1
$\begingroup$

I am working through an example on the MNIST dataset, and was just curious, if your image input data were missing some pixels, how would you encode it. Since the values are always positive, and normalized between 0 and 1, would it make sense just to encode it as -1 or something.

Thanks

$\endgroup$
1
$\begingroup$

You'd substitute a statistical aggregator, such as the median or mean, for the missing data. Calculating the aggregator for a neighborhood region would "smooth out" the missing pixels.

You wouldn't set it to -1 because lost pixels are effectively noise, which you want to suppress, not exaggerate.

$\endgroup$
2
  • 1
    $\begingroup$ Thanks! What if it was a situation where you legitimately want the machine to learn 0 information from that missing pixel - basically ignore the pixel. Actually nvm, it's probably the same. $\endgroup$
    – rollback
    Aug 4 '20 at 4:11
  • 1
    $\begingroup$ I have never heard of such a method. I believe that aggregating the neighbors is as close to ignoring it as you can get. Even if you had a variable number of inputs, the mere omission of data is, in and of itself, information that will be learned. $\endgroup$ Aug 4 '20 at 4:14

Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.