Application of Machine Learning

Machine Learning is the field of study that gives computers the capability to learn without being explicitly programmed. ML is one of the most exciting technologies that one would have ever come across. Here introduces the application of machine learning.

Supervised learning
The task of supervised learning description is how to predict the output y by training the model on the data with labeled input and output when the input x is given. From a statistical point of view, supervised learning focuses on how to estimate the conditional probability p (y|x). In practical situations, supervised learning is the most commonly used. For example, given a patient's CT image, predict whether the patient has cancer; given an English sentence, predict its correct Chinese translation; given the company's financial report data of this month, predict the company's stock price in the next month.
regression analysis
Regression analysis is perhaps the simplest type of task in supervised learning. In this task, the input is any discrete or continuous, single or multiple variables, while the output is a continuous value. For example, we can extract some characteristics of the company's financial report data of this month, such as total revenue, total expenditure and whether there are negative reports, and use regression analysis to predict the stock price of the company in the next month.
If we define the difference between the output value predicted by the model and the real output value as the residual error, the loss function of common regression analysis includes the sum of squares of the residuals of the training data or the sum of the absolute values. The task of machine learning is to find a set of model parameters to minimize the loss function. We will introduce the regression analysis in detail in the following chapters.
classification
It is worth mentioning that the prediction concerned by regression analysis can often solve the problem that the output is continuous. When the output of the prediction is a discrete category, this supervised learning task is called classification. Classification is very common in our daily life. For example, we can extract some characteristics from the company's financial report data of this month, such as total revenue, total expenditure and whether there are negative reports, and use classification to predict whether the CEO of the company will leave next month. In the field of computer vision, a picture is recognized as one of many categories of objects, such as cat, dog and so on.
Classification of animals
Given several features extracted from an instance as input, our classification model can output the probability of each class, and take the category with the largest probability as the classification result.
tagging
In fact, some seemingly classified problems are difficult to be classified in practice. For example, there's something wrong with classifying the following picture into either a cat or a dog.
As you can see, there are both cats and dogs in the picture above. Actually, it's not finished yet. There are grass, tires, stones and so on. Instead of just classifying the above figure into one category, it is better to mark out all the categories that we are concerned about. For example, given a picture, we want to know whether there are cats, dogs, grass, etc. Given a category whose input and output are not quantitative, this is called tagging task.
This kind of task is sometimes called multi label classification. Imagine that people might tag multiple tags on a technical blog post at the same time, such as "machine learning", "technology", "programming language", "cloud computing", "security and privacy" and "AWS". In fact, the tags are sometimes related to each other, such as "cloud computing" and "security and privacy". When an article may be marked with a large number of people, it is very difficult to label. This requires the use of machine learning.
Search and sort
Search and sorting focuses more on how to sort a bunch of objects. For example, in the field of information retrieval, we often focus on how to sort a pile of documents according to their relevance to the search items. In the Internet age, due to the popularity of search engines, we pay more attention to how to sort web pages. In the early Internet era, there was a famous web page ranking algorithm called PageRank. The sorting results of the algorithm do not depend on the specific user search items. These sorting results can better sort the web pages containing the search items.
Recommendation system
Recommendation system is closely related to search ranking, and is widely used in shopping sites, search engines, news portals and so on. The main goal of the recommender system is to recommend what the user may be interested in. Recommendation algorithm uses a variety of information, such as user's self description, reaction to previous recommendations, social network, preferences and so on. The following figure shows the results of amazon.com's recommendation for one of the author's books on deep learning.
Search engine search term automatic completion system is also a good example. According to the first few characters entered by users, it can automatically complete the items that users may search for in real time. In one of the works of one of the authors, if the system finds that the user has just opened a sports mobile application, when the user spell out "real" in the search box, the system will recommend "Real Madrid" (Real Madrid, football team) to the "real" which is searched more frequently than usual "Estate" is more advanced than it is in the picture below.
Sequential learning
Sequence learning is also a kind of machine learning problem which has attracted much attention recently. In such problems, input and output are not limited to a fixed number. This kind of model can usually deal with any length of input sequence, or output any length of sequence. When the input and output are variable length sequences, we also call this kind of model seq2seq, such as language translation model and speech transcription text model. Here are some common sequential learning cases.