Title from: Vision China
Talking about artificial intelligence, I believe that readers will not be unfamiliar, and the meaning of the words is easy to understand.
In some people’s eyes, artificial intelligence is already a very mature computer technology, which can help us do a lot of things, to predict the weather of tomorrow in local areas, analyze the fluctuations of stock trading in financial markets; We do very small things, small stickers to automatically follow the face, push the news or goods I am interested in, these can be achieved with artificial intelligence.
But in other people’s eyes, artificial intelligence is still a new thing in the laboratory. It is still far away from our life. The artificial intelligence we are now exposed to is at most a personal mental retardation. The concept of intelligence is just the gimmick of the capital game. In fact, it has not changed anything.
Artificial intelligence is not a master key, not everything can be solved with artificial intelligence. But again, artificial intelligence is not a vain new technology, and its presence has appeared in every aspect of life.
In my daily work, I found that many of my friends and Internet practitioners also have a lot of misunderstandings about artificial intelligence. I tried to combine my past experiences and talk about my views. Let’s talk about it first, the concept of artificial intelligence is suddenly so hot.
01 Why did artificial intelligence suddenly rise?
Many people mistakenly believe that artificial intelligence is a new concept that has been created in recent years.
In fact, the concept of “artificial intelligence” was first proposed at an academic conference at Dartmouth College in 1956. Although the meeting lasted only one month and did not make any substantial progress, the conference officially proposed the term “artificial intelligence” for the first time and it is still in use today.
Although at the time, research on artificial intelligence was very slow, the classic sci-fi movie “2001 Space Roaming” expressed the illusion of artificial intelligence at that time. After nearly 50 years of development, artificial intelligence has gradually entered the daily life of people from the fantasy of movies, and has become a powerful assistant in various fields.
This process is not always smooth.
In the 1980s, the Japanese studied a computer system that could simulate the decision-making ability of human experts, called an expert system. This expert system is actually a huge knowledge base, and through some inference rules, the system can find answers based on questions.
This expert system is able to provide answers based on input questions that are representative of artificial intelligence technology at the time, and to a certain extent, the performance of the computer “intelligent.” Therefore, this project has received great attention from the Japanese government, and has invested a lot of human and material resources research, hoping to create an expert system with faster calculation speed and higher knowledge reserve. Stimulated by the Japanese, the United States and many European countries have also entered this track.
It is foreseeable that the initial success of the expert system was limited because it was unable to self-learn and update the knowledge base, which was extremely costly to maintain. Just like the car navigation system that was not connected before, the map needs to be updated every year. Otherwise, the system will be abolished after one year and the correct guidance cannot be given.
The failure of the expert system has also caused people to have a huge crisis in the trust of artificial intelligence, the collapse of the hardware market and the confusion of theoretical research, and the cessation of the government and institutions to artificial intelligence Funding in the research field has led to years of lows.
When capital no longer focuses on artificial intelligence, the theoretical research on artificial intelligence is still slow. In 1988, American scientist Judea Pearl introduced probabilistic methods into the process of artificial intelligence, which had a major impact on the development of artificial intelligence. In 1989, Yann LeCun and the team at AT&T Bell Laboratories used convolutional neural network technology to implement artificial intelligence to identify handwritten zip code digital images.
In the last two decades, artificial intelligence technology has gradually integrated with computer technology and the Internet. The development of four major catalysts, such as massively parallel computing, big data, deep learning algorithms and human brain chips, and the reduction of computing costs have made artificial intelligence technology soar.
It benefitsWith the development opportunities of computers and the Internet, it is called business intelligence, data analysis, informationization, automation, etc., and penetrates into every corner of social development. On the one hand, the promotion of the Internet has created a lot of application scenarios for artificial intelligence, reflecting the true value; on the other hand, the upgrade of computer hardware and software provides powerful computing power for artificial intelligence, which was previously theoretically The implementation of the algorithm has come to the fore, allowing artificial intelligence to create miracles in more and more events, even beyond humans.
In 2011, Watson defeated human players in the natural language quiz competition. The accuracy of image recognition algorithms surpassed humans in the ImageNet Challenge. In 2016, AlphaGo defeated Li Shishi and became the first AI robot to defeat the world Go champion.. .
02 Artificial intelligence or artificial mental retardation?
In the past two years, artificial intelligence has been criticized most by people: artificial intelligence does not reflect intelligence.
Many people’s perceptions of artificial intelligence are split. On the one hand, the media constantly reports on artificial intelligence and what new achievements have been made. The foreign big coffees make people wary of the development of artificial intelligence, artificial intelligence is also included in the planning of China’s development, etc.;
On the other hand, there is often an automatic driving in the news and an accident. The smart furniture in the home is like a mental retardation. The information platform always pushes the same type of news in a foolish way. These phenomena make us wonder. Where is artificial intelligence?
Before answering this question, we need to figure out the difference between strong artificial intelligence and weak artificial intelligence.
In the beginning, there was no strong or weak point in the word Dartmouth at the Dartmouth meeting. It is widely believed that artificial intelligence is to let the machine have ideas and be able to make decisions like humans. At that time, the research of various algorithms was also aimed at this goal, hoping to simulate the way human decision-making gives the machine real intelligence.
But some people soon discovered that the artificial intelligence implemented in this way is not really intelligent, but a simulation of human intelligence. The American philosopher John Searle proposed a thinking experiment: Chinese room , which is like this:
Imagine a person who can only speak English in a room. This room is closed except for a small window on the door. He carried a book with a Chinese translation program. There is enough manuscript paper and pencil in the room.
Paper written in Chinese is sent to the room through a small window. People in the room can use his book to translate the text and reply in Chinese. Although he does not speak Chinese at all, Searle believes that through this process, people in the room can let anyone outside the room think he is fluent in Chinese.
It’s worth noting that this book is only a grammatical correspondence and does not involve any semantic description. The people in the room only need to follow the corresponding answers and put together the corresponding Chinese characters to hand out. In the process, he did not understand the question and what the answer he wrote was.
Searle believes that artificial intelligence works like this. He thinks computers can’t really understand the information they receive, but they can run a program, process the information, and give a smart impression.
For example, image recognition technology works by changing the color into a digital code, then finding features from these digital codes, finding a dictionary, finding the corresponding explanation and then displaying it. In fact, the computer does not know whether it is an airplane or a rabbit, but the dictionary tells it that this feature has a large probability corresponding to the word “aircraft”.
Most of the algorithms are essentially games that play probabilities. The different ways are just the different information required for model training, and the way in which the corresponding “airplane” is calculated is different.
< /p>
All current well-known and well-known models use matrix computing training data to obtain a certain probability distribution. The probability distribution of complex models is usually high-dimensional, and various mathematical methods are derived here, but the essential idea is still to describe the characteristics of training data through probability distribution. With these, for the same kind of data, the same probability distribution can be used to describe, so as to achieve the so-called “identification” or “prediction”.
In fact, it’s not that the model really understands what an “aircraft” is like humans, but in this way, the model can recognize images that look like airplanes with great probability.
The industry has generally recognized this. Therefore, the concept of artificial intelligence is further divided into strong artificial intelligence and weak artificial intelligence.
Strong artificial intelligence genre is still pursuing the computer to have the human mind and consciousness, with independent choice behavior. Just like Memph, which evolved from the curing process to self-awareness in the Western world. However, the research of strong artificial intelligence is difficult, and there is no mature application on the market.
When weak artificial intelligence is more like a tool for solving specific problems. The characteristic of such problems is that statistics can be used to summarize experiences and form solutions, and this method of solving problems is called “machine learning.”
The most basic approach to machine learning is to use algorithms to parse data, learn the rules of the data, and then make decisions about events in the real world. Different from the traditional programming method, machine learning is trained with a large amount of data, and learns how to complete the task from the data through various algorithms.
Quantitative transactions, face recognition, and AlphaGo are all machine learning models that excel at a single aspect. When training the model, we only taught AlphaGo the skill of Go, so it can only play Go. If you throw a math problem to AlphaGo, obviously it is impossible to start.
All machine learning models can only perform specific tasks, and many times we can meet more scenarios by combining them. For example, a smart speaker is essentially a speech recognition model combined with NLP (Natural Language Processing) model, it is not really able to understand what we mean by what it means, just to convert the received information into the input of the model, find the corresponding output in the dictionary. .
From the characteristics of machine learning, if you want to summarize experience through statistics, the quantity and quality of data is the decisive condition. Without data, there is no artificial intelligence.
In other words, artificial intelligence can’t make judgments when you don’t make the same type of behavior, or there are fewer people who are close to your behavior. This is also an important reason for artificial intelligence to become artificial mental retardation. As behavior increases, data becomes more and more, and data quality gradually rises, you will find that predictions are more and more accurate, and artificial intelligence can truly “think what you think” through big data.
03 What problem is suitable for machine learning?
Before we talked about weak artificial intelligence like tools, specifically to solve a specific problem. But is all the problems suitable for machine learning to solve? Obviously the answer is no.
There are three basic conditions for solving problems that are suitable for machine learning.
(1) You can learn regularly. This type of problem must have commonality, and there is an inherent law waitingBe discovered;
(2) Programming is difficult to achieve. There is a complex relationship between data, and it is difficult to list the rules in an exhaustive way;
(3) There are enough data to learn the rules. Without data support, machine learning is like setting up a house with less brick structure.
Have a chestnut.
We are familiar with spam detection as a classic scenario using machine learning. The most common types of spam are various types of marketing messages, and the senders of such messages are usually various types of websites that have been registered with a mailbox. In this scenario, we found that marketing emails must contain certain product information or promotional information, so such emails have certain rules.
But because of the different product types, it is difficult to write all the rules in a programmatic way. Even if it can be written, the sender will design various rules to avoid the detection of the system. At the same time, we can easily find a lot of spam and normal mail as sample data. So this scene is very suitable for machine learning.
But if we want to determine how many characters a new message contains, I’m afraid it won’t work. Although this problem is equally difficult to solve with programming and has a large amount of historical mail support, the question of how many characters are included is too random and has no rules to follow, so it is not suitable.
It can be seen that machine learning is not a panacea, not all problems can be solved with it. Machine learning is good at finding the law through known experiences to solve problems. If there are no rules to follow in the face of the problem, it is completely a random event, so even if you use more complex machine learning algorithms, it will not help.
It is worth noting that many problems seem to be irregular. In fact, because humans can’t handle the situation where the amount of data is too large, it seems that the messy data masks the behind-the-scenes. Such problems are not really unseen. Just need to use the right method.
We can use machine learning to analyze large amounts of data to obtain rules and use rules to predict unknown data. Not only can we see the laws that humans can see from the data, but more importantly, we can discover the laws that humans cannot see in a shorter period of time. I think this is the greatest application value of machine learning.
In the medical field, through image recognition technology, it has been realized that the computer can automatically recognize tumor cells and help doctors to quickly diagnose medical diagnosis;The way of learning automatically detects product defects and improves the production rate, helping enterprises to speed up the production cycle and reduce production costs. In the financial field, traditional networked software can avoid traditional programmatic transactions because it cannot adjust the algorithm according to real-time market changes, resulting in asset loss. risks of. There are also a wide range of applications in machine learning in the fields of retail, security, aviation, internet, etc. It has changed tremendously in our lives.
In the end, we must realize that the current artificial intelligence is not really intelligent, just an intelligence that simulates human behavior. The true intelligence is still far away from our lives. Fortunately, only the intelligence that simulates human behavior has brought such great convenience to our lives. I believe that with the development of technology, we can make more scenarios beyond imagination.