MATLAB: Neural network sensitivity analysis inputs

neural networkneural networks

Is there a command to the neural network, which tells me that input variable is more important in the equation generated?

Created a network of 10 neurons, and analyzing the weights is very difficult. Exist a command that MATLAB make me a sensitivity analysis ?

XOXO

Best Answer

No. However, I have gotten very good fast suboptimal results by starting with polynomial models of standardized (zero-mean/unit-variance) variables because polynomials are linear in the coefficients. Typically, I start with backward and forward selection of linear and quadratic models. Relatively useless variables are often readily discarded at this stage before continuing with backward and sometimes forward selection with neural models.

Hope this helps.

Thank you for formally accepting my answer

Greg

P.S. Sizes of MATLAB EXAMPLE regression, time-series and classification datasets can be obtained from NEWSGROUP threads 337914, 338208 and 339984, respectively.

Related Solutions

MATLAB: How to train a feedfordward neural network with error weights

 Use the command window
 help mse
 doc mse

Hope this helps.

Thank you for formally accepting my answer

Greg

MATLAB: Sensitivity analysis ANN

Please clarify your terminology. What are some examples of input parameters? Ditto for output parameters.

For sensitivity analysis of input variables and weights:

Standardize (zero-mean/unit-variance) training inputs and use the original means and standard deviations to normalize the validation and test sets.

Repeat the process for regression net outputs.

For pattern recognition nets with c classes, use one-of-c binary normalization (each column of the target matrix is a column of the c-dimensional unit matrix).

The easiest way to test variable sensitivity is to randomly reorder the row for the variable in question and obtain the resulting summary statistics (e.g, min/median/mean/std/max or quartiles) of the chosen MOP (i.e., measure of performance ; e.g., normalized mean-squared error, NMSE or the coefficient of determination R^2). Training data MOPs should be degree-of-freedom adjusted to mitigate the optimistic bias. Validation and test set estimates are assumed to be unbiased.

The easiest way to test weight sensitivity is to repeatedly randomize the weight in question and obtain the resulting summary statistics.

These techniques have been used iteratively to eliminate ineffective variables and/or weights. As usual for greedy searches, the results are suboptimal but are obtained relatively quickly.

Hope this helps.

Greg

Best Answer

Related Solutions

MATLAB: How to train a feedfordward neural network with error weights

MATLAB: Sensitivity analysis ANN

Related Question