Developing an integrated
computing environment
Computing Japan visits the Motoda Research Group, Advanced Research Laboratory, Hitachi Ltd.
Through a combination of research in artificial intelligence, probability theory, linguistics, and computer architecture, a group of computer scientists at Hitachi's Advanced Research Lab is working to develop a new type of integrated computing environment.by Steven Myers
Founded in 1985, in commemoration of Hitachi's 75th anniversary, the Advanced Research Laboratory (ARL) carries out a diverse assortment of long-term, basic research projects. The primary research areas of the lab include radiation/electron beam physics, biotechnology, material science, and software science.
Hitachi ARL puts strong emphasis on multidisciplinary research, a policy that is reflected in the diversity of projects and collaborations underway there. The laboratory is also highly active in hosting foreign visiting researchers (commonly 10 to 15 at any given time), and also sends its own scientists to prestigious universities and research institutions overseas.
The facilities
The laboratory itself is located in a scenic wooded area in Hatoyama-machi, Saitama prefecture. The physical environment is beautiful and inspiring, and the lab facilities are excellent.Computing Japan's visit to Hitachi ARL was arranged by senior engineer Hiroshi Motoyama. He explained that, because ARL was established "before the bubble era," the company was able to provide for a large physical area and scenic surroundings that would be almost unimaginable today. Motoyama went on to say that the lab currently operates on an annual budget of about ¥4 billion (1% of Hitachi's ¥400 billion total), an amount that is "just right, in that it's enough to ensure the researchers have everything they need, but not enough that they can afford to be wasteful or inefficient."
We were given several detailed presentations and viewed demonstrations of work carried out by a group of computer scientists led by Dr. Hiroshi Motoda, a senior research scientist at the lab. Computing facilities for the software science labs consist primarily of Sun SPARC servers and workstations, with scattered NeXT machines. Although all of the scientists are given ample working space, only the senior researchers such as Dr. Motoda have their own rooms.
The software science section of the lab is organized into four major groups. One focuses on programming and computation, while the other three investigate such AI-related themes as machine learning, neural networks, and natural language processing.
The Motoda Research Group
Dr. Motoda's group at ARL is actively investigating a wide range of AI-related topics, with the aim of developing an integrated computing environment capable of accurately filtering, extracting, and translating information according to individual preferences. The underlying theme of this research, explains Dr. Motoda, is "the support of intellectual activities by computers."The system envisioned by the group would have the ability to search huge collections of documents (such as those found on the World Wide Web), extract pertinent articles, highlight relevant sections, translate the information into the user's desired language, and adjust future search criteria according to the documents the user selects. Additionally, the user should be able to search for documents not only by keyword, but also by graphics -- such that a search on one picture would return documents containing similar or related pictures.
Obviously, such a system requires highly advanced machine translation and inductive learning functionality, as well as an extremely high-performance hardware platform. Motoyama's team consists not only of AI researchers, but also computer architecture experts in charge of building a large-memory personal computer with the power to carry out these computation-intensive tasks.
The machine translation aspect of this project is particularly intriguing. Special emphasis has been given to developing numerical techniques for examining English words with multiple meanings, and choosing the correct meaning based on the context in which the word appears. This process is sometimes referred to as "disambiguation."
An example of the disambiguation technique is shown in the accompanying figure. When the system encounters the word "suit," for example, it first examines the surrounding terms and stores information about each word based on its position within the sentence. This information is then compared with that from many "training contexts" that have been stored for all possible meanings of "suit." Thus, the appearance of words such as "claims" or "court" within the same sentence would signal the system that the suit in question is a lawsuit, or soshou (ëiè), rather than an item of clothing.
System demonstrations
After a thorough presentation by Dr. Motoda covering the various facets of the project, we were given a demonstration of the parts that had been implemented thus far. First, Dr. Makoto Iwayama showed the document retrieval system, which produced impressively accurate results searching on "key documents" rather than key words. The system is able to scan a document, assign it to one or more categories, then search through other documents from the same categories in order to find related information.By using this type of classification scheme as a "preprocessing" mechanism, the search space is significantly reduced, resulting in much faster retrieval. As one might expect, searches on articles related to popular topics falling into clearly-defined categories produced much better results than those on topics about which little has been written. Iwayama explains that documents in the database are continuously re-categorized according to what the user has previously selected as being relevant to his or her interests.
Next, Dr. Ken-ichi Yoshida gave a brief presentation on his work with implementing an intelligent and adaptive user interface. Yoshida's research centers around using a directed graph to model relationships and dependencies between various UNIX commands and their I/O operations. By assigning numerical weights to these relationships, probabilities are calculated for predicting subsequent commands used in carrying out a certain task, based on the initial command. Yoshida believes that this type of prediction method will be useful in creating a user interface that can make suggestions to a user for accomplishing various tasks, based on that user's previous actions.
Finally, we were shown to the Motoda group's hardware lab, which is headed by Dr. Atsuo Kawaguchi. In this lab, work is underway to create a personal computer with 2GB of main memory -- the type of power necessary to implement the integrated computing environment envisioned by the group. Kawaguchi explained that he has built the computer from the ground up, and thus has developed an intimate understanding of hardware and operating system issues. The group has acquired the source code for BSD UNIX, and they recently succeeded in getting this OS to run on their unique PC.
Future directions
Over the past two years, the Motoda group has published ten technical papers detailing the various facets of their work. With so many different areas of research integrated into a single project, however, it is difficult for them to outline a detailed plan for the future.Dr. Motoda says he expects the project to continue for another 10 years, and notes that some of the technologies are already being implemented in research labs. The most difficult parts of the project to implement, he explains, are those related to autonomous learning and image-understanding capability. Over the next year, though, the group plans to devote itself primarily to developing and integrating extensive help functionality into their system.
Contact Information:
Hitachi Advanced Research Laboratory
Hatoyama, Saitama 350-03
Japan
Phone: (+81) 492-96-6111
Fax: (+81) 492-96-6006
email: motoyama@harl.hitachi.co.jp
WWW: http://hatoyama.hitachi.co.jp
