Whenever AI (artificial intelligence) is mentioned, the topic of deep learning also often enters the conversation. Although it's one of those keywords in the realm of technology, it may be difficult for many to imagine how it relates to practical solutions to the problems facing society. Here, we introduce the foundational knowledge upon which deep learning is built, and a case study involving Kawasaki Geological Engineering.
Fujitsu Insight 2017 Special Presentation Report on the Cutting Edge of AI and IoT Application
What is Deep Learning?
The Difference Between Machine Learning and Deep Learning
There are two methods in AI learning: machine learning and deep learning.
In machine learning, humans set algorithms and make machines learn them. For example, to make a machine recognize images of humans in machine learning, human features such as details of the head or the position of the eyes are extracted and designed, and the machines decide whether or not it is human. In contrast, in deep learning, the machine automatically performs the extraction and design, and makes the judgement on its own.
This is the major difference between machine learning, in which humans define the features, and deep learning, in which the machine extracts those features on its own.
What is the Biggest Advantage of Deep Learning?
The major advantage of deep learning is processing speed. Let's compare using this example of a system that was built utilizing image analysis technology for application in wildlife protection measures for wild boar.
With conventional image processing technology, it took from six months to a full year to make the machine learn to decide if an animal that entered the cage was a wild boar or not. In comparison, with deep learning, the machine achieved the same result in just three days of learning. By utilizing deep learning, the manpower and hours it takes for conventional machine learning to achieve its objective can be greatly reduced. This is the power of deep learning.
Deep Learning Case Study―Using AI to Solve the Problem of Aging Infrastructure
Labor Saving in Massive Image Data Processing
Kawasaki Geological Engineering is implementing Fujitsu's AI system, which uses deep learning, called Zinrai Deep Learning.
There are about 3,300 annual cases of roads collapsing in Japan. The problem is being dubbed as "the aging of infrastructure." Today, in order to prevent incidents like the road collapse that occurred in front of Hakata Station in November 2016, we are focusing our energy on developing surveying technology that will visualize cavities under road surfaces using echolocation technology.
During surveys, a car equipped with probing radar that sends sound waves underground is driven across the road and continuously captures the echoes that bounce back. Up until now, multiple engineers specializing in processing this image data performed cross-checking to decipher them.
The image data is massive, and it takes over 16 hours for one person to decipher it with just human power. Furthermore, as it usually takes about five to six people for cross-checking, the total time this task requires is about 100 hours. This is an exhausting task for engineers. This is where, with Fujitsu's help, we decided to lessen the workload and shorten the hours by developing an AI system that can automatically decipher image data.
Cutting the Analysis Time of Image Data to One-Tenth
The cavity surveys we conducted using Fujitsu's Zinrai Deep Learning yielded surprising results. After having engineers and the AI system survey the same road, the abnormal responses detected by the AI that indicated the locations of cavities matched with the locations that the engineers pointed out. Furthermore, among the abnormal responses detected by the AI, some were images that were determined to not be cavities by the engineers.
As a matter of fact, the process of locating cavities using images by humans is not 100 percent accurate. With AI, many of the elusive abnormal responses can be detected from various perspectives. The AI detects abnormal responses that can't be located by human observation alone, and by having human eyes review that data to determine if they indicate cavities, a much more precise method of surveying is made possible.
By using Zinrai Deep Learning, the time it took for image analysis has been cut to one-tenth. By reducing the amount of work done by humans, operational costs can be cut, and more extensive surveys and periodic surveys can be carried out.
In addition, by applying the results that were gained to further advance AI learning, this technology can evolve into an even more reliable cavity detection system.
Fujitsu is continuing to expand the new possibilities created by AI, with the following three leading-edge technologies: deep learning, supercomputers and quantum computing technology. As we continue to consider new applications for AI to confront the various problems facing companies today, we aim to provide solutions for some of society's fundamental challenges.
- Shigeharu Yamada
Director at Kawasaki Geological Engineering Co., Ltd.
General Manager, Maintenance Division, Tokyo Metropolitan Headquarters
Section Chief, Research and Development Office, Research Group
- Hirofumi Nagai
Head of AI Frontier Division
AI Platform Business Unit