Tech Off Thread

5 posts

Forum Read Only

This forum has been made read only by the site admins. No new threads or comments can be added.

Neural Network - Basics?

Back to Forum: Tech Off
  • User profile image
    Unoriginal​Guy

     

  • User profile image
    DoomBringer

    This is an fairly basic but OK explanation I suppose.  I mean, one thing you don't seem to mention is that each node should weight individual inputs.  Changing these weights is what makes a NN adapt or learn.  (By weight I mean value one input more than another)

  • User profile image
    Rossj

    Sven posted a sample NN application (with source) a while back, I believe he used them quite a lot during his PostGrad ...

  • User profile image
    Frank Hileman

    http://en.wikipedia.org/wiki/Artificial_neural_network

    In my own work with neural networks they were not "fired" in the sense you describe. They instead exist in degrees of on and off. They use floating point numbers for inputs, outputs, and the internal node state, so it can be "fully off" to "fully on" to anywhere in between.

    I see no point in having a neuron with only one input and output.

    What happens is the neruon sums its inputs, a weighted sum according the to strength of the connection (determined during training), then applies a function to this sum. Assume that the sum can only result in a number between 0 and 1 -- it is weighted by each connection, as well as normalized to 0 to 1. The function that is applied is usually S-shaped (sigmoid) so that the neuron tends to turn on or off more fully. After applying the function you have the degree to which the neuron is on or off, this value then goes out to all outputs, which then do the same thing.

    The only thing that makes this interesting is the training, which allows you to create a kind of pattern matching machine, with certain inputs producing certain outputs, as trained, but in-between inputs, producing outputs that are interpolated.
     
    For example I once trained a neural network to recognize time signatures based on the accent pattern of a rhythm. The network could then determine the degree with which a random percussive performance matched various signatures. This output was fed into a similar neural network trained in reverse (signature to rhythm) and the second network produced an accompaniment.

  • User profile image
    qswitch426

    Hey Guy...

    I wrote a series of articles on Algorithmic Intelligence in Scientific Computing Magazine back in 2004.  The March 2004 column is available in the archive section of the site.

    http://www.scimag.com/PRArchivebyIssue.aspx?RELTYPE=DA&YEAR=2004&MONTH=03

    It is written for computer professionals, but is more of a concept piece and does not obscure things with math.

    ...I hope it helps!  -Bill  =]

Conversation locked

This conversation has been locked by the site admins. No new comments can be made.