Opportunities, Challenges and the Road Ahead for Field-Coupled Nanocomputing: A Panel Discussion - Field-Coupled Nanocomputing: Paradigms, Progress, and Perspectives - page 382

try to do. I was wondering if I could make a really useful and attractive random
number generator that could seed other things.
Porod: I have a slightly different take perhaps. Maybe taking inspiration from
biology might be a good thing. The secret behind the CPU is that it is so general;
you just know how to crunch bits and you have the software to tell what this bits
mean. You have hardware that can solve a large class of problems. So I think the
issue is commercial; designing a chip is very hard, designing hardware is hard
and expensive, and it has to justify your investment.
So I can imagine special purpose hardware for suciently large classes of
problems that justifies this investment. Everything does not necessarily have to
be general purpose. A class of applications that I think would map onto QCA is
vision applications. If you think about how visual processing is done - Sanjukta
[Bhanja] mentioned it this morning - it essentially involves near neighbor inter-
actions or near neighbor computations. You extract features from that with very
little movement of data and with very little communication overhead. I am famil-
iar with some of the work that is going on in Berkeley at the Vision Research
Lab suggesting that the early visual processing in the retina takes place in about
then different layers in the retina. Each layer represents essentially a different
connectivity between processing elements you can think about it as neurons.
Each of these connections essentially extracts different kinds of features. So this
maps onto cellular neural networks - or CNN - and some people are working
on that. I could see that if one could come up with a QCA-like implementation,
where some of these connections that you want for image processing were real-
ized by physical interactions - improving power and speed and all that - that
there might be a suciently large market. That might be a killer app. So I am
keeping my eyes open for vision types of applications.
Bhanja: Professor Porod, do you have CNN implementations of processing low-
level features for image processing?
Porod: Well, to some extent. We had a MURI Project associated with the
vision research I mentioned at Berkeley. Part of this project involved a group in
Budapest that is working on cellular nonlinear networks, trying to implement the
kinds of templates they extract from the retina into hardware. We were a part
of that project, looking to see whether or not this hardware could be naturally
implemented in a QCA-like fashion. That was a few years back, and we were not
far enough along with the magnetic implementations to have something reliable
enough that you could actually do some engineering with. But if one had reliable
building blocks, then [one could implement] something based entirely on near-
neighbor computations where you do not have to think about threads, about
moving data from memory into some processing element so that the processing
element can crunch on it a bit. So I think vision applications might be a good
target. But, again, I think we need some reliable building blocks first. We need
some bricks to start constructing buildings.
Lent: Were they implemented in analog?