I've developed a neural network which takes in n inputs returning m outputs.

I want to see which inputs contribute most with each output. One idea I had is for all inputs/output combinations, lock the other inputs to some values, and vary the target input, analysing the target output. I'd then have a list of outputs (the same amount as the sample of inputs). I don't know if this is correct but I was thinking I could use the standard deviation of the outputs to measure how much this input affects the output. But then I remembered that all the other inputs are only locked to a very specific value. So I would then have to redo this whole process with every combination of other inputs. And then (again I don't know if this is right) average all of the standard deviations for an overall measure of how much this target input affects the target output

I..e. Let's say I use 1000 samples for my target input, I would have to do this for 1000^(n-1) input combinations. Essentially, this process has a complexity of s^n where s is the sample count. This seems very high, is this necessary? Also would this whole process, in general, really provide a good way of seeing which inputs affect each output?


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.