In this article, there will be no code, no theory, and no knowledge. All are personal views and insights that are more macroscopic and abstract, and will be as simple and understandable as possible except for the fifth part. Can understand.
The main contents are:
1. The current limitations of artificial intelligence and machine learning
(also known as: the media hyped the threat theory of artificial intelligence, the engineer smiled)
2. Use a simple metaphor to explain the essence and value of "learning"
(aka: the essence of learning is fitting)
3. Continue to use the above analogy to explain exactly what machine learning is doing
(aka: machine learning comes with a gene for brute force method)
4. Continue to use the above analogy to explain why machine learning needs an ecosystem
(aka: What are primary school students doing now?)
5. Propose a preliminary architecture of a machine learning ecosystem named "Dual Decision Core Learning System".
(Aka: Put humans and artificial intelligence in an equal position in the system)
1. The media vigorously promoted the threat theory of artificial intelligence, the engineer smiled
Recently, since the battle between Alfa Dog and Li Shishi, artificial intelligence has really caught fire, and then the artificial intelligence threat theory has regained its trend once again every few decades like a certain retro fashion. Below is a vivid and somewhat ironic picture.
But as anyone who has an understanding of artificial intelligence, it is known that artificial intelligence is currently in a huge bottleneck rather than threatening humanity.
Why do you say that? For any normal elementary school student, they can learn mathematics in this lesson and learn pinyin in the next lesson. If they are lucky, they will also be taken by a Chinese teacher to watch the plants in the small garden to understand the world. The proposition of "the corner of the world".
In other words, real intelligent individuals (primary students) use the same set of decision cores (brains) to complete the various tasks mentioned above. If you let an engineer train a neural network that can recognize pinyin and learn arithmetic and writing, he will make you look. Elementary school students themselves will never know what they have in common in learning arithmetic and learning pinyin, so that they can use the same system to complete these two very different tasks, not to mention that when they grow up, they may face Complex scenarios.
If human intelligence is likened to a human body with complex structural systems composed of various cells with different degrees of differentiation, the current "artificial intelligence" is at most a highly differentiated cuticle. People would think that there is a threat to the cuticle cells, which is really ridiculous, especially when they are completely alone, unable to communicate and cannot be combined, the biggest threat is nothing more than letting yourself fall down to create some dandruff for you.
So if you want to get rid of the bottleneck and create real artificial intelligence, you need a model with a lower degree of differentiation, a model with a high degree of differentiation that requires different functions in addition to the "cuticle", and most importantly, a "model center". Models with different degrees of differentiation and different functions need to rely on this hub for communication and collaborative work.
Second, the essence of learning is fitting
Personally, the essence of learning is fitting, whether it is human learning or machine learning.
Why do you need to learn, why do you need to fit? Let me put a conclusion that does not speak human words:
"Learning" is the process of human beings' understanding of the world from a higher dimension.
Next, I will explain this sentence in two steps:
1. What is a higher dimension?
2. Why do I need to know the world from a higher dimension (= Why do I need to learn)
What is higher dimension
Speaking of high dimensions, people who have studied linear algebra may think of matrix dimensions; people who have seen Interstellar Crossing may think of high latitude spaces; and those who have done data mining will think of the dimension of features. In fact, there is a deep close relationship between these three dimensions, and I wo n’t elaborate here. Interested people can imagine and think by themselves after reading this article. This will be a very interesting process. . Here I choose to explain from the third angle, the dimension of the feature, what is high latitude.
As an example, this example may have been used in the third and fourth parts of this article.
First look at the Forbidden City. First of all, the Forbidden City stands on the ground as an object with a certain volume. Suppose your family lives near the Forbidden City, or more directly assume that you live in the Forbidden City. You can go to and from work every day. You may have seen the Forbidden City many times since you were young. The Forbidden City may be a very familiar thing in your heart. Too.
But this is of no use. Your familiarity with the Forbidden City is just an illusion. Because in simple terms, the Forbidden City is just a very flat and rigid existence in your mind at this time. Why do you say that? If a foreign friends to Rob, he asked you in broken Chinese: "wow you live in the Forbidden City so powerful, that the Forbidden City covers an area of much ah?"
This time you will look ignorant force, the presence of so many years in front of you In the Forbidden City, you have seen it countless times but you will not know its length, width and height because you did not perform the action of "learning".
The length, width, height, and floor area are just a few of the basic characteristics of the Forbidden City. At the same time, there are many foreign friends, and you can ask countless questions. You are very angry, you are determined to start "learning" trying to master enough knowledge to meet all foreign friends.
So you know how many rooms there are in the Forbidden City, what is the area of each room; how many corners are there in the corridor, how many different colors of paint are used, what does each color mean, and how much is placed Antiques, what is each, what is the name, what is the meaning and story ...
And all this knowledge is characteristic in the eyes of engineers. How many of these features can there be? That is, how many dimensions? Thousands of dimensions, hundreds of millions of dimensions, even infinite dimensions. Of course, your energy is limited. You have learned the characteristics of 500 dimensions in order to deal with foreign friends. It is very powerful. It is already the "Forbidden City."
Therefore, the Forbidden City is no longer a physical existence that you pass through every day. All this knowledge, plus your original stereotyped impression of the Forbidden City, forms a brand-new 501-dimensional Forbidden City in your mind. In other words
Forbidden City in your head, into existence five hundred and one-dimensional.
In other words, the Forbidden City you know at this time is a 501-dimensional one-dimensional mapping of the real Forbidden City from your perspective (a certain angle).
What kind of existence is the Forbidden City? This is a classic philosophical question, and I dare to give an answer here. The Forbidden City is one, which exists in the infinite dimensional space. What is the concept of the existence of infinite dimensions? ——The idea of ordering the original existence of the ruler in the Forbidden City, every insect that was accidentally sealed in the gap between wood and wood during the construction process, and every bit of a trace left by every person who lived there Everyone who visits it, every opinion it produces, every legend and impression about the Forbidden City circulated ... These are all in the infinite dimensional feature, except for the length, width and height.
Through learning, you use human intelligence to successfully fit a 501-dimensional one-dimensional palace in your mind. But you can't fit the most original and true forbidden dimensional forbidden city through learning.
(Private goods are simply entrained here, and there are other ways to learn things that cannot be done. For example, meditation, it is completely out of reach here, stop.)
Why do we need to know the world from a higher dimension
Let's start with the simplest, most direct and violent point of view, why we need to understand the world from a higher dimension. Back to the foreign friend named Rob, and the question he asked, "What is the area of the Forbidden City?" As a person who lives in it every day, you can't answer it, but someone can certainly answer it, such as the Land Bureau, or cultural relics. Bureau.
Why should they know? Because only by knowing this knowledge can they do something, such as improving the underground drainage system of the Forbidden City and renovating the Forbidden City.
To give another simple example of living close to life, high school biology has to learn the cell model. Even if you put a real cell cut like a goose egg in front of you to let you see and touch at will, you will still be ashamed. You will not know the name, structure, function of the various organelles of the nucleus, cytoplasm, mitochondria, chloroplast, and Golgi apparatus, and how they collaborate. About equal to know nothing.
Recognizing whether the world has many or few dimensions, multiple angles or a single angle, wide or narrow, is a different mode of thinking. It is reflected in the current narrow machine learning process (that is, mathematical modeling), which is feature engineering, which determines the upper limit of your model's ability; reflected in human intelligence, it is a person's view of everything, which determines his thinking. Realm and what he will do.
3. Machine learning comes with a gene for brute force cracking
Learning is something that humans have done for thousands of years using their own brains and some simple or complex tools (paper and pens, experimental equipment ...). And now, human beings have developed a new tool to assist themselves in learning the world. This is the machine.
In a nutshell, this is machine learning.
So how do machines help humans learn? Back to Rob, the forbidden lover who asked questions, this time he asked you: "How many load-bearing walls are there in the Forbidden City?" Well, this feature is not in the one hundred and one dimension you have learned, you tell him Tell him the answer when he meets next time.
After returning home, you basically have two options:
one, find a way to get the drawing of the Forbidden City, and then get the answer through the information in the drawing;
second, go through each wall of the whole Forbidden City by yourself, and count the answers.
The first method is the "expert" method. That is the typical way that humans have known the world for a long time-to figure out the principles and truth behind things, and then to figure out the answers to specific questions. Once you find the drawings, you can not only know "how many load-bearing walls are there in the Forbidden City", but also easily grasp the "how many rooms", "what structure is inside" and so on. And this method also has an advantage. Once you get the drawings, basically your answer will not be wrong. However, if you want to get a drawing, it is a difficult problem for the boss.
The second method is the brute force method. It is time-consuming and labor-intensive, and there is a high possibility of errors. Generally, only an approximate result can be obtained. However, as long as you are willing to work hard, you can definitely get an answer. There is no possibility of getting the drawings. But how big is the Forbidden City, it is too tired to go through it by yourself. So you hired 50 elementary school students who have no money and no money to do this for you.
The 50 elementary school students do not understand anything, neither you understand the Forbidden City, nor the load-bearing wall, and they are like scattered sand. They will only execute the instructions you have given clearly enough, and they are the helpers humans have invited—machines.
The question of how to dismantle the "how many bearing walls are there in the Forbidden City" turns it into concrete executable steps, how to organize and lead 50 pupils, how to arrange them to traverse the Forbidden City with high efficiency and low error, 50 Elementary school students may develop hierarchical and hierarchical relationships, how to make them adaptively adjust to the specific terrain encountered to improve efficiency, which can also run through some reward and punishment mechanisms. These are all algorithms.
With the advent of machine learning, people who might have known nothing about the Forbidden City, but those who are very good at leading elementary school students have the opportunity to solve the problem of "how many load-bearing walls does the Forbidden City have?"
And "fitting" is almost synonymous with violent dismantling.
4. What are primary school students doing now?
In the first part, it was mentioned that artificial intelligence is currently in a big bottleneck period. Why does this bottleneck exist and how to solve it. A more abstract solution will be given in this section, and a slightly more specific solution will be given in the last section.
Before that, in order to avoid conceptual confusion, first distinguish between the three concepts of "data mining", "machine learning", and "artificial intelligence".
"Data mining" is the most specific definition. As the name implies, it is the process of extracting the answers to questions from a bunch of data collected by a group of elementary students, and using another group of elementary students to perform various operations on the data. Data mining and "intelligence" are not much related, both intelligent and unintelligent, can be used for data mining.
"Machine learning" is a more abstract concept than data mining. "Mining" emphasizes that the answer to the question or new knowledge is finally obtained, while "learning" emphasizes the ability to give the machine adaptive and human-like learning.
"Artificial intelligence" is one of the three concepts that emphasizes "intelligence". "Manual" refers to "machine" to a large extent, and "learning" is a part of "intelligence". So personally think that it can be said that "machine learning" is part of "artificial intelligence". In addition to "learning", "intelligence" should also have other abilities, such as "entertainment (self-satisfaction)" and "imagination (creation)". Quite fun.
There is one point here, not to mention "self-satisfaction" and "creation". At present, artificial intelligence is far from achieving the function of "learning", and at present, the main focus is on "learning", so in this article In other less rigorous articles, there is no strict distinction between the concepts of machine learning and artificial intelligence, and it is often used in which scenario to look at which one is used (in general, machine learning is much more specific than artificial intelligence).
Back to the title of this chapter "What are the primary school students doing now?" They are busy counting how many load-bearing walls, how many rooms, how many pillars, and how many people can live in the Forbidden City. Almost all of them are connected to the shaft and are busy solving various specific problems. In fact, they only need to fit the drawings of the entire Forbidden City once to get answers to many questions.
Why is there a situation in which elementary school students are currently busy with a high degree of differentiation? Primary school students are expensive, and those who can get answers to questions by manipulating them are also expensive and difficult to find. The main purpose of Party A who has the ability to build such an expensive system is to make money, and getting the answer to a certain question is only a step that is a bit important or even less important in their complicated process of making money. So for those who bring their short-sighted genes (not derogatory), drawing the entire drawing is actually a low-cost method.
The second reason is that for real-world problems, such as a recommendation system problem, a click-through rate prediction problem, and an image recognition problem, what exactly is a "drawing"? This problem is still very blank in theory and practice.
And because as mentioned in the preface, all the content of this article does not refer to any literature, it is entirely my own understanding and intuition. And the lack of intensive work on theoretical knowledge has left the author who has explained the problem to this point in a state of sincerity and fear. But even so, I feel that it is valuable to articulate my ideas more clearly (even for me).
Personally, I believe that the knowledge graph, or the broad knowledge graph, has the potential to become a drawing.
Personally, the similarity comparison of complex structures in high-latitude space is related to the essence of learning.
The first point may not need any explanation. The knowledge graph has the potential to become a "central structure" for communication and collaboration between models, because the knowledge graph itself has both the knowledge and the genes of the model, or it is a kind of knowledge Structured representation. I can't say anything more specific and less learned, and confirming and realizing this conjecture is a relatively long-term goal for individuals.
The second point is more abstract. I have been looking for a model to learn (master) knowledge efficiently. Compared with the structured knowledge that has been formed in terms of structure, the knowledge expansion on this basis is one of the most efficient learning methods that I can obtain through my own learning process. Here is a deeper step, and it is a bit super-level, so I won't go into it here.
V. Putting humans and artificial intelligence in an equal position in the system
Here simply throws a definition, or a very basic prototype.
Double-decision core learning system = human intelligence + model hub + model + data collection mechanism (reptile)
System composition:
1. Model hub: Responsible for coordination and communication between models and models, as well as communication with human intelligence. It is one of the optimization goals of the continuous evolution of the entire ecosystem. It can be roughly regarded as artificial intelligence. In the end, what exactly serves as the model center, different systems can have different choices. (At present, I will consider using some kind of decision engine developed by knowledge graph)
2. Model: If the model center is the brain of artificial intelligence, then the model is the hands and feet. It is responsible for implementing various functions of differentiation. It accepts the input of external data and the input of internal data (model hub), and the input of human intelligence to the model is often as common as air.
3. Data collection mechanism: equivalent to reptiles, may be more structured reptiles. In different system designs, it can accept the control of human intelligence or the control of the model center, or it may be controlled by two core intelligences at the same time.
4. Human intelligence: it can serve different kinds of functions in different system designs. In the primary system is the existence of the leader, while the higher-level system is more independent of human intelligence. The ultimate goal may not be to completely retire human intelligence (it may fall into a certain extreme). Instead, it allows human intelligence and artificial intelligence to work together as efficiently as the human right and left brains.
Core features
(The difference between an ecosystem and a common machine learning process)
Equalization: Let the model center and human intelligence jointly determine the direction of system (model) optimization.
Self-adaptation (tentative) : It is not optimized to solve specific problems, but to optimize itself.
If you must consider the current lack of structured, very common, from crawler to modeling output process as the "machine learning ecosystem" defined in this article. As a result, in this type of process, human intelligence permeates every trace of the system, and the role of artificial intelligence (model hub) can only see a shadow in the steps of feature selection and model evaluation. Second, as mentioned above, the purpose of the entire process is to solve specific problems, and the significance of the machine learning ecosystem is to self-evolution.
The value and significance of the dual-decision core learning system
In the short term, this structured design is expected to break the current bottleneck of artificial intelligence and allow the pace of artificial intelligence to move forward. In the long run, the purpose of the evolution of the entire ecosystem is to coordinate and optimize the left and right brains of the system of human intelligence and artificial intelligence. Even for me personally, what I most want to see is the optimization of artificial intelligence to human intelligence. This statement looks a bit scary, to be more specific. For example, I think artificial intelligence is very promising to help human intelligence discover the connection between any two things. Let human intelligence be liberated from the left brain, or even replace the left brain, and more develop the potential of the right brain, which is also possible for artificial intelligence .
For the holders of artificial intelligence threat theory, what I want to say to them is. Artificial intelligence is essentially just a mapping of human intelligence. How humans manage the evil of human intelligence, artificial intelligence will not exceed this model. I do n’t mean that the mischief of artificial intelligence will not happen, or even think it will happen, but since humans are said to be able to control the mischief of human intelligence, there is basically no need to worry about not being an opponent of artificial intelligence.
But to be honest, I personally think that all human development is moving unstoppably on the path of self-destruction. The content of this article may be "expected" to become a factor to accelerate this process.
The reason why I wrote this blog is because some people in the working department suggested that everyone share their most familiar parts in the field of machine learning. I reflected on it. From the sophomore year of contacting machine learning to almost two and a half years now, what can I say? It seems that I ca n’t do anything, because it seems to me that I have n’t mastered a relatively high-value skill in a certain field of machine learning. If you have to talk about something, it may be feature engineering. But I also feel that for the past two and a half years, I have been able to do my best for machine learning. I cannot do anything without it. So I spent a whole day sleeping and forgetting food all day on Saturday, and it ’s no exaggeration to say that at the same time, I exhausted my physical and mental energy. I extracted the most valuable content in my mind for machine learning in these two and a half years . It is also an account of my introductory stage in the process of machine learning, and it is also my record and guidance for my future direction. Due to energy reasons, the last chapter is relatively hasty. I didn't express my views very well, but so far I can only do so.
finish！
2017.08.19
