The first step to utilize AI (artificial intelligence)

There are more opportunities to hear and see phrases such as AI (artificial intelligence) and Deep Learning in the world.
However, it is also a fact that it is difficult to know exact knowledge such as what is AI.
Therefore, based on the story of the "Deep Learning" project that DeepMind Co., Ltd., an engineer, Mr. Wakasaka, will give a story of "AI" which can have the power to move the next generation society.

What is AI (artificial intelligence)?
What is Deep Learning ?


SMEs with advanced technologies are also experiencing problems of lack of human resources such as lack of successors and productivity, but if they can make good use of AI and machines can do that part, probably a clue to solve the problem is born.

AI tends to be called Artificial Intelligence, but there are many other definitions besides that. However, in the world, I often see and hear the word "AI". Does it mean that technology has advanced at least for practical use in the business situation?

First of all, please tell me what kind of initiatives are being done with DeepMind using AI.


We are focusing on factory automation※1 using AI. What we are developing is "Deep learning" which is one of the types of machine learning in" AI "technology."

Machine learning is a mechanism for artificial intelligence programs to learn by themselves. In order for a machine to learn, it is necessary to judge data "Yes" or "No" because computers can only distinguish between 0 and 1. This discrimination is referred to as "the division method", and machine learning automatically acquires this "division method" while the computer processes a significant amount of data.

There are various methods in this "division" process. The method used for deep learning is a method called "Neural networks".

"Neural network" is a technique hinted at by the cerebral nerve circuit of a human being and expresses the connection of the brain neuron ※2 in a pseudo program by the program. Learning by learning the ties in many layers is called "deep learning".

Deep learning makes it possible to handle complex data such as image data, voice data, natural language data and so on. Various information is contained in these data sets. Until now all human beings thought out "meaningful" features from the information. If the human beings were able to deliver the data "characteristics" properly, the machine worked well. Otherwise it did not work.

However, deep learning extracts "features" by the computer itself based on a significant amount of data. By doing so, they can acquire proper knowledge by themselves, according to the circumstances and purposes, and be able to use it.

※1 A system that aims at automating production processes in factories. It is abbreviated FA (f · a)
※2 "Neurons" constituting the brain. Nerve cells are cells that transmit and receive information by emitting electric signals.

TrendCompass talk image 01

Possibility of deep learning
= What is the fusion of people and machines?


what is the technology part called factory automation that you mentioned earlier, and what kind of usage is being done?


In factory automation, deep learning is used in technology to detect defective products such as anomaly detection.

For example, we use data obtained from sensors and images to find inferior goods when we mass-produce screws, but we use deep learning to find abnormal parts with high accuracy from data accumulated enormously Yes.

To create a learning model in which the AI itself searches for the optimal instruction, it is necessary to have a server that can process a significant amount of data, such as the time and power to process a massive amount of data and GPU (image processing unit). For that reason, we are accelerating the computation process to deep learning, but reducing the running cost of power to use for a long time to realize power saving software development and small proprietary hardware that can effectively advance the network model we are developing.

Complex and significant amounts of computation is required by making deep learning compact and straightforward, it can be used for various purposes.


Until now, what is required is a large-scale facility such as a supercomputer, it will become smaller, and will be able to incorporate "AI" everywhere. Why are you advancing technologies for energy saving and speeding up?


We aim to "smooth the boundary between people and machines and make the world better". There is a machine as a role to expand human ability, but we would like to be able to use it further near the skin feeling.

For example, we are still using smartphones without much consciousness at the moment, but we aim to make it more sensually possible to use the expanded capabilities of machines than users are aware of it.
The image is a world ※3 like the animated "Ghost in the Shell". To reproduce such a futuristic world with reality, it is necessary to count servers with sophisticated GPUs. However, our goal will be farther with large equipment. To reach that area, software that realizes speeding up and power saving and light and compact hardware are indispensable.

※3 A world where science and technology has advanced dramatically, and many people can directly access the Internet by the cerebrum


Isit that "AI" exists as a function of connecting humans and machines?


That's right. To make it "AI" as a connecting role, the high recognition performance and classification performance of "Deep learning" become important. It is because we can extract "characteristics" of data, which could not be understood by conventional machine learning. Looking at the image, extracting "features", distinguishing it is equivalent to human vision. If that precision goes up further, you can substitute machine parts that were dependent on sight and hearing. By combining sensors also, it will be possible to let AI learn to smell and taste.

If a lot of devices using deep learning are created, and they move autonomously, the future that can be used without being conscious of the invention will be more realistic.


What will happen if human senses are interchanged and the boundary between people and machine becomes liquid?


For example, in the manufacturing industry, to raise the quality and precision, AI itself will be judged autonomously in various ways. Then, the time, which we used, for quality control until now can be drastically reduced.


If AI comes in for product inspection, product accuracy will grow dramatically.
SMEs may be interested in seeing the introduction of AI for the first time from parts like quality control. Will it be possible to use machines as human hands? Can you brush up something, for example?


If there is a "model", I think it is possible. The efficiency of learning is different if there is a "model". Although there is a way to learn even if there is nothing, it is not efficient because the accuracy is still low.


Furthermore, if the technology of AI advances, human beings will be trying hard to be those who make models as models. After modeling, it is one of the ideals that we can leave it to machines and computers.

TrendCompass talk image 02

The first step to utilize AI


What should we start with when SMEs think to utilize AI from now on?


What is a" role model "of each company? You can use it for AI by conscious of it and making it data. Data that becomes "role model" is data with "meaning" to understand the good or bad of products and services.

For example, if you take the "defective item" of the screw as a photograph, you can let AI learn the defective item. If it is only one, it is difficult to find a different "defective item," so please leave image data of various "defective items." It will be able to find many defective products in the future as much as 100 pieces, 1000 pieces, and incomplete item data gather. By classifying the gathered incomplete product data and the data of a good screw passed and let AI learn, it will be possible to produce highly accurate products stably.


SMEs can introduce AI and deep learning so that we can extract and learn a lot of "role model" data. If you can make good use of the results, you may be able to bring in innovation.

But if we had a "role model," would it mean that artificial intelligence never exceeds humans? "


The current deep learning has the drawback of being unable to do logical things, rather than thinking things in stages, it is a feature of this technology to distinguish it sensually.

For example, let's say there was a sign with "Kenya" and a picture with a yellow animal. In human beings, looking at the land pattern of Kenya and the color and shape of the animal, this yellow animal will easily answer "lion." But deep learning makes it difficult to associate the features in the image step by step, so it is difficult to answer "lion" with just one image.

On the contrary, deep learning is good at sensuously judging which one is the "lion" from among the many image data.

The technique to logically correlate in stages is studied in another "machine learning." There is a possibility that artificial intelligence that thinks logically thinking and intuitively by multiplying with other technologies can be born, but thinking beyond humans is not realistic at this stage.

TrendCompass talk image 03

The future of deep learning


In the future, how will deep learning be utilized?


We are developing deep learning which allows the machine to autonomously operate in space, deep sea, volcanos etc where human activity is restricted. Although it is a story of a personal dream, even AI's first time to see the moon autonomously scraps the rock, makes a lot of ground, builds a building, etc.


It means that you can do a lot of things and understand just by the machine does. The story that AI autonomously builds a building on February also feels the possibility. Furthermore, if we combine the technologies of SMEs, the opportunities will expand more. In the future, I would like to expect the dream of "a convenient world where the boundary between people and machines became smooth."

TrendCompass talk image 04


A problematic little story has also appeared, but it seems that it can not be said that you understand AI in a real sense directly regarding artificial intelligence, and deep learning by AI that they are aiming (LeapMind) It seems to bring high potential to the company's production site.
Mr. Wakisaka says "I would like people in small and medium-sized enterprises to come to consult," I think how to make use of AI and how to adopt it will be one key to the innovation of SMEs.

Takuya Wakisaka

Takuya Wakisaka Vice president of engineering

The full-stack engineer who provides technical support such as interactive projection, mainly on designing, building and developing web services · applications as a freelance to retirement after a part-time job at SIer (System Integrator) from Yokohama National University studying electronic information department.


LeapMind is a Deep Learning special startup.
LeapMind accelerates DoT (Deep Learning of Things) applied to everything, using technology "Deep Learning" representing the 21st century as important as "Internet." We are developing a Deep Learning solution "JUIZ" which can affect Deep Learning even in scenes where GPU and cloud computing cannot be relied upon. Toward its development, LeapMind conducts research and development in both software and hardware areas of Deep Learning.

Yoshiaki Nagasaki

Yoshiaki Nagasaki Navigator of New Value Creation NAVI / CEO of paragraph

He is an editor in charge of editorial planning such as "Begin" and "Men's EX", editor in charge when re-issuing "WIRED Japan version". Also experienced branding direction such as various fashion brands. Currently, he does numerous companies including WEB direction such as UNIQLOCK and Nikon, and he is doing various creative activities such as efforts by the editor of WEB magazine "TOKYOWISE" as well.

Related Company