[infurl]

https://www.digitaltrends.com/news/new-style-ai-learns-things-differently/

a new research paper from the University of Waterloo in Ontario describes a potential breakthrough process called LO-shot (or less-than-one shot) learning. This could enable machines to learn far more rapidly in the manner of humans. That would be useful for a wide range of reasons, but particularly scenarios in which large amounts of data do not exist for training.

https://www.researchgate.net/publication/336316939_Soft-Label_Dataset_Distillation_and_Text_Dataset_Distillation

This article goes into more detail about how it works.

https://techxplore.com/news/2020-10-math-idea-dataset-size-ai.html

[LOCKSUIT]

Hmm, seeing a cat once and recognizing it in other images with different rotations etc.

And seeing a horse, and a bird, and recognizing a Pegasus because someone told you it is halfway between.



Well, yes we have various in between features, so highlighting one, then seeing it, you will trigger its name most. What this is is a feature given, someone saying it is a Pegasus, needs to be "shown/said", this itself is a observation seen.

As for the other problem, dependless of rotation or missing parts or rearranged features, it still gets activated some. Is there various keys to make this happen though? Time Delay acceptance, translation (you see fish near a hairless cat and normal cat - so you think hairless cat is cat despite having less internal parts matching - mostly external contexts as clue), recency - i just saw a cat - this may be a cat - it IS 8% cat and that doesn't always happen bruh! So yes then, it depends on various methodologies, to see 1 example and recognize a cat in example 2, but wait, that means, can we do translation? Yes, 1shot, um, well, not helpful but at least we can a bit! Word2Vec needs lots of examples to discover cat is similar to dog. Man, with 1 image of a cat, not showing its face but only its tail, but its face only in a 2nd image, you can get the face from the first by using reflections of objects in the house even though not mirrors perfectly, there is a algorithm that kinda does this but is in early stages.

[MikeB]

I know nothing about this, but I'm amazed there's no vector drawing involved yet (dot points joined by curved lines). If you show a child a cat/dog, then ask them to draw it, it'll be a crude vector drawing...

Then match images to a % of how close it is to the vector

[ivan.moony]

I might be wrong, but I believe intelligent vector graphics generation would require a combination of ANN with GOFAI. Not sure how, though, but the idea is just flashing in my mind.