The world is undergoing rapid technological evolution. Many technologies are changing at a dizzying pace. Above all, artificial intelligence, which has been attracting attention in recent years, is evolving rapidly as it enables us to do things that we could not do before. Regarding artificial intelligence, we would like to introduce some reference materials this time.
A short introduction to the contents of popular keywords such as "artificial intelligence" and "AI". From the outline to the definition, major classification, typical cases, and main history.
Artificial intelligence (Jinkou Chino, English: Artificial Intelligent, abbreviation: AI) is a part of "intelligent behavior" performed by human beings artificially reproduced using a computer program, or its research field. Point to. Intellectual behavior is the general activity of thinking and executing with the head (strictly speaking, the brain), and this is all human behavior such as "drawing a picture", "recognizing words", "playing a game", etc. Applies to.
Machine learning has been attracting attention in recent years as a type of artificial intelligence. Machine learning refers to the artificial realization of appropriate intellectual behavior by learning from data, or its research field. It is thought that human beings will be able to perform appropriate intellectual behavior by learning from experience. For example, by looking at fruits such as apples and mandarins many times, it becomes possible to distinguish the types of fruits. The same is true for machine learning, where learning from data can help identify fruits.
Deep learning is attracting a great deal of attention as a method of machine learning. It has rapidly gained popularity since the 2010s, creating a new artificial intelligence boom called the "third". Deep learning has begun to be applied to various fields such as image recognition (for example, judgment of dogs and cats from images), voice recognition, and autonomous driving technology, and is expected to develop further in the future. There is.
Below, artificial intelligence is explained as briefly as possible from several points of view.
The meaning of artificial intelligence that is generally recognized is as described above, but there is no sentence that strictly defines it. Different people have different definitions. Table 1 summarizes some of the most obvious definitions.
Artificially created human-like intelligence or technology to create it Yutaka Matsuo (Associate Professor, Graduate School of Engineering, The University of Tokyo)
An artificially created, intelligent reality. Or Hideyuki Nakashima (President of Future University-Hakodate), a field that studies intelligence itself by trying to create it.
Makoto Nagao, a system that simulates human brain activity to the limit (Professor Emeritus, Kyoto University, Former Director of the National Diet Library)
Main definitions of artificial intelligence (cited: "Deep learning textbook Deep Learning G test (generalist) official text", source: "Journal of the Japanese Society for Artificial Intelligence")
1.2 Difference between robot and artificial intelligence
As a general image of artificial intelligence, it is often confused with a robot. Let's clarify the difference again.
It can be said that artificial intelligence is equivalent to the "brain" in humans. Robots, on the other hand, can be said to be equivalent to the "body" of humans.
Of course, there are cases where artificial intelligence is used as a part of a robot, just as the brain is a part of the body. However, this is not the case, and artificial intelligence that does not have a body = robot is being used more.
Please note that artificial intelligence is a field centered on "intelligent behavior", and whether or not it is used in a robot is another matter.
2. Major classification of artificial intelligence
Artificial intelligence can be broadly classified into two categories, "strong AI" and "weak AI" (John Searle, 1980), in terms of ease of realization. The meaning of each will be explained.
Strong AI is an AI that can completely imitate the intellectual activities performed by humans, and is also expressed as "general-purpose AI" (AGI: Artificial General Intelligence). The antonym is weak AI.
For example, "Doraemon" that can think and act like a human being is exactly "strong AI". In other words, it is artificial intelligence with a mind.
However, the current technology has not yet advanced to the level where "strong AI" can be realized. AI realized by current machine learning and deep learning is called "weak AI".
Weak AI (Narrow AI) is an AI that realizes only specific processing, and is also expressed as "specialized AI". It is artificial intelligence that has no mind and is used as a useful tool. There is a strong AI in the antonym.
For example, the process of identifying defective products from the products manufactured at the factory is "weak AI". This process only identifies defective products, and cannot do anything as versatile as humans do.
The current technical level is at the stage of realizing this "weak AI". Research is continuing with the expectation that "strong AI" can be realized by increasing versatility by combining weak AI.
Let's introduce an example where artificial intelligence has produced results. Especially in deep learning methods, there are great success cases in fields such as "image recognition / generation", "speech recognition / generation", "natural language processing", and "robotics (reinforcement learning)". Among them, the application examples of "image recognition" are abundant and the most successful, and it is expected that further progress will be made in the future. Please note that each case is not necessarily one AI technology, and is often a combination of complex advanced technologies.
Object recognition: Identification of dogs and cats, verbalization of the meaning of objects, etc.
Defective product detection: Inspection of parts made in the manufacturing industry (for example, car screws)
Detection of foreign matter: Foreign matter detection on food manufacturing lines, etc.
Lesion detection: Finding lesions from medical images such as X-rays
Self-driving: Currently, competition is intensifying in industry
3.2 Main examples of voice recognition / generation
Smart Speaker: Products such as Google Home and Amazon Echo that give some answers when asked by voice
3.3 Main examples of natural language processing
Chat bot: When a question is posted on a website, an appropriate answer is automatically returned to establish a conversation. For example, high school girl AI "Rinna"
3.4 Main cases related to robotics (reinforcement learning)
Automatic game match: AI became famous for defeating professional human Go players in 2015. For example, AlphaGo
4. Main history
Artificial intelligence research, which began around 1950, has a history of alternating "boom" and "winter era" every 10 to 20 years (Fig. 2). It should be noted that there is no clear division between booms and winter times, so the following is a rough division of times.
Figure 2 History of artificial intelligence
Let's briefly introduce the contents of each boom.
In 1956, at the Dartmouth Conference, Mr. John McCarthy presented the name "artificial intelligence", and it is said that a new research field was born as a result. After that, in 1957, a concept called "Perceptron", which is the basis of neural networks, was proposed, and research on artificial intelligence, especially neural networks, became active.
This is the "first artificial intelligence boom". The focus of AI research in this era was mainly on "inference / search (mazes, puzzles, proofs of mathematical theorems, etc.)".
However, in the 1960s, it was easy for Perceptron to show the limits of simple Perceptron, specifically the inability to identify linearly inseparable patterns (Marvin Minsky & Seymour Papert, 1969). It became clear that only the problem could be solved (called the "toy problem problem"). For this reason, from the first half of 1970, we will enter the "winter era" when the popularity of AI research will decline.
In the 1980s, artificial intelligence called an "expert system" was born, which accumulated knowledge in specific fields such as medical care and answered questions, and became popular with companies around the world. Around this time, the government-led "fifth generation computer" project was also promoted in Japan. Many Japanese companies participated in this, and the adoption of artificial intelligence was extremely prosperous in Japan as well.
This is the "second artificial intelligence boom". The focus of AI research in this era was mainly on "knowledge." By the way, it was around this time that neural networks began to use an algorithm called "backpropagation (inverse error propagation)" (David Rumelhart et al., 1986).
However, in the 1990s, it became clear that expert systems needed to accumulate a huge amount of knowledge, and that it would cost a huge amount of money. With that being recognized as a limit, the adoption of artificial intelligence in industry has also slowed down. The second long winter era has arrived.
In 2006, a study (Geoffrey Hinton et al.) Was announced that would break the winter era. This is a method called "deep neural network" that deepens the hierarchy of neural networks, and is the beginning of the current "deep learning".
Since 2010, the performance and capacity of computers have increased, the Internet has become widespread, and the cloud that can manage huge amounts of big data has developed, and the environment related to data has been improved and advanced. This has made it easier to study deep learning, which requires vast amounts of data.
At the "ILSVRC" (ImageNet Large Scale Visual Recognition Competition) held in 2012, artificial intelligence using deep neural networks (Alex Kryjevsky et al.) Recognized other artificial intelligence. Greatly exceeded the accuracy. This triggered a worldwide enthusiasm. This is the beginning of the "third artificial intelligence boom."
This boom is still ongoing. The current focus of AI research is mainly on "machine learning and deep learning." One of the immediate goals or milestones of this technological innovation is "fully autonomous driving."
Shinichi Asakawa, Arisa Ema, Ikuko Kudo, Yusuke Negago, Keisuke Seya, Takayuki Matsui, Yutaka Matsuo, & Others (2018) "Deep Learning Textbook Deep Learning G Test (Generalist) Official Text"
Takashi Onoda (2016) Let's get started with machine learning in the 3rd artificial intelligence (AI) boom! --Deep Insider
Masahiko Isshiki, & Cooperation: Akio Abe (2018) Introduction to Machine Learning & Deep Learning (Overview) --Deep Insider
Ministry of Internal Affairs and Communications (2016) 2016 White Paper on Information and Communications | History of Artificial Intelligence (AI) Research
Journals of the Past – The Japanese Society for Artificial Intelligence
Marvin Minsky and Seymour A. Papert (1969) "Perceptrons"
John R. Searle (1980) "Minds, brains, and programs" "Behavioral and Brain Scie"