Hidefumi Kobatake
President, Tokyo University of Agriculture and Technology
Intelligence and Modeling
Intelligence is the ability to think, understand, and judge. In Chinese characters, intelligence "
" is depicted by two parts: an arrow "
" and mouth "
", which together mean "to grasp the essence of things" just like an arrow shooting the bull's eye. The English word "intelligence" comes from "intelligentus" in Latin, which derives from the verb "intellego" consisting of inter (among) and lego (understand and choose). This means "the ability to choose an appropriate one from among various options," and is exactly equivalent to "" meaning to grasp the essence of things. This similar origin of "intelligence" in English and Chinese is surprising and inspiring, given that both languages developed independently through entirely different language systems.
To make an object intelligent, be it a system or equipment, means to give it a mechanism to respond and react as if it had intelligence or wisdom. The object requiring this intelligence is not completely isolated from the external environment. It always interacts with the external environment. In other words, to make an object intelligent is to provide it with functions to act with appropriate responses or reactions for each occasion through this interaction.
Recently, many medical image processing systems are being made intelligent. Twelve years have passed since the computer-aided diagnosis system (CAD system) using medical images was introduced and practically applied. At first, the system was used to analyze mammograms taken during medical checkups, detect breast cancer, and prevent radiologists from overlooking any lesions. Such CAD systems are now widely used to detect cancers in the lung, liver, and colon, or diffuse pneumonic diseases. In addition, image processing technologies are also evolving from the usual two-dimensional radiograms to three-dimensional X-ray CT or MR imaging. Although the performance of CAD systems has been steadily improving, their ability to detect cancers is still inferior to that of human specialists. When the systems were first introduced, doctors felt as though they were intelligent, but since then, this favorable impression has faded. The progress of their performance is slowing down and they come with various problems. One of the reasons is that people work hard only to improve the ability to detect lesions without understanding the structure of organs before diagnosis. Doctors carefully observe the images, analyze not only the areas of interest of target organs of each patient but also their internal structure, and detect any abnormalities compared to their normal structure. On the other hand, current CAD systems spend more time detecting abnormal parts than understanding organ structures intrinsic to patients. Some normal structures show pseudo-abnormalities, and so the systems tend to misinterpret them as abnormalities. To overcome this problem, the necessity "to first understand the body structure intrinsic to the patient based on images" is increasingly being realized.
In the area of the computer vision, the importance of models describing targets has long been recognized. Such models are important in medical image processing as well. However, it took a long time to introduce the idea of modeling organs using mathematical statistics because organs have complex contours which are difficult to depict with straight lines and planes and they are too diverse to deal with mathematically. As for the trunk (chest and abdomen), a systematic approach has been attempted in Japan since 2009 in a major five-year project "Computational Anatomy" Note 1) subsidized by the Grant-in-Aid for Scientific Research from the Ministry of Education, Culture, Sports, Science and Technology. It attempts to construct a model describing the interior of the human body using mathematical statistics, understand the organ structure intrinsic to the patient using the images based on it, and provide truly intelligent diagnostic and computer-aided surgery.
Many drivers do not always follow the directions given by their car navigation system, including myself. I ignore instructions especially when I am driving through favorite narrow back streets with no traffic lights, off the main route. Drivers use these routes because there are some benefits. Whenever I use such streets, I feel bad to my car navigation system for ignoring its instructions and hope that it would eventually learn the routes I like to take. I have thus been thinking of a learning function that can be added to car navigation systems, and found the exact function in the car navigation system of a car I bought recently. On discovering it, I felt the new system had some intelligence and was satisfied with the progress of technology. The shortest path is uniquely decided on a roadmap. However, the actual traveling time depends on the traffic condition, which affects the selection of the optimal route. The ideal car navigation system should be able to adapt to the driver's values and preferences as well. The precondition for realizing true intelligent navigation is to precisely provide information on the current situation including the traffic conditions of possible routes and the driver's preferences. It should take quite a while for such an intelligent navigation system to be released because the social infrastructure required for grasping the situation of vast two– dimensional road networks is incomplete, and there still exists numerous technical problems in determining driver preferences. However, technologies for realizing this system are steadily advancing.
Modeling technologies, which are necessary for managing factories efficiently, have advanced greatly. According to the principle of feedback control, when simplified by ignoring conditions such as maintaining stability, system outputs can be maintained close to target values at any time by merely ensuring sufficiently large gain of controllers irrespective of the dynamic characteristics of the controlled objects. However, a model that reflects the dynamic characteristics of controlled objects is crucial to satisfy more elaborate control than simple feedback. Furthermore, optimal management of the entire factory is required nowadays, including not only optimal management of local subsystems but also interaction among those many subsystems. At present, the control and management of the factory should be considered based on the model that can properly describe the factory as a whole. It is necessary to appropriately grasp the situation of the entire factory just as a mechanism is needed to accurately understand the objects handled by CAD systems and car navigation systems.
Previously, sensing targets were simple physical values such as velocity, flow rate, temperature, chemical values and so on. In recent years, indicators such as safety, maintainability of instruments or systems and energy-saving efficiency, which explain the situation of systems more comprehensively compared with conventional measurements, are regarded as more important. However, it is difficult to measure them directly, as they are somewhat elusive and can be obtained only by combining several data. Models need to be able to describe the interaction among subsystems in the factory and those compound indicators. In addition, there is also a need to grasp the current state of the entire factory based on such models, and management and control of individual local systems aiming at optimal management are required. For this reason, the growing demands for field instruments and analyzers which can obtain intelligence and be networked using fieldbuses are only natural. If information collected through networks is successfully integrated and used to grasp the current state of the entire factory, then this will enable proper actions towards the desired situation. This is a condition for realizing intelligence and optimal management in a broad sense.
Factories that are intelligent and more sophisticated will increasingly become important. Modeling technology serves as the basis for achieving such factories, and commitment to continuous research and development in this area are essential because its scope and depth are boundless requiring extremely high technology. Yokogawa Electric Corporation is a global company with advanced instrumentation and control technologies including intelligent field instruments and analyzers as described in this special issue. I expect Yokogawa will continue to develop and lead the field.