As reported in Part 1, AI is gaining attention around the world and was a key topic of discussion at the G20 Osaka summit held at the end of June. Though few companies are incorporating AI into their business operations, there is no doubt that in the near future the technology will be commonplace.
As the use of AI spreads, we must get better at understanding why our AI programs make specific judgments. Speaking on this subject, Dr. Setsuo Arikawa, the 22nd President of Kyushu University and a pioneer in discovery science, has said that "until now, we have been in a period in which people were satisfied merely if AI produced newsworthy results, but this period is already over. Going forward, people will not be convinced unless AI presents both its judgments and the reasons for them at the same time."
As if to prove Dr. Arikawa’s statement, many studies to render AI’s judgments explainable – a subject referred to collectively as Explainable AI or XAI – have been conducted around the world. One of the outcomes of this is a concept called Wide Learning™: a technology that is completely different from that of deep learning, which has been the driving force behind the third AI boom. The most prominent characteristic of Wide Learning™ is that its learning models reveal how judgments are made.
Wide Learning™ Tests Every Combination
Wide Learning™ Tests Every Combination Interview with project members at Fujitsu Laboratories
Deep learning enables computers to learn from an enormous amount of data and find answers autonomously. Because deep learning makes it possible to perform tedious training tasks automatically (so long as data is available), it has come into wide use around the globe. However, in deep learning, judgments appear to be made within what is known as ‘a black box’ because it is not always clear why or how the computer has come to its specific judgments.
On the other hand, Wide Learning™ is characterized by the way in which it clarifies how AI makes judgments. Since Wide Learning™ indicates how conclusions have been drawn, people can more easily understand how the computer reached its decision. Because of this characteristic, Wide Learning™ is referred to as ‘white box AI’ in contrast to ‘black box AI’.
So, how does Wide Learning™ make judgments?
Wide Learning™ requires multiple fields as input data. When given a problem to solve, the computer will generate all possible combinations of fileds emerging from these data, and then extract highly relevant ones to the problem. Considering all of the possible combinations in this way makes it easy to deduce clear reasons for the final judgments made.
For example: assume that there are 1,000 customers’ data records that indicate whether or not customers have purchased a specific product. Each customer record contains three data fields: the customer’s sex, license status, and marital status. These data fields can be combined in various ways, including two-field combinations, such as "sex and license status" and "license status and marital status," or in a three-field combination, namely "sex, license status, and marital status." Wide Learning™ extracts the combinations that support purchasing or non-purchasing after taking into account all possible combinations.
Assume, then, that there are 100 men with licenses, 60 of whom have purchased the product. In this case, the purchase rate is 60%. Let’s add the final data field and assume that 10 married men have licenses. If nine of these 10 men have purchased the product, the purchase rate is 90%.
Each of these combinations is used as a hypothesis, and an important hypothesis that attains a certain hit rate is known as a ‘knowledge chunk.’ In the example above, the combination of "male, holding a license, and married," which has a hit rate of 90%, is a more important hypothesis compared to the combination of "male and holding a license," which has a hit rate of 60%. In addition, hypotheses can also be discovered from non-purchaser data. If there are 20 single men without licenses and 18 of them have not purchased the product, this also becomes an important hypothesis regarding who doesn’t make a purchase.
As described above, Wide Learning™ examines every hypothesis and extracts all of the knowledge chunks, thereby discovering new hypotheses that people could not otherwise have imagined. This judgment process makes it easy for people to explain which combinations have a strong influence.
Figure: What is Wide Learning™?
Wide Learning™ Originated from Discovery Science
During the second AI boom, which occurred in the 1980s, the rules that governed decision making were formulated by experts and data input by them. Since these rules were based on the expertise of these consultants, they could still be explained by humans. However, rule-making was a costly, time-consuming hurdle that led to the end of the boom. Wide Learning™ aims to pick up where this boom left off, and provides an approach to AI that reveals how judgments are made. The key to applying Wide Learning™ in everyday cases has been the development of a sophisticated algorithm to rapidly compute a multitude of hypotheses.
Figure: Ultra-fast computation of an astronomical number of combinations
For example, if there are only three data fields (as in the case outlined above), the number of hypotheses is only 2+4+8. But what if there are 100 fields? For 100 fields, the number of combinations is two to the power of 100 – an unimaginably large number that far exceeds a quadrillion. Yet even such a vast number of combinations can be examined within a few seconds using an algorithm that Fujitsu has developed.
Wide Learning™ is an advanced application of the field of study known as ‘discovery science’. New discoveries have been studied by scientists since as early as the 1900s, in the pre-computer era. However, historically, it was impossible to examine every possible combination; thus, hypotheses could only be tested in somewhat limited domains.
Dr. Arikawa, an advocate of discovery science, has commented on the effects of Wide Learning™ in this context. "In the past," he says, "we had to stop examining hypotheses at some point because it was impossible to examine every hypothesis. Since some hypotheses had to be discarded, we researchers had no choice but to hesitantly let them go. Wide Learning™ eliminates such stress because it makes it possible to test every hypothesis."
Dr. Setsuo Arikawa, chairperson of the board of the Open University of Japan and professor emeritus at Kyushu University
Making AI possible, even with little data
Explainability is not the only characteristic of Wide Learning™. Another characteristic is that Wide Learning™ can make judgments fairly accurately even with small amounts of data. As described above, when there are but a few data records, so long as there are multiple data fields, Wide Learning™ can generate all possible combinations of these data fields and use them to make judgments. Thus, even if there are only dozens of data records, hypotheses can still be generated and judgments can still be made.
Since Wide Learning™ can be introduced even when there is little data, it is especially suited to early introductions of AI. Deep learning requires the collection of a massive amount of data in order to find answers autonomously, and accordingly can only be introduced after collecting vast amounts of well-formatted data. Conversely, Wide Learning™ can be implemented with small amounts of data that have already been accumulated. Then, once more data has been accumulated, more accurate testing can be performed.
How to increase sales? How white box AI helps with forward planning
Wide Learning™ can be deployed across an array of use cases. A prime example is marketing applications. Analyzing result data on whether or not customers have purchased a specific product makes it possible to predict which segment contains the most likely purchasers. For example, in the scenario outlined previously, such analysis indicated that it is important to target the "married males with a license" segment because it offers a high purchase rate.
It is also possible to go a step further and move from prediction to action. Assume that there is a knowledge chunk that consists of a "direct mail has been sent" field and a combination of the "male, holding a license, and married" fields (remember, the purchase rate of this segment is 90%). Now assume that the purchase rate of the combination of the "male, holding a license, married, and direct mail has been sent" fields reached 100% in the historical data. In this case, it’s fair to predict that the product purchase rate for this segment can be increased by proactively sending direct mail.
Figure: Image of the new technology
In September 2019, Fujitsu extended Wide Learning™ and developed a technology to automate ‘action extraction’. While a good illustrative case, the direct mailing example above is very simple. Because, in practice, there are always an enormous number of knowledge chunks, it is impractical for people to manually consider segments and actions based on each and every knowledge chunk. Fujitsu’s technology makes it possible to automatically analyze an enormous number of knowledge chunks, to detect prospective segments, and to present the optimal actions for such segments.
Beyond marketing, Wide Learning™ is also effective in healthcare and manufacturing.
Going forward, AI will also be used more often to make medical judgments. Although at this time medical workers can’t leave treatment to AI’s judgment, they can use such judgments as a reference. However, even if an AI concludes that "a patient should take drug A," that patient and their doctor won’t be able to accept the AI’s judgment without first understanding the reasons behind it.
Wide Learning™ enables drug analysis based on treatment history. For instance, if results indicate that administering drug A to patients of "male, in their 40s, and smoking" improved their symptoms dramatically, the doctor can confidently advise a patient to take the drug and help boost the patient’s understanding.
Elsewhere, in the manufacturing industry, application of Wide Learning™ is effective for identifying the causes of product defects. Generally, various manufacturing processes contribute to the creation of an individual product. Product defects inevitably arise in initial production, but the defect rate should be reduced as quickly as possible. By analyzing each process and looking for defects using Wide Learning™, it’s possible to more quickly and accurately identify defect factors. If the AI can identify that multiple defects arise when, for example, the temperature of a machine used in process A is 50℃ or more and the pressure of the pressing machine used in process B is at least 5% higher than usual, then it’s possible to alter the process to minimize defective outcomes. Wide Learning™ makes it possible to identify the cause of defects based on just a small number of samples.
Figure: "Wide Learning™" usage image
Over the course of this two-part series, we’ve explored ‘explainable AI’. We’ve learned that when AI makes judgments and presents the reasons in a more transparent, understandable way, it can cover more domains and use cases.
However, even with the ability to understand an AI’s reasoning, people will not always accept its judgments without doubt. When vital medical judgments are required, as in the example described above, or judgments on large investments are needed, it may be difficult to trust AI’s explanations and instructions. In such cases, experts will be the ones to provide explanations.
Thus, it may be better to consider AI’s judgments as something that merely supports human judgment – for example, by inviting an AI to join a three-member decision-making group as the fourth member. This kind of application of AI is likely to prove preferable, and Wide Learning™ is ideal for such casual usage.
Senior Research Officer
Nikkei BP Intelligence Group
Mr. Kimura joined Nikkei BP after graduating from the Department of Mechanical Engineering, Keio University, where he received a master’s degree in 1990. He covered and wrote articles mainly on cutting-edge processing technologies and production management systems for Nikkei Mechanical, a manufacturing industry magazine. Afterwards, he was involved in the publication of the first issue of Nikkei Digital Engineering, a magazine on utilizing computer techniques in the manufacturing industry, such as CAD/CAM and SCM. In January 2008, he assumed the position of chief editor of the Tech-On! website (the present Nikkei xTECH). In July 2012, he assumed the position of chief editor of Nikkei Business and its digital edition, Nikkei Business Digital (the present Nikkei Business Electronic Edition). He designed business-related content and led website operation and app development. In January 2019, he organized and established a project for Nikkei Business Electronic Edition. He has served in his present post since April 2019. He is in charge of content consulting services for leading enterprises and designing content for websites such as Beyond Health.