Background
The project was initiated by professor Shinji Tomita (Kyoto
University, Japan), that invited me to visit his lab as a research
student. His interests are in the areas of computer architecture and
parallel processing. The course part of my doctor studies at NTH had
mainly been on neural networks and computer architectures. It was
quickly established a good relation to all the professors (Shinji
Tomita, Hiroshi Nakashima and Shin-ichiro Mori) in the group of prof.
Tomita. Together we decided that implementation of neural networks on
parallel architectures was an important and interesting research topic
to be done in my project. This research could easily be continued when
I returned home and compared with research to be done on the parallel
reconfigurable neurocomputer (RENNS), designed by the Computer Group, IDT.
The research project
The lab employs a highly parallel computer with 64 processing
elements (PE), Fujitsu AP1000, on which the research was done.
Moreover, a larger version of the computer, consisting of 512 PEs,
was made available by Fujitsu Parallel Computing Research Facilities,
Kawasaki, Japan. Thus, we could measure performance speedup using a
large number of PEs.
I was taken well care of from the first day at the university. Although
different culture and language, all the helpful people made it easy for me
to get started. After I had given a introductory talk about my research, a
discussion followed about my coming research. Then, I regularly spoke with
my advisor, prof. Tomita. He was very interested in my research and all the
time, he gave me advice and commented my results in a constructive manner.
The students in the lab were helpful to answer my questions regarding
practical problems like language, computers, etc. The research in the lab
was presented and discussed in usually weekly seminars. I gave 4
presentations during my period in the lab (each of 1 hour in length).
Results
The programming of the AP1000 computer went well and several different
algorithms were implemented. The performance obtained was higher than
implementations on other general purpose computers.
The results show that using a highly parallel computer can drastically
reduce the training time of neural networks.
Moreover, that a general purpose computer can be an alternative to
specifically designed neurocomputers. The research has suggested methods to
obtain high speedup ratio for a large number of processors.
During this research a large parallel computer was used for free. It would
not have been possible at NTH to get a similar access. Moreover, the
results can be used in my future research, to compare them with
implementation on other (smaller) parallel computers.
Conclusion
This project successfully gave good results in parallelising neural
networks on a parallel computer. I'm very satisfied that language and
culture didn't become any problem in this research. To me it was a
very good motivation to live in Japan and do parts of my dr.ing.
research there. Japan is doing much research, both on neural networks,
parallel processing and computer architecture. The way of doing
research is also slightly different from at home. Thus, I feel that I
have learned both more about my research field and more about
how to do research. To live in Japan is also very interesting, giving
me many memorable experiences.
Acknowledgement
I have very much appreciated the good
support and help I received from the Research Council of Norway in this
project. Due to the strong Yen currency, it would have been impossible to
make this project come through, without your financial support.