At present, attempts in deep learning are being made in all fields. In deep learning, it is a method of processing in three or more layers, for example, processing in the first layer and processing the data that appears in the next layer. Please refer to the article about deep learning this time.
In 2016, Go AI "AlphaGo" developed by Google's AI startup company "DeepMind" defeated Go world champion Lee Se-dol (Korea). At that time, it was said that it would take more than 10 years for AI to beat professionals in Go, and the achievement was reported in the media all over the world with the keywords "AI (artificial intelligence)" and "Deep Learning (deep learning)". I did.
And just the other day, on October 18, 2017, "Deep Mind" announced the latest Go AI "Alpha Go Zero". In "Alpha Go", we learned the striking lines of professional players in advance, and from there, we became stronger in the battle between AIs. However, the latest version of "AlphaGo Zero" is characterized by learning the rules of Go and improving the playing power only by self-learning (reinforcement learning). "AlphaMaster", an improved version of "AlphaGo", won all 100 battles against "AlphaGo" in 3 days by repeating self-plays without the need for a database of striking lines that humans have thought out through thousands of years of ingenuity It seems that he has grown to win 100 races and 89 wins.
"Alpha Go Zero" is an example, but attention is focused on "AI (artificial intelligence)", which has undergone remarkable evolution, and "Deep Learning", which supports its evolution. Therefore, this time, I will explain the basic idea and meaning of "AI (artificial intelligence)" and "Deep Learning", and the future.
What is AI (artificial intelligence)? What is Deep Learning?
First, I will introduce the basic concepts of "AI (artificial intelligence)" and "Deep Learning (deep learning)". The first thing to keep in mind is that "AI (artificial intelligence)" is a comprehensive concept and technology, and "Deep Learning" is one of the methods that support AI (artificial intelligence).
For example, when a human sees an animal, it instantly determines whether it is a dog or a cat. The mechanism is realized by comparing the information obtained from the eyes and ears with experience and knowledge and guessing "whether it is an animal" or "what kind of thing". The basic concept of AI (artificial intelligence) is the same, and it is to imitate the "guess" made by the human brain with a computer.
At that time, "learning" is important. AI (artificial intelligence) cannot be guessed without experience and knowledge, and it is not possible to derive an appropriate answer. It is necessary to learn the rules and rules necessary to make a judgment there. The learning method is called "Machine Learning", and there is a method called "Deep Learning".
Humans cannot distinguish animals among babies, but they will be able to distinguish them as they grow up. AI (artificial intelligence) also grows by learning like humans. And "Deep Learning" is attracting attention as a method closer to the human brain.
AI is an acronym for "Artificial Intelligence" and is translated as "artificial intelligence" in Japanese. Wikipedia says, "It refers to an attempt to artificially realize intelligence similar to humans on a computer, or a series of basic technologies for that purpose" (quoted from Wikipedia), but the interpretation varies depending on the researcher. There is no separate and strict definition. As a general interpretation, remember "concepts and techniques for artificially imitating human intelligence."
AI (artificial intelligence) requires computers to "learn" like humans and "guess" based on knowledge, which requires complex platforms and algorithms. AI (artificial intelligence) is used in various places such as voice recognition of smartphones, automatic driving to avoid obstacles, Internet image search and web page search, robot control and image processing in the industrial field.
Also, AI (artificial intelligence) has the image of the latest technology, but in fact research has been ongoing since the 1950s. The development of AI (artificial intelligence) that utilizes the current big data and deep learning is also called the "third artificial intelligence boom."
Machine learning is a method of finding regularity and relevance from a large amount of data, and making judgments and predictions. For that purpose, it is necessary for humans to specify features (features) that should be noted, such as "Be careful about color and shape."
Deep Learning is an advanced method of machine learning. Using a multi-layered neural network modeled on the human brain neural circuit, AI (artificial intelligence) decides the settings and combinations of features by itself. In machine learning, it was necessary to specify the point of interest such as "Be careful about color and shape", but in the case of Deep Learning, learning is done automatically without any instruction. However, a large amount of data is required to improve accuracy, and the direction of learning changes depending on the data to be read, so it is necessary to select carefully.
About types such as "specialized AI and general-purpose AI" and "strong AI and weak AI"
AI (artificial intelligence) is classified into "specialized artificial intelligence (Narrow AI)" and "general-purpose artificial intelligence (AGI)" depending on the application. There are also cases where it is classified into "weak AI" and "strong AI" depending on the level of functionality. Here, we will explain the classification of AI (artificial intelligence).
"Specialized artificial intelligence (Narrow AI)" refers to something that demonstrates performance in a specific work or area, such as the Go AI "AlphaGo". On the other hand, "Artificial General Intelligence (AGI)" refers to those that perform as well as or better than humans without limiting the work or area. As an image, general-purpose artificial intelligence corresponds to a robot program that is close to life and that thinks for itself and acts independently, as in science fiction movies. However, AI (artificial intelligence), which is as versatile as humans or more, is currently impossible. As far as AI (artificial intelligence) is put into practical use, it can be called specialized artificial intelligence.
Weak AI and strong AI are classified according to the degree of function, etc., and the criterion is how close they are to humans. There is no clear standard such as how weak AI is and how strong AI is, but in general, those who do work mechanically without being conscious like human beings are learned as if they are conscious. Those that can make decisions are called strong AI. AI that utilizes Deep Learning can be said to be a strong AI.
What do I need to use AI? About learning models and datasets
When actually using AI (artificial intelligence), it can be divided into a "learning phase to create artificial intelligence" and a "prediction / recognition phase using artificial intelligence". Of particular importance is the learning phase of growing a baby's AI (artificial intelligence), which requires a "data set" and a "learning model" for learning. In general, regularity and relationships are extracted from a "data set" and learning is repeated to create a "learning model".
Especially in the case of Deep Learning, a large amount of data is required to improve accuracy. Data accuracy is also an important factor. If you learn with incomplete data, you risk making the wrong decision and AI will not learn in the direction you are looking for.
What is the computational cost of building a trained model?
To build a trained model, design after clarifying "answers you want AI to derive," "learning required to obtain those answers," "data sets required for learning," etc. .. At that time, it is necessary to collect raw data, create a database, create a data set, and repeat learning, so it takes a huge amount of time to organize and learn the data.
In addition, AI (artificial intelligence) needs to be tuned to learn in the desired direction. Therefore, when introducing AI (artificial intelligence), specialists such as AI (artificial intelligence) consultants and data scientists are indispensable.
Deep Learning has made it possible to handle complex data that was not possible with conventional machine learning. From here, AI (artificial intelligence) will evolve further. If AI (artificial intelligence) learns more advanced things by itself, it is expected that the burden on humans will be reduced and the work will be dramatically more efficient.
In a report released by Accenture Research and Frontier Economics, "AI will increase economic growth by an average of 1.7% in 16 industries by 2035." "AI technology will increase productivity by more than 40% by 2035." There is a possibility ”and so on.
In addition, IDC Japan announced the domestic market forecast of "cognitive / AI (artificial intelligence) system" on November 15, 2017. From 2016 to 2021, the average annual growth rate is expected to be 73.6%, and the market size in 2021 is expected to be 250,109 million yen, which is about 16 times that of 2016. AI is expected to grow rapidly in Japan as well. The driving force behind this is Deep Learning. AI (artificial intelligence) will be used not only in the IT field but also in various places such as education, service industry, agriculture, and manufacturing industry.