"It only takes a few lines of code to build a regression model." Programmers recognize the performance of Google AutoML and think that the model designed by AutoML is comparable to that designed by machine learning experts. A few days ago, Google engineers introduced the Google AutoML project in China and Silicon Valley respectively.
Confused doubts followed — — AI evolved again? ! Already self-developed? Can control their own evolution? Is it to get rid of humans?
How many steps does evolution take?
AI has indeed evolved, and it can do more and more things and achieve remarkable results. Behind it is the "triple jump" of AI implementation path — —
Zhao Zhigang, a researcher at the Big Data R&D Department of Jinan Center of National Supercomputing, said: "At first, we used mathematical formulas and ‘ if… … then’ Wait for the statement to tell the computer what to do in the first step and what to do in the second step, teach it by hand, and then give the machine N groups of inputs and outputs. The rules or laws in the middle are learned by itself. "
"Before, many smart minds spent their whole lives studying how to extract effective features." Mo Yu, CTO of Smart Point, which focuses on intelligent shopping guide dialogue robots, explained, "The invention of neural network algorithm and the emergence of deep learning technology have made AI evolve to 2.0, and the work of extracting features is carried out by AI itself, and our work has also changed accordingly."
It is easy to explain the transition from "1.0" to "2.0" with the model of mathematical function: if the tasks of image recognition, semantic understanding and chess playing are all regarded as different Y=f(X), that is, the input picture, voice or chess move of "cat" is "x", and the output cat, answer and chess move is "y". Before deep learning, people find the formula corresponding to function F through their own analysis and tell it to AI. After deep learning, people input a lot of correspondence between x and y, and AI finds the formula corresponding to function f by itself.
"The specific content of the function F found by AI may be better than that found by humans, but humans don’t know it, just like a black box." Mo Yu said, "But the form of F is designed by AI researchers through research. If deep neural network is used, the modules in the network and the organization between modules are also designed in advance."
With the maturity and popularization of deep learning technology, there has been a specific pursuable experience in model construction. "The release of various common neural networks has made the threshold for employment lower and lower. Some ordinary model construction and optimization, just graduated students can get started in online learning tutorials. " Zhao Zhigang said.
When modeling becomes a learnable skill, AutoML appears. What it can do is the model design work of AI researchers. "It will help different companies build artificial intelligence systems, even if they don’t have extensive expertise." Google engineers recommend this. AI successfully evolved to 3.0.
In fact, AutoML still replaces the work that human beings can extract experience. "If people described a set of ‘ Road network ’ With the technical assistance of deep learning, the machine can find the optimal path as quickly as possible; Then AI can now design its own road network. " Zhao Zhigang was concise.
It can be seen that both deep learning and AutoML only replace the work that some groups of human beings have studied thoroughly. "What machines can do, try not to work by hand" is the life creed of many programmers, which gave birth to AutoML. Based on the same creed, Microsoft developed DeepCoder. "It can be used to generate programs that satisfy given input and output." Mo Yu said, but its performance is not satisfactory at present, and only some simple programs can be written.
Who is "God"
There is no doubt that the answer is human.
Since AI has evolved to a higher level of model design, what changes have taken place in the "hand of God"?
"An alchemist", Mo Yu vividly talked about his work in two words. "Smart is a professional intelligent customer service. The work of R&D personnel mainly focuses on problem modeling (how to turn actual problems into problems solved by artificial intelligence technology) and algorithm optimization (how to improve the effect of artificial intelligence algorithms)."
"refining" means constant debugging and improvement. "For a specific person, the more you throw your temper, the better. The more accurate the answer, the better." Mo Yu said, "Our X is the customer’s question, Y is the reply of the robot customer service, and the middle function F needs training."
This is not an easy task. If we divide the experience of human society into three categories: definite rules with formulas, knowledge that can be expressed in words, and feelings that can only be understood but cannot be expressed. The last category is the most difficult to ponder.
"Therefore, we try to build a perfect closed-loop feedback, understand the preferences of specific users, and finally achieve what they like through the expression of emotions and interests." Mo Yu said, "At present, it is in the stage of man-machine collaboration, but the acquisition of more and more samples will help our intelligent customer service to give accurate and pleasing answers."
It can be seen that not all fields are suitable for AI’s self-development. For example, in the aspect of problem modeling, how to abstract practical problems into machine learning problems can not be completed by AI. In the stage of AI2.0, R&D personnel also need to design the form of function F manually.
So, what will be the future of AutoML popularization?
"Humans are liberated from lower-level jobs." Zhao Zhigang said, "If the model design can be done by AI, then AI researchers will explore more the design of the basic modules that make up the model."
"Developing AI models with AutoML is similar to children playing with ‘ Lego ’ Toys. " Zhao Zhigang put it in simple terms, and the designers of Lego disassembled the whole world into detailed modules, so that everything could be used and then combined into complex models. The higher level of human work is to find basic units, that is, modules, for AI in different fields. For example, in the field of image recognition, humans have designed various modules such as convolution and pooling. "AutoML can build a model based on this, and constantly adjust the module combination to obtain a more reasonable output. The finer the module, the more versatile it can be, and the more self-developed AI can be used. " Zhao Zhigang said.
Extended reading
Where are the talents needed by the industry?
Alleviating the shortage of talents is the main selling point of AutoML. "AI systems are blooming everywhere, but AI talents are far behind." Google explains why AutoML is indispensable. What is the current status of AI talents?
In 2017, the Global AI Talent Report and the BAT Artificial Intelligence Talent Development Report were released one after another. "The shortage of AI talents is real." Xu Wenjuan, vice president of Shengshi Investment Group, said, "The shortage of talents in start-up and development enterprises is particularly serious. From the current global perspective, the United States has the largest number of AI talents, and China’s AI talents can’t be compared with it in terms of number of people and experience. "
Zhao Zhigang has the same feeling: "China’s AI field is short of veterans, experts, generalists and top masters." The optimization and debugging of the model requires experience, the exquisite design of the model requires superb skills, and the application of AI in various industries requires compound talents. In addition, there are only a handful of top talents leading the development direction of AI, and most of them are abroad.
Xu Wenjuan introduced that China has the most AI talents in BAT (Baidu, Ali, Tencent). Generally, there are several kinds of background experiences of such talents, such as returnees, BAT work experience, or from universities or research institutes.
"AI self-development should not be able to replace people’s work in the short term, and there is still a long way to go." Xu Wenjuan said. Zhao Zhigang analyzed from an academic point of view: "Only when human beings design AI models in different application fields and further decompose them into a series of general modules, such as the periodic table of elements in chemistry, DNA and RNA in biology, can this self-development have more applications."
