On 10/14/2011 9:29 AM, karl ramberg wrote:
Interesting article :
http://www.itnews.com.au/News/276700,ibm-eyes-brain-like-computing.aspx
Not much details, but the what they envisions seems to be more of the
character a autonomic system that can be quarried for answers, not
programmed like today's computers.
I have seen stuff about this several times, with some articles actively
demeaning and belittling / trivializing the existing pre-programmed Von
Veumann / stored-program style machines.
but, one can ask, but why then are there these machines in the first place:
largely it is because the human mind also falls on its face for tasks
which computers can perform easily, such as performing large amounts of
calculations (and being readily updated).
also, IBM is exploring some lines of chips (neural-net processors, ...)
which may well be able to do a few interesting things, but I predict,
will fall far short of their present claims.
it is likely that the "road forwards" will not be a "one or the other"
scenario, but will likely result in hybrid systems combining the
strengths of both.
for example, powerful neural-nets would be a nice addition, but I would
not want to see them at the cost of programmability, ability to copy or
install software, make backups, ...
better IMO is if the neural nets could essentially exist in-computer as
giant data-cubes under program control, which can be paused/resumed, or
loaded from or stored to the HDD, ...
also, programs using neural-nets would still remain as software in the
traditional sense, and maybe neural-nets would be stored/copied/... as
ordinary files.
(for example, if a human-like mind could be represented as several TB
worth of data-files...).
granted, also debatable is how to best represent/process the neural-nets.
IBM is exploring the use of hard-wired logic and "crossbar arrays" /
memristors / ...
also implied was that all of the neural state was stored in the chip
itself in a non-volatile manner, and also (by implication from things
read) not readily subject to being read/written externally.
my own thoughts had been more along the lines of fine-grained GPUs,
where the architecture would be vaguely similar to a GPU but probably
with lots more cores and each likely only being a simple integer unit
(or fixed-point), probably with some local cache memory.
likely, these units would be specialized some for the task, with common
calculations/... likely being handled in hardware.
the more cheaper/immediate route would be, of course, to just do it on
the GPU (lots of GPU power and OpenCL or similar). or maybe creating an
OpenGL-like library dedicated mostly to running neural nets on the GPU
(with both built-in neuron types, and maybe also "neuronal shaders",
sort of like "fragment shaders" or similar). maybe called "OpenNNL" or
something...
although potentially not as powerful (in terms of neurons/watt), I think
my idea would have an advantage that it would allow more variety in
neuron behavior, which could likely be necessary for making this sort of
thing "actually work" in a practical sense.
however, I think the idea of memristors is also cool, but I would
presume that their use would more likely be as a type of RAM / NVRAM /
SSD-like technology, and not in conflict with the existing technology
and architecture.
or such...
___
fonc mailing list
fonc@vpri.org
http://vpri.org/mailman/listinfo/fonc
