doshisha_commonheader
  1. Home
  2.  > Programs & Courses
  3.  > Graduate School of Science and Engineering : Information and Computer Science (Laboratories : Co-Creation Informatics Laboratory)

Graduate School of Science and Engineering
Information and Computer Science

Co-Creation Informatics Laboratory

Research on pattern recognition, data mining, signal processing, and computer-supported cooperative works

Staff

Prof.Katagiri

Shigeru KATAGIRI
[Professor]
MD
Telephone : +81-774-65-7567
mail
Office : KE-110
Database of Researchers




Miho OHSAKI
[Professor]
M
Telephone : +81-774-65-6468
mail
Office : KE-111
Database of Researchers

Research Topics

<1>
Developing remote collaboration support technologies
  • Developing highly realistic sound-field reproduction technologies in multimedia communications
  • Developing video echo canceling technologies
  • Developing a natural interface with t-Room using body motion
  • Developing t-Room’s user interface
  • Developing technologies for improving "the feeling of being in the same room" using cross-media information
  • Developing technologies for improving "the feeling of being in the same room" by controlling time delay
  • Researching evaluation criteria for "the feeling of being in the same room" improvements in t-Room
<2>
Knowledge discovery from medical data
(Discovering knowledge useful for medical treatments from data accumulated in hospitals)

  • Developing time-series clustering methods
  • Developing knowledge discovery assistance systems
  • Data modeling by multi-dimensional spectral analysis and using it for knowledge discovery
<3>
Researching pattern recognition technologies based on the minimum classification error training method/generalized probabilistic descent method
  • Developing an ensemble minimum classification error training method
  • Defining geometric margin control in the minimum classification error training method
  • Developing “DISCERN,” discriminative training library software for education and research
  • Developing feature representation for recognition using genetic algorithms

Research Contents

Research background and goals
We are currently facing many challenges that must be quickly resolved such as environmental problems, the depletion of fossil fuels, and the reduction of industrial might brought about by an aging, low-birthrate society. “Reduce the movement of people and things, and create much value from little energy” – the solution for these challenges can only be highly-efficient value creation using advanced technologies. For us the decisive factor for creating this value is the utilization of computers. We believe that co-creation between computers and we humans will allow humanity to overcome these challenges and is the trump card for creating a truly affluent society.
With these beliefs, in the Co-Creation Informatics Laboratory we are researching remote collaboration support technologies that connect distant people utilizing computers, and researching the technologies to create value from large amounts of complicated data that humans cannot handle by making computers themselves smarter.
Connecting people with the power of computers
With the appearance of the Internet and cellular telephones, telecommunications technologies such as telegraphs and telephones have produced an information society where you can communicate “anytime, everywhere, and with anyone.” Without a doubt, these new communication tools are quite convenient. However, we are stuck with small screens and keyboards, it is by no means easy to fully express our thoughts. Now in the present where the ability of computers and the Internet has improved, we feel we must change the goal of our technology development from the “small, convenient” way that has been pursued up until now to a “large, genuine” way that conveys our entire communication scene.
We in the Co-Creation Informatics Laboratory are doing advanced research of the “Future Telephone t-Room” proposed by NTT Laboratories to bring these ideas to a realization. By controlling multimedia devices like multiple cameras, displays, microphones, and speakers with multiple computers, it connects distant people as if they were right next to each other.
However, it’s not good enough to just simply convey video and sound. Current technology has many issues that must be improved such as image/sound reflections (echoes), video blind spots, skewed points of view, unnatural video/sound-field reproduction, and video/sound going out of synchronization. Our current goal is to resolve these issues and work to make t-Room more advanced by utilizing digital signal processing, pattern recognition, and computer communication technologies.
Mining knowledge with the power of computers
In recent years, computer calculation and storage performance has been making spectacular progress. If simply repeating the four arithmetic operations, it’s no exaggeration to say that computers have already surpassed humans. Computers’ storage power is the same. The amount of text and video data the computers connected by the Internet is not an amount that can be memorized by a single person. However, we humans have many kinds of high intellectual powers that even powerful computers cannot imitate. One of these intellectual powers is the power to mine for knowledge. This is called the power of data mining or knowledge discovery.
To also provide computers with this power to mine, we in the Co-Creation Informatics Laboratory are aiming to establish medical data mining technologies in particular to discover valuable knowledge from the massive and complicated time-sequence data acquired from the medical field, and we are researching and developing those basic technologies. Our approach places an emphasis on a signal processing approach for expressing knowledge with statistical meaning from massive and complicated data. We are also performing simultaneous modeling of tens of dimensions of time-sequence data that no human could possibly perform to bring to a realization of knowledge mining that fully utilizes the power of computers.
Making computers smarter
One more intellectual power that humans can easily perform but computers cannot is pattern recognition. We humans can listen to sounds and understand visual scenes that we see, and we can accurately judge in an instant what is being talked about and what we can see. We can also read text and easily understand its content. However, as an example computers now can search for items to see if they are present such as whether the word “computer” is entered in a database, but they cannot easily judge whether there are similar items with the searched item, pattern recognition in other words. Please say “good morning” out loud. Everyone says this in a different manner with a different voice. By no means do different people speak with exactly the same voice pattern. Even a single person’s voice will be different each time they say “good morning.” We humans hear this differing pattern as the same words, “good morning,” without any problem. But to make a computer listen like this is not easy at all.


We in the Co-Creation Informatics Laboratory aim to advance these pattern recognition technologies by researching and developing new recognition system design methods with the cutting edge technique called the minimum classification error training method (or the generalized probabilistic descent method) as the foundation. The basic concept is simple. The basis of recognition is in comparisons. The “good morning” pattern the computer is trying to recognize is compared with a number of patterns stored on a computer .If the stored “good morning” pattern is clearly more similar to the “good morning” pattern to be recognized rather than other patterns like “good evening,” there are no problems. The pattern is correctly recognized. However, let’s make the stored “good morning” pattern that of an adult male. And then let’s make the stored pattern for “good evening” a child’s voice. At this time, if the “good morning” to be recognized is a child’s voice, this “good morning” may be judged more similar to the child’s “good evening” rather than the adult’s “good morning.” Depending on whether the computer judges the similarity in voices or words, we can understand that these kinds of variations or errors can occur as a result. In order to prevent these kinds of errors, our technique is to repeat changes in the stored “good morning” and “good evening” patterns to achieve accurate recognition, or learning in other words.
We ourselves have been involved in the development of the minimum classification error training method. With this background, we are advancing research to further improve and develop minimum classification error training while competing at an international standard.

Keywords

  • Remote Communication and Collaboration
  • t-Room
  • Multi-media Signal Processing
  • Pattern Recognition
  • Discriminative Training
  • Minimum Classification Error

  • Generalized Probabilistic Descent
  • Data Mining
  • Knowledge Discovery
  • Clinical Data
  • Time-series Data

.