hi !

i want to reduce the number of inputs to my neural network in order to
reduce computational time...
data : 1000 points (10x10x10) (contains alot of zeros also)

method 1 (i ve tried this)

change the points to a 2 dimension matrix apply pca.. input into
neural network

any one has a better approach for me ?
kapil

What about simple feature selection?

-Will Dwinnell
http://will.dwinnell.com



k_taneja@softhome.net (kapil) wrote:
"i want to reduce the number of inputs to my neural network in order
to reduce computational time...
data : 1000 points (10x10x10) (contains alot of zeros also)

method 1 (i ve tried this)

change the points to a 2 dimension matrix apply pca.. input into
neural network

any one has a better approach for me ?"

simple feature selection ? how would i do that ?

thanks

Will Dwinnell wrote:
>
>
> What about simple feature selection?
>
> -Will Dwinnell
> <http://will.dwinnell.com>
>
>
>
> k_taneja@softhome.net (kapil) wrote:
> "i want to reduce the number of inputs to my neural network in
> order
> to reduce computational time...
> data : 1000 points (10x10x10) (contains alot of zeros also)
>
> method 1 (i ve tried this)
>
> change the points to a 2 dimension matrix apply pca.. input into
> neural network
>
> any one has a better approach for me ?"
>

kapil <k_taneja@softhome.net> wrote:
"simple feature selection ? how would i do that"

Throw predictor variables away.  There are many ways to go about it:
forward selection, backward selection, stepwise selection and any
number of other heuristics (genetic algorithms, etc.).  Probably the
simplest to implement is forward selection:

-Build single-predictor models, trying each candidate input variable
-Keep the predictor which performs best
-Build two-predictor models, trying each of the remaining predictors
with the best one
-Keep the best of the second predictors
-Etc. until some stopping criterion

Feature selection is often effective and has the advantage of not
mutating the variables into something uncomprehensible (as is often
the case with solutions like PCA, ICA, clustering, etc.).

-Will Dwinnell
http://will.dwinnell.com