| View previous topic :: View next topic |
| Author |
Message |
Artificial Neural Network
Guest
|
 Posted: Tue Nov 11, 2008 4:03 pm Post subject: More Questions About neural network........... Please, Thank |
 |
|
Hi all,
When I read about the subject I saw that it's very common to use the
function -
V = e^v / ( 1 + e^v )
(the sign ^ means "power of...")
to decide if the neural unit Should or Should Not fire a signal in
it's output, in what way is this function better than a random
threshold that only above it the unit will fire? the input of this
function can be 20, or 274, or 28520...... can the function handle
with any value?? it's sounds a little strange.... when I use this
function which values will make it 'Fire' and which values will not?
And another question, can you explain please in a simple words as much
as possible what is the meaning of the Bias parameter? when I first
give it a random value (in the Init functio) what are the limits for
this value? should it be 0-100 ? should it be 0-10,000 ? how do I know
what are the limits for this value? (if it's too small the unit will
'Fire' for any signal, if it's too high the unit will never
'Fire'..... so how do I know what value to give there?
I realy want to start building my application, so please don't send me
to go learning, try to explain it in simple words, later on I'll try
to read more about it, but now I'm really looking for some
answers.....
Thanks! |
|
| |
|
 Back to top |
Greg Heath
Guest
|
| Posted: Tue Nov 11, 2008 8:59 pm Post subject: Re: More Questions About neural network........... Please, T |
|
|
On Nov 11, 11:03 am, Artificial Neural Network 3000
<artificialneuralnetwork3...@gmail.com> wrote:
| Quote: | Hi all,
When I read about the subject I saw that it's very common to use the
function -
V = e^v / ( 1 + e^v )
|
= logsig(v) % logistic sigmoid
| Quote: | (the sign ^ means "power of...")
to decide if the neural unit Should or Should
Not fire a signal in it's output,
|
No.
It is a continuous function that always "fires".
However if v is a linear combinations of other
variables, it can be scaled so that V makes a quick
transition between "very small" and "almost one".
| Quote: | in what way is this function better than a random
|
replace random with "specified"?
| Quote: | threshold that only above it the unit will fire? the input of this
function can be 20, or 274, or 28520...... can the function handle
with any value??
|
It is better because it is continuous. Therefore,
it can handle slow as well as abrupt transitions
and approximate smooth functions better.
More importantly, it can be trained directly
using the chain rule for derivatives.
| Quote: | it's sounds a little strange.... when I use this
function which values will make it 'Fire' and
which values will not?
|
It learns the values from training with labeled
samples.
| Quote: | And another question, can you explain please in a simple words as much
as possible what is the meaning of the Bias parameter?
|
Tne negative of the threshold T when
v = w1*x1 + w2*x2 - T.
Therefore, it controls when v changes from "very
small" to "almost one".
| Quote: | when I first
give it a random value (in the Init functio) what are the limits for
this value? should it be 0-100 ? should it be 0-10,000 ? how do I know
what are the limits for this value? (if it's too small the unit will
'Fire' for any signal, if it's too high the unit will never
'Fire'..... so how do I know what value to give there?
I realy want to start building my application, so please don't send me
to go learning, try to explain it in simple words, later on I'll try
to read more about it, but now I'm really looking for some
answers.....
|
You don't "give" weights and thresholds final
values. Instead, you give them very small random
initial values and use numerical optimization to
find final values which minimize the sum of the
squared differences between network output and
target output values obtained from labelled
training data.
See the introductory text references (e.g.,
Fausett) in the FAQ.
Hope this helps.
Greg |
|
| |
|
| Back to top |
Artificial Neural Network
Guest
|
| Posted: Tue Nov 11, 2008 9:16 pm Post subject: Re: More Questions About neural network........... Please, T |
|
|
Thanks very much Greg, and yes it's realy helps!! :-)
Right after I wrote my questions I start reading the following FAQ -
http://www.faqs.org/faqs/ai-faq/neural-nets/part1/preamble.html
and it also answered lot of my questions.
Thanks man! |
|
| |
|
| Back to top |
|
