/
/
Want to know what else AI can do? Must have a sober "Bayesian consciousness"

Want to know what else AI can do? Must have a sober "Bayesian consciousness"

  • Categories:Industry News
  • Author:
  • Origin:
  • Time of issue:2017-10-23 15:56
  • Views:0

Want to know what else AI can do? Must have a sober "Bayesian consciousness"

Information

Many readers have talked to us about being surprised that artificial intelligence can do so many unexpected things. From playing Go to autonomous driving, from recognizing ancient characters to mapping the universe, it seems a bit omnipotent. There are also readers who want us to do a special feature, a list of "you never thought AI could do this".
 
Very, very sorry, I can say very responsibly that this list is too easy, but the problem is that there is no way to list it all. Even without considering the possibility of technology, it is almost impossible to calculate what AI can do in new papers and research reports every day.
 
But that's not to say AI is omnipotent. In fact, AI is not as efficient as traditional computers in the vast majority of basic work, and it is impossible to replace it at all. AI always pops up in unexpected fields because it adopts a completely different underlying strategy, focusing on solving problems that classical computing and humans cannot solve.
 
The so-called fish is worse than fishing. Instead of sighing together "AI can do this!", let's think about "how to know what else AI can do". To gain this capability, you need to know what the basis for today’s mainstream AI solutions is. This has to mention a name that is crucial to the development of AI: Bayes.
 
Want to find God, but found artificial intelligence
 
If there is no Bayesian old man, it is difficult to imagine what artificial intelligence will be like today. Not only artificial intelligence, but statistics, applied mathematics, surveying and mapping, medicine, and even criminology will be greatly affected.
 
But a man so important to academia is not a scientist at all. Thomas Bayes, who lived 300 years ago, was a priest in England. Of course, he was also an amateur mathematician.
 
It may be because of the consideration of combining hobby and work. Bayesian proposed the method of applying induction to probability and statistics, hoping to prove to the world that God exists. But unfortunately, three hundred years later, the basis for the existence of God has still not been found, but Bayesian decision-making has blossomed in countless scientific fields.

The so-called Bayesian induction, in fact, the basic principle is very simple. For example, if a person does a good deed, is he a good person? Obviously not. But if a person does good deeds every day, is he a good person? In fact, it is not necessarily. It is also possible that he is heinous and conscientious behind his back, but if there is no evidence that he does bad things, it is obviously very likely that he will be a good person if he does good deeds every day. This is the core logic of Bayesian induction: it is not necessary to obtain all evidence before making a judgment, but to make judgments based on known conditions, and then continuously verify, adjust, and modify the judgment through data to make it endlessly rationalized.
 
This logic sounds quite simple, and it seems a little unreliable, and Bayesian's ideas were not even published before his death. Even in the more than two hundred years after his death, Bayesian induction has received little attention. Because classical statistics starting from the rigor of data can obviously better touch the essence of things, rather than using "guess" to start the calculation like Bayesian induction.

It was not until the 1970s and 1980s that Bayesian theory, which had been silent for more than a hundred years, began to be re-emphasized in statistics.
 
Although classical statistics is reliable, it needs to rely on a complete data model, which is often too inefficient to meet actual needs. For example, when searching for ships in distress on the sea surface, classical statistics requires meteorological data, monitoring data, and passing ship data in each sea area, and then comprehensively calculates these factors to accurately locate. But in fact, it is impossible to complete these data immediately, and even if it is possible, the search and rescue work where every second counts cannot wait. To solve this problem with Bayesian theory, experienced experts will first make subjective judgments about the area where the ship was wrecked, and then use the continuously obtained data to revise the judgment of the experts little by little, and try to solve the problem in the shortest time. This is the famous The wreck of the USS Scorpion in 1968 was also a sign of the application of Bayesian theory.
 
In its work, Bayesian theory emphasizes that starting from human prior knowledge, making fuzzy judgments on the target, and then continuously learning to make judgments and proofreading, this has become the starting point for the birth of a large number of artificial intelligence technologies.
 
Is it a sad story that 300 years ago, the theory that was supposed to prove the existence of God has become the basis of artificial intelligence 300 years later?
 
Or has the Bayesian priest actually found the right answer?

Bayesian Consciousness: The Foundation of All Learning AI

 

Unlike classical calculations consisting of 0s and 1s, Bayesian calculations do not need to be based on complete data to obtain answers. This incomplete data reasoning ability is very similar to the cognitive and judgment process in the human mind. As a result, a large number of Bayesian theory and AI have been combined to apply to technical models of incomplete information derivation.
 
For example, Bayesian networks, Bayesian classifiers, and Bayesian logic are all very basic AI tools today. Bayesian network can be regarded as the basic condition for the self-validation of machine learning theory. The Bayesian method is also widely used in NLP, machine vision, knowledge graph and other fields, and has become the support of optimization result-based algorithms and technologies.

In our life, Bayesian can be said to be everywhere. For example, do you feel that the video recording or live broadcast of your mobile phone is getting clearer and clearer? A big reason for this is that the video optimization algorithm under the blessing of Bayesian logic is used in the camera algorithm, and a clearer and more natural shooting effect is obtained through confrontation generation, which is the so-called video beauty.
 
The importance of Bayesian to artificial intelligence can also be demonstrated from another perspective. In the 1970s, there was a relatively brief AI renaissance. At that time, knowledge representation and expert systems became the protagonists of artificial intelligence, and people hoped to use supercomputers to summarize all human knowledge and network. In the end all questions can be answered there.
 
This model won a lot of money and attention at the time, but it went bankrupt in just a few years. Because human knowledge and data are too complex, collecting all knowledge is only a theoretical possibility, and practical operation is far away.
 
Today, the second revival of AI, which is mainly characterized by machine learning, benefits to a large extent from the transformation of thinking brought about by Bayesian consciousness: human beings do not need to collect a lot of knowledge, but only need to start from a part of the existing knowledge, and let the machine keep going. It is enough to learn and verify your own abilities and keep improving. After all, what humans need is not an omniscient and omniscient existence, as long as the intelligent body can be stronger than humans.
 
It can be said that Bayesian consciousness has become the basis of all learning AI today, relying on its high sensitivity to changes in reality. After understanding the logic of starting from incompleteness and gradually moving toward completeness in Bayesian consciousness, we also understand what else AI can do in the future.
 
The human brain and the future of technology have the same goal
 
Some scholars believe that Bayesian consciousness may be the closest applied mathematical logic to the thinking mode of the human brain. Just like letting a child know a dog, you don't have to teach him the type, family, and habits of the dog, but also use the dog's ears, nose, and limbs into data to make him understand. Children will immediately know that this is a dog, and then strengthen their understanding of dogs in their continuous learning, know that there are different kinds of dogs, know the difference between wolves and dogs, and so on.
 
Therefore, when we used to think about many problems, we would actively let the brain imagine like a computer. In the era of mobile Internet, we are accustomed to thinking about everything with the mobile phone as the center. But in the age of artificial intelligence, Bayesian consciousness tells us that perhaps it is time for humans to think like humans themselves.
 
After certain technological singularities, machines have been able to recognize, reason and judge complex problems through local knowledge in local areas like humans. The best example is probably the emergence of AlphaGo ZERO. In fact, the logic of AlphaGo is a kind of Bayesian thinking. You must know that for Go, classical calculations cannot exhaust all changes, and violent exhaustion can only bring about crashes.
 
What AlphaGo adopts is to let the agent learn the rules of Go, and then learn a lot of human chess scores. This is the goal of continuous absorption of data calibration in Bayesian induction. In actual combat, the internal algorithm of the agent will also self-verify the rationality of each step of the prediction, and finally obtain the optimal solution.

After a certain accumulation, the Bayesian system of AlphaGo can no longer rely on the data provided by humans, but through self-learning of high-quality data, thus generating the ability to kill the previous generation in a short time. This can be seen as the Bayesian system has abandoned its reliance on primary data and entered a process of further self-calibration.
 
There is reason to believe that there will be more of these phenomena in the future. Because the learning ability of the human brain has several limitations, but Bayesian agents do not.
 
Through fuzzy knowledge, continuous learning will lead to the generalized unknown. It is likely to be the common direction of the future of the human brain and technology. At least for now, this kind of technical logic has been used more and more in forward-looking science. Such as quantum Bayesian, Bayesian genetic algorithm and so on.
 
On the other hand, wondering where else AI can do amazing things. Might as well think like Bayes: First of all, are there problems such as efficiency, reliability, cost ratio, blindness, etc. in this field, and if so, is it necessary to introduce AI. Secondly, look at whether there is prior knowledge in this field as the basis for the agent. Again, look to see if the field continues to generate the data and knowledge to feed machine learning.
 
If these conditions are established, then AI is not far away.
 
After a long education from PC to mobile phone, we may have become accustomed to the operation of the digital world. However, artificial intelligence may break the classic combination of human and computer, and use human perception and learning ability and computer computing power to open another way. Maybe it's time to change the Internet thinking and let the brain have a little more tacit understanding with artificial intelligence.

硕和

Industry application        Product results        After-sales service        News        About SEOKHO        Contact us

Copyright ©  Zhejiang SEOKHO Robot Technology Co., Ltd.   浙ICP备20002940号-1    WWW.300.CN

MENU

这是描述信息
搜索