Seeing Chen Meiping: The past and present lives of the three schools of artificial intelligence

As an emerging discipline, due to people's different understandings of artificial intelligence, different thinking about artificial intelligence has arisen. Throughout the development of artificial intelligence, researchers with different disciplinary backgrounds have different understandings of artificial intelligence, so three major schools of artificial intelligence have emerged. Traditional artificial intelligence is called the symbolist school, and symbolism is mainly based on intelligent simulation methods of logical reasoning; Others believe that it can be achieved by simulating the neural network structure of the brain, that is, the connectionist school; Others believe that the answer can be found in the patterns in which organisms interact with the environment, known as the behaviorist school.

1

Symbolism

Symbolism, also known as logicism, psychology or computer science, was mainly represented by two professors at Carnegie Mellon University: Allen Newell and Herbert Alexander Simon. In the 50s of the 20th century, they collaborated with another famous scholar at the time, John Cliff Shaw, and developed a set of heuristic programs—logical theorists. This set of programs proved Turing's diagnosis that machines can be intelligent for the first time, and opened the prelude to the use of computers to explore human intelligent activities. The symbolism they founded made outstanding contributions to the development of artificial intelligence. Symbolism bases the intelligence generated in the abstract thinking of the brain on the assumption of the physical symbol system, and realizes the corresponding intelligent behavior through other symbolic structures produced by physical pattern establishment, modification, duplication and deletion operations composed of conforming entities. Symbolism has been prominent for a long time, making important contributions to the development of artificial intelligence, and is the mainstream school of artificial intelligence in the early days. The main achievements of symbolism are: machine theorem programs (LT, GTM, GPS...). ), heuristic algorithms, expert systems, etc.

In the 80s of the 20th century, symbolism began to decline and decline. Symbolism faces four main challenges: 1. automatic acquisition of knowledge; 2. Automatic integration of diverse knowledge; 3. Knowledge-oriented representation learning; 4. Knowledge reasoning and application. Although symbolism realizes artificial intelligence by simulating human thought processes, it is difficult to achieve breakthrough results in the above four problems.

2

Connectionism

Connectionism, also known as biomimetic or physiology. The pioneers of connectionism were the American psychologist W.S. McCulloch and the mathematician W. Pitts. In 1943, Home McCulloch and Home McCulloch jointly proposed the McCulloch-Pitts model, an early model of neuronal networks. The principle of connectionism is neural networks and the connection mechanism and learning algorithm between neural networks, through the research of bionics and biology, human brain models and neural network models are created from neurons, and electronic devices are used to simulate the structure and function of the human brain. Connectionism believes that intelligence is an introspective thinking process, that is, the interaction between brain neurons and the mutual communication of information, and artificial neuron networks can be established on the basis of simulating the structure of the brain's nervous system, so as to achieve corresponding intelligent behavior. Connectionism emphasizes information exchange, self-organizing behavior, competition, and collaboration, and it not only looks at the world from the perspective of matter and energy, but also values the existence of "order" in matter.

After the introduction of the formal neuron model (M-P model) in 1943, the connectionist school began its ups and downs. The perceptron was invented in 1957, after which connectionism fell silent. In 1982, the Hopfield network, in 1985 the limited Boltzmann machine, in 1986 the multilayer perception machine was invented, in 1986 the backpropagation method solved the training problem of the multilayer perceptron, and in 1987 the convolutional neural network began to be used for speech recognition. Since then, connectionism has gained momentum, from models to algorithms, from theoretical analysis to engineering implementation, laying the foundation for neural network computers to go to market. In 1989, backpropagation and neural networks were used to recognize numbers written by bank checks, enabling the first commercial application of artificial neural networks. As human beings enter the era of the Internet and big data, after the three elements of artificial intelligence, algorithms, computing power, and data are complete, the connectionist school has begun to shine. In 2009, multi-layer neural networks made a major breakthrough in speech recognition, in 2011 Apple worked to integrate Siri into the iPhone 4, in 2012 Google's driverless car began road testing, in 2016 DeepMind defeated Go champion Lee Sedol, and in 2018 DeepMind's Alphafold cracked a 50-year-old protein molecule folding problem. The connectionist school has achieved brilliant results in the field of artificial intelligence, so much so that the artificial intelligence that the industry now talks about basically refers to the technology of the connectionist school, and relatively speaking, symbolism is called traditional artificial intelligence.

While connectionism is so strong today, hidden dangers that could hinder its future development have emerged. The limitation of connectionist thinking is that good results can be obtained under differentiable, strongly supervised learning, closed static system tasks, and the results obtained by training are also limited to the problem given conditions.

3

behaviorism

Behaviorism, also known as evolutionism or cybernetics. Cybernetic thought was influenced by behaviorism, and Wiener pointed out in his 1948 "Cybernetics": "Cybernetics is developed on the basis of self-control theory, statistical information theory and biology, and the adaptive, self-organizing, self-healing and learning functions of machines are determined by the input and output feedback behavior of the system." In the early research on artificial intelligence, it mainly focused on simulating the intelligent behavior and role of humans, linking cybernetics, self-organizing systems, engineering cybernetics, biological cybernetics, information theory, etc., tearing off the seeds for the birth of intelligent control and intelligent robot systems in the 80s of the 20th century.

American robot manufacturing expert Rodney A. Brooks, as a representative of behaviorism, believes that intelligence depends on perception and action, and proposes the "perception-action" model of intelligent behavior. Behaviorism believes that intelligent behavior arises from the interaction process of perception and behavior of complex external environments, and decomposes complex behaviors into many simple behaviors and studies them one by one. Perception is a certain response to an environmental stimulus, and behavior is a statement of that response. This rapid feedback can be adapted to complex, non-systematic, and non-modeled objective environments. As the basis of behaviorism, cybernetic thought links the working principle of the nervous system with information theory, control theory, logic, and computers, and simulates the intelligent behavior reflected in the control process. The main results of the behaviorist school are: hexapods, Japanese asimo, large dogs.

The behaviorist school had a strong sense of purpose at the beginning of its birth, which also led to its advantages and disadvantages being obvious. Its main advantage is that behaviorism attaches great importance to results, or the performance of the machine itself, and is very practical. The disadvantage of behaviorist thinking is that it oversimplifies the process of human behavior, ignores the internal process of human psychology, and ignores the importance of consciousness.

To sum up, we can briefly think of symbolism as studying abstract thinking, connectionism studying figurative thinking, and behaviorism studying perceptual thinking. Symbolism focuses on mathematical interpretability, connectionism favors humanoid brain models, and behaviorism favors application and body simulation. Symbolism relies on artificial intelligence to give machines, connectionism relies on machines to learn intelligence on their own, and behaviorism obtains intelligence in relation to environmental interactions and feedback. Both connectionism and behaviorism are trained using reinforcement learning methods. The advantages and disadvantages between the three are obvious, and in the field of artificial intelligence, a more mature intelligence theory has not yet been formed. With the development of artificial intelligence, these three universities are gradually moving from early fierce debates and separate research to complementary and comprehensive research, and work together to create more powerful artificial intelligence. It is believed that with the deepening of artificial intelligence research, these three universities will integrate and play a role in the practical application of artificial intelligence, and will also find the final answer for the theory of artificial intelligence.