Neural-Network-Experiments icon indicating copy to clipboard operation
Neural-Network-Experiments copied to clipboard

have a non-identifiable class.

Open madhaven opened this issue 2 years ago • 2 comments

Hey, I think having a class that is "non-identifiable" helps avoid classifying images that it hasn't seen?

madhaven avatar Aug 30 '22 05:08 madhaven

As far as I'm aware, the program gives a confidence value for each thing it is aware of. A "non-identifiable" type is implicit if it can give low confidence to every possible type.

Eg:

80% confidence on both tractor and cruise ship

  • AI thinks it could be either

80% confidence on tractor, but 12% on cruise ship

  • AI thinks it's a tractor

12% confidence on both tractor and cruise ship

  • AI doesn't think it's either

MaddyGuthridge avatar Sep 02 '22 04:09 MaddyGuthridge

it just never does that because the the answer is never neither

if you wanted it to be unsure more you would have to add random images to the data sets

(I think, don't quote me on this, I don't really know what I'm doing)

Lachy-Dauth avatar Jan 31 '23 09:01 Lachy-Dauth