NEURAL NETWORK MARKUP LANGUAGE

[posted to comp.ai.neural-nets]

Hey all,

I went hunting today for an SGML/XML based language that
could describe the structure of a neural net.

In reading the FAQ for this newsgroup, I found references to three
languages (non XML) which discribe the layout of a network:
Aspirin, NeuDL, and nn... All of these seem to have their uses,
but none are quite what I'm looking for.

Anyway, http://manifestation.com is working on an open-source AI
program called Sabrina, which uses neural networks, and since its
open source, I want whatever we do to be useful to the most number
of people.

From a programming standpoint, it's very useful to think of neural
networks as a bunch of matrices and vectors, but when we diagram
them out, it's so much easier to think of the as a physical machine.

Anyway, here's a thought about what a Neural-Net Markup language
would look like. The goal is to define the *structure* of the
net - including initial weights and values - in a simple format
that can be read by both humans and machines.

My question is, what's missing? What would you want in a language
like this that's not already there? (Besides, of course, conversion
programs to port it to other neural net languages, compile it,
visualize it, etc.)

- Michal Wallace
(sabren@manifestation.com)

------------------------------------------------------------------

<neuralnet name="A Simple Neural Network">

<layer name="N" class="input">		
	<neuron name="X1" value="1"/>
	<neuron name="X2" value="0"/>
	<neuron name="X3" value="0"/>
	<neuron name="X4" value="1"/>
	<neuron name="X5" value="0"/>
</layer>

<neuron name="Outside1" value="1"/>

<layer name="M" type="hidden">

	<!-- NET refers to the sum of the inputs times 
	their respective weights... -->

	<neuron name="M1" function="1/(1+(e^(-NET)))">

		<input name="X2" weight="-.24"/>
		<input name="X1" weight=".4"/>

		<input name="M1" weight=".04"/>

	<!-- Note that inputs can come in any order, from
	any layer. This one, for example, feeds back on
	itself. Inputs that are ommitted are assumed to
	have a weight of 0. -->

		<input name="X4" weight="1.4"/>
		<input name="X5" weight="1"/>
		<input name="Outside1" weight=".3"/>
	</neuron>
</layer>
</neuralnet>