Aiming to create an AI that understands human language essentially. Approach to the development of ”AEI” and its strengths
2020.11.18
Mr. Nagata, Executive Vice President of pluszero, will talk about AEI, which pluszero is focusing on, and the future that pluszero wants to realize by using AEI.
About "AEI," the practical AI that pluszero is working on
"AEI (Artificial Elastic Intelligence) is a term we coined to describe AI that understands the meaning of essential words as well as humans.
The word "elastic" reflect the word "flexible," and it implies that it can understand the meaning of words and respond flexibly.
Our development of AEI is based on our company's vision of "Expand the human potential."
In order to expand human potential, we need to expand what AI can do. For this reason, AEI is an indispensable area for us.
The possibilities of AI are evolving day by day, and industrial applications are increasing. On the other hand, there are also limitations. If we cannot have the necessary data to solve a problem, we may not be able to fully utilize the performance of AI.
For example, in order to be able to understand the meaning of human words, it is necessary to capture the essential meanings contained in language and images.
In general, there is not enough data that expresses the intrinsic meaning, and the extension of current AI technologies, mainly deep learning, may not be able to return the expected results.
The term "AGI" is used as a contrast to AEI. This is a human equivalent performance in any task.
It will have a huge impact if it is perfected, but at present, there is no technological prospect.
On the other hand, in order for current AI to be able to understand the meaning of human words, it is necessary to spend a very large amount of money on each finely divided task.
In other words, it becomes cost-prohibitive to create an engine that understands the meaning of human words over a wide range.
Therefore, we are seeking a development process that can solve all issues at a realistic cost, and we are trying to build a general-purpose infrastructure that can handle more issues with partial customization.
That is the "AEI" we are trying to build.
Today's AI has very high performance for each individual task, but it also lacks flexibility.
Through the development of the "AEI" technology, we aim to create a more flexible AI and spark the fourth AI boom. Japanese and English systems will be developed first,and other languages would be following.
The word "elastic" reflect the word "flexible," and it implies that it can understand the meaning of words and respond flexibly.
Our development of AEI is based on our company's vision of "Expand the human potential."
In order to expand human potential, we need to expand what AI can do. For this reason, AEI is an indispensable area for us.
The possibilities of AI are evolving day by day, and industrial applications are increasing. On the other hand, there are also limitations. If we cannot have the necessary data to solve a problem, we may not be able to fully utilize the performance of AI.
For example, in order to be able to understand the meaning of human words, it is necessary to capture the essential meanings contained in language and images.
In general, there is not enough data that expresses the intrinsic meaning, and the extension of current AI technologies, mainly deep learning, may not be able to return the expected results.
The term "AGI" is used as a contrast to AEI. This is a human equivalent performance in any task.
It will have a huge impact if it is perfected, but at present, there is no technological prospect.
On the other hand, in order for current AI to be able to understand the meaning of human words, it is necessary to spend a very large amount of money on each finely divided task.
In other words, it becomes cost-prohibitive to create an engine that understands the meaning of human words over a wide range.
Therefore, we are seeking a development process that can solve all issues at a realistic cost, and we are trying to build a general-purpose infrastructure that can handle more issues with partial customization.
That is the "AEI" we are trying to build.
Today's AI has very high performance for each individual task, but it also lacks flexibility.
Through the development of the "AEI" technology, we aim to create a more flexible AI and spark the fourth AI boom. Japanese and English systems will be developed first,and other languages would be following.
The challenge is to "Make the machine understand the original meaning of a word naturally" from two approaches.
There are two main points of our approach.
The first is to convert natural language into a unique representation, and the second is to use a variety of algorithms.
What is a "unique semantic representation of natural language"?
When natural language is handled by a machine, it must first be converted into a form that the machine can handle.
However, in the first step of converting natural language, the general meaning that humans have when dealing with language is lost.
As a result, no matter what kind of algorithm is used after that, the final problem to be solved will be limited. This is the major issue of current natural language processing.
In response, we are developing a "unique semantic representation" that can be understood by machines while retaining the essential meaning of natural language.
What is the "use of various algorithms"?
Today, AI is mainly based on machine learning technologies such as deep learning.
Machine learning is very useful and has a wide range of applications, but there are issues that cannot be solved by itself.
That is why we take the approach of optimally combining machine learning and other methods according to the objectives. This is the use of a variety of algorithms.
Research in natural language processing does not produce immediate results overnight, but requires persistent efforts over a very long period of time. Therefore, it is difficult to continue research just because you are intellectually curious.
At pluszero, we have many members who can persevere in such research.
In general, it is said that it is difficult for AI to be able to read text, but we are working on it with the belief that it will eventually become possible.
The first is to convert natural language into a unique representation, and the second is to use a variety of algorithms.
What is a "unique semantic representation of natural language"?
When natural language is handled by a machine, it must first be converted into a form that the machine can handle.
However, in the first step of converting natural language, the general meaning that humans have when dealing with language is lost.
As a result, no matter what kind of algorithm is used after that, the final problem to be solved will be limited. This is the major issue of current natural language processing.
In response, we are developing a "unique semantic representation" that can be understood by machines while retaining the essential meaning of natural language.
What is the "use of various algorithms"?
Today, AI is mainly based on machine learning technologies such as deep learning.
Machine learning is very useful and has a wide range of applications, but there are issues that cannot be solved by itself.
That is why we take the approach of optimally combining machine learning and other methods according to the objectives. This is the use of a variety of algorithms.
Research in natural language processing does not produce immediate results overnight, but requires persistent efforts over a very long period of time. Therefore, it is difficult to continue research just because you are intellectually curious.
At pluszero, we have many members who can persevere in such research.
In general, it is said that it is difficult for AI to be able to read text, but we are working on it with the belief that it will eventually become possible.
Secure a pool of talent in a wide range of advanced fields
At pluszero, we have a broad pool of talent in advanced fields. We have a total of 150 full-time and part-time employees.
With this pool of talent, we are able to assemble optimal teams for projects in all areas.
In the development of natural language processing, in order to use a variety of algorithms, it is necessary to cover a cross-section of fields generally referred to as "humanities" and "science".
In the humanities, we use logic, linguistics, psychology, and philosophy. On the other hand, science fields include machine learning, computing, science, mathematics, and neuroscience.
At pluszero, members with expertise in both the humanities and the sciences bring their knowledge to each other and promote research by generating emergence of knowledge.
With this pool of talent, we are able to assemble optimal teams for projects in all areas.
In the development of natural language processing, in order to use a variety of algorithms, it is necessary to cover a cross-section of fields generally referred to as "humanities" and "science".
In the humanities, we use logic, linguistics, psychology, and philosophy. On the other hand, science fields include machine learning, computing, science, mathematics, and neuroscience.
At pluszero, members with expertise in both the humanities and the sciences bring their knowledge to each other and promote research by generating emergence of knowledge.
Profile
Executive Vice President
PhD (Information Science and Technology)
Motoki Nagata
Graduated from the Department of Mathematical Engineering and Information Physics, Faculty of Engineering, the University of Tokyo, and completed the doctoral course in Mathematical Informatics, Graduate School of Information Science and Technology, the University of Tokyo.
D. in Mathematical Informatics from the Graduate School of Information Science and Technology, the University of Tokyo. He specializes in business development that combines mathematics and technology, such as project manager for system development and supervising the operation of a mathematical solution database.
In graduate school, he conducted stability analysis of electric power systems and financial systems as a JSPS Research Fellow (DC2).
He has led the theoretical development of projects such as the Value Appraisal Project, algorithms for Real Time Bidding, and logic for pricing and inventory strategies.
In order to realize a society in which all people do not have to worry about food, clothing, and shelter, and can focus on human activities free from simple labor, he is trying to tackle the fields of education, food, energy, medical care, and nursing care.
PhD (Information Science and Technology)
Motoki Nagata
Graduated from the Department of Mathematical Engineering and Information Physics, Faculty of Engineering, the University of Tokyo, and completed the doctoral course in Mathematical Informatics, Graduate School of Information Science and Technology, the University of Tokyo.
D. in Mathematical Informatics from the Graduate School of Information Science and Technology, the University of Tokyo. He specializes in business development that combines mathematics and technology, such as project manager for system development and supervising the operation of a mathematical solution database.
In graduate school, he conducted stability analysis of electric power systems and financial systems as a JSPS Research Fellow (DC2).
He has led the theoretical development of projects such as the Value Appraisal Project, algorithms for Real Time Bidding, and logic for pricing and inventory strategies.
In order to realize a society in which all people do not have to worry about food, clothing, and shelter, and can focus on human activities free from simple labor, he is trying to tackle the fields of education, food, energy, medical care, and nursing care.
