Matt Grover - 2002-09-29

Logged In: YES
user_id=171453

Rate coding is better handled by perceptron networks because
it requires averaging spikes over time before computations
can be done. That would be very difficult to do with a
spiking network but perceptrons are based on rate coding
principles. The other coding schemes can be done and it is
a good idea to provide ways to easily set up networks that
utilize them.

Population coding should be largely a matter of setting up
the correct network topology. That won't be all that
difficult. Phase coding is a bigger problem, though,
because it requires background activity that creates a
reference for the phase. The obvious methods of
implementing background activity are not efficient and would
lead to very slow simulations. A possible solution to this
is to use statistical methods to simulate noise
(http://www.cns.unibe.ch/~michi/pub/rgf02.pdf). This can
probably be used to generate a background signal by choosing
an appropriate statistical distribution (I'm not completely
sure yet). This method will also require a different type
of neuron model than what we are already using. I have been
experimenting with this already but it is not yet clear if
it will fit into the Amygdala design without making
significant changes.