[Article 2] Ever Evolving Artificial Intelligence--AI Keeps Learning and Growing Each Day

AI Special Feature: Learning System

Human Centric AI Zinrai is the product of Fujitsu’s extensive technical expertise in the AI field. Article 2 features a learning system that enables AI to grow and develop by extracting valuable principles and patterns from vast amounts of data through everyday interactions.

Leading AI Technology Extracts Valuable Knowledge from a Variety of Data

The present age is referred to as the third AI boom. AI technologies are used in various fields such as finance, manufacturing, transportation, logistics, and cyber security, where new forms of value are created. The learning system is a core technology for artificial intelligence.

The learning system is a technology to realize the human ability to learn from experience on computers. One of the key features of AI is to be able to learn, decide, and grow every day like humans. Specifically, AI automatically finds rules, patterns, knowledge, and the like from a variety of data such as images, voices, numbers, and texts, and applies them to understand current situations, predict the future, and make decisions.
This learning system is a basic technology to support perception and recognition, knowledge processing, and decision-making and support functionality of Zinrai, Fujitsu’s AI technology, as well as to expedite continuous growth of its AI.

The learning system can be classified into “Learning with a Teacher” and “Learning without a Teacher.” “Learning with a Teacher” is to learn regularities by using data—such as correct/incorrect answers labeled in advance by humans—as a teacher. On the other hand, “Learning without a Teacher” is a system in which computers learn regularities by themselves by reading large amounts of data.

However, in most cases, rules must be found in a situation where no data serving as a teacher exist. For example, in the field of cyber security where customers always have to deal with new attacks and viruses, they need to detect attacks without teacher data. “Learning without a Teacher” is used in this field.

Let’s take a look at the application fields of the learning system by dividing it into“Learning with a Teacher” and “Learning without a Teacher.”

Fujitsu Develops Handwritten Chinese Character Recognition Technology with 96.7% Accuracy

The handwritten Chinese character recognition technology is an example of “Learning with a Teacher.” This technology learns and assimilates the defining characteristics of various character patterns, which serves as a teacher.
However, handwritten characters vary from writer to writer due to different writing habits. For AI to recognize these variations of a character as the same character, it was necessary to develop mechanisms to learn variations in character shape.
So Fujitsu developed “Handwritten Character Recognition Technology” enabling automatic generation of a wide variety of deformed character patterns for learning by using a unique deep learning technology. This allows for much more precise character feature learning, resulting in achieving the handwritten Chinese character recognition rate of 96.7%, exceeding human levels. With this, it is expected that handwritten slips can be processed more efficiently.

Fujitsu is focusing its efforts on the study of deep learning. Deep learning is the most advanced neural network technology, which simulates the mechanisms of the human brain.*1 So far the technology has only been applied to the analysis of limited data such as images and voices.
Now Fujitsu has developed deep learning technology that can classify volatile time-series data with a high degree of accuracy by extracting geometric features from time-series data. Time-series data can be subject to severe volatility, making it difficult for people to discern in the data. This technology could be used to accurately detect equipment anomalies or forecast breakdowns in factories using IoT devices or could be used to analyze vital-signs data to assist with medical diagnoses and treatment.

*1: A model simulating the mechanisms of the human brain, such as neurons (nerve cells) and synapses that connect neurons to transmit signals

Outlier Learning Technology: Extracting Unknown Cyberattacks in a Short Time

Then, in what fields is “Learning without a Teacher” being used? In the following, we will introduce a case of detecting cyberattacks. Corporate networks are exposed to new attacks and the threat of viruses every day. Also, there are many cases where even known viruses intrude into networks using clever tricks, such as targeted attacks. It has been difficult to detect these unknown attacks and tricky attacks only by human monitoring.
As conventional technologies have classified attacks based on the features frequently observed in data, it has been difficult to extract unusual small-scale attacks hidden in a large number of other attacks.

Therefore, Fujitsu has developed the outlier learning technology which extracts small-scale (outlier) data groups that share rare features by focusing attention on features that do not frequently appear in data. The challenge was the amount of calculation required to extract an appropriate size of data groups from a very large number of combinations of data sets.
This technology efficiently searches data groups that share rare features by repeatedly separating and combining data groups. By applying this technology to the detection of cyberattacks, it has become possible to extract new attacks in a short time that used to be extracted by humans spending three months.

Ever Evolving Learning System—Focus Going Forward Is Application in Business Scenes

This learning system has a broad potential of applications in various fields, such as predicting the risk of disease in the medical field.
How to analyze huge amounts of data accumulated daily and generate knowledge useful for humankind—Fujitsu will continue promoting technical development so that AI that learns, decides, and grows by itself can serve as an excellent guide for humans in the future.

*Featured Article 3 “AI Interpreting Human Emotions from Human Behaviors and Responses—Mood Media” is to be released on Tuesday, March 22.