Outline and current status of AI research
Overview of AI research
There is no definition for AI, various technologies are called AI, and there is no unified definition for AI research.
Furthermore, AI research is expanding, making it difficult to get a bird’s-eye view of the whole. In the future, it is expected that many systems that use the results of AI research will be used more and more in the real world.
Therefore, it is important to grasp the correspondence between AI systems and AI technology.
Therefore, the Japanese Society for Artificial Intelligence AI Map β 2.0 has been released as a guide targeting beginners in AI research who will be active in the future, as well as researchers and practitioners in different fields who aim to utilize AI.
AI is mainly divided into two types, general-purpose artificial intelligence and specialized artificial intelligence, and research on specialized artificial intelligence is actively being carried out, especially in companies.
| General purpose artificial intelligence : artificial intelligence that can handle not only specific tasks but also various tasks like humansSpecialized artificial intelligence・・・Artificial intelligence that processes problems in a limited area |
In the research of specialized artificial intelligence, research that utilizes machine learning technology such as deep learning is particularly active, and research that uses machine learning to recognize and generate images, voices, and texts is also active. has become
International situation of AI research
Yutaka Matsuo, a leading AI researcher, said that in AI research, “Both the United States and China are at the top, followed by the United Kingdom, Canada, Germany, Singapore, etc., and Japan is still below. ‘ said.
It can be said that Japan is lagging behind in AI research. In AI research, the United States and China are competing for hegemony.
Among the world’s top 50 (industry-government-academia) institutions that lead AI research, well-known companies such as Google, Facebook, and Microsoft dominate the top three.
America appears to be leading. However, recently there is a view that China has overtaken the United States.
In 2020, China overtook the United States for the first time in citations for research quality.
Nicholas Sharan, who served as the Pentagon’s first chief software officer (CSO) , resigned in October 2021 over protests against America’s slow pace of innovation.
And in the race to develop artificial intelligence (AI) technology, he said:
“We (America) are not likely to win the race against China for the next 15 to 20 years. At this point, the game is already over. In my opinion, the battle is over.”
Many people probably thought that the United States was leading China in AI research around the world, and that the United States could continue to lead China in the future.
However, progress in AI research and development in China is remarkable, and it is believed that the battle between the United States and China over “AI hegemony” will intensify in the future.
Main AI research fields
Here are seven major areas where AI research is advancing.
- algorithm
- voice recognition
- image recognition
- machine learning
- natural language processing
- Inference/Search
- data mining
- robot
I will explain each.
(1) Algorithm
An algorithm is a method or procedure that a computer uses to perform some calculation.
A computer composes a complex program by folding many simple choices, and it can be said that an algorithm corresponds to this choice.
AI processes and analyzes vast amounts of data. Algorithms build models to express the patterns and features of that data.
We can see that algorithms play a very important role in the process of AI learning.
② Voice recognition
Speech recognition is a technology that allows computers to recognize speech. In other words, it is a technology that measures human speech as sound (air vibration), analyzes the waveform data obtained from it, and converts it into character data.
Many people have probably used smart speakers such as Alexa, Google Assistant, and Apple’s Siri. The technology used in them is exactly voice recognition!
Currently, we are developing research to enable recognition outside of limited situations such as inside a car, and to identify who is speaking.
(3) Image recognition
Image recognition is a technology that allows computers to understand images.
It is one of the pattern recognition technologies that read features such as color and shape from an image and put the features into a learning machine so that new images can be recognized.
As a familiar example, it is used for face recognition of smartphones. It is also used in various other industries, and it can be said that it is a research field with room for growth.
④Machine learning
It is a technical research in which computers learn by themselves using collected data and try to find consistent rules.
Specifically, it learns relationships between variables and the importance of variables from data, and discovers important hidden variables (factors). This allows us to classify data, make predictions, etc.
In this way, AI research has become indispensable.
⑤Natural language processing
This is research that makes computers understand the words, contents, and meanings of text data.
Specifically, we perform morphological analysis, syntactic analysis, contextual analysis, semantic analysis, etc. on the data, and analyze it from a superficial point of view rather than the content.
It is used in translation software, search engines, and other things that we usually use, and natural language processing is essential to realize AI that replaces human roles.
It is also applied to speech recognition and information retrieval, and it can be said that this is a field where research is still progressing.
(6) Inference/Search
Inference is a method of deriving a consistent answer using a large number of rules.
The most basic is the syllogism. This is a three-step conclusion: ” Socrates is a man. Man dies. So Socrates dies. “
Exploration is a technique for finding items that meet certain conditions from a collection of data. It is the underlying technology for inference.
With these technologies, AI can instantly divide patterns, making it possible to solve puzzles and mazes overwhelmingly quickly.
I’m sure many of you are aware that AI is challenging professional shogi players and playing a good game.
⑦ Data mining
It is a technology that combines database technology and machine learning, and is researching a method to find useful information from a large amount of unorganized data.
Let’s use the retail industry as an example.
You can find useful information to achieve your goals, such as optimizing campaigns and improving the accuracy of sales forecasts, from the customer database that is accumulated daily. This will allow you to take action that will increase your sales.
In this way, this research has brought various benefits to various industries and is in demand.
⑧Robot
It is a research that combines mechanical engineering and artificial intelligence research.
We decide how to move the robot by applying methods in each field of AI. Routine work that repeats simple work is gradually being replaced by robots.
By entrusting tasks that can be done by robots to robots, it will lead to a solution to labor shortages. At the same time as the attention of AI, the introduction of robots is also progressing.
Therefore, robot x AI is one of the hottest fields.
Lab for AI research
There are many laboratories doing AI research. Here, I would like to introduce the Matsuo Laboratory at the University of Tokyo.
The Matsuo Laboratory is the laboratory of Professor Yutaka Matsuo, a leading AI researcher. Starting from Hongo , where the University of Tokyo is located, the laboratory has a vision to create an ecosystem where globally active researchers and entrepreneurs gather.
At the Matsuo Laboratory, we conduct basic research centered on deep learning (especially world models).
In addition, he has received the Makoto Nagao Memorial Special Award, which is a representative of Japan in the field of web mining, and students have also received the Dean Award of the School of Engineering.
Furthermore, the Matsuo Laboratory conducts joint research with many companies.
The purpose is to bring great value to society by applying the latest research and technology to business, and to brush up research and development by utilizing actual data.
For example, joint research with Mizuho Bank on sophistication of foreign exchange transactions using AI.
The aim is to refine the execution of foreign exchange transactions by analyzing foreign exchange transaction data using AI, speeding up transactions, increasing transaction volume, minimizing exchange risks, and maximizing profits.
There are many other labs doing AI research.
This article summarizes the laboratories conducting AI-related research by region, so please refer to it.
6 domestic and foreign companies conducting AI research
From here, we will introduce 6 domestic and foreign companies that have many affiliations with the “2000 Most Influential AI Scientists” .
- Microsoft
- NEC
- cyber-agent
- NTT
I will explain each.
Google has many excellent AI researchers , but 135 researchers were selected as “2000 Most Influential AI Researchers”. This is the top number of educational institutions such as universities.
Google is a world leader not only in the use of AI for services, but also in research and development, and has invested a huge amount of money in AI research and development.
In 2014, it acquired British AI company DeepMind for £400 million.
DeepMind is the company that developed the Go program AlphaGo, which became a hot topic in 2015 when it won a match against a professional Go player.
The biggest feature of AlphaGo is that it uses neural networks. The program itself learns by repeating the game against itself tens of millions of times.
Another feature of AlphaGo is the combination of a search algorithm called Monte Carlo tree search.
Due to a different approach from previous AI efforts , AlphaGo had high accuracy. After purchasing DeepMind, which develops such ultra-high-precision AI, Google’s AI research has gained momentum.
Since then, we have established AI laboratories in many regions such as Africa, China, India, and the United States, and have continued to lead the world in AI research by focusing on more global AI development.
②Microsoft
At Microsoft, 80 researchers have been named to the 2000 Most Influential AI Researchers. This number also exceeds that of other educational institutions such as universities.
In 2017, Microsoft announced the creation of a new AI lab, Microsoft Research AI, a research and incubation hub within its research organization focused on solving AI’s toughest challenges.
A team of scientists and engineers was established to work with Microsoft research and product groups.
The primary reason for its establishment is to integrate the research fields of AI, which have developed as separate research fields such as machine learning , cognition, and natural language processing.
This integrated approach has led to advanced understanding and the development of tools that can assist people in complex, multifaceted tasks.
Also in 2019, we announced a $1 billion investment in OpenAI to support the development of AGI (general purpose AI). OpenAI is an American non-profit organization established in December 2015 with the goal of creating the world’s first general-purpose AI.
It can be said that Microsoft has invested a huge amount of money in AI research and development and is leading the world in AI research.
Fifty-one people were named to Facebook’s 2000 Most Influential AI Researchers list.
Facebook has an AI research group called Facebook AI Research (FAIR). The purpose of setting up FAIR was to research cutting-edge AI technology and raise the level of AI technology throughout society.
Developed by FAIR, PyTorch is an open source machine learning library written in the Python language.
It is one of the most popular machine learning libraries. A major feature of PyTorch is that it closely resembles the basic operation method of Numpy.
Numpy is a must-have library for machine learning in Python. User-friendly design is one of the reasons for its popularity. Another feature is that it uses a dynamic computational graph.
Many of the staple products in libraries for building neural networks use static computational graphs. By introducing a dynamic computation graph, it becomes easier to implement complex networks due to its flexibility.
In this way, Facebook’s AI research and development is characterized by user-like progress.
④NEC
NEC has defined the technological areas of “recognition AI,” “analysis AI,” and “control AI,” and is working on research and development to create new social value.
NEC has won the top spot in the world five times in face recognition technology evaluation tests conducted by the National Institute of Standards and Technology (NIST).
In particular, large-scale “1:N authentication”, which is mainly used in the FinTech area such as cashless payment, public transportation, digital ID, etc., which requires higher accuracy, can capture still images of 12 million people. Performance was evaluated with an authentication error rate of 0.22%.
Face recognition is an authentication method that identifies the face information captured by the camera by matching it with a database.
The process detects “where the face is” in the image, finds “where the characteristic points of the person’s face are,” such as eyes, nose, and mouth, and detects the feature points of the face It requires advanced analysis to determine “who is” from.
In order to improve the performance of important elements in face recognition,
- Face detection technology that finds the position of a target face in an image with high speed and high accuracy
- Facial feature point detection technology that stably analyzes facial features regardless of aging or changes in facial expressions
- High-precision face matching technology aiming to eliminate false matching in all situations
is focused on research and development.
As a result, compared to other companies, we have achieved a minimum error rate, authentication accuracy that is less affected by aging, and authentication accuracy that does not depend on the number of registered people.
⑤ Cyber Agent
CyberAgent established “AI Lab” for the purpose of research and development of a wide range of AI technologies related to digital marketing in general.
We have researchers specializing in machine learning , econometrics, computer vision, natural language processing , and HCI.
In addition, we are strengthening industry-academia collaboration with universities and academic institutions that have advanced AI research technology and are actively working on practical application, and are working on research and development aiming not only to solve business problems but also to contribute academically.
The AI Lab is strengthening its research by field. Most recently, in October, we established the “Digital Human Research Center” to study the generation of CG avatars using AI.
In order to automatically generate various human movements and create high-definition digital humans, this organization has expanded its research subjects to include motion and effects.
We will create CG models that do not feel out of place compared to real people, use them in advertisements, TV commercials, and other videos, and develop technologies for various businesses such as online customer service. We aim to realize a world where people can thrive.
In October of the same year, we established the “Completely Automated Dialogue Research Center,” which specializes in research in fields such as speech synthesis and speech recognition.
In addition to speech synthesis and speech recognition, we will strengthen research in areas such as natural language processing , voice quality conversion, emotion recognition from voice, and speaker identification.
By 2026, we aim to establish a “fully automated conversation” that can understand the user’s intention and the context of the conversation.
⑥ NTT
NTT conducts AI research and development based on the following four concepts.
- “Agent-AI” that interprets information emitted by people and understands their intentions and emotions
- “Heart-Touching-AI” that reads the unconscious mind and body of people and understands their deep psychology, intellect, and instincts.
- “AmbientAI” that reads and understands the universe (objects, people, environment), predicts and controls instantly
- “Network-AI”, in which multiple AIs organically connect and grow to optimize the entire social system
The future of AI research
AI has the potential to change society in the future. Therefore, AI research is very important for future development, and we can predict that its importance will increase.
Also, from the perspective of AI advanced countries, Japan has an extremely small number of AI researchers and is in short supply. In that sense, the demand for AI researchers who have a correct understanding of the possibilities of AI is increasing.
AI is mainly divided into two types: general-purpose artificial intelligence and specialized artificial intelligence.
Since the beginning of AI research in the 1950s, the goal has been to develop robots that behave like humans, in other words, general artificial intelligence. However, due to technical constraints, the development of “specialized artificial intelligence” limited to specific functions has been promoted.
Since the advent of deep learning, highly versatile technology has developed rapidly.
In the near future, it is believed that multimodal AI, which learns by combining multiple pieces of information, will increase.
“Multimodal AI” is an AI that learns by combining various types of input information (images, voice, text, etc.) rather than having the AI recognize each of them individually.
Humans perceive and act on various information in a multifaceted manner.
When considering how to bring AI closer to humans, it is necessary to combine and process such multifaceted information and create an AI model for the recognition function.
There is a lot of research going on right now about how best to handle multiple types of information.
The results of research on “multimodal AI” provide hints for “general- purpose artificial intelligence .” In this way, we are steadily moving toward the development of general artificial intelligence.
A point of contention in considering the realization of AGI is the “technological singularity ”. Singularity is a concept that indicates a turning point where AI surpasses human intelligence and a change in the world brought about by AI.
According to Ray Kurzweil , an authority on AI research, it is now said that the singularity will arrive in 2045.
Introduction of AI research paper search site
Finally, I would like to introduce three AI article search sites for researchers.
| aiXiv : A site where you can search for related papers by keyword search. You can download the pdf file for free.Google Scholar : A thesis search engine operated by Google. You can also search by keyword using the search window.CiNii : A database that allows you to search all kinds of academic information, such as articles, books, and journals. |
Summary
This time, we introduced the field of AI research, laboratories and companies in Japan and overseas, and convenient tools for research.
We hope that it will be of some help to those who are interested in AI research and those who want to work on AI research.
