How to implement feature selection on a fitting ANN?

I want to use feature selection on my neural network model. The sequentialfs function seems to be the perfect tool to do so, but I don't know how to incorporate fun the function handle of MSE. So my question is, is it possible to use MSE as the function that defines the criterion used to select features and to determine when to stop. And if so, how can I achieve that? FYI: MY ANN model takes as input a 6 columns matrix with 5700 observations. I use MSE normalized (percent) to quantify its performance.

[Greg Heath]

WOEFULLY INSUFFICIENT EXPLANATION:

Is this a regression problem or a classification problem?

What are the inputs and outputs?

size(input) = [ I N ] = [ 6 5700 ]

size(output) = [ O N ] = [ ? 5700 ]

Need more details

Greg

[Mohammed Hasan Goni]

I am also having the same Problem. my Problem is Neural Network Regression Problem.Where I have Inputs Inputs 12*156106 Output is 1*156106, where I am trying to remove some the Inputs where the prediction error(RMSE)is less than some amount. Or itcan be said that, I am trying to find something which can be used in random forest PredictoirImportance Option.

Answer 1

[Greg Heath]

Although there are fancy ways to rank variables for neural networks, most are (in my opinion, too complicated to waste time over).

What I have found is that sufficiently good variable rankings can be obtained by using simple 1st and 2nd order polynomial models with MATLAB variable selection algorithms.

In doing so I first look at linear and linear with squares before considering cross products.

Those 3 results tend to yield good enough information for rejecting variables with low prediction ability.

Hope this helps.

Thank you for formally accepting my answer

Greg

[Xinzhe Yuan]

Hi Greg, I know it's a long time after your answer. But I am really new to the feature selection topic and found your answer is propaply helpful to my problem. My question is can you be more specific on the MATLAB variable selection algorithms? Like what are they? Any insight is very appreciated.

[Greg Heath]

Sorry, it's been too long.

Greg