Teach robots to learn.

If, one day, in every family, there is a robot that washes dishes and cooks, cleans the room and undertakes all the housework; In every company, most of the work is done by robots, and human beings can completely get rid of coolies. Does this future appeal to you? But in order to enter such a dream era of artificial intelligence, we must first teach machine learning.

Let the machine learn by itself

Just as everyone must learn and train before acquiring skills, machines can only become intelligent through learning. Machine learning originated from a branch of artificial intelligence. In this field, computer science tries to create computer intelligence similar to human beings.

Real machine learning is fundamentally different from what we think of as traditional programming. When it comes to computer programs (or algorithms used in programs), we usually think that engineers give computers a series of instructions, telling them how to process a series of inputs and then produce corresponding outputs. The browser will track the visited web pages and then respond to the user's input in a predictable way. But these are all coded by human beings in advance, not the result of active learning by machines.

Machine learning means that the machine programs itself. These machines can be programmed like humans after training. In 20 15, Google released a picture recognition application software called "Deep Dream", which can not only recognize images, but also create some unexpected fantasy scenes with images. For example, if you present a landscape map, the software will analyze your picture and output a landscape map in the eyes of a computer.

How is this done? The working principle of Google Deep Dream is to establish a computer's own neural network system by simulating human neural network, and obtain the information of objects through neurons for analysis. The computer's neural network system contains thousands of interacting neurons to achieve accurate mathematical operations.

Of course, in order for machines to recognize the information of objects, researchers have been training computer neural networks with a large number of pictures for the past four years, for example, showing many pictures to Deep Dream software and telling the theme of each picture. Once you have seen hundreds of dog heads from hundreds of angles for a thousand times, you can learn to output images by yourself.

In the experiment, the "deep dream" produced a blurred image mixed with the outline of birds, eyes and dog heads. Although they are not so lifelike, they also reveal the creativity of computer image processing. It has learned to recognize the faces of kittens and puppies without human supervision and guidance.

The machine structure of neural network imitates the human brain and gives full play to the super memory function of computer, so it is more widely used in life. Google's search engine, Amazon's recommendation directory, Facebook's friend dynamics and spam filtering, as well as military, financial, scientific research, and autonomous driving, which is more reliable than human driving, are all concrete applications of neural network computing.

Nowadays, machine learning has been successfully applied in more fields, from data mining programs to detect credit card transaction fraud, to information filtering systems to gain users' interest in reading, and to cars that can drive themselves on expressway. It can be said that our intelligent life is the result of machine learning.

Research on robot learning algorithm

Machine learning can be traced back to around the end of World War II in the 1940s. Due to the painstaking efforts of the scientific elites of the warring parties during the war, computer theory has developed by leaps and bounds during that period. At that time, cybernetic researchers conceived a neuron computer model, which could roughly simulate biological neurons and could be expressed in concise mathematical form.

However, in the face of an uncertain and diverse world, it is simply more difficult to cope with the mathematical form formulated in advance. In other words, the world of manual programming is too far away from the real human world, and the real world will not be so orderly and disciplined.

For example, the computer mathematically expresses in advance that a horse has four legs, but this will cause two problems. First of all, how does the computer learn to understand this fact? Secondly, what should be done to those horses who lost a leg in the accident? These seemingly stupid questions are the biggest obstacle to manual programming. This is why search engines can't answer questions, but only search for keywords.

But if machine learning creates a self-programming system, it can respond to its own mistakes and constantly update its internal state. Vulnerabilities in manual programming need to be detected before release, and machine learning algorithms can constantly correct errors in the process, which is more flexible and intelligent than artificial intelligence.

But how does the machine learn? This involves the algorithm. It can be said that the algorithm is the core of building the Internet. Nowadays, many online search and communication methods are based on set mathematical formulas, such as Google search engine, Apple voice system and Facebook. In the information age, our life is actually guided by some mathematical formulas. In medicine, there are formulas for calculating diabetes and malaria. Today, we also asked the formula learning machine to check the X-ray quantity during chest X-ray.

Connecting learning, symbolic deduction, Bayesian learning and analogical learning are four contemporary computer learning paradigms. Among them, connection learning is to imitate the human brain nervous system and establish a computer artificial neural network; Symbolic deduction is to express a problem or knowledge as a logical network and learn through symbolic deduction. Bayesian learning theory is a process of learning and reasoning through probability rules; Analogical learning is learning by comparing similar things. These four learning paradigms are still too complicated. Therefore, Pedro Domingos, a professor at the University of Washington, provided a bolder assumption that the existing algorithm formulas can be unified into a general algorithm in the future.

The world under general algorithm

Domingos conceived such a general algorithm, which can unify the theories that have been found in physics and biology into a standard model or central rule, and at the same time can discover all the knowledge, all the existing knowledge and all the future knowledge from the data. For example, the universal algorithm can be obtained from Tycho? In Bula's space observation, Newton's law is introduced, even though it has no relevant basic knowledge.

The cerebral cortex may be a typical example of this general algorithm. Some neuroscientists believe that in all fields, the cerebral cortex can constantly adjust the functions of the subcortical brain and spinal cord by using only one formula, so as to constantly learn to adjust according to the environment and hear, see or understand the meaning of the surrounding world.

In the information age, general algorithms will also play a role similar to the cerebral cortex. It can learn and use information on the basis of data cloud, change the current rigid and passive execution mode of computers, actively improve functions and improve output, which will bring revolutionary changes to human information life.

For example, the Internet has a huge amount of information, and you can get millions of web pages by typing a few keywords, which often makes people have no choice. But with the general algorithm, the computer will become an encyclopedia. Just ask a few questions and you can give an accurate answer quickly.

At the same time, today's recommendation system will be completely updated. Now, everyone will encounter a lot of recommendation information. Based on the fragmented data left by everyone, the million recommendation system will recommend different things for you every day: Storm Video will recommend movies for you when you start watching their movies; Amazon recommends books for you according to what you bought and what you didn't buy; Sina will recommend hundreds of interest groups when you register. But what most people may need more is a more intelligent system, which can provide more targeted recommendation services based on all the data and information you generate online. For example, it can recommend corresponding things at every stage of your life, not only books and movies, but also houses and jobs. In order to achieve this effect, you first need data from daily life, but on the other hand, you also need general algorithms, because in the face of a large number of data, there is no formula that can't be processed.

If we can successfully find a general algorithm, artificial intelligence will be truly realized. However, for this future intelligent picture, it is inevitable to worry. For example, machine learning can handle most jobs, and there will be a large number of unemployed people in the world. How will they survive? Will it become the source of social instability? If artificial intelligence is used by politicians with ulterior motives, will the world be peaceful?

Pedro Domingos, the advocate of general algorithm, is not worried about these problems, but very optimistic. He believes that computers will not have the ability of biological evolution, nor will they invent things themselves, so they will not pose a threat to mankind. On the contrary, boring jobs will be undertaken by robots, and humans will do more interesting jobs. The environment of the earth will be better and better, and human beings will live longer, be happier and be more creative. At the same time, people will not appear on the battlefield, and robots will come out in person, which will prevent human beings from being disabled by war.

(This article comes from the article number 1, 20 16, New Theory of Big Science and Technology * Encyclopedia).