What is machine learning?

Speaking of machine learning, we should give an accurate definition of machine learning. On the intuitive level, if computer science is a science about algorithms, then machine learning is a science about "learning algorithms", or, unlike ordinary explicit programming, machine learning is a field that studies how to make computers learn when explicit programming is impossible. It should be noted that whether they are explicit to humans-whether humans can clearly understand each decision-making step, there are no codes and instructions that constitute different algorithms for computers.

More precisely, the definition of machine learning is as follows:

If the performance of a computer program on T (measured by P) improves with experience E, it is said that the program has learned from experience E about a certain task T and a certain performance measurement P.

A (machine learning) program is an algorithm that can learn task T from empirical data E, and its performance metric P on task T will become better with the learning of empirical data E..

Because machine learning must make use of some experiences, they often exist in the form of data, which we call data sets, and each data in them is called records. For example, we can predict whether a person has a common disease by his gender, age and height. The following data is available:

(gender: male; Age:18; Height:174; Are you sick: no)

(gender: female; Age:17; Height:164; Are you sick: yes)

(gender: male; Age: 20; Height:181; Are you sick: yes)

(gender: female; Age:16; Height:161; Are you sick: yes) ...

This can be called a data set, in which everyone's data is called a record. In records, descriptive data about objects are called attributes, because there are often many attributes, such as age and height, which can form attribute vectors, and the space for these vectors to grow is called attribute space. And the quantity that our algorithm needs to predict is called a label-on it, it means "sick or not". In some datasets, there are tags, while in others, there are no tags. The space formed by labels is called label space, also called output space.

Obviously, because we can only get a part of the total data, that is, the training sample, the model obtained by our program can not only adapt to this training sample, but also have a good prediction effect on the total data. This means that our model must have generalization ability.

The model we train is called hypothesis, and all the models together form a hypothesis space. Obviously, there may be various hypothetical spaces consistent with the training data-just like for a classroom study with few knowledge points, many people can get high marks, but for the whole data, different learning modes obviously have different effects-a real test of many difficult knowledge points and a test that separates the superficial top student mentioned above.

Every hypothesis, that is, the training model, must have its inductive preference, that is, which model will be chosen if it has not been seen in the training set, or both. Inductive preference is the ability basis of model generalization.