As we all know, human intelligence mainly includes induction and logical deduction, which respectively correspond to the two main schools of artificial intelligence, connectionism and symbolism. The typical methods of the former include neural network, statistical learning and deep learning. The latter includes formal language, logical reasoning, expert system and so on. These two schools have ups and downs in the history of artificial intelligence development. With Geoffrey Hinton, a professor at the University of Toronto in Canada and a master of machine learning, publishing a paper on deep learning in Science in 2006, deep learning continues to heat up in academic and business circles, and deep learning-related applications such as speech recognition, image recognition, natural language processing and search advertising are hot. As a representative of Connectionism, deep learning has become the hottest field of artificial intelligence.
In this regard, Professor Lin Zuo Quan of Peking University thinks: "The rapid development of deep learning since 2006 has not made much contribution to the basic principles of artificial intelligence. The essence of deep learning is to calculate grades through a large amount of data. The first level becomes another representation after learning, and then feature extraction becomes the second level. The more levels, the better the effect. In addition, deep learning or the calculation of each layer of so-called learning is actually the application of mathematical problems, that is, solving an information function, but in principle these nonlinear functions are difficult to calculate. Therefore, it brings two major problems: first, the significance of increasing the depth level of deep learning network; The other is the theoretical problem of each layer calculation. Computational mathematics cannot be solved, and deep learning cannot be solved. "
Qi Guilin, a professor at the School of Computer and Engineering of Southeast University, believes that it is a difficult problem in the field of artificial intelligence to make machine learning have cognitive and reasoning abilities. "Not all companies have big data capabilities like Google. Google runs deep learning very well. But another company may not have such a good effect. How to improve the learning ability of a machine on a small amount of data actually requires it to have cognitive and reasoning abilities. At present, several authoritative experts in the field of deep learning have expressed the need to introduce human rule reasoning into neural networks on different occasions to make neural networks more explanatory. "
"People's understanding of deep learning has not reached the level of artificial intelligence we want, and it is very difficult to reach human wisdom through deep learning models." Li Wenzhe, chief data scientist of inclusive finance, added.
For symbolism, it is believed that artificial intelligence originates from mathematical logic, the core idea is to apply the law of logical reasoning, and the embodiment in artificial intelligence is the proof of machine theorem. Symbolism holds that knowledge is a form of information and the foundation of intelligence. Knowledge representation, knowledge reasoning and knowledge application are the core of artificial intelligence. Knowledge is represented by symbols, cognition is the reasoning process of symbols, and the reasoning process can be described by formal language. It also advocates the establishment of a unified system of artificial intelligence through logical methods. Professor Lin Zuo Quan said: "The core goal of symbolism is still to explore the basic principles of artificial intelligence, which belongs to basic research. One of the original purposes of artificial intelligence is to simulate human intelligent behavior through computers and explore the basic principles of intelligence. This goal is far from being achieved. "
At present, all kinds of artificial intelligence have attracted wide attention. In addition to hot events like Watson's participation in dangerous games and Google's AlphaGo's battle against the world champion of Go, more applications based on artificial intelligence have been widely used, such as the application of automatic fraud detection system in the banking field, the sales pricing of retailers, smart home robots, face recognition systems, automatic speech recognition and so on. So for the company, is the direction of artificial intelligence choosing connectionism or symbolism?
In this regard, Li Wenzhe, chief data scientist of inclusive finance, took the financial industry as an example to show that both directions are very useful. He said: "The characteristic of the financial field is that the company will not have a lot of data when it is first established, so it will not try Connectionism, because deep learning definitely needs a lot of data to get better results. When the amount of data is small, the experience of experts is the most important, which belongs to symbolism. For example, the analysis of fraud and the evaluation of credit risk are based on the previous experience of experts. When the company has accumulated a large number of data samples after years of development, you can try the connectionist algorithm. " In Li Wenzhe's view, the biggest consideration in adopting symbolism or connectionism is the amount of data. "Considering connectionism or symbolism in the company's business is the quantity and complexity of the company's data. A lot of symbolism depends on experience, and a lot of logic is artificially defined. When the amount of data is very large and the attributes are very complex, it is difficult to define it in this way. At this time, deep learning is needed. "
However, in Li Wenzhe's view, deep learning is still in the primary development stage, and users are still doing many attempts and experiments. He said: "Deep learning has gradually become popular since 2006, but it is still in its infancy. Many people who do deep learning are doing and trying different methods. At a certain stage, someone will study the theoretical level. "
Usually artificial intelligence is often divided into three levels, namely weak artificial intelligence, strong artificial intelligence and super artificial intelligence. AlphaGo like Google is a typical representative of weak artificial intelligence, which has powerful artificial intelligence programs in a single field; In addition, robot writing, Siri, Microsoft Xiao Bing, etc. All belong to this level; Usually, weak artificial intelligence has no self-awareness, and calculates according to a fixed structure to get the answer. With the popularity of big data and computing power, weak artificial intelligence is basically realized. Then, when will strong artificial intelligence that can actively find problems, build problem models and solve problems come? How far is the era of super artificial intelligence even surpassing human beings from us?
Many experts believe that the era of strong artificial intelligence or super artificial intelligence will come in the near future. Ray. Kurzweil, an American futurist and engineering director of Google, even predicted in his book Singularity approximation: "Around 2045, artificial intelligence will come to a singularity. Crossing this critical point, artificial intelligence will surpass human wisdom and human history will be completely changed. " However, Professor Lin Zuo Quan does not agree with this view. He said: "The topic of artificial intelligence threatening human beings has been going on for decades. I don't agree that 2045 will be the time singularity when artificial intelligence surpasses human beings. In recent years, the strong artificial intelligence that some people call general artificial intelligence has basically disappeared, and it is difficult to see the possibility of realization in a short time. "
"Artificial intelligence has been applied in many mature methods and has become an important part of infrastructure. Historically, artificial intelligence has also been hotly debated several times. This artificial intelligence craze has promoted the development of artificial intelligence. Although there may be a bubble, this process is actually helpful for the development of artificial intelligence. " Professor Lin Zuo Quan finally said.