MATLAB: How to estimate sensitivites of a neural net model with exogenous variables

Deep Learning Toolboxneural networks

Is there a way to estimate the slope (first derivative) of the response function of a neural net model with respect to various inputs around a point? The measure I am looking for is similar to the beta coefficients in a linear regression.

Best Answer

Differentiate the corresponding equation. For example, differentiate w.r.t. x:

 y = b2 + LW * tansig( b1 + IW * x )

Related Solutions

MATLAB: Names of variables

IF those input arguments had names in the calling workspace of this function (which is not guaranteed) it is possible to retrieve those names using the inputname function.

However, I would be hesitant to use this information other than perhaps to use in an error message or text that you display as some sort of status. If you're thinking to use it with assignin to modify variables in your caller's workspace, I STRONGLY recommend against that unless there is no other way to do what you need (and even then, post about what you're trying to do on Answers to see if others can think of a better solution.) Doing that is likely to make debugging very difficult later on, and is an example of common coupling or perhaps even a form of content coupling. Prefer data coupling instead if possible.

MATLAB: Dynamic comments on variables

This sounds like a magic meta programming. I recommend to avoid such smart tricks and keep comments in the source code as usual. Bot of course it is possible:

obj.pmExpo.iter_comment = ['this is a comment indicating the use of this variable, ', ...
                           'so I do not need to go back to the code to know what it does'];

And now the function to display a field:

function MagicDisp(S, Field)
sprintf('Field: %s\n', Field);
disp(S.(Field));
Magic = [Field, '_comment'];
if isfield(S, Magic)
   fprintf('%s\n', S.(Magic));
end
end

Now call it as:

MagicDisp(obj.pmExpo, 'iter')

Well, the name "MagicDisp" implies, that I would not use it for productive code. The function could be enhanced, but I would not do it, because variable which do explain their meaning actively are too tricky to be useful. What would happen if you run a larger code, which uses 200 of such variables and process them in a loop?

If a variable should do anything actively, that it would be starting the daily backup. This is more important, forgotten too often, and the explanations can appear in the code instead.

Best Answer

Related Solutions

MATLAB: Names of variables

MATLAB: Dynamic comments on variables

Related Question