human brain learns things at a drop of a hat. single exposure, lesson learned. that's the goal. - where we're heading now is the OPPOSITE of that, we're getting further and further away from the right way of solving this problem.
Well, strictly speaking human brain need more than a single exposure. But yes, a single example/item/iteration could still be enough.
But I disagree with the fact that AI are heading the opposite. It's way more simple than this: They
are the opposite of us, period.
Take any average 4yo, he know how to draw, how to sing, how to tell a story. It's mostly innate. The instant he understand how to hold a pencil, he will draw. The instant he get aware of his own voice, he'll starts to sing. And the instant he'll starts to have some vocabulary, he'll starts to tell stories.
What humans need to learn, is the capability to reproduce something. Because child drawing are, well, what they are... Their stories mean nothing, and when they starts to sing, you want to die.
But, as I said, AI are the opposite of this. Reproducing something is their nature, what they need to learn, is how to draw, how to sing and how to tell stories.
There's probably hundreds different way to tell a software, and therefore an AI, to reproduce something. For a drawing, it can goes from a basic "copy bit after bit", to something more complex that would involve selections and masks. Imagine Photoshop, where you use the auto-select tool to keep only the girl in the image, and then copy/paste her into another image. A software don't need much to be able to auto-select all by itself, by example based on contrast or gap between the colors.
But a software wouldn't be able to draw all by itself. It don't just need the right algorithms to be coded, it also systematically need instructions for that. Because the algorithm can only tell it how to draw a geometric figure, how to fill it with color, plain or gradient, not where and when to do all this.
And, of course, on top of this there's the main ability that differentiate between humans and machines, an ability so well known that it's precisely used to make that difference: figures and patterns recognition.
Once again it's something that is innate for humans, that need really few exposure to be able to always recognize a figure in the future, even when a bit distorted, blurred, or when it come into a different form; think about cars by example, whatever if it's a Ford Model T, a Ferrari or a Cybertruck, you recognize it as being a car.
But it's something that software, and so AI, have to learn. And need to learn it again for every single figure, and for any variation of those figures. Once an AI recognize a Ford Model T as being a car, it still have to learn that a Ferrari and a Cybertruck are also cars.
So, as I said, AI don't head at the opposite, they are already fully at the opposite, because it's from where they starts.
Strictly speaking, AI should be trained three times. Firstly to gain figure/pattern recognition abilities. Then a second time to gain the capability to draw, sing, and tell stories (to limits to those three). Then finally to reproduce through those two capabilities. And in between each training, it should only keep in memory the knowledge related to what it was trained for.
But for this, it need that we understand what happen inside the black boxes... And, while there's apparently few progress, it's still far to be the case.
