Backpropagation

Artificial Intelligence Programming

Started by nICO80 September 24, 2004 06:36 AM

5 comments, last by fup 19 years, 7 months ago

122

Author

September 24, 2004 06:36 AM

I'm trying to write a simple neural network with the backpropagation algorithm. After reading a thousand papers :) I think I have understand it quite good, but I miss something in the backpropagation step. I can calculate all the deltas and update weights, but what about the neuron bias? It will always be fixed or I should update it as if it were a weight? Thank you nICO

fup

463

September 25, 2004 12:31 AM

you can think of the bias as an additional input into each neuron, with a constant value of 1 (or -1, it makes no difference). Each neuron therefore, will have an additional weight associated with the bias input. It is this weight that the backprop algorithm should update.

My Website: ai-junkie.com | My Books: 'Programming Game AI by Example' & 'AI Techniques for Game Programming'

nICO80

122

Author

September 25, 2004 01:58 AM

Ok, thank you! It works!!!

Btw, can someone give me a link to some explanation on unsupervised networks?

nICO

evolutional

1,397

September 25, 2004 04:32 AM

If you look on AI Junkie, Mat uses A GA to self-train the network - is that useful for you?

nICO80

122

Author

September 26, 2004 12:07 PM

Yes, thank you, that's useful.
I also found useful information in the sites in the links section of the site. (I'm currently reading NEAT documentation... seems really interesting.)

bye
nICO

Anonymous

September 26, 2004 08:01 PM

Quote:Original post by fup
you can think of the bias as an additional input into each neuron, with a constant value of 1 (or -1, it makes no difference). Each neuron therefore, will have an additional weight associated with the bias input. It is this weight that the backprop algorithm should update.

Isnt each link weight feeding into the nodes Summation to be corrected by the backprop??