AI Action Plan Formulation Committee Report | Natural Language Understanding, Manufacturing, Simulation!

What is the AI ​​Action Plan Committee?

Currently, artificial intelligence (AI) technology development is actively invested in research and development worldwide, mainly in the United States and China, and each country continues to evolve to obtain cutting-edge technological capabilities.

Japan also needs to build a system for AI technology development that can respond to these trends of the times.

It is also undeniable that Japan is lagging behind the United States and China in the utilization of AI technology through big data.

Therefore, from January 2021, NEDO (National Research and Development Agency New Energy and Industrial Technology Development Organization) will conduct a “survey for formulating a comprehensive research and development action plan and extracting business in the field of artificial intelligence (AI) technology ” . It was started.

The AI ​​Action Plan Formulation Committee , which promotes this survey, has been held since February.

The AI ​​Action Plan Formulation Committee will form a committee of experts and consider a clear action plan based on overseas cases and domestic and overseas institutional policies.

The committee will be held six times between February 2021 and June 2021 in the following flow.

In addition, the committee consists of the following members.

ASCII.jp and AINOW will serve as media partners of the AI ​​Action Plan Formulation Committee to strengthen the utilization of AI in Japan.

This article introduces the discussion at the 2nd AI Action Plan Formulation Committee meeting held on April 27 (survey to formulate an action plan for overall R&D in the field of artificial intelligence technology and extract business). I will continue. ASCII.jp and AINOW will serve as media partners of the AI ​​Action Plan Formulation Committee to strengthen the utilization of AI in Japan.

5th AI Action Plan Formulation Committee

In February 2021, the AI ​​Action Plan Development Committee formulated 20 future societies based on the “Survey on future predictions by science and technology conducted by governments and institutions ” conducted by the Institute of Future Engineering, a public interest incorporated foundation in February 2021. We will list up important events and proceed with discussions based on them.

AI technical keywords that are highly relevant to each social event are listed.

In the 5th AI Action Plan, based on the “AI elemental technologies to be implemented,” “technical goals,” and “social phenomena to be implemented,” which were discussed up to the 4th, a detailed action plan was developed. and made suggestions.

In this article, I would like to introduce three topics that were particularly important during the discussion.

About natural language understanding

First, the committee members explained the importance of natural language understanding related to various themes such as well-being, manufacturing, daily life, safety and security.

Even before Google’s Japanese translation and Amazon’s Alexa came out, I thought that the field of natural language understanding was important as a strategy for Japan. However, since Japan has not developed the technology to understand Japanese, it is not possible to obtain natural language data of Japanese at present. If we don’t do that at least from now on, we will be in a terrible situation where the Japanese language will no longer belong to Japan. For that reason, I think it is necessary to have the attitude of “doing anything, even if it’s too late, anyway”.

There was an opinion.

The discussion of natural language understanding also touched on embodied and linguistic ‘double-decker brains’.

The two-story brain is thought of by dividing the functions of the brain into the first floor, which controls movement and perception, and the second floor, which controls language and logic. Several ways of dividing it, such as System 2, have also been proposed.

System 1 is actually up to the point of hearing and understanding language. System 2 is to think about various things after that.

However, when considering the realization of AI that understands meaning, all committee members agreed that the key point would be how to connect them.

In the first place, an important point for discussion is how far AI has been able to achieve this point.

Regarding the 1st floor, or system 1, there was an opinion that deep learning has achieved most of it, but that is not necessarily the case. “There are many things that can be done, but there are also (intuitive) cerebellar movements”, so it is not easy, and not all of them have been realized.

As for the 2nd floor, it is said that symbolic processing in AI is similar, but there is also research aiming at the 2nd floor with deep learning, and it can be said that this is still at the stage of exploring various methods.

About manufacturing (AI x science)

In manufacturing, the point of discussion was “Should the current manufacturing part be divided into each research area, such as organic, inorganic, and drug discovery?”

Regarding this, the committee members

I think that the specific processes differ greatly depending on the target field, but basically they have something in common in that they use AI to automatically search a large space.

However, if it is to be applied to a NEDO project, each area should be divided.

There was an opinion.

In the wide-ranging field of manufacturing, when we say, “Let’s optimize the parameters of each process,” we are currently doing the same thing as manufacturing in other domains.

On that point, the previous committee member said,

For example, elsewhere there are teachers talking about optimizing processes in agriculture and manufacturing. It is important to know what to do around here and how to dig deeper.

Also, even in the simulation story I talked about before, there is another aspect of “manufacturing” such as design.

I feel that the message is whether this approach is really appropriate, and whether I personally want to select and concentrate on something here. Also, in terms of the future, organic chemistry has made considerable progress, so it may be a good idea to cut it out in the future and make it a NEDO project.

If it is used as a general-purpose machine learning technology in various places, I think that it will work as it is, and if it is limited to organic chemistry, it will progress in various places without NEDO taking the lead.

said.

About simulation

In the simulation, the first point of discussion was whether or not the story was based on specific technology.

Regarding that, from the committee

In the case of material search, we talked about making training data in a simulator and speeding it up, and I think it was an example of Preferred Networks, making a simulator to make training data for image recognition in factories. There are many stories of I think simulators are indispensable for the future development of machine learning.

The story came out.

Also, there was a question, “If you can make a simulator, can’t you learn the rules there?”

There is one pattern that must be done with machine learning when the simulator can perform forward calculation but not backward calculation. The other is a pattern that can be calculated positively, but it takes time and “it can’t be helped”.

In addition, there are cases where the parameters to be assumed are not known even if the calculation can be performed. I know such a large strategy, but the simulator is useful in the part of how to combine it with accuracy. I think that NEDO and the Ministry of Economy, Trade and Industry should definitely do it

I answered.

Another committee member commented on the simulation

In the monodzukuri research area, I think that cross-sectoral interaction between AI and simulation is a high priority. Individually, there are fields that are progressing and fields that are not.

I don’t know if the simulation as an action plan is an optimization of AI and simulation or planning, but while making predictions with as few times as possible, design based on that, try doing it in a small space based on that, I think it’s like deduction, induction and optimization going round and round.

said the opinion.

Although there are issues with the data set, simulations seem to be organized in the form of “simulator and machine learning” or “deduction + induction” as an action plan.

summary

This time, we have introduced natural language understanding, manufacturing, and simulation from the five topics that were the points of discussion at the 5th AI Action Plan Committee.

The content of the discussion this time was to prioritize and dig deep into the social implementation visions that have come up so far, so I got the impression that the formulation of the action plan has become quite tangible.

At the next 6th meeting of the AI ​​Action Plan Formulation Committee, all the discussions that have taken place so far will be summarized, and the action plan will be approved. We plan to incorporate it into our action plan.