Information Technology Reference
In-Depth Information
Porod:
They were implemented in CMOS chips, but there were issues competing
with general purpose processors. Some of these chips are actually fairly sophisti-
cated, but they were made by university groups. Computers get faster and faster
and faster, so even though they had parallel hardware, it was hard to compete
with the new computer you buy at Christmas! Another issue was that they were
working with a general purpose CNN processor, where they had to load in dif-
ferent templates encoding how each pixel looks to its nearest neighbors. There
was a significant overhead in this template matching operation. But I think if
you could identify - just like in the retina - say ten different connections, and
you implement those in hardware in ten chips, the processing could be very fast
and you just mix and match the information from these ten chips. Even though
it would be special purpose in some sense, something like this would be relevant
for a very large class of applications and, I think, worth some investment.
Wo lkow:
You know, another related thing, the European Union just announced
a funding on a billion-Euro brain simulation project. They are trying to explore
what kind of models begin to represent a brain. Maybe we should watch for what
emerges there and see if we could offer, like you said, localized, special functions
that might emerge as attractive and useful. Maybe we could deliver something
that's just what they need.
Porod:
So, just this one comment, since you mentioned the brain. I would like
to make a distinction between the retina and the brain. The retina actually does
a lot of computation. I do not think that this is generally appreciated. So the
world around us presents this huge flux of information that we need to process.
It's not processed in the brain. This huge flux of information is broken down
into relevant pieces of information - which are the features - which are then
handed on as a trickle of information back into the brain where some of the
higher level cognitive functions occur. It seems to me that the higher cognitive
parts could be done by a CPU-based computer, but that there are opportunities
for specialized hardware that do what we do: We do the same thing over and
over and over again, just breaking down this flow of information into lines or
edges or whatever.
Wo lkow:
You know, some of the most advanced cochlear implants use that same
knowledge. They pre-filter and they process information in almost mechanical
ways, with waveguides and so on, so that minimal computing is needed after-
wards. So you do not use more bits then you need and you do not general-purpose
process anything. You try and cleverly handle information by pre-processing it.
Graziano:
I have a comment here. Look at the work going on around the con-
cept of dark silicon; that seems to be the direction. It is impossible to have bigger
and bigger and bigger CPUs even with parallel organization. The trend seems
to be that you have a big circuit with a lot of blocks then you switch them on
only when you need them, and then you specialize some of them. So you have
some general purpose part where some part of the computation is done, and,
depending on the application, you have some blocks that are specialized. You
only switch them on when you need them. This is clearly the case in CMOS.
Search WWH ::
Custom Search