Diederik,
Thanks for your reply -  I agree that I left out the details. I'll try to
give a brief description - please bear in mind that this is, as you say,
just vapour, subject to change.
The sort of network I'm looking for is based on the 'fire together, wire
together' principle - it's classical in that respect, rather than being any
kind of designed hierarchy. One change is that new connections between
neurons can form according to some rule favouring short connections over
long.
I'm being innovative in two ways: Firstly, each neuron has its own,
individual genome, which controls the 'chemistry' of the neuron, and its
immediate surroundings. So, for instance, it can indirectly control the
activation potential, or alter synaptic plasticity, or cause the release of
'neurotransmitters'. Thus, I need quite a complex neuron, similar to a real
biological neuron.
Secondly, the neurons can spread their genes to other neurons (along the
axons!) They can do this most effectively by anticipating the spiking
activity of the other neurons.
Under these conditions of Darwinian neural competition, I hope to find that
the mutual information of connected neurons is minimized. I would also like
to think that the networks which arise in this fashion will be 'preadapted'
to learning temporal data.
As to how the network will be trained to produce the right responses, I'm
not sure - I want to use a method that has some biological parallel. The
problem is that in the evolved networks we have prior experience of, the
reward mechanism evolved in parallel. I don't see any way of doing that.
The training set should have problems of (at least) two forms to test my
hypotheses:
(1) after 'hearing' a sequence of pulses, reproduce them, and (2) after
being presented with several images (e.g. red circle, red square, red
triangle, green circle), indicate the odd one out. Being able to do either
of them should show I'm on to something.
Is that any help? I was reluctant to give too much away because it's a
rather far fetched concept, but as you can see, the neurons have to be
capable of doing a lot. I think I can justify taking this one of many
options in neural networks, if only because no-one seems to have let the
neurons themselves compete before.

Nathan Cook

On 12/26/06, Kingma, D.P. < [EMAIL PROTECTED]> wrote:

Dear Nathan,
Your description of the kind of neural-net scheme needs more detail before
I can give any more particular direction. It leaves a huge amount of
possibilities.
For example, is your neural net similar to hierarchical network of
abstract facts, being build by agents, whereas the agent's own functionality
(parameters) being altered by a (re)engineering agent guided by expectation
maximization over the training samples?
Or is it more like a classical neural network, being the difference that
your neurons perform a more exotic function of their input?
I've been thinking about 'evolving neural networks that are good at
learning spatial and temporal patterns' as well, but I have to say the
theoretical possibilities are endless, but scientifically vapour, if not
tested on some *real* data. To be taken seriously, you need a good
specification of your problem, and a straigth-forward way is to select some
representative training set, preferably used by previously published papers.

Diederik Kingma



On 12/26/06, Nathan Cook < [EMAIL PROTECTED]> wrote:

> Hello list,
>
> I have an idea for evolving neural networks that are good at learning
> spatial and temporal patterns. However, to implement it, I need a model for
> a spiking neural net, such that each neuron has several different parameters
> affecting its operation (I will then use genetic algorithms to modify these
> parameters). The richer the behaviour of the neurons the better. Can someone
> direct me to a suitable resource? I would be pleased to provide more details
> if anyone's interested - my goal is to create networks which can learn more
> than one sort of information (say, visual and audio) at once, and even do
> some form of induction on this information.
>
> Nathan Cook
> ------------------------------
>


-----
This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/?list_id=303

Reply via email to