Machine learning is changing our work and life

Machine learning is a discipline in which computers build models based on data and use models to simulate human intelligence activities. With the rapid development of computers and networks, machine learning is playing an increasingly important role in our lives and work, and is changing our lives and work. Machine learning in everyday life We often use digital cameras in our daily lives. You may not know that the face detection technology on digital cameras is based on machine learning technology! I know three great scientists and engineers. They are Robert Schapire, Paul Viola, and Lai Shiyi. All three of them are related to this. Together with Yoav Freund, R obert invented the very efficient machine learning algorithm A da B oost. P aul successfully applies the A da B oost algorithm to face detection. Lai Shiyi and his Omr on team made the A da B oost face detection algorithm on the chip. It is said that the face detection on the digital camera of 60% or 70% in the world is now using the Omr on chip.

In our work and life, this kind of example has emerged one after another. Internet search, online advertising, machine translation, handwriting recognition, spam filtering, etc. all use machine learning as the core technology. Not long ago, the InternaTIonal Conference on Machine Learning (ICML 2011) was held in Bellevue, Washington, USA. More than 700 researchers, professors and students participated in the event, setting a record high. The three keynote speeches of the conference introduced the application of machine learning in Microsoft's Kinnect game console user sensing system, Google's G oggles image search system, and IBM's Watson automated question answering system. These facts make people feel that a new era of machine learning is being used more widely. Machine learning and artificial intelligence intelligence are the inevitable trend of computer development. The various intelligent activities that human beings engage in, such as mathematics, art, language, music, sports, learning, games, design, research, teaching, etc., make computers work, and now it is still very difficult. This is the conclusion of artificial intelligence research for decades.

In the study of artificial intelligence, people have tried three roads. I refer to them as appearance (extrospecTIon), introspection (introspecTIon), and simulation (simulaTIon). The so-called appearance refers to observing the working condition of the human brain, exploring its principles, and clarifying its mechanism to "realize" the function of the human brain on a computer. For example, computational neuroscience research is based on this motivation. However, the complex information processing of the human brain is difficult to observe and model. Just as we only observe the signal transmission process in a computer, it is difficult to judge what kind of calculation it is doing. Introspection is to reflect on their own intelligent behavior, record the reasoning and knowledge that they realize on the computer, and thus “reproduce” the human intelligence, such as the expert system attempt. The biggest problem with introspection is that it is difficult to generalize, that is, to make a difference. No matter what kind of picture, or even in abstract painting, people can easily find the face.

This ability is called generalization ability. It is difficult to make a computer generalized by introspection. The principle of intelligence of oneself is very likely to be agnostic to human beings.

The mouse in the cage may think that touching the handle is the "cause" of getting food, but it can never understand the food delivery mechanism of the entire cage. Simulation is to record the input and output of human intelligent operation, and use the model to simulate, so that the model gives similar performance to human input and output, such as statistical machine learning. Practice has shown that statistical machine learning is the most effective means to achieve the goal of computer intelligence. The biggest advantage of statistical learning is that it has generalization ability; the disadvantage is that it always gets the optimal solution in statistical sense (for example, face detection). Now when people talk about machine learning, it usually refers to statistical machine learning or statistical learning.

The advantages and disadvantages of machine learning look at a simple example. This example illustrates the basic principles of statistical learning and the advantages and disadvantages that result. Suppose we observe that the output of a system is a series of 1's and 0's, and we want to predict what its next output is. If 1 and 0 are half of the observed data, then we can only make predictions with an accuracy of 0.5. However, if we observe that the system has inputs, it is also a series of 1s and 0s, and the ratio of the output to 0 when the input is 1 is 0.9, and the ratio of the output to 1 when the input is 0 is also 0.9. In this way, we can learn the “model” from the given data, predict its output according to the input of the system, and increase the prediction accuracy from 0.5 to 0.9. The above is the basic idea of ​​statistical learning, especially supervised learning. In fact, this is the world's simplest statistical machine learning model! The conditional probability distribution P(Y|X), where the random variables X and Y represent the input and output, taking values ​​1 and 0. It can be assumed that all supervised learning models are complex versions of this simple model. We use this model to predict the possible output based on a given input characteristic. The biggest advantage of statistical learning is that it has generalization ability, and it can predict the corresponding Y for any given X. Vapnik's statistical learning theory can also analyze predictive power and give generalization upper bounds. But from this example, we can also see that the prediction accuracy of statistical learning is not guaranteed to be 100%. For example, face detection will be wrong and Chinese word segmentation will be wrong.

Statistical learning is a "countryman" approach. There is a joke. A countryman enters the city, eats at a restaurant, and does not know how to eat at a restaurant, imitating the people next to him. What others do, he also learns what to do. One of the neighboring tables deliberately teased him, rolled the candle on the table in the cake, threw the candle on the ground when the countryman did not pay attention, and then took a bite of the rolled cake. The countryman followed the school and took a bite of his own cake. Statistical learning is based only on the input and output of observations. "Imitating" human machine learning is a discipline in which computers build models based on data and use models to simulate human intelligence activities. With the rapid development of computers and networks, machine learning is playing an increasingly important role in our lives and work, and is changing our lives and work. Machine learning in everyday life We often use digital cameras in our daily lives. You may not know that the face detection technology on digital cameras is based on machine learning technology! I know three great scientists and engineers. They are Robert Schapire, Paul Viola, and Lai Shiyi. All three of them are related to this. Together with Yoav Freund, R obert invented the very efficient machine learning algorithm A da B oost. P aul successfully applies the A da B oost algorithm to face detection. Lai Shiyi and his Omr on team made the A da B oost face detection algorithm on the chip. It is said that the face detection on the digital camera of 60% or 70% in the world is now using the Omr on chip.

In our work and life, this kind of example has emerged one after another. Internet search, online advertising, machine translation, handwriting recognition, spam filtering, etc. all use machine learning as the core technology. Not long ago, the International Conference onMachine Learning (ICML 2011) was held in Bellevue, Washington, USA. More than 700 researchers, professors and students participated in the event, setting a record high. The three keynote speeches of the conference introduced the application of machine learning in Microsoft's Kinnect game console user sensing system, Google's G oggles image search system, and IBM's Watson automated question answering system. These facts make people feel that a new era of machine learning is being used more widely. Machine learning and artificial intelligence intelligence are the inevitable trend of computer development. The various intelligent activities that human beings engage in, such as mathematics, art, language, music, sports, learning, games, design, research, teaching, etc., make computers work, and now it is still very difficult. This is the conclusion of artificial intelligence research for decades.

In the study of artificial intelligence, people have tried three roads. I call them extrospection, introspection, and simulation. The so-called appearance refers to observing the working condition of the human brain, exploring its principles, and clarifying its mechanism to "realize" the function of the human brain on a computer. For example, computational neuroscience research is based on this motivation. However, the complex information processing of the human brain is difficult to observe and model. Just as we only observe the signal transmission process in a computer, it is difficult to judge what kind of calculation it is doing. Introspection is to reflect on their own intelligent behavior, record the reasoning and knowledge that they realize on the computer, and thus “reproduce” the human intelligence, such as the expert system attempt. The biggest problem with introspection is that it is difficult to generalize, that is, to make a difference. No matter what kind of picture, or even in abstract painting, people can easily find the face.

This ability is called generalization ability. It is difficult to make a computer generalized by introspection. The principle of intelligence of oneself is very likely to be agnostic to human beings.

The mouse in the cage may think that touching the handle is the "cause" of getting food, but it can never understand the food delivery mechanism of the entire cage. Simulation is to record the input and output of human intelligent operation, and use the model to simulate, so that the model gives similar performance to human input and output, such as statistical machine learning. Practice has shown that statistical machine learning is the most effective means to achieve the goal of computer intelligence. The biggest advantage of statistical learning is that it has generalization ability; the disadvantage is that it always gets the optimal solution in statistical sense (for example, face detection). Now when people talk about machine learning, it usually refers to statistical machine learning or statistical learning.

The advantages and disadvantages of machine learning look at a simple example. This example illustrates the basic principles of statistical learning and the advantages and disadvantages that result. Suppose we observe that the output of a system is a series of 1's and 0's, and we want to predict what its next output is. If 1 and 0 are half of the observed data, then we can only make predictions with an accuracy of 0.5. However, if we observe that the system has inputs, it is also a series of 1s and 0s, and the ratio of the output to 0 when the input is 1 is 0.9, and the ratio of the output to 1 when the input is 0 is also 0.9. In this way, we can learn the “model” from the given data, predict its output according to the input of the system, and increase the prediction accuracy from 0.5 to 0.9. The above is the basic idea of ​​statistical learning, especially supervised learning. In fact, this is the world's simplest statistical machine learning model! The conditional probability distribution P(Y|X), where the random variables X and Y represent the input and output, taking values ​​1 and 0. It can be assumed that all supervised learning models are complex versions of this simple model. We use this model to predict the possible output based on a given input characteristic. The biggest advantage of statistical learning is that it has generalization ability, and it can predict the corresponding Y for any given X. Vapnik's statistical learning theory can also analyze predictive power and give generalization upper bounds. But from this example, we can also see that the prediction accuracy of statistical learning is not guaranteed to be 100%. For example, face detection will be wrong and Chinese word segmentation will be wrong.

Statistical learning is a "countryman" approach. There is a joke. A countryman enters the city, eats at a restaurant, and does not know how to eat at a restaurant, imitating the people next to him. What others do, he also learns what to do. One of the neighboring tables deliberately teased him, rolled the candle on the table in the cake, threw the candle on the ground when the countryman did not pay attention, and then took a bite of the rolled cake. The countryman followed the school and took a bite of his own cake. Statistical learning simply "impersonates" people's intelligent behavior based on the input and output of observations. Sometimes it can be very intelligent. But it is essentially data-based and is an "imitation" in the statistical sense. If no key features are observed, it will go to "cakes that bite the candle."

Machine Learning and Internet Search I am working with my colleagues on Internet search related research. According to the survey, 60% of Internet users use search engines at least once a day, and 90% of Internet users use search engines at least once a week. Search engines have greatly improved the quality of people's work, study and life. In the basic technology of Internet search, machine learning occupies an important position. In my opinion, Internet search has two major challenges and one big advantage. Challenges include scale challenges and artificial intelligence challenges; the advantage is primarily scale advantage. Scale challenge: For example, search engines can see trillion-level URLs. There are hundreds of millions of billions of users per day, and thousands of machines need to crawl, process, and index web pages to serve users. This requires technology development and innovation in many aspects such as systems, software, and hardware. Artificial Intelligence Challenge: Search is ultimately an artificial intelligence issue. The search system needs to help users find information as quickly, as accurately as possible. This essentially requires "understanding" of user requirements (such as query statements), as well as text, images, and video on the Internet.

Today's search engines can help users find information to a large extent through keyword matching and other "signals." However, it is still not enough. Scale advantage: There is a large amount of content data on the Internet, and search engines record a large amount of user behavior data. This data can help us find information that seems to be difficult to find. For example, “What is the population of New York City?” “What is the population of the city?” “Who is the author of the Spring River and the South Bank of the Green River?” Note that these data are distributed according to a power function. They can help with Head (high frequency) requirements, and the demand for tail (low frequency) is often difficult. Therefore, the challenge of artificial intelligence for tail is even more significant. Today's Internet search can meet some basic needs of user information access to a certain extent.

This is due to the successful development and application of many cutting-edge technologies, including machine learning techniques, such as sorting learning algorithms, web page importance algorithms, and more. These machine learning algorithms can, to a certain extent, take advantage of scale to address artificial intelligence challenges. However, today's Internet search distance is "a questionable, accurate, fast, all, good" ideal still has a certain distance. This requires the development of more and better machine learning techniques to solve the challenges of artificial intelligence, especially in the tail. Looking to the future, the research and development of machine learning technology will help us make tomorrow even better! Intelligent behavior. Sometimes it can be very intelligent. But it is essentially data-based and is an "imitation" in the statistical sense. If no key features are observed, it will go to "cakes that bite the candle."

Machine Learning and Internet Search I am working with my colleagues on Internet search related research. According to the survey, 60% of Internet users use search engines at least once a day, and 90% of Internet users use search engines at least once a week. Search engines have greatly improved the quality of people's work, study and life. In the basic technology of Internet search, machine learning occupies an important position. In my opinion, Internet search has two major challenges and one big advantage. Challenges include scale challenges and artificial intelligence challenges; the advantage is primarily scale advantage. Scale challenge: For example, search engines can see trillion-level URLs. There are hundreds of millions of billions of users per day, and thousands of machines need to crawl, process, and index web pages to serve users. This requires technology development and innovation in many aspects such as systems, software, and hardware.

Artificial Intelligence Challenge: Search is ultimately an artificial intelligence issue. The search system needs to help users find information as quickly, as accurately as possible. This essentially requires "understanding" of user requirements (such as query statements), as well as text, images, and video on the Internet. Today's search engines can help users find information to a large extent through keyword matching and other "signals." However, it is still not enough. Scale advantage: There is a large amount of content data on the Internet, and search engines record a large amount of user behavior data. This data can help us find information that seems to be difficult to find. For example, “What is the population of New York City?” “What is the population of the city?” “Who is the author of the Spring River and the South Bank of the Green River?” Note that these data are distributed according to a power function. They can help with Head (high frequency) requirements, and the demand for tail (low frequency) is often difficult. Therefore, the challenge of artificial intelligence for tail is even more significant. Today's Internet search can meet some basic needs of user information access to a certain extent.

This is due to the successful development and application of many cutting-edge technologies, including machine learning techniques, such as sorting learning algorithms, web page importance algorithms, and more. These machine learning algorithms can, to a certain extent, take advantage of scale to address artificial intelligence challenges. However, today's Internet search distance is "a questionable, accurate, fast, all, good" ideal still has a certain distance. This requires the development of more and better machine learning techniques to solve the challenges of artificial intelligence, especially in the tail. Looking to the future, the research and development of machine learning technology will help us make tomorrow even better!