SANS annual report 1994
Previous 
Up 
Next
Software Development and Parallelization
Software Development and Parallelization
Most simulation work in the field of ANN:s today is done on standard
von Neumann machines. The tremendous development in RISC
microprocessor performance has been beneficial to NN research. Even
so, it has been pointed out many times that such sequential computers
are quite inefficient for execution of neural networks, as the
networks themselves are inherently parallel. For this and other
reasons it is important to implement the neural network algorithms and
architectures on parallel machines. Above all, it enables high
computational capacity. This allows us to investigate the scaling
properties of our algorithms as problem size increases and opens up
the possibility of escaping from ``toy problems'' to some real world
applications. A second, perhaps less obvious reason, is that the use
of parallel architectures puts relevant constraints on algorithm
development. In this way we avoid inadequate solutions in our
algorithms with lots of sequential sections embedded. The fact that
the SANS group has convenient access to massively parallel computers
(a 16K Thinking Machines Connection Machine CM200 and a 55-node IBM
SP2) at the
Center for Parallel Computers (PDC)
further enhances our possibilities for conducting the sub-projects
described below.
The Connection Machine CM200 has also been used in some of the
sub-projects described under "Computational Neuroscience"
.
The IBM SP2 is also beginning to be used in these projects.
Subtopics
Next