Marvin Minsky <minsky at media.mit.edu> wrote:
> "Lance Sherman" <lancesherman at insightBB.com> wrote in message
news:<V0I0d.23947$MQ5.12109 at attbi_s52>...
> > I'm not sure I see the question -
> > by "biological neural network", do you mean a biologically plausible
> > artificial NN?
> > "Harry Erwin" <herwin at theworld.com> wrote in message
> > news:1gjy8r1.1ov7fmr5a7cl2N%herwin at theworld.com...> > > I've been exploring neural implementations of displacement
> > > representations. I'm suspicious that Minsky and Papert (1969) implies
> > > that it is very difficult to use a biological neural network to do
> > > vector operations. Any pointers?
> > >
> > > --
> > > Harry Erwin <http://www.theworld.com/~herwin>
> > > My neuroscience wikiwiki is at
> > > <http://scat-he-g4.sunderland.ac.uk/~harryerw/phpwiki/index.php>
>> Well, It is easy to make analog networks that do matrix-like
> operations, but non-linear networks (e.g., with thresholds) would make
> that rather difficult! However, what we showed in 1969 was
> non-looping neural networks could not recognize topological features
> of things in was that woulkd be scalable. And those theorems still
> are true, regardless of improvements in how fast such networks can
> learn the things that they can learn.
I was trying to do it with simplified neocortical columns to model
something bats do in the lab. You can do it with a hidden layer of
coincidence detectors, but the number of cells required for the task is
biologically unrealistic. I need anywhere between 1000 and 1000000
individual cells to represent the target state and the displacement
(depending on how much of the target volume is covered by the receptive
field of each cell), and the hidden layer scales as the square of that.
Harry Erwin <http://www.theworld.com/~herwin>
My neuroscience wikiwiki is at