What is the best way to deal with “None of the Above” in “Image Classification”?

Question

This seems to be a fundamental question that some of you should know. I have an image classifier implemented in CNTK with 48 classes. If the image does not correspond to any of the 48 classes, then I would like to conclude that it was not among these 48 types of images. My initial idea was that if the highest performance of the final Softmax layer were low, I would be able to conclude that the test image did not match anyone. Although sometimes I see this happening, in most tests Softmax still produces a very high (and erroneous) result when transmitting an "unknown image type". But maybe my network is “full”, and if it’s not, my original idea will work fine. What do you think? Any way to determine the 49th grade,called "no-top"?

Answer 1

In fact, you really have these two options - probability thresholds (softmax values) and adding a garbage class.

In my area (speech), both approaches are their place:

" " "" (, ), . , . - < 50% , , . , , , , .

" " , (, 10 ) (, ), . " ". , , . . , - , , , , , , softmax .

, , . ( , ), , , . , , , (, , , ). , . .

Answer 2

, , .

Zhang LeCun, , , . " ". , , "" , . , , " " .

- -hoc, softmax, .