Aiming at developing AI that is trusted by people and progresses society with over 30 years of know-how and cutting-edge technology

In the last few years, development of AI using deep learning has exploded, as distinguished from conventional machine learning technologies, as a driving force. In addition, services that use advanced AI and that are applied in business have increased sharply. However, various problems with applying AI to the real world and business have become apparent, including in terms of performance. The Japanese government has proposed stipulating the seven international rules for the use of AI, which include accountability, fairness, and transparency.

Fujitsu Laboratories Ltd. has worked on research and development of AI since the 1980s and has just established the Artificial Intelligence Laboratory as a dedicated R&D organization to accelerate research and development of fully-fledged AI technologies. While applying more than 30 years of AI related expertise and know-how, more than 100 researchers work on the research and development of state-of-the-art technologies with the aim of solving the existing problems of AI. The most important theme set by the Artificial Intelligence Laboratory is to realize “AI that is trusted by people and progresses society”.

AI which uses deep learning technology to learn has the advantage that it can automatically extract features of data for better prediction and obtain knowledge by recognizing and classifying data by itself. However, it is difficult to explain how AI gets answers it obtains autonomously by learning a huge amount of data. For this reason, an expert must inspect the answer obtained by AI.

This kind of AI is called “black-box AI,” and it presents the problem that it is difficult to apply it to mission critical business fields for which reliability and convincingness are required. To solve this problem, the Artificial Intelligence Laboratory which was established within Fujitsu Laboratories Ltd. is working on research and development of “Explainable AI,” that can earn people’s trust. As a result of the research, the laboratory developed industry first technologies called “Deep Tensor” and “Wide Learning” for which it is possible to explain the process by which AI get answers. Combining these technologies with “Knowledge Graph” enables more detailed explanations using peripheral knowledge which is not included in training data to be provided.

“Deep Learning is very good at learning fixed-length and fixed-structure data such as images. Deep Learning is not very good at learning and accurately classifying data for which the length or the structure is not fixed or which has a significantly large amount of variables as in the case of graph data (*note). While some special deep learning technologies to deal with such data have been developed, they were typical “black box AI,” and it was difficult to explain how forecasts were obtained. In the light of this fact, we have developed a new technology to convert such data into a unified expression called a “tensor” and extract important features. This unified expression simplifies data, enabling inference factors that can become reasons for forecasting to be presented. A new technology for optimizing unified expression by extending back propagation of deep learning is also applied to reduce classification errors. This is Deep Tensor.” (Maruhashi)

*Note: Graph as referred to here is a graph in graph theory mathematical field. Such a graph consists of vertices and edges and abstractly expresses the relation between, for instance, people, objects, and things.

The Artificial Intelligence Laboratory has been working on research and development of Deep Tensor since 2015, and it is said that evaluation experiments have already started for its application to various fields.

The security field including network intrusion detection is one of the fields to which Deep Tensor is expected to be applied. Whether an attack is included in time-series data of a vast amount of logs is forecasted based on the relationship between IP addresses and ports. It has been proven that Deep Tensor is effective in attendance management and health management at general companies, judgment of financing risks by financial institutes, and development of new drugs in the medical industry.” (Maruhashi)

While Deep Tensor is technology that presents prediction results and inference factors, which are the reasons for the results, Knowledge Graph is technology that finds the basis for supporting the prediction results. According to Katsuhito Nakazawa, Research Manager of the Knowledge Technology Project, Knowledge Graph is a vast knowledge base in which meaning is assigned to various knowledge data (documents, research papers, open data, and web information, etc.) and combined and configured in graph format.

Knowledge Graph basis finding technology associates inference factors that affect prediction by Deep Tensor with prediction results and shows the basis by which Deep Tensor gets the prediction results. With this system, when you trace the association of a certain inference factor on a knowledge graph, you can find a series of basis that extend across multiple knowledge data.” (Nakazawa)

It is said that cases based on the combination of Deep Tensor and Knowledge Graph were tested considering the efficiency of inspection by experts in the genome medical field. As a result of this testing, it was verified that the basis for supporting events of which the relationship is known only partly can be found and associated by utilizing training data using public databases in the bioinformatics field and medical document databases and Knowledge Graph.

In other words, explainable AI is AI that is capable of explaining its results. Explainable AI enables people to understand the process with which AI get results and correct the process if it is wrong. With such AI, people come to trust new findings and insights from AI.

Furthermore, another new technology called Wide Learning was developed through the advancement of AI that differs from Deep Learning. In recent times, implementations of AI in society have advanced, and the fields to which AI is applied have migrated and extended from image/sound recognition to decision support in business. Deep Learning is very good at image/sound recognition, and since creation and collection of data necessary for learning is easy, a large volume of training data can be used to enhance accuracy. Meanwhile, in decision support in business, it is difficult to prepare sufficient and balanced data necessary for learning. Ensuring there is sufficient quality and quantity of training data is a critical issue. In addition, there are many cases where decisions made by AI directly affect management, and therefore it is essential to ensure transparency in the decisions made by AI. Wide Learning has been developed to solve these issues.

“Wide Learning combines all data items and extracts “all” hypotheses that can be valid as I/O relationships from huge combined patterns. Each extracted hypotheses is important knowledge for business and is called a “knowledge chunk.” Decisions made using a large number of obtained knowledge chunks are more accurate than those using conventional AI even when the data is imbalanced. Even at locations where there is a small amount of target data for decisions, such as purchasing of new products and machine failures, it is expected that using AI will create new value.

In a test of Wide Learning using marketing data, the accuracy was reportedly improved by 10 to 20 % compared with Deep Learning.

“Knowledge chunks, which are the essential elements of Wide Learning, have a logical expression form to explain the reasons for decisions made by AI. Not only can reasons be explained, but also new findings in business can be also expected by analyzing knowledge chunks. In fact, there are cases where Wide Learning found knowledge which even experts had not noticed and led to decision support in business.” (Ohori)

In this age where all things are connected, efforts to realize Society 5.0*1have started at the national level. In a Society 5.0 world in which various information is connected, AI is essential technology for enhancing human creativity and realizing solutions to social issues across various fields by generating knowledge. You can say that explainable AI realized by a combination of Deep Tensor, Wide Leaning, and Knowledge Graph developed by the Artificial intelligence Laboratory at Fujitsu Laboratories Ltd. is essential for expanding the scope of application of AI. It is expected that these technologies will be implemented in society as technologies to support “FUJITSU Human Centric AI Zinrai,” Fujitsu’s AI technology, and make contribution to various fields in the world. Fujitsu Laboratories will keep making efforts to develop and evolve the advanced AI technologies demanded by society.

^*1Source: Cabinet Office “Society 5.0”