The Defects of Artificial Intelligence

Press: For the recent development of artificial intelligence has been some thinking, take time to sort it out, is a throw a brick to attract jade bar. Welcome to the discussion.

A New Paradigm of Artificial Intelligence

The research paradigm of AI is quite different from that of mathematical science. This paradigm is data-oriented, highly engineering, methodologically developed and epistemologically inadequate. First, let's talk about data orientation. Teacher Li Feifei's ImageNet project is a very typical example. There is a huge amount of work in collecting, collating and labeling a large number of pictures. On the basis of these high-quality data, a worldwide image recognition competition has been conducted. By building a larger depth learning network, the team has rapidly outperformed traditional algorithms, which has become a landmark achievement of depth learning. Its basic workflow is to accumulate large-scale data, get a deeper network model, and achieve better performance. Although we now have more different kinds of deep networks to deal with different tasks, this basic framework has not changed substantially. In this framework, data collection almost occupies the most important position. If the quality of data is not good, it is easy to cause the so-called garbage in - garbage out. Data set size is an important factor affecting the efficiency of deep learning. Generally speaking, when the data set is small, deep learning is not necessarily superior to traditional machine learning methods. With the increase of data set size, the performance of traditional machine learning algorithm is easily saturated, while the performance of deep learning can be more excellent with the increase of network size.

In the traditional field of mathematics and physics, we do not rely so much on data for model building. Perhaps because of the limitation of human brain's thinking ability, people are more accustomed to construct a set of model system based on simple hypothesis and using logic deduction or formula deduction to interpret actual data. Mathematical and physical sciences use a lot of data: for example, Kepler has found a specific structure from the Tycho observation data, but this structure is not described in the way of neural networks, but in a more mathematical concise form. The law becomes a natural result only when it is incorporated into the framework of Newtonian mechanics. Now the means of observation are more advanced. It is said that the LHC can produce 100 trillion bytes of collision data per second. But these data are not intended to derive the model, but mainly to verify the standard model constructed artificially with only dozens of parameters, and the amount of information it contains does not exceed the model. In many cases, theory can be put forward or developed with little or no data. Galileo's kinematics should be constructed on the basis of very few observations. When Taylor put forward the atomic theory, I believe there was no observational data. The charm of theory is that it can be produced from people's experience, thinking or intuition. It has inherent simplicity but can explain a large number of practical observations.

The intrinsic pursuit of data leads to a lot of hard work in artificial intelligence research. Combining with industry, the entry of capital further enlarges this effect. With the emergence of AI start-ups in China, data annotation has become a new profession. There have been many related reports, such as the article "Artificials Behind Artificial Intelligence" written by Jiazi Lightyear. These people are paid about 4,000 yuan a month, mechanical click on the mouse in front of the computer to do picture labeling, the data obtained is ultimately used for driverless projects. Many go down to villages to collect face recognition data for laundry powder or soybean oil. Even today, with the gradual capitalization of scientific research, it is hard to imagine that a discipline can directly promote new labor relations like this. This is partly due to the second issue that we are going to talk about, that is, AI research is very engineering.

If you have taken Professor Wu Enda's in-depth learning course, you will find that in-depth learning is very advanced in engineering and has a very clear working mode. For a deep network model, those parameters have the greatest impact. People have rich experience on how to judge the direction of the adjustment parameters through performance. Ultimately, data quality, scale, and feature selection are the key factors that determine performance. This leads to work in the field of artificial intelligence, which is very easy to scale-up, i.e. scale-up. It's also Silicon Valley's favorite development model - rapid technological leaps and innovations through capital catalysis. Similar models have created the familiar facebook, amazon, and now the mobile Internet world of the internet.

In the field of mathematics and physics, we cannot simply innovate by accumulating users or data. Generally speaking, the degree of development of theory corresponds to the degree of development of Engineering category. New engineering practice may require theoretical innovation, and theoretical progress can expand people's ability to create engineering. Without Einstein's equation of mass and energy, I don't think people will be able to build an atomic bomb with any data for many years. So artificial intelligence can be regarded as a wonderful flower, thanks to the tremendous development of computing power, the degree of its engineering development has been disconnected from the theory. This lack of theory or epistemology can lead to serious consequences, which we will discuss later.

Artificial Intelligence Hasn't Promoted the Progress of Physiological Ideas

The biggest criticism of AI lies in its interpretability. It is difficult for people to understand how the ever-larger deep network plays its role and what functions each node has. For image recognition tasks, early depth networks

Please read the Chinese version for details.