View Full Version : Neural network implementation?

07-12-2008, 05:01 AM
Hi guys, with all this conjoined intelligence wandering around here, i thought i'd ask a question that's been bugging me for a while now.

I've this thing in my head about 'neural networks' as i see them, a cell with multiple inputs, one output, and a variety of weight factors on each input.

Kind of like an OR port with multiple inputs, only each input doesn't have a digital signal, but an analog one.

So, say input 1 has a weight of 3, input 2 a weight of 5.

A certain threshold level is used to trigger the output, say at an input level of 6 the output is triggered.

This means that input 2 will affect the output stronger than input 1.

Now the question is, how does one implement this in software?

Is it a simple matter of comparing (Input1*weight) + (Input2*weight) with the trigger level?
Or am i missing something here?

07-12-2008, 09:59 AM
A fair bit... Error propogation for one biggie. Picking the right topology for a neural network is such a complex question that we frequently write genetic algorithms just to analyse the problem against the topology. :)

Do you code in C? I have some old book code that made following along easier.

07-12-2008, 10:27 AM
Here's a pretty readable example of what goes into the simplest form - backpropogation.


07-12-2008, 11:02 AM
Thanks, that'll keep me busy for the next few weeks :wink:

We were given only the slightest intro to neural networks in school (ie. somewhat what i described above) but hadn't delved into it much further

07-12-2008, 11:15 AM
I have a bunch of old stuff from the late 80's and early 90's that was from a much simpler time. Most of it should still build on Linux with prehaps a little tweaking. Let me know if you want it...

07-12-2008, 11:18 AM
that'd be much appreciated, if it doesn't involve too much trouble though!

07-12-2008, 11:19 AM
I'll package it up today and toss it in the member uploads.

07-12-2008, 11:25 AM
Great, thanks Adrenalynn!

On a side note, my sister just gave birth to her first daughter, I'm an uncle now.
Wiiiiiiiii !! :veryhappy:

07-12-2008, 11:44 AM
Congrats! When you going to register her here on the forum and get her started in robotics? :)

You need to build a robotic nanny now.

07-12-2008, 11:56 AM
Well, the bioloid kit should arrive on monday, so uhm... tuesday? :veryhappy:

No matter what amazingly complicated robots are built by a team of hundreds of brilliant engineers, nothing can leave a person in awe like a newborn grasping your finger..

07-12-2008, 12:02 PM
That's one side. The other side is that no matter how much toxic waste is dumped into rivers and streams, no matter what the half life of a nuclear warhead, nothing can leave a person in awe like what a baby puts into its diapers. :D

07-12-2008, 12:18 PM
Lol, all the more reason to wait a lil longer with my first-born..

07-12-2008, 03:03 PM
I can't let Adrenalynn have all the fun!
I've had my graduate classes in AI about 2 years ago. Most of the concepts have stayed the same with Neuro nets. (except one change)
My former brilliant instructor has left her homepage up... she has some great links regarding AI. I would suggest downloading items because it may not be there for long. Don't forget to scroll down and check out her links on Neuro Nets. And if you have questions, let me know (But be warned... it's been a while since I've had to use it!) Most of my uses were involving computer vision instead of robotics.

07-12-2008, 03:06 PM
No matter what amazingly complicated robots are built by a team of hundreds of brilliant engineers, nothing can leave a person in awe like a newborn grasping your finger..
For me, it was my two year old saying "I want a robot!" :tongue:

07-13-2008, 10:34 AM
Information overload.. I'm gonna look in to plugging a harddrive into my brain or something..

Thanks Eric!!

12-10-2008, 01:50 PM
I know this is an old thread... but like me someone might come along.

We use neural nets in our music software... to classify and guess which song the user might like based on his choices.

The best free neural net out there is FANN http://leenissen.dk/fann/

This sucker is so easy to implement and its so so so good, ya and fast.

Anyway, if you are talkling NNs then you need to check out FANN. Comes in 16 flavors!

laters :)