervised learning, structured learning, and reinforcement learning. The same color and different squares are of the same type. The blue color refers to the learning situation. Usually the learning situation is something you have no control over. For example, because we don’t have data for supervised learning, we do reinforcement learning. The red ones refer to your task, the problem you want to solve. The problem you want to solve depends on the equation you use, including regression, classification, and structured. Therefore, under different situations (blue square), it is possible to solve this task (red square). There are different models in different tasks, represented by green squares.
Three major tasks:
The regression problem
is the function found by the machine. Its output is a scalar, such as distance length. This is called regression. For example, in the picture below, we will ask you to predict PM2.5 (for example, predict PM2.5 tomorrow morning). In other words, you need to find a function whose output is PM2 at a certain time in the future. A value of 5, this is a regression problem.
The difference between regression and classification problems
is that the types of things we want the machine to output are different. In Regression, the machine outputs a numerical value, and in Classification, the machine outputs a category. Classification problems are generally divided into two types, one is called binary classification and the output is yes or no (Yes or No); the other is called multi-class (Multi-class). In Multi-class, the machine is asked to do a multiple-choice question, which is equal to yes Give him several options, each of which is a category, and ask him to choose the correct category from multiple categories.
Structured learning
means that the output of the machine must be structured. For example, in speech recognition, the machine input is a sound signal and the output is a sentence. Sentences require many words to be put together to complete. It is a structured object. Many textbooks directly say that machine learning consists of two major types of problems, regression and classification. This is like telling you: there are only five continents in the world we know, but this is only a small part of the real world. The real world is as shown in the picture. The real world should also include structure learning, and there are still many issues that have not been explored.
five love types
The first step in supervised learning
is to choose a function set. Choosing different function sets means choosing different models. There are many types of Models. The simplest is the linear model. Among the nonlinear models, the most familiar one is Deep learning. In addition, there are many nonlinear models such as SVM.
Supervised learning requires a large amount of labeled training data. The training data tells us the relationship between the input and output of the function we are looking for. Based on these data, select different models to complete the corresponding tasks. In addition, the output in training data often cannot be obtained in a natural way, and artificial power is required to label it (labeling data is pure bullshit, everyone knows it).
Semi-supervised learning
: Is there any way to reduce the amount of labels required? This is semi-supervised learning. For example, you have a small amount of labeled data of cats and dogs , but at the same time you have a large amount of Unlabeled data of cats and dogs , but you do not have the strength to tell the machine which ones are cats and which ones are dogs. You can use semi-supervised learning technology to make use of these data without labels.
Transfer learning
Transfer learning means: Suppose we want to classify cats and dogs, and we only have a small amount of labeled data. But there is a large amount of data that may or may not have labels. But this data has nothing special to do with the problem we are considering now, which means that what we want to distinguish is the difference between cats and dogs, but the data contains a lot of pictures of other animals and a bunch of Haruhi Suzumiya, Misaka Mikoto, you With this bunch of irrelevant pictures, how exactly can it help. This is what transfer learning is about.
Unsupervised learning
is the hope that machines can learn without a teacher. If there is no label at all in the training data, what kind of things can the machine learn? For example, if we let the machine read a large number of articles, what can it learn after letting the machine read a large number of articles?
Reinforcement Learning
When we compare reinforcement learning with supervised learning, in supervised learning we tell the machine what the correct answer is (i.e. label). But in reinforcement learning, we do not tell the machine what the correct answer is. All the machine has is a score, that is, whether it did well or not. If we want to use the reinforcement learning method to train a chatbot now, the training method will be like this: you send the machine offline, let it talk to the customers who come in, and then think about it for a long time, and finally Still furious, he hung up the phone. The machine then learns that it did something wrong just now. But he didn't know what was wrong. He had to go back and figure it out on his own. Shouldn't he have said hello in the first place? There shouldn't be any swearing in the middle or something like that. It doesn’t know, and no one tells it what it didn’t do well. It needs to go back and reflect on what step it did poorly.
Summary:
To put it simply, it introduces what machine learning is, what is the process of solving problems (train-test), what are the problems to be solved (three major tasks), in the face of these problems, according to the different characteristics of the data, etc. (there are Labels without labels or anything), we can discuss what situations can be divided into (five major situations), over.
Quote:
Li Hongyi Machine Learning Video of Station B
Li Hongyi Machine Learning Notes (LeeML-Notes)