One of these IS like the other

So, you’re using #MachineLearning to do image classification (of course you are ). The issue at hand is, How do you differentiate the important stuff from the background?.
It’s a bit of a trick question, because, well

• • What is “important”? That it’s a Cat (next to piglets)?
• • What is “background”? (You’re looking for piglets!)
  And so on.

In #DeepLearning, we call these “core” features vs “orthogonal” (or “style”) features, and differentiating between the two can be, well, hard. 

In a paper by Heinz-Demel and Menshausen (°), the authors come up with a neat trick to deal with this, one that requires a lot less effort than good old #SupervisedLearning — they assume that somewhere in the image is the uniquely identifiable thing you are looking for. For example These images have my piglets in them.
That’s pretty much it. Now that the model “knows” that my piglets are in there somewhere, it’ll happily ignore all the extraneous stuff including image quality, brightness, color, movement. Also, Cats .

It’s a seriously cool thing, and, like most awesome things, obvious in retrospect — check it out!

(°) “Grouping-By-ID: Guarding Against Adversarial Domain Shifts” — https://arxiv.org/abs/1710.11469v2