1. Learning with Different Output Space Y
The example we introduced in the last lesson that the bank judges whether to issue a credit card to a user based on his personal circumstances is a typical binary classification problem. In other words, there are only two outputs. Generally, -1 means no credit card is issued (negative type), and +1 means credit card is issued (positive type).
The problems of binary classification are very common, including credit card issuance, spam identification, patient disease diagnosis, answer correctness estimation, and so on. Binary classification is a very core and basic problem in the field of
machine learning. Binary classification has linear models and non-linear models. Different models are selected according to the actual problem.
In addition to binary classification, there are also Multiclass Classification problems. As the name implies, there are more than two outputs of multiple classification, y={1, 2,…, K}, K>2. General applications of multiple classification include number recognition, image content recognition and so on.
Both binary classification and multivariate classification are classification problems, and their outputs are discrete values. Second, for another situation, such as training models, predicting house prices, stock returns, etc., the output of this type of problem is y=R, that is, the range is continuous in the entire real number space. This kind of problem, we call it regression (Regression). The simplest linear regression is a typical regression model.
In addition to classification and regression problems, in natural language processing and other fields, a
machine learning problem is also used: Structured Learning. The output space of structured learning contains a certain structure, and some of its solutions are usually extended from multi-classification problems and are more complicated.
To briefly summarize, machine learning is divided according to the output space, including different types such as binary classification, multiple classification, regression, and structured learning. Among them, binary classification and regression are the two most basic and core types, and they are also the main part of our course.
2. Learning with Different Data Label
If the training samples we get have both input features and output, then we call this type of learning Supervised Learning. Supervised learning can be binary classification, multi-class classification or regression. The most important thing is to know the output label. The opposite type of supervised learning is unsupervised learning. Unsupervised learning has no output label yn. Typical unsupervised learning includes: clustering problems, such as automatic classification of news on web pages; density estimation, such as traffic analysis; anomaly detection, such as user network traffic monitor. Under normal circumstances, unsupervised learning is more complicated, and many unsupervised problems can be implemented using some algorithmic ideas of supervised learning.
What lies between supervised and unsupervised learning is called semi-supervised learning. As the name implies, semi-supervised learning means that part of the data has output labels, while the other part of the data has no output labels. In practical applications, semi-supervised learning is sometimes necessary. For example, when a pharmaceutical company tests certain drugs, taking into account issues such as cost and experimental population restrictions, only part of the data has output tags.
Supervised, unsupervised, and semi-supervised learning are the three main types of machine learning. In addition, there is a very important type: Reinforcement Learning. In reinforcement learning, we give some input to the model or the system, but cannot give the true output we hope. According to the output feedback of the model, if the feedback result is good and is closer to the real output, we will give it a positive incentive. , Deviate from the true output, give it reverse excitation. Continuously using the form of "feedback-correction" to make the model learn better step by step, which is the core of enhanced learning. Reinforcement learning can be compared to the process of training a pet. For example, we have to train a dog to sit down, but the dog cannot directly understand our command "sit down". During the training process, we signal to the dog that if it performs well, we will reward him, and if it does actions that are completely unrelated to sit down, we will give it a small punishment. Such continuous correction of the dog's movements will eventually allow it to act according to our instructions. In real life, there are many examples of enhanced learning, such as an advertising system that is continuously improved based on user clicks and choices
To summarize briefly, machine learning is classified according to data output labels, including supervised learning, unsupervised learning, semi-supervised learning, and reinforcement learning. Among them, supervised learning is the most widely used.
3. Learning with Different Protocol
According to different protocols, machine learning can be divided into three types:
Batch Learning
Online
Active Learning
Batch learning is a common type. The training data D obtained by batch learning is a batch, that is, the entire D is obtained at one time, and the learning model is performed on it to obtain our final machine learning model. Batch learning is the most widely used in practical applications.
Online is an online learning model. The data is updated in real time. According to the data one by one, we update our algorithm synchronously. For example, an online mail filtering system judges whether it is spam based on the content of each mail and the current algorithm, and then updates the current algorithm in time based on user feedback. This is a dynamic process. Both the PLA and reinforcement learning that we introduced earlier can use the online model.
Active learning is a new type of machine learning that has emerged in recent years, which allows machines to have the ability to actively ask questions, such as handwritten digit recognition, where the machine generates a number by itself or actively asks questions about its uncertain handwriting. One of the advantages of active learning is that when it is difficult to obtain sample labels, it can save time and cost, and only put forward requirements for some important labels.
To summarize briefly, according to different protocols, machine learning can be divided into batch, online, and active. These three types of learning can be compared as follows: duck-filling, teacher teaching, and active questioning.
4. Learning with Different Input Space
The machine learning classification introduced in the previous sections is based on output classification, such as classification according to the output space, classification according to the label of the output y, and classification according to the method of obtaining the data and the label. In this part, we will talk about the types of inputs.
The first type of input is concrete features. For example, the size and weight of coins in coin classification problems; such as specific features such as patient information in disease diagnosis. Concrete features are the easiest to understand and use for machine learning.
The second type is raw features. For example, the mxn-dimensional pixel value of the picture where each digit is located in handwritten number recognition; such as the frequency spectrum of the voice signal. Raw features are generally abstract and often require humans or machines to convert to their corresponding concrete features. The process of this conversion is Feature Transform.
The third type is abstract features. For example, when a shopping website makes purchase predictions, the participants are provided with abstract encrypted data numbers or IDs. These features X are completely abstract and have no actual physical meaning. Therefore, it is more difficult for machine learning, and more conversion and extraction of features are required.
To summarize briefly, according to the input X type, it can be divided into concetet, raw, abstract. Converting some abstract features into concrete features is a very important part of the machine learning process.
