Fujitsu Wide Learning

2018: Announcement of technology

On October 4, 2018 the Fujitsu Laboratories company announced development of Wide Learning, machine learning technology capable to make exact decisions even in the conditions of lack of necessary amount of data.

The Artificial Intelligence (AI) is used in different areas, but the accuracy of AI can be rather low when the volume of the analyzed data is insufficient or unbalanced. The Wide Learning technology, according to representatives of Fujitsu company, allows to make more exact decisions in comparison with the developments used earlier, and learning process becomes more uniform even then when the analyzed data are not balanced. Good results of work are achieved because the technology takes hypotheses with the high level of importance, collecting a big set of the hypotheses formed by all member combinations of data and then controls extent of influence of each hypothesis on the basis of the blocked close hypotheses. Besides, since hypotheses register in the form of logical expressions, specialists can also understand the reason of this or that solution.

The Wide Learning technology, according to the developer, allows to use AI even in such areas as medicine and marketing when this, necessary for decision-making, are absent in the necessary volume.

As noted in Fujitsu, there is one problem which interferes with development of potential of AI technology. It consists in imbalance of data. Depending on specific industry, there can be a situation when it is difficult to receive necessary amount of data for training of AI in the purposes on the basis of which he will make the decision. It, in turn, leads to the fact that many technologies cannot report results of data processing with the high level of accuracy for the subsequent practical use. Moreover, the basic reason on which development of AI restrains is that even if it provides rather exact recognition, experts and even developers cannot explain why AI provided this or that answer. AI technologies based on deep training, as a rule, do high-precision solutions due to training at a basis of large volume of data. However in actual practice there is a set of examples when necessary amount of data is absent. In such cases of AI technology it is difficult to make the exact solution. Moreover, the model of machine learning for the existing AI based on deep training represents model of a black box which cannot explain the reason of solutions AI that it creates a problem with transparency of solutions.

Considering these requirements, Fujitsu Laboratories developed Wide Learning, machine learning technology capable to do exact solutions even in the conditions of imbalance of data.

Fujitsu noted two main advantages of Wide Learning technology.

1. She creates member combinations of the hypotheses given for extraction of large volumes. This technology reviews all examples of member combinations of data as hypotheses and then analyzes the level of importance of each hypothesis on the basis of coefficient of hits. For example, in the analysis of trends who buys certain products a system integrates all types of examples of data members for those who made or did not make a purchase, for example, the feme sole aged from 20 up to 34 years with driver's licenses, and then analyzes how many she is available hits in data of those who made purchases when these examples of combinations are followed as hypotheses. Hypotheses which have coefficient of hits above a certain level are considered as important hypotheses and receive the name "knowledge array". It means that even if amount of data is insufficient, a system can take all hypotheses which deserve attention that can promote opening of earlier not considered explanations.
2. It regulates the level of influence of arrays of knowledge for creation of exact model of classification. A system creates classification model on the basis of several taken arrays of knowledge. During this process if array cells of knowledge are often blocked with the elements creating other arrays of knowledge, a system controls influence level to reduce their influence on classification model. Thus, a system can train the model capable to execute accurate classifications even if the data mentioned as correct are not balanced. For example, in case the man who did not make purchase creates the majority of a data set of purchases if AI is trained without control of level of influence then the array of knowledge which includes existence or lack of driver's licenses irrespective of a floor, will not have a great influence on classification. Using the provided method the level of influence of arrays of knowledge, including a male as a factor, is limited because of overlapping of this element whereas influence of smaller quantity of arrays of knowledge which include existence of driver's licenses becomes more when training, creating model which can correctly categorize both men, and existence of driver's licenses.

Fujitsu Laboratories tested this technology in such areas as e-marketing and medicine. As a result of carrying out testing using the reference data used in the field of marketing and medicine which were provided by Repository UCI (UCI Machine Learning Repository) this technology increased accuracy by 10-20% in comparison with technology of deep training. It successfully reduced the probability that a system will pass buyers who with a high share of probability can sign on service, or patients with medical indications approximately for 20-50%. In marketing data (in tests about 5,000 records of buyers were used) only about 230 records concerned the buying customers that did this data set unbalanced. This technology reduced the number of the potential buyers who are not included in marketing actions with 120, an analysis result using technology of deep training, up to 74.

As arrays of knowledge which form a basis of this technology have a format with logical expression, capability to explain the reason of adoption of this or that solution also is useful to specialists. Even if it is defined that it is necessary to make changes based on new data to model, there is an opportunity to make more corresponding changes since specialists can understand the reason of these results, the developer emphasized.

According to the company, Fujitsu Laboratories will continue to use this technology for processing of tasks which require the indication of the reasons for the decisions made by systems based on AI including financial transactions and medical diagnoses, and the tasks connected with seldom appearing phenomena including fraud and failure of the equipment. The company set for itself the task to begin commercial use of the provided development as machine learning technology with support of the Fujitsu Human Centric AI Zinrai project in 2019 financial year.

Robotics