preface
Most of us know ai mainly from science fiction movies, in which robots have human minds and extraordinary abilities. However, the current reality of artificial intelligence is only very weak artificial intelligence, it is only a single function of a program.
The 60-year development of artificial intelligence has experienced several booms and busts. Although it has made good progress, there is still a big gap between reality and ideal, and the road ahead is tortuous.
Before the Dartmouth conference
The Dartmouth conference in 1956 has been called the birth of AI, but there was already research on AI before that. At the International Congress of Mathematicians in 1900, mathematician Hilbert presented mathematical Problems of the Future, some of which were related to artificial intelligence.
Alan Turing designed the Turing machine, the theoretical prototype of the modern computer. In 1950, he published Computational Machines and Intelligence, which defined machines and thinking, as well as the Turing Test, a standard by which a machine is considered intelligent.
This famous test consisted of two enclosed rooms, one with a person in one room and a machine in the other. Outside the room was a tester who could only communicate by wire. If it was impossible to tell which room was occupied and which room had a machine, the Turing test was passed.
A program named “Eugene” passed the Turing Test in 2014, 64 years after it was first proposed. Alan Turing made significant contributions to artificial intelligence and is also known as the “father of artificial intelligence”. If the genius hadn’t killed himself by eating a poisoned apple at the age of 42, perhaps there would have been more progress in artificial intelligence.
Another genius known as the “father of computers” was Von Neumann, a scientific generalist in the fields of modern computing, game theory, nuclear and biological weapons. He designed the modern computer, the von Neumann architecture. The advent of modern computers allowed mathematicians to translate their research into practical results.
Dartmouth Conference
With the theory of mathematicians and the computer of engineering as the foundation, artificial intelligence really emerged. In August 1956, in the quiet town of Hannoth, Dartmouth College, John McCarthy, Marvin Minsky (experts in artificial intelligence and cognition), Claude Shannon, The founder of information theory), Allen Newell (a computer scientist), Herbert Simon (a Nobel laureate in economics) and others are getting together to discuss an entirely uninspiring topic: using machines to mimic human learning and other aspects of intelligence.
The conference lasted for two months, and although there was no general consensus, it did come up with a name for its discussion: artificial intelligence. So 1956 became the first year of artificial intelligence.
At that meeting Arthur Samuel developed a checkers program that was able to learn by itself, learning from games to improve, and three years later it was able to beat him.
Artificial intelligence took off after Dartmouth, and in 1956 Oliver Salfred developed the first character recognition program, breaking new ground in pattern recognition. In 1957, Rosenblat invented perceptron, and in 1960, GPS, a general problem solving system. In 1965, artificial intelligence reached a bottleneck, which was not a small setback. It entered the winter, and related funding was constantly reduced. In 1968, the first successful expert system DENDRAL came out, and various expert systems emerged continuously, forming a new branch of knowledge industry.
After 1977, AI got stuck. It could only accomplish small tasks, and the biggest problem was that no one knew how to acquire knowledge, how to learn from such a wealth of information.
The three major schools of
For the bottleneck of knowledge acquisition, machine learning has become the focus of attention since 1980s. Whereas traditional AI is spoon-fed, machine learning is heuristic, letting the machine learn on its own. One group thinks it can be done by mimicking the brain’s structural neural networks, known as connectionism. Others think the answer can be found in patterns of interaction between simple organisms and their environment, behaviorism. Traditional AI is known as the symbolic school. From the 1980s to the 1990s, there were three schools of thought.
The symbol school
The representative of the symbolic school is John McCarthy, one of the founders of artificial intelligence. His view is that any system that can operate some patterns or symbols of physics and transform them into other patterns or symbols may produce intelligent behavior. The physical symbol here can be composed of high and low potentials, or it can be the electrical impulse signal of the neural network.
Semiotics focuses on the higher levels of human intelligence, such as reasoning, planning, knowledge representation, etc.
Computer games put the symbolic school on the map, and in 1988 IBM developed deep Thought, a chess-playing intelligence program with the speed of 700, 000 moves. In 1991 Deep Thought II drew with the Australian chess champion, and in 1997 deep Thought upgraded Deep Blue defeated the world chess champion.
In 2011, IBM’s Watson supercomputer beat human contestants in the quiz, a free-form quiz on current affairs, history, literature, science, sports, geography, culture and more.
The symbolic school was losing momentum, and the dominance of ARTIFICIAL intelligence began to give way to other schools.
Connect the school
Just as human intelligence comes from the brain, every human brain has a trillion neurons, which are intricately interconnected. It was natural to wonder if the brain’s frontal intelligence could be simulated with a large number of neurons. The connection school holds that advanced intelligent behavior spontaneously emerges from the connections of a large number of neural networks.
The development of the connectedness school has been bumpy. In 1957 Frank Rosenblatt extended the computational model of single neurons, adding learning algorithms and calling it a perceptron. It adjusts weights to accomplish learning based on the error between the output of the model and the desired output.
In 1969, artificial intelligence authority Minsky pointed out through theoretical analysis that perceptrons could not learn all the problems, even the simplest one: whether a two-digit binary contains only 0 or 1. The death blow nearly killed neural network research.
In 1974, Jeff Hinton, the saviour of the connectionist school of artificial intelligence, proposed that neural networks would be powerful if they were numerous. By connecting multiple perceptrons into a layered network, it could solve Minsky’s problem successfully. However, multi-neurons also bring more complex network training problems, which may have hundreds or thousands of parameters to be adjusted. Hinton et al. found that the back propagation algorithm proposed by Arthur Bryson et al. could solve the training problems of multi-layer networks.
Soon the connectionists ran into trouble again because there was no theory to back them up, because neural networks could solve problems, but why they repeatedly failed at some of these problems is not clear. Moreover, people do not know how to improve the efficiency of neural network because of the ignorance of its operating principle.
Around 2000, statistical learning theory was put forward to point out that our model must match the problem to be solved. If the model is too simple and the problem itself is complex, the expected accuracy cannot be obtained, while if the problem is simple but the complex model is used, overfitting will occur.
behaviorism
The starting point of the behavioral school is completely different from the other two schools. They do not focus on the human beings with advanced intelligence, but on low-level insects. Insects are intelligent because they can move around and react quickly.
Robotic insects don’t have sophisticated brains. They don’t need any brain intervention, just coordination of their limbs and joints to adapt to their environment. In complex terrain, they can avoid obstacles intelligently, and this intelligence comes not from complex design, but from interaction with the environment.
The most famous is Boston Dynamics’ robot dog, which can walk, climb, run and carry loads over all kinds of difficult terrain.
Over time, adaptation forces evolution, and John Holland has published genetic algorithms for computer simulations of evolution. It abstracts biological evolution in nature, extracting two links: variation and selection. In computers, organisms are simulated by binary strings, and the selection of nature is abstracted as fitness functions.
There is no need for unity
Around 2000, artificial intelligence entered the new century. The development of artificial intelligence did not solve problems, but also introduced new questions one after another, which were more and more difficult to answer and involved more and more profound theories. So they simply ignored theoretical problems and focused on application. Practice was the only criterion for testing truth, and no matter what school of thought, it was a good school that could solve practical problems.
In this context, artificial intelligence is further divided into many independent disciplines. Examples include automatic theorem proving, pattern recognition, machine learning, natural language understanding, computer vision, and automatic programming. Unified AI makes no sense and has no need to exist.
Deep learning makes its debut
In the second decade of the 21st century, deep learning became the most prominent research in artificial intelligence. In 2011 the Google X lab took 10 million images from YouTube and fed them to a Deep learning-enabled Google Brain, which found the cat on its own three days later without human help. In 2012 Microsoft used deep learning to do real-time speech recognition and translation for speakers, which means simultaneous translation.
Although deep learning emerged in the 1980s, it was not effective due to the lack of hardware capabilities and data resources at that time. It wasn’t until 2009 that Hinton and his students continued to work in the field that they found unexpected success, when they applied deep learning to speech recognition and broke the world record for 25% fewer errors than before. Deep learning is catching on.
The reason for the big performance boost in deep learning is that it resembles a deep neural network like the human brain, which better mimics the work of the human brain.
The development of ARTIFICIAL intelligence is full of twists and turns. What will the future hold?
————- Recommended reading ————
My 2017 article summary – Machine learning
My 2017 article summary – Java and Middleware
My 2017 article summary – Deep learning
My 2017 article summary — JDK source code article
My 2017 article summary – Natural Language Processing
My 2017 Article Round-up — Java Concurrent Article
Talk to me, ask me questions:
The public menu has been divided into “reading summary”, “distributed”, “machine learning”, “deep learning”, “NLP”, “Java depth”, “Java concurrent core”, “JDK source”, “Tomcat kernel” and so on, there may be a suitable for your appetite.
Why to write “Analysis of Tomcat Kernel Design”
Welcome to:
