ANNs

[sylvain kepler]

Would a ANN thread be a good point in the forum ?
I'm opening the topic by considering ANNs a part of models of computation.
ANNs are one of these among which one may find :
Systolic algorithms (sorting networks), perceptrons (ANNs), Markov algorithms (universal model of computation).

[mobilegazer]

i think this topic probably fits better in the bernsee/dspdimension forum, don't you think so neuronaut?

cheers,
d

[neuronaut]

Well, it would depend on what exactly we are going to discuss here, I'm not quite sure that I understand yet what the thread is all about. If it relates to the Neuron or music/audio processing then I would probably leave it in here, if it discusses ANNs on a highly abstract level then the DSPDim section might be a better place, I agree.

--smb

[sylvain kepler]

If it discusses ANNs on a highly abstract level then the DSPDim section might be a better place, I agree.

Let's put it in DSPDim..
Update : the thread has been moved here. I see that there were a very few threads. I presume this is a good reason and an obvious related Neuron topic to afford.

[neuronaut]

Sure. So, what do we discuss?

--smb

[sylvain kepler]

Yes. Maybe this can start by a basic question for you Neuronaut. Has Neuron core resulted from models and other stuff using Matlab for instance, and then coded in C++ ? Might these matlab function been C++ compiled to make some shared libraries featured in higher level C++ code etc...? As what concerns ANNs, I 'm thinking straight away about Matlab toolboxes, one of which is the ANN one...In electronics industry, most of time the Physical level is modelled with Matlab. And as for what concerns the coding, ie for circuits, code used is hardware design language stuff...but C++ as well for one of these HDLs you might know about eventually.

[neuronaut]

The technology was developed using an ANN CASE tool on the SGI Irix which was later ported to MacOS. It is a proprietary ANN/DSP toolkit that allows generation of ANSI C (and now C++) code from topologies and learning rules that can be user defined and include an arbitrary number of linear and nonlinear elements and DSP processes. Practically everything developed by Prosoniq comes from this tool (which is (c)1990-2006 myself ).

User interface etc. for the products is a different story, of course. The generated code doesn't provide a user interface (expect for some rudimentary debug log capabilities).

I used Matlab quite a bit in my "early days" but for my day-to-day work I find it unneccessarily complicated as well as counter-intuitive (1-based array indexing and such). Also, while there's an X11 version which kindof runs on MacOS X it is very unstable and just a p.i.t.a. to use, and I don't want to do my development on a PC if there's no good reason for it. Nothing against the PC as such (I do software development for Windows, too), but Macs are just easier for me to work with if I have the choice.

--smb

[sylvain kepler]

Now as for the Neural networks :

There a couple of network achitectures commonly used in the programming context which are :

-Backpropagation neural Networks
-BAM
-Adaline
-Self-Organizing Neural Network.

Were Some of these or mainly one of them used when programming Neuron Core ?

[neuronaut]

Well, I really can't talk too much about this here, but I can tell you that none of this list was used.

--smb