NOTE: this is an archived page and the content is likely to be out of date.
AI Zinrai Supporting Digital Innovation
Vol. 55, No. 2, 2019
Fujitsu announced FUJITSU Human Centric AI Zinrai as its AI technology brand in November 2015 and began providing Zinrai Platform Service in April 2017. This issue introduces AI initiatives at Fujitsu from the three viewpoints of "examples leveraging AI technology," "AI solutions," and "cutting-edge technologies that support Zinrai."
New AI technologies and services based on them have developed at a fast pace in recent years, bringing major changes to businesses and our private lives. With more than 30 years of research in AI, Fujitsu addresses this major transitional era with FUJITSU Human Centric AI Zinrai, a solution that offers ICT with human perspectives at its heart. Zinrai integrates our cutting-edge technologies, such as machine learning, sensory media, knowledge technology, and mathematical techniques, to make ICT evolve in ways that are more human-friendly. This paper describes Zinrai from the viewpoints of software, services, infrastructure, and hardware. It also gives an account of the initiatives at Fujitsu for developing AI businesses.
Points of contact for customers such as reception counters and call centers tend to have high staff turnover. Therefore, they must handle customer inquiries and leverage knowledge with competence without having the expertise of experienced staff. There are two ways to respond to the knowledge utilization needs: by means of FAQ and by leveraging structured knowledge information. Fujitsu has recently had an opportunity to offer the Osaka City Government in Japan its AI technology for domain-specific semantic search function. With this technology, necessary information can be identified from a large volume of data concerning family register administration and work manuals. Then it enables staff members without extensive expertise to efficiently handle their administrative work regarding family registers. It also eliminates the need to prepare the enormous learning data necessary to leverage FAQ. This paper explains the domain-specific semantic search function and describes the case in which it was introduced at the Osaka City Government.
In the distribution industry, supply chains are formed by manufacturers, wholesalers, and retailers. Their business activities, such as product development, production management, logistics, and sales, are based on demand forecasting. Therefore, business reforms and improvements by capturing demand more accurately have become critical for companies. In recent years, however, consumer demand and products have diversified, and product life cycles have become shorter. It is thus increasingly difficult to accurately forecast future demand. Moreover, the aging population and declining birthrate make it difficult to secure sufficient labor. These situations raise expectations for advanced technology, leveraging AI, to automate and enhance demand forecasting. Fujitsu Laboratories has developed the dynamic ensemble forecasting technology, an AI-based demand forecasting. Using a model integration method, this technology flexibly handles various product characteristics and is reliable in executing highly accurate forecasting. It also lessens the operational burden by means of automatic tuning. We have also developed an attribute decomposition model for forecasting demand of brand-new products with no past results data. This paper describes the dynamic ensemble forecasting technology and the attribute decomposition model, a solution based on these technologies, and its application.
Wind power generation is a fast-growing type of renewable energy that is highly recognized for its potential. Turbine blades are a crucial part of power generation systems, and they bear substantial loads in operation. Therefore, quality inspection standards during production are extremely strict in order to meet critical to quality (CTQ) requirements. Quality inspections include ultrasonic non-destructive testing (NDT) methods, for instance, which involve an engineer's detailed examination of ultrasonic testing (UT) scan image data covering a length of approximately 75 meters to identify possibly defective parts of a few centimeters. This is a time-consuming process, and human error is always a risk. Fujitsu Laboratories of Europe Ltd. has developed a unique technology, Imagification, and combined it with an image-recognition deep learning engine to develop a system for automating defect detection, helping to reduce the inspection load and prevent human errors. We have further developed the system by digitizing and integrating customer knowledge and expertise in blade structure, ultrasonic NDT procedures and know-how in defect characterization to build a new automated quality inspection solution. This paper explains the newly developed defect detection system and describes the process of digital co-creation with customers to realize commercial applications of the technology.
With low-cost IP cameras readily available across the world, the number of monitoring cameras installed in major cities has seen a dramatic rise in recent years. Smart City Monitoring based on AI is attracting attention as a means of making efficient use of these cameras. In March 2018, Fujitsu launched its new smart city monitoring solution FUJITSU Technical Computing Solution GREENAGES Citywide Surveillance V2. Citywide Surveillance is based on deep learning technology, and in order to apply this suite to a business operation, there are challenges to address in terms of accuracy, speed, and cost as fundamental features. Furthermore, as deep learning alone cannot deliver the value customers want, it is necessary to develop auxiliary technologies at the same time. This paper explains the Fujitsu technologies that bring customers the needed accuracy, speed, and cost performance as well as our co-creation initiatives to deliver value to our customers. It also introduces examples of Citywide Surveillance applied in business contexts.
With Zinrai Platform Service, Fujitsu offers sensing and recognition, knowledge processing, and decision and support functions covering each major AI element technology, as well as various other functions and knowledge that customers can combine as suitable for their business needs, with the goal of facilitating AI use by customers. Functions such as image and voice recognition, and knowledge discovery from text-based information are provided as application programming interfaces (APIs). These can be used on their own, but their application requires some effort, such as data preparation and parameter tuning. To cope with this problem, the APIs provide functions and configurations ready-made and tuned to solve particular issues. They are designed based on knowledge and functions that allow customers to combine component technologies according to their use scenarios and facilitate easy introduction of AI to their businesses. This supports the adoption of AI by enterprises. This paper presents the major functions of Zinrai Platform Service and describes its characteristics. It also explains the main concepts behind the platform feature configuration and discusses, based on some application cases, the value of this service for businesses.
There are combinatorial optimization problems in our society-selecting the best option from combinations of various factors, such as finding the best procedures in disaster recovery efforts and optimizing investment portfolios. In combinatorial optimization problems, the number of combinations increases exponentially as the number of factors increases, which makes it extremely time-consuming for general-purpose computers to solve certain problems within a realistic time frame. Research and development in quantum computing is underway in order to solve combinatorial optimization problems quickly. However, the current state of quantum computing is limited in terms of stable operation and the size of problems it can handle. Furthermore, quantum computing requires the conversion of a combinatorial optimization problem into an Ising model to solve it. Against this background, Fujitsu launched its Digital Annealer Service in May 2018. This is a new architecture inspired by quantum computing. This paper explains the technology to employ Digital Annealer to solve customers' real combinatorial optimization problems, namely, the formulation of real problems and the conversion to quadratic unconstrained binary optimization (QUBO). It also describes the initiatives at Fujitsu to leverage Digital Annealer in creating a new global computing market.
With the Japanese government's current policy to accelerate the research and development of AI technology and its social implementation of research results, supercomputers designed for AI processing (hereafter, AI supercomputers) are rapidly being developed. Unlike conventional supercomputers, AI supercomputers have the following new requirements: system sizing based on standard performance evaluations that focus on learning accuracy and learning speed in deep-learning processes, and a software arrangement where frequent updates to deep-learning frameworks can be dealt with in a timely manner. Fujitsu and Fujitsu Laboratories have devised new performance evaluation methods and operational technologies for AI supercomputers. We combined these with the technologies for conventional supercomputers and deployed AI supercomputers to the National Institute of Advanced Industrial Science and Technology (AIST) and RIKEN Center for Advanced Intelligence Project (AIP). These systems will be a reference point in terms of future AI supercomputers. This paper first outlines AI supercomputers and their trends, then describes the approach to operating issues and the performance evaluation methods required by AI supercomputers. Finally, it describes the application of the operating technologies to AI supercomputers and their effects.
One of the most significant advancements made in AI in recent years is the greatly enhanced accuracy of machine learning through deep learning. However, because deep learning deals with huge volumes of data and involves vast neural networks in the learning process, it is often difficult to explain how or why an output was reached even if the inference was correct. This issue impedes the application of AI technology to such business areas as finance and medicine, which demand absolute reliability. To address this issue, we have developed an AI technology that combines Deep Tensor, Fujitsu's original machine-learning technology based on enhanced deep learning, and another Fujitsu-developed machine-learning technology based on knowledge graph, a knowledge base represented by graph data taken from past documents and databases. This has enabled us to logically explain the reasons and bases in which Deep Tensor reaches its inference output. This paper explains the technology that makes Explainable AI possible in terms of application cases in network intrusion detection and genomic medicine.
The commercialization of AI technology has accelerated in recent years, with a growing interest in various machine-learning technologies such as deep learning. However, machine learning is based on statistical data analysis, and it is known today that certain information contained in such data is lost through analytical processes. To make the most of such information, we have developed a new machine learning technology based on topological data analysis (TDA) that focuses on and analyses the "shapes of data." This paper explains TDA as a new data-analytical method. As applied cases of TDA, it also describes the time-series deep learning for analyzing time series data and anomaly-detection technology, with an account of a bridge deterioration assessment in which the latter was applied.