[PD] ann_som users

[B. Bogart]

I Read about those CGC but not the neural gas.

Looks like the new SOM/CGC stuff in FANN is nice, I'll have to look into
that as it could make ann_som a very clean implementation.

Any idea if the network saving stuff was fixed? I would need that
working for my application.

Thanks for the heads up.

.b.



Jamie Bullock wrote:
> Hi,
> 
> On Mon, 2008-02-25 at 15:26 -0800, B. Bogart wrote:
>> Hey all,
>>
>> I just wanted to get a sense of the ann_som users out there.
>>
>> I would like to make some changes/additions, in particular:
>>
>> * Gaussian neighbourhood function to increase convergence with many sensors.
>>
>> * built in function for randomizing sensors (my network has about 4
>> million weights, which is pretty anoying to use "init" for.)
>>
>> Maybe more as things move along.
> 
> Nice to see someone working on this. I haven't used [ann_som] in
> earnest, but I have used [ann_mlp], which is great, and am generally
> interested in the progress of AI things in Pd. 
> 
> One thing it might be worth looking at is the SOM and Growing Neural Gas
> support that was added to the FANN library as part of GSoC last year.
> There is some discussion about it on this thread:
> 
> http://leenissen.dk/fann/forum/viewtopic.php?t=46&start=30
> 
> The code is also the FANN CVS under the gsoc2007 branch. I did get it to
> compile at one point but I haven't really tested. It may be in some way
> informative ?
> 
> Anyhow, I look forward to the [ann_som] improvements.
> 
> Jamie
>