I. Overview of "data"
Although I don't know the specific situation of artificial intelligence, I generally know that artificial intelligence at this stage needs to be based on big data. The first chapter of "Intelligent Times" introduces the related issues of data in detail.
First of all, "before the advent of computers, the text content in general books was not regarded as data." Today, not only the words in books, but also our activities, our daily behaviors, preferences and so on have been regarded as some kind of data.
In the early days of human society, we observed phenomena, summarized data, extracted useful information from data, and formed knowledge (knowledge is systematic) on the basis of information, thus guiding our behavior. Our standard process for using data is as follows:
In the past, the role of data was not valued. "There are two reasons. First, due to the lack of data in the past, the accumulation of a large number of data takes too long, so that the effect is not obvious in a short time. Secondly, the connection between data and the information you want to obtain is usually indirect and can only be reflected through the correlation between different data. " This correlation needs to be explored. Teacher Wu Jun cited the correlation between Wang Jinxi photos and Japanese bids, the correlation between Google users' search volume of a program and its ratings, and the correlation between search trends and influenza epidemic to illustrate the importance of correlation.
As for the first reason, "before the advent of the Internet, it was not easy to obtain a large number of representative data. Of course, it is no problem to do some statistics within the allowable range of error, but only in rare cases can we solve complex problems simply by relying on data. Therefore, before the 1990 s, the whole society did not pay much attention to data. "
We should make better use of the correlation of data by establishing a suitable mathematical model. "To establish a mathematical model, we must solve two problems, one is what kind of model to use, and the other is what the parameters of the model are." "... if you don't choose a model at first, it will be difficult to mend it later. Therefore, in the past, whether it was theory or engineering, everyone focused on finding models. "
"With the model, the second step is to find the parameters of the model, so that the model can at least match the previously observed data. In the past, this point was far less concerned than looking for a model. But today he has a more fashionable and profound word-machine learning. " (Cheng Press: Input data and constantly adjust the model, similar to today's machine learning method)
"Back to the mathematical model, in fact, as long as there is enough data, you can replace a complex model with several simple models. This method is called data-driven method, because it first has a lot of data, not a preset model, and then uses many simple models to fit the data. Although a set of models found by this data-driven method may have some deviations from the real model in the case of insufficient data, the results are equivalent to the exact model within the allowable range of errors, which is mathematically reasonable. In principle, this is similar to Chebyshev's law of large numbers mentioned earlier.
Of course, if the data-driven method wants to succeed, in addition to the large amount of data, there is also a premise that the sample must be very representative, which is a sentence in any statistics textbook, but it is difficult to do in real life. ……"
Second, big data and machine intelligence.
"After 2000, due to the emergence of the Internet, especially the later mobile Internet, the amount of data not only increased sharply, but also began to be interrelated, and the concept of big data appeared. After 2000, due to the emergence of the Internet, especially the later mobile Internet, scientists and engineers found that using the method of big data can make a leap in the intelligence level of computers, so that computers will gain higher intelligence than human intelligence in many fields. It can be said that we are experiencing a technological revolution brought by big data, the most typical feature of which is the improvement of computer intelligence, so we might as well call this revolution an intelligent revolution. When the intelligence level of computers catches up with or even surpasses human beings, our society will undergo earth-shaking changes, which is the terrible thing about big data.
So why does big data eventually lead to such a result, and what is the relationship between big data and machine intelligence? To make this clear, we must first explain what machine intelligence is. "
"1946, the first electronic computer ENIAC was born, which made human beings rethink the question of whether the machine can be intelligent."
"The real scientific definition of what is machine intelligence or the founder of the electronic computer Dr. alan turing. 1950, Turing published a paper entitled "Computing Machine and Intelligence" in Thought magazine. In the paper, Turing did not talk about how computers get intelligence, nor did he propose any intelligent methods to solve complex problems, but only proposed a method to judge whether machines are intelligent. " Let a referee sit in front of the tomb and communicate with the "people" behind the scenes. There is a machine and a person behind the scenes. If the referee can't judge whether he is communicating with a person or a machine, then it means that this machine has the same intelligence as a person.
"This method is called Turing test by later generations. Computer scientists believe that if a computer achieves one of the following things, it can be considered as having the kind of intelligence that Turing said:
1. Speech recognition
2. Machine translation
Automatic summarization or writing of text
4. Beat the human chess champion
Answer questions automatically
Today, computers have done the above things, and sometimes they have overfulfilled their tasks. For example, in chess, it has not only defeated the world champion of chess, but also defeated the world champion of Go, and the latter is 6 ~ 8 orders of magnitude more difficult than the former. Of course, it is not smooth sailing for mankind to come to this step, but it has taken a detour for more than ten years. "
Bird school: artificial intelligence 1.0
"According to records, in the summer of 1956, Shannon and a group of young scholars held a brainstorming seminar at Dartmouth College. ..... In fact, this is a brainstorming seminar. This 10 young scholar discussed the problems that computer science had not solved or even studied at that time, including artificial intelligence, natural language processing and neural network. The statement of artificial intelligence was put forward at this meeting. "
"Strictly speaking, the term artificial intelligence has two definitions today. The first refers to machine intelligence, that is, any method that can make a computer pass the Turing test, including the data-driven method that we will often talk about in this book. The second is the concept in a narrow sense, that is, the specific methods of studying machine intelligence in the 1950 s and 1960 s. Today, almost all textbooks with the words "artificial intelligence" in their titles (including the book "Artificial Intelligence: A Modern Method" co-authored by Stuart Russell and Norweg, which is the best seller in the world) still mainly introduce those "old and good artificial intelligence".
Later, in order to draw a clear line between themselves and traditional methods, scholars who used other methods to generate machine intelligence emphasized that they were not using artificial intelligence. Therefore, academic circles divide machine intelligence into traditional artificial intelligence methods and other modern methods (such as data-driven, knowledge discovery or machine learning). Of course, when people outside the computer field talk about artificial intelligence, they often refer to any machine intelligence and are not limited to traditional methods. Therefore, in order to facilitate the distinction, we try to use machine intelligence to express concepts in a broad sense in this book. When expressed by artificial intelligence, it usually refers to the traditional artificial intelligence method, and even we sometimes emphasize artificial intelligence 1.0.
So what is the traditional artificial intelligence method? Simply put, it is to understand how human beings produce intelligence first, and then let the computer do it according to human thinking. Today, almost all scientists don't insist that "machines should think like people to get intelligence", but many laymen still imagine that "machines think like us" when talking about artificial intelligence, which makes them both excited and worried. In fact, when we go back to the origin of Dr. Turing's description of machine intelligence, we can find that the most important thing of machine intelligence is that it can solve problems that the human brain can solve, not whether it needs to adopt the same method as people.
Why are the ideas of early scientists as naive as those of laymen today? This truth is simple, because thinking according to our intuition is the easiest way. In the history of human invention, the early attempts in many fields were to imitate the behavior of people or animals. For example, humans dreamed of flying thousands of years ago, so they began to imitate birds. There are similar records in the east and the west, where a bird's feather is tied to a person's arm and jumped down. Of course, the results of the experiment can be imagined. Later, people called this methodology "bird flying school", that is, by observing how birds fly, they imitate birds to build airplanes without understanding aerodynamics. In fact, we know that the Wright brothers invented the plane by aerodynamics rather than bionics. Here, we should not intuitively laugh at the naive ideas of our predecessors, which is a universal law of human understanding.
When artificial intelligence was first put forward, this research topic was very popular all over the world, and everyone seemed to think that computers would be smarter than people before long. Regrettably, after more than ten years of research, scientists have found that artificial intelligence can't solve any practical problems except making a few simple "toys", such as letting robots pick bananas like monkeys. By the end of 1960s, other branches of computer science had developed very rapidly, but the research on artificial intelligence could not go on. Therefore, the American computer science community began to reflect on the development of artificial intelligence. Although some people think that the intelligence level of a machine is limited because it is not fast enough and its capacity is not large enough, some people of insight think that scientists have gone the wrong way. If they go that way, computers can't solve the intelligence problem no matter how fast they go. "
Minsky cited an example that Bashir used in semantic information processing: the pen is.
"In the Box" and "The Box in the Pen" illustrate the limitations of artificial intelligence at present.
"These two sentences will get the same parsing tree. It is impossible to judge which sentence pen should be used as a fence and which sentence should represent pen according to the two sentences themselves or even the whole article. In fact, people's understanding of these two sentences does not come from grammatical analysis and semantics itself, but from their common sense or world knowledge, which cannot be solved by traditional artificial intelligence methods. Therefore, Minsky gave his conclusion: the' current' (referring to 1968) method can't make computers really have human-like intelligence. Because Minsky enjoys a high reputation in the field of computer science, his paper led the US government to cut almost all the funds for artificial intelligence research. In the next 20 years or so, the research on artificial intelligence in academic circles around the world is at a low tide. "
Another way: statistics+data
"In the 1970s, human beings began to try another development path of machine intelligence, that is, adopting data-driven and supercomputing methods, and this attempt began in industry rather than universities.
At that time, IBM was in a desperate situation in the world computer and even the whole IT industry. ..... At this time, IBM can no longer consider how to occupy a larger market share, but how to make computers smarter.
1972, Fred Jelinek, a professor at Cornell University (1932-20 10) went to IBM for academic leave. Just then, IBM wanted to develop a "smart computer", and Jarinik was in charge of the project temporarily. As for what an intelligent computer is, at that time, everyone knew that it could either understand people's words and translate one language into another, or win the world chess championship. According to his own specialty and IBM's conditions, Jarinik chose the first task, that is, the computer automatically recognizes human voices. "
Jarinik believes that speech recognition is an intelligent problem, but a communication problem, that is, human speech is a process of brain coding. After the coding is completed, it is transmitted to the listener's ear. The listener's acceptance and understanding is a decoding process, and the speech recognition problem can also be handled in this way. He "used all kinds of mature digital communication technologies at that time to realize speech recognition, and completely abandoned the set of methods of artificial intelligence (referring to traditional methods,
"When studying speech recognition, Jarinik and his colleagues inadvertently created a method to solve intelligent problems by statistical methods. Because this method needs a lot of data, it is also called data-driven method. The biggest advantage of this method is that with the accumulation of data, the system will get better and better. Compared with the past artificial intelligence methods, it is difficult to benefit from the improvement of data. "
"After speech recognition, scientists in Europe and America began to consider whether data-driven methods can be used to solve other intelligent problems. Jarinik's colleague Peter Brown applied this data-driven method to machine translation in 1980s. However, due to the lack of data, the initial translation results are not satisfactory. Although some scholars agree with this method, others, especially those who worked in this field in the early days, think that it is not enough to solve intelligent problems such as machine translation through statistics based on data. From the early 1980s to the mid-1990s, there has been a controversy in the computer field, that is, whether the data-driven method is suitable for all fields and whether speech recognition is just a special case. To put it simply, scholars who were engaged in speech recognition, machine translation, image recognition and natural language understanding were divided into two distinct factions. One group insists on solving problems with traditional artificial intelligence methods and simply imitates people, while the other group advocates data-driven methods. These two schools have different strengths in different fields. In the field of speech recognition and natural language understanding, the school that advocates data-driven has gained the upper hand relatively quickly. In image recognition and machine translation, for a long time, the data-driven school was at a disadvantage. The main reason for this situation is that in the field of image recognition and machine translation, the amount of data in the past is very small, and the accumulation of such data is very difficult. Needless to say, image recognition, before the advent of the Internet, no laboratory had millions of pictures. In the field of machine translation, in addition to general text data, a large number of bilingual (even multilingual) data are needed. Before the advent of the Internet, similar data could not be found except for the Bible and a small number of UN documents. " However, with the rise of the Internet, data acquisition becomes easier. In 2005, Google beat all the machine translation research teams in the world with data-driven method. It won because it used thousands or even tens of thousands of times more data than other research institutes.
"Nowadays, in many research fields related to" intelligence ",such as image recognition and natural language understanding, if the methods adopted cannot take advantage of the greatest advantages of data, they will be considered outdated.
Data-driven method began in 1970s, and developed slowly but steadily in 1980s and 1990s. After entering the 2 1 century, due to the appearance of the Internet, the amount of available data has increased sharply, and the advantages of data-driven methods have become more and more obvious, and finally the leap from quantitative change to qualitative change has been completed. Now computers can do many things that need human intelligence, benefiting from the increase of data.
Data in various fields all over the world has been expanding outward, and gradually formed another feature, that is, a large number of data began to cross, and data in various dimensions gradually changed from points and lines to networks, or the correlation between data was greatly enhanced. In this context, big data has emerged. "
"Before the emergence of big data, computers were not good at solving problems that needed human intelligence, but today these problems can be solved by changing ideas. The core is to turn intelligent problems into data problems. As a result, the world has started a new round of scientific and technological revolution and intelligent revolution. "
Although computers can do more and more things in recent years, which gives people the feeling that they are "fast but not smart enough", when we have enough data, we can turn intelligent problems into data problems, and machines no longer need to think and solve problems like people. As long as we input enough data and cooperate with the appropriate algorithm (model), the machine can make the best judgment. Even though Deep Blue defeated Caspar, "but behind this seemingly smart appearance, it is actually a combination of a large number of data, rather than complex algorithms and super computing power-Deep Blue has never thought like humans."
"Computer playing chess and answering questions reflect the decisive role of big data in machine intelligence. We will see many kinds of robots in the future, such as Google self-driving cars, computers that can diagnose cancer or write articles for newspapers. They don't need to be humanoid like robots in science fiction movies, but they are all smarter than humans in some ways. Behind these robots is a powerful server cluster in the data center. In terms of methods, the way they get information is not through reasoning like us, but by using big data to learn information and knowledge from data. Nowadays, this world-changing revolution triggered by big data has quietly happened, and we will introduce it in more depth in the next chapter. This technological revolution is characterized by the intelligence of machines, so it is not an exaggeration to call it an intelligent revolution. "