Computational intelligence that incorporates the human element — There is a vast number of things that cannot be done without cooperation between humans and computers

Computational intelligence that incorporates human factors — There is a vast number of tasks that cannot be solved without cooperation between humans and computers

By Professor Hideyuki Takagi,
Faculty of Design and (working concurrently for) Research Center for Applied Perceptual Science

Industrial robots can be designed with just technologies in control engineering and robotic engineering. However, the sales points of home robots — such as their cuteness or the safety impression in their motions — cannot be measured, and the only thing that counts is how humans feel about them. There are several well-established technologies for retrieving knowledge automatically from a database, but retrieving well-prospective knowledge requires evaluations of domain experts. It is the same in signal processing. There are many signal processing tasks that cannot be evaluated except human visual or auditory inspections.

Currently, our research focuses in particular on interactive evolutionary computation (IEC) that combines human subjective evaluations based on human experience, knowledge and KANSEI and the optimization capability of computers, and supporting techniques for human awareness.

Typical example is hearing-aid fitting, which was the dissertation topic of my first doctoral student. Currently, we are conducting joint research on cochlear implants fitting with Faculty of Medicine of Fukuoka University and Kyushu University. We also conducted a joint research with Fukuoka Dental College on designing image enhancement filters based on visual inspections of a medical doctor whose expertise is medical image diagnostics.

Geophysicists at an Australian research laboratory had designed a system that find the initial geophysical conditions resulting current geological phenomena surveyed by geologists. It was a tough job, and their geologists were unable to find solutions through their simulation solution even after two weeks of trial and error. However, they could find them after we provided them our suggestion of IEC use in 2000, which caused quite a sensation at their lab.

In joint research on designing MEMS (circuits in which electronic and mechanical elements have been incorporated on a chip like an LSI) with UC Berkeley, multi-objective optimization was necessary to satisfy multiple different design requirements. IEC-based design could obtain better MEMS design than that obtained by only computer optimization; MEMS design experts can evaluate what was good and bad when they look over MEMS circuit as a whole, which is gotten incorporated into the optimization through IEC.

There are many artistic IEC applications, such as 3-D computer graphic (CG) lighting design and indoor lighting design for given design concepts.

Although majority of IEC research is biased to application-oriented research, research on reducing user's fatigue must be conducted to make technologies for carrying out coordinated work between tireless computers and humans practical

Developing algorithms in also necessary for this research. Some of them are improving human interface, developing techniques for accelerating optimization, developing human user model using machine learning and predicting which solutions will be likely to receive high evaluations from human users.

Since the number of research on these topics is fewer than that of application-oriented research, our laboratory is focusing its research on these topics rather than that application-oriented research which is easier to understand.

When a system that has been optimized based on human evaluations is in turn subjected to analysis, we may be able to know the characteristics of the humans who carried out the evaluations. This approach can be used as a tool for human sciences. I think that there are no laboratories who take this approach in the world except our laboratory.

We found that the happy–sad range for schizophrenics is significantly narrower than that for mentally healthy persons using IEC and psychological scaling. Behavior checklist is the only way to determine the level of schizophrenics' symptoms so far, but this discovery holds out the possibility to be a new diagnosis method.

By analyzing the characteristics of hearing-aids fitted by IEC mentioned in the above, we found three unknown new facts: (1) hearing-aids characteristics fitted using pure tones, which is a conventional fitting method, and those fitted using speech are different, (2) those fitted using speech are almost the same regardless speakers or with/without noise, and (3) those fitted using speech and those fitted using music are different. The ideal hearing-aids or cochlear implants in the future will probably have a function that adapts to the change of sound environments.

Throughout the world, cochlear implants are fitted based on two hypotheses: (a) that the more channels giving electrical stimulus the better (i.e., that frequency-resolution capability improves), and (b) that each channel should be fitted to be able to hear sounds from the user’s minimum level up through his/her maximum comfortable level. When a French team applied IEC to the fitting of cochlear implants, however, their optimized fitting characteristics are quite far from these hypotheses, nevertheless their word recognition rates were higher than those fitted based on conventional hypotheses. This fact implies that there must be unknown facts in our auditory system. Now, we are tackling to find out the unknown fact through joint research with Faculty of Medicine of Fukuoka University.

Professor Hideyuki Takagi

Faculty of Design and (working concurrently for) Research Center for Applied Perceptual Science

SoCPaR2015 : 7th Int. Conf. on Soft Computing and Pattern Recognition

November 13th–15th, 2015, Ohashi Campus