AI does evolve, it can do more and more things, the achievements are remarkable. Behind it is the "triple jump" in the path of AI realization. But "AI self-development should not be able to replace human work in the short term, 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 for different application areas and further decompose a series of universal modules, like the periodic table in chemistry, DNA and RNA in biology, this self-development can have more applications."
"It only takes a few lines of code to build a regression model." Programmers recognize the performance of Google AutoML's work, believing that the models designed by AutoML are on par with those designed by machine learning experts. A few days ago, Google engineers highlighted the Google AutoML program in China and Silicon Valley, respectively. Unexplained doubts followed closely - AI has evolved again? Can it develop itself already? Can manipulate its own evolution? Is it going to get rid of humans?
Zhao Zhigang, a researcher at the Big Data Research and Development Department of the Jinan Center for National Supercomputing, said: "At first we used mathematical formulas and statements such as 'if?then' to tell the computer what to do in the first step and what to do in the second step, hand-in-hand, and then we gave the machine n sets of inputs and outputs, with the rules or laws learned by itself. of the rules or laws it learned on its own."
"Before that, a lot of bright minds spent their lives studying: how to extract valid features." Mo Yu, CTO of Smart Point, a company that specializes in intelligent guided conversational robots, explained, "The invention of neural network algorithms, and the emergence of deep learning technology have caused AI to evolve to 2.0, where the work of extracting the features is carried out by the AI itself, and our work has changed."
It's easy to explain the shift from "1.0" to "2.0" with a model of a mathematical function: if you think of the achievement of tasks such as recognizing an image, semantic comprehension, and playing chess as different Y=f(X), i.e., input of "cat" picture, sound or chess move is "X", and the output "cat", answer or chess move is "Y Y". Before deep learning, the person finds the formula corresponding to function f through his own analysis and tells it to AI. while after deep learning, the person inputs a large number of correspondences between X and Y, and the AI discovers the formula corresponding to function f by itself.
"The specifics of the function f found by the AI may be better than those found by the human, but the human doesn't know it, just like a black box." Mo Yu said, "But the form of f is designed by the AI researcher through research, and if a deep neural network is used, the modules in the network and the way the modules are organized with each other are also designed ahead of time."
As deep learning techniques mature and generalize, specific traceable experiences in model building emerge. "The release of various ****ty neural networks has made the barriers to the profession lower and lower. Some common model construction and optimization, fresh graduates can get started by learning tutorials online." Zhao Zhigang said.
When building models became an acquirable skill, AutoML came along. It can do exactly what AI researchers do in model design. "Will help different companies build AI systems, even if they don't have extensive expertise." Google engineers pitched it this way.AI successfully evolved to 3.0.
In fact, AutoML replaces what is still a job where humans can distill experience. "If before the human depicts a set of 'road network' to find the function f, with the technical assistance of deep learning, the machine can find the fastest optimization path; then AI can now design its own road network." Zhao was succinct.
It can be seen that, whether it is deep learning, or AutoML, are only replacing a part of the human group has drilled through the work. "What a machine can do, try not to manual labor," is the life creed of many programmers, this creed gave birth to AutoML. in the same creed, Microsoft developed DeepCoder. "It can be used to generate programs that satisfy a given input and output. " Mo Yu said, but its performance is currently unsatisfactory, and it can only write some simple programs.
The answer to the question of who is "God" is, without a doubt, human beings.
Since AI has evolved to a higher level of modeling, how has the "hand of God" changed?
"Alchemy", Mo Yu used two words to describe her work, "Smart One is specialized in intelligent customer service, the work of R&D staff mainly focuses on problem modeling (how to transform actual problems into problems solved by AI technology) and algorithm optimization (how to improve the effectiveness of AI algorithms). the effect of artificial intelligence algorithms)."
"Refinement" means constant debugging and refinement. "The more you throw your temper at a specific person, the better, and the more precise the answer, the better." Mo Yu said, "Our X is the customer's question, Y is the robot customer service response, and the function f in between needs to be trained."
It's not an easy task. If you divide the experience of human society into 3 categories: defined rules with formulas, verifiable knowledge, and feelings that can only be understood. The last category is the hardest to figure out.
"Therefore, we want to find ways to build a perfect closed-loop feedback to understand the preferences of specific users, through the expression of emotion and fun, and ultimately to do what they want." Mo Yu said, "Currently in the human-machine collaborative work stage, but more and more access to samples will help our intelligent customer service to give accurate and pleasing answers."
It can be seen that not all areas are suitable to be given to the AI self-development to do, such as problem modeling, how to abstract the actual problem converted into a machine learning problem, the AI can not yet be done autonomously. In the AI2.0 stage, developers also need to manually design the form of the function f.