Turing's design of computer differed from Von Neuman's design in that
Turing's computer had processing associated with memory locations.
When one wanted two numbers added essentially one transferred the two
numbers to some particular storage location and the adder circuitry
there added the numbers. Programs for Turing's computer then largely
consisted of moving data to the particular address that performed the
desired operation.

It is clear that the human brain is not set up like a Von Neuman
computer. In some way not yet well understood the cells of the brain
perform memory *and* processing. Impressively, the cells of the human
brain operate both digitally and in analog fashion at the same time.
We have no machines which are able to do that directly, although
programs can be written which simulate it. The difference there is
that the simulation requires some number of bits of data to represent
each neuron and there is much data transfer to and fro to perform the
elementary operations which constitute the simulation.

So long as our computers are designed with the Von Neuman design
computing will engender huge numbers of memory accesses/transfers in
order to get anything done at all. The current rage is to have several
Von Neuman design computing engines, called cores, in a cluster on the
processor chip. Currently the core count is four and soon there will
be eight core processors in the commercial marketplace. There is the
view by the manufacturers that because of heat problems with increased
clock frequencies that the direction to go is to have more and more
cores.
The problem with that is there is not a well developed science of
parallel programming and scheduling which can be used to program these
increased core count machines. Because the entire past of commercial
computing has been uni-processor there is rather little software
around that can effectively utilize the multi-core computer. This will
of course have to change so that progressively greater core count in
machines is a viable business model.

We see that artificial neural networks (ANN) which are simulated in
computers require many memory transfers and a whole host of elementary
operations to effect them. In order that a network size be large
enough to compare with the capabilities of even an ant's brain or a
house fly's brain one needs much real computer memory to store all
those bits and large amounts of transfers to feed the logic and
arithmetic circuitry of the Von Neuman architecture.
Multi-core machines have the promise of getting
us more computing power but we need a way to effectively perform
parallel processing in the software. Because so many elementary
operations must be performed to effect (perform) the ANN simulation of
the ant's or house fly's brain we need a fast computer to do it. One
with lots of memory. While we can construct a neural net equivalent in
size (neuron count) with the ant of fly brain we have no idea how to
program/operate that net so as to be able to do what an ant or fly can do.

We have a paint box filled with colours and a blank canvas. We need to
know how to put the paint on the canvas to effect a scene or picture.
As yet we do not know how to make a neural net capable of doing what
the ant's or fly's brain allows them to do. Our present (ANN) neural
network capabilities are a great deal more modest than that of an ant
or fly. It is clear that while we have some idea about the importance
of the connectivity of networks in general we dont have a well
developed idea as to what
activity across that network will produce results as impressive as
that done by nature with real neurons. Perhaps it is that our current
ANNs are too schematic and don't yet adequately capture all the
salient aspects of real neurons. [continued next posting]