[LOCKSUIT]

I have a question.

Can modern computer vision (or any know algorithm ex. a simple one that uses no backpropagation) see 1 image of ex. a cat and then if shown 10 dummy images - one of which does have an unseen cat - recognize which image has a cat - which is the cat is saw before but blurred, brighter, noisy, rotated, stretched, flipped, inverted brightness? This requires great accuracy at recognizing something it knows but that is very distorted.

I do have some beliefs of what the answer is but I really don't know, there could be some algorithms that do ace it but are not scalable for example. Or maybe Google's aces it but you just don't know what algorithm they use. Fill me in on what you know.

[LOCKSUIT]

And not allowed to use Data Augmentation i.e. when flip/ rotate/ brighten/ etc your own images you know already. Yes you do do that when generate internally probably true discoveries and save those memories, you need to store some things. Of course we don't store every Data Augmentation possible ex. every rotation, brightness, patches of such. We only store most important, the patterns. But this doesn't mean we store a few diverse views of each image feature, when you have another way that recognizes all the not saved and the ones in Data Augmentation...which i believe now i do have in sight.

[LOCKSUIT]

may have to read this better tomorrow:
https://towardsdatascience.com/a-friendly-introduction-to-siamese-networks-85ab17522942

[MikeB]

This might not help, but I would write a program that turns everything to vector lines / cartoons... that would remove any blur, and you can much more easily apply known vector shapes of cats and dogs over the top and see if they align. That's probably pure pattern matching though.