However, most scientists who have been trained in scientific research for a long time have unconsciously turned "all phenomena can be simplified into strict descriptions" into axiomatic assumptions, but have never thought about whether this assumption applies to intelligence itself. An extreme example of this idea should be Sir roger penrose of Britain, who has written three books in succession, The Emperor's New Mind, The Shadow of the Mind and The Consciousness of the Big, the Small and the Man. Small and human mind), to clarify the explanation of consciousness (I think there is no definition difference between "consciousness" and "intelligence" here) requires a complete theory of quantum gravity, and it is impossible to create consciousness based on the current classical physics technology by using Turing machine stop and Godel's incompleteness theorem. However, the complete theory of quantum gravity does not exist today, otherwise there would not be such a beautiful movie as Interstellar. Scientists represented by Penrose only made some guesses in this respect, which is far from proving that they are right. In addition, Penrose's use of Turing machine downtime and Godel's incompleteness theorem is aimed at machines, turning people into a third-party perspective. However, from the perspective of the third party, these paradoxes do not necessarily exist for the subject. If these paradoxes want to appear, every human individual should consider himself rather than other logical systems. Maybe one day aliens will see us humans, and then apply Cantor's diagonal principle to a certain model of the human brain, and then say, "Look, humans are not smart because they have paradoxes"-this is obviously the wrong way of thinking. In fact, there are many people in this school, such as Shi and Rao Yi (related reference: Summit Dialogue: Appreciation, Popularization and Sublimation of Life Science), Professor Zhu of the discipline, and most scholars of the theoretical school (note: only a few people here actually think that this kind of intelligent theory needs quantum gravity theory as a premise, but they all think that intelligent phenomena can be simplified into rigorous descriptions). But they may not realize that they use this assumption to look at intelligence.
The characteristic of this school is the pursuit of a perfect explanation of everything, and it would be best if all intelligent phenomena could be reduced to a formula. However, at present, such an attempt has not been completely successful. An example is the traditional statistical machine learning theory, which deviates from the actual application when applied to the actual model, so there is no way to refer to it, but this is because these theories are the "upper limit" of the worst case and naturally cannot describe the general situation of actual use. To some extent, the development of deep learning is a process of jumping out of this "upper limit" limitation, and of course it also benefits from the progress in data, calculation and model. But personally, I think the research in this school is actually very useful. Even if intelligence can't be reduced to several formulas in the end, it is valuable to figure out what can be clarified in this process.
On the other hand, some other researchers believe that intelligence, as a phenomenon, is actually only an intuitive description of the complexity of human biological system (especially nervous system). Because complexity is its basic requirement, it cannot be simplified to strict rules. This irreducibility even includes the question of "how to define intelligence". Turing Test was put forward by Turing in his famous philosophical paper Computing Machinery and Intelligence published in 1950, and it is an intelligent definition based on this concept. That is, as long as the intelligent reference objects (i.e. people themselves) unanimously recognized by the testee and the judge are different in general (i.e. statistically), they can be considered as intelligent (in my opinion, this description is actually an early and imprecise prototype of PAC learning theory). Personally, I think this view that intelligence is complex and irreducible is simple and useful. Geoffrey Hinton, the father of deep learning, said at the 20 16 IEEE/rse james clerk maxwell Medal Award Ceremony that Turing and von Neumann did not admit that intelligence could be created based on logic. If we contact history, I guess they rejected the whole idea that intelligence can be reduced to strict laws, and engineering methods like neural networks may be the source of progress. More importantly, we recognize each other that every human individual is an agent, but we never ask each of us to reduce him to a theory or understand every state of his neurons in order to understand another individual-this is physically unrealistic. With the improvement of computing power of artificial machines, we may have to pay attention to the behavior of the system, and its internal operating mechanism can only be at an irreducible level. This of course also depends on the material basis of human intelligence. If our biological brain becomes extremely powerful in the future, at that time, we may see the current neural network or the mechanical abacus of the past. In addition, even if individual intelligence can be reduced to a rigorous theory, there is still a bigger phenomenon than human individual intelligence, that is, the evolution of human society and human beings (how to create machines to realize society and evolution? )。
At present, the research on deep learning is mostly based on this idea, mainly in practical aspects such as model design, optimization algorithm and application field. Personally, I feel that since such an attempt has not yet seen the end (for example, new tasks are constantly made possible by deep learning methods), there is no need to aim at "inventing a theory to describe all deep learning", because we don't know where the boundaries of deep learning are in practice, let alone whether it is fundamentally feasible to restore intelligent phenomena to strict laws.
Before I end, I want to explain that the distinction between the above two factions is not absolute and independent. Many researchers have put forward very good theories to understand these models while completing many excellent practical work. Although these theories are not as reducible as physics, they are still very important.