Not my idea, heard it somewhere. That the crucial difference between a human being and AI is that if you show a 3 year old kid one picture of a cat, a kid can recognize all other cats. Was it a lion or a tiger.
You can feed ML 5000 pictures of cats and it can recognize a cat in a picture with something like 95% confidence.
Have you done this test for real? My nephew calls everything which moves but not a human a "dog". In his world there are flying dogs and swimming dogs. Probably if he would see an elephant that would be a big dog, while a giraffe would be a tall dog. Now obviously he will learn the customary categories, but it is definitely not a one-shot thing.
> You can feed ML 5000 pictures of cats and it can recognize a cat in a picture with something like 95% confidence.
This is an area of active research. One term of art is "one-shot learning". The general idea is that you show 5 million things to the AI, but none of them are a harumpf, and it learns how "things" are. And then you show it a single image of a harumpf and tell it that it is a harumpf and it will be able to classify things as harumpf or not from then on.
How great do these things work? They kinda work, but you can still get a phd for making a better one. So they are not that great. But I wouldn't pin my hat on this one "crucial difference between a human being and an AI", because you might get surprised once humans teach the AIs this trick too.
As long as he can tell a "tall dog" (giraffe) apart from a "swimming dog" (say, a duck) that's still compatible with what the parent says.
It's about recognizing them as distict, and assigning them to the same class of things, the rest is just naming, that is, its at the language and vocabulary level, not at the recognition level.
For comparison GPT has like 1TB of text, and they're hundreds of thousands of books, articles, wikipedia, and so on. So already 3 orders of magnitude more.
And of course the "16 hours x 150 words per minute x 3 years" is totally off by a few orders of magnitude itself.
As a related example, 3 year old children often cannot reliably recognise basic shapes (letters from the alphabet), and certainly not after a single example. I daresay an ML model would outperform a child in OCR even with significantly less exposure to letter forms in its training.
When a child looks at a picture of a cat at the age of three, they have already learned to recognise animals, faces, fur, object, depths in photographs, the concept of a cat being a physical thing which can be found in three dimensional space, how physical things in three dimensional space appear when represented in 2D… the list goes on.
It’s simply not within our capabilities at the moment to train ML models in the same way.
I don't understand what you mean with this, as that was certainly not me as a child. As a child I thought cats and dogs might be differently sexed animals of the same species. I also thought that the big cats could be related to each other, though how any of them was related to the housecat was beyond me, given the size difference.
Sounds irrelevant to the parent's point. Which isn't that you knew what a cat was (with regards to taxonomy or whatever), but that you could tell one from a dog or a fire hydrant.
With HDC, a network learns to encode features it sees in a (many-dimensional, hence the name) hypervector.
Once trained (without cats), show it a cat picture for the first time, it will encode it. Other cat pictures will be measurably close, hopefully (it may requires a few samples to understand the "cat" signature, though).
It reminds me a bit of "LORA" embeddings (is that the proper term?), or just changing the last (fully-connected) layer of a trained neural network.