S Tian Yuandong
Alphago’s upgraded version – alphayuan has the ability to learn by themselves, only three days to learn defeated alphago; The kingdom of Saudi Arabia has just granted citizenship to Sophia, a “female” robot, making her the first ever robot to be granted citizenship… Will the development of artificial intelligence threaten human society? Is the day when robots control humans ever closer?
Most of the time, panic stems from mass ignorance of technology. In the eyes of ai scientists, current AI is as different from the human brain as a few simple symbols are from a novel.
The success of the Alpha yuan is just a coincidence
Alphago is not as influential in the industry as it is in the public. Because in the field of ARTIFICIAL intelligence, every method has limitations and no one method is a panacea. When every part of the emerging ai is clearly understood, its mystique will dissolve, along with the current doubts and fears.
The genius of AlphaGo is that it combines both engineering and scientific research to push AI to the extreme in one direction with massive computing resources and engineering optimization. It also draws on people on the go, and in 10 years of progress in computer vision, such as go and intensive study direction in monte carlo, since the game tree search (commonly known as “and” around), random walk disk valuations, with characteristics of artificial shallow network to quickly move, weigh the breadth and depth search and weigh from exploration and a priori knowledge, Convolutional neural network (CNN), residual network (ResNet), rotary flip sample enhancement and so on. None of this was first thought up by the DeepMind team, but piecemeal from past experience. It’s just that the little advances of the past have not made it into the public eye, and the Afar dog has achieved this final step.
There are several prerequisites for this success.
One is that the model and the problem match. For Alphago, the match between the model and the problem is very important. Convolutional neural networks are a good match for Go, so with just 4.9 million samples it can learn more than humans can.
For humans, this match is found through a lot of intuition and experimentation, while machines can quickly try and error, and quickly implement human ideas. If the model does not match the problem, it will not be of much use, because the problem of “dimension disaster” in machine learning, which requires violence to fill the high-dimensional space, is really beyond any amount of computing resources at present.
Convolution neural network can go so congenial with, or because the rules of the game itself is pure, there is some inner beauty, and CNN is very suitable for this kind of beauty, once adapted to the computer with high speed computing power can be launched comprehensively, even starting from scratch, not by humans, move and around alone, can also be successful. With the right path, the stability and accuracy of the computer, combined with computation speeds billions of times faster than humans, are the key to enabling the Alpha element to travel in three days far more than the human race as a whole can travel in 3,000 years.
The second is to live in an ideal world of complete information transparency, and after making decisions, have an absolutely correct understanding of the rules, changes and development of this world.
Go players know exactly what the situation is going to be with each move, and they know exactly how to win or lose. Alphayuan doesn’t use a human board, but it has learned the rules of the go world from humans and knows exactly what moves have been made in the past seven moves. This makes it very easy for it to extrapolate from the situation, which is much better than humans in the sky. Going from scratch to surpass the human race is impressive, but only in the highly regulated world of Go.
The third requirement is to have a lot of computing resources. The two Alpha papers, published in Nature, invariably highlight the number of resources used in wartime, while inadvertently omits the number of resources used in training. The latter can be calculated from articles (and experiments) in the tens of thousands or more. And that’s just the amount of resources needed to reproduce the results. As a pioneer, there is no successful route in a hundred, and the combined cost of all kinds of exploration, trial and error, the final use of resources is even astronomical.
Alphago’s amazing ability to deal with go is due to satisfying these three conditions, while AI’s ability to deal with other problems is not so easy. Intelligent conversation systems are a good example. Human language has both basic grammatical rules and plenty of exceptions that break them; Both at this time and under the circumstances of the convention, and with The Times of the total changes in Lin Lin. An intelligent system can dialogue with others, both from the vague interaction deeply understand the basic rules for now, and should constantly update their knowledge system, to abstract out talking about plan, speculated that others hide status from time to time and the development trend of the world – these are now the difficulty of artificial intelligence.
So AlphaGo is not as influential in the industry as it is in the public. Because convolutional neural networks, which produced surprising results on Alphago, had previously been a great success in computer vision, fitting advertising data with lots of discrete features was notoriously ineffective. Similarly, left/right interaction works quite well for other games, but it may not be as good as manual data for optimizing machine translation.
In contrast, the foreign view of Alphago is much more rational, and there are more in-depth discussions on technology on major forums. When we have the strength to take AlphaGo apart and have a clear understanding of its parts, the mystery will disappear, along with the current doubts and concerns.
Artificial intelligence is not yet comparable to the nervous system in the brain
Both brain science and artificial intelligence are essentially about finding an algorithm to efficiently model the world. The brain has evolved to operate in its own way, while ARTIFICIAL intelligence can develop independently based on mathematical principles and big data, just like birds and planes, to find their own rules.
The real neuron is a continuous signal processor described by differential equations, with a finely timed structure, electrophysiology, ion channels and neurotransmitter receptors, energy supply and immune system. Due to measurement limitations and the complexity of biological systems, we do not yet fully understand individual neurons in the brain, let alone systems in which large numbers of neurons are combined.
Existing ai algorithms are largely independent of brain science. Statistical learning methods that are effective in various problems, such as linear and nonlinear fitting methods, nearest neighbors, decision trees, random forests, and support vector machines, have little correspondence in brain structures. Even the so-called neurons in deep learning are nothing more than a combination of linear superpositions and non-linear operations, a huge simplification of neurons in the real brain, so much so that brain scientists don’t even recognize them as neurons.
However, the two disciplines can be instructive to each other. CNN, for example, works in a similar way to neurons in the brain that have limited visual field. Recently, Jeffrey Hinton, a professor of computer science at the University of Toronto and the father of neural networks, tried to improve the performance of existing neural networks using the brain’s cylinder idea. Other researchers have used statistical learning and neural networks to model the behavior of certain neurons in the brain.
My personal opinion is that whether it’s brain science, whether it’s artificial intelligence, it’s essentially a question of finding an algorithm that effectively models the world. The brain has evolved over a long period of time, but it is also burdened by a large amount of incremental evolution. It does not need to be worshipped as the only template for intelligence. Artificial intelligence can develop independently on the basis of mathematical principles and big data, just like the difference between birds and airplanes, and find its own rules.
Human beings have been specialised in textbooks that have overestimated their abilities. In fact, in the long history, human beings are not so special. As an intelligent body, human beings follow their own development laws and realize their own transformation on the basis of natural evolution. And the strengths, weaknesses, and weaknesses that evolution has bestowed upon us will condense into a page in history that will become a footnote to the planet’s future civilizations.
We just got a glimmer of artificial intelligence
People don’t need to fear machines, because machines need people. The real world is so complex that it implies a future that cannot be accurately predicted. In the face of the real world, which is a behemoth, machines are as small as humans and will inevitably go hand in hand.
Standing in the industry’s perspective, artificial intelligence still has a lot of problems to solve. For example, we could vaguely say that CNN and Go are a good match, or that CNN and computer vision are a good match, so together they get unexpected results. But what exactly is a match? What is the quantitative and precise definition of the word “match”? For other questions, what is the best model and what data is needed? What exactly does a neural network model, and how does it differ from our cognition? None of these fundamental questions has been answered, and at present, it seems, both in terms of tools and ideas, we are far from touching the edge of the answer.
At the beginning of the 20th century, the edifice of physics was largely built, with the exception of two clouds overhead; In contrast, artificial intelligence can be said to be still in the lead clouds, snow filled the long night, people can only vaguely see a little moonlight in the dark clouds overhead. We’ve just been in the dark too long to get used to seeing light — the sun’s not up yet.
However, the first thing that is basically certain is that the future world is bound to be a world of cooperation and gradual integration of human and machine. In just ten years, we have been inseparable from computers and mobile phones. In the future, if there are artificial intelligence products with stronger memory, faster reasoning and more accurate prediction, which can help humans see farther and hear more clearly, people will be eager to use them and make better products on this basis. Imagine a chip in your brain that could double your reaction time. Who would say no? No one can resist such a tide.
So instead of being afraid, fit in. Artificial intelligence you don’t do, I don’t do, someone always do, and do it, there will be a great competitive advantage. The advertising system based on machine learning is already like this — written program, lying can make money, and people have to do is how to improve it. This and the traditional industry needs to be maintained to have a stable profit margin compared, is already a high realm. In the future, when AI gets better, it can automatically improve the algorithm, which is another level. And it’s not hundreds of years away. It’s happening right now. The train of history is speeding forward. Before, it might have trotted along, but later it could only hitch a ride.
Finally, I don’t think people need to fear machines, because machines need people. The real world implies a future that cannot be accurately predicted, complex connections that are intertwined, an unimaginably large space for action, and billions upon billions of coherent goals and meanings. Go’s balance of states far exceeds the total number of atoms in the universe, yet it is nothing more than a coffee table in the backyard, a few stones on a warm spring day, and a sip of tea drifting in the wind.
In the behemoth of the real world, machines and people will inevitably go hand in hand.
(The writer is a researcher at Facebook’S AI Research Institute)