Sign In
Create Account
Follow Us on Twitter
Subscribe on Youtube
OK..some very basic questions here... so please indulge me
Firsly.... the delta W that is computed for each weight... it is simply added to it right? The delta rule takes care of the sign I am assuming?
second: the weights can be negative or positive right?
Thanks!!
4 replies to this topic
Back to top|
|
|
#2
Posted 07 November 2009 - 12:08 PM
WingedPanther
-
- Moderators
-
- 16,831 posts
A spammer's worst nightmare
You'll have to provide a little more context before anyone is likely to be able to answer your question.
#3
Posted 07 November 2009 - 12:39 PM
ibad
-
- Members
-
- 39 posts
Learning Programmer
Ok..context....
delta W is delta-weight... the change in weight...calculated with the delta rule.
I am using sigmoid function neurons, it's a simple MLP with each layer fully connected to the next one. It's going to be used for letter recognition (very simple OCR).
The net is shown a 32x32 graphic file which has a black and white charcter in it from a to b (lower case).
It is feed-forward. erm....what else... yeah the outputs are binary encoded. 7 output neurons, so the output is supposed to be ascii.. output neurons are also sigmoid. Any output value above 0.9 is considered a 1, any below 0.1 is considered a zero.
And... the weights are reals, well doubles ... supposed to simulated real valued weights. So I am asking if the weights can be negative or not (I am a beginner...so this might be very simple for you).
I am also asking if the delta-W that is calculated with the delta rule produces is already sign corrected... in which case it is simply added to the current weight that is being adjusted.
I am thinking that the answer to both questions is yes..the weights can be negative, and the delta-W is just added... but I don't really have any guidance, am teaching myself... so I just wanted to be sure.
delta W is delta-weight... the change in weight...calculated with the delta rule.
I am using sigmoid function neurons, it's a simple MLP with each layer fully connected to the next one. It's going to be used for letter recognition (very simple OCR).
The net is shown a 32x32 graphic file which has a black and white charcter in it from a to b (lower case).
It is feed-forward. erm....what else... yeah the outputs are binary encoded. 7 output neurons, so the output is supposed to be ascii.. output neurons are also sigmoid. Any output value above 0.9 is considered a 1, any below 0.1 is considered a zero.
And... the weights are reals, well doubles ... supposed to simulated real valued weights. So I am asking if the weights can be negative or not (I am a beginner...so this might be very simple for you).
I am also asking if the delta-W that is calculated with the delta rule produces is already sign corrected... in which case it is simply added to the current weight that is being adjusted.
I am thinking that the answer to both questions is yes..the weights can be negative, and the delta-W is just added... but I don't really have any guidance, am teaching myself... so I just wanted to be sure.
#4
Posted 07 November 2009 - 05:42 PM
WingedPanther
-
- Moderators
-
- 16,831 posts
A spammer's worst nightmare
I'm thinking the same thing, but haven't done any work on Perceptron's (or neural nets).
#5
Posted 04 March 2010 - 02:08 AM
Predictor
-
- Members
-
- 20 posts
Newbie
ibad said:
OK..some very basic questions here... so please indulge me
Firsly.... the delta W that is computed for each weight... it is simply added to it right? The delta rule takes care of the sign I am assuming?
second: the weights can be negative or positive right?
Firsly.... the delta W that is computed for each weight... it is simply added to it right? The delta rule takes care of the sign I am assuming?
second: the weights can be negative or positive right?
Yes and yes.
