In Week 3
2. Survey of major AI techniques (optional),
@4:38: " .For those of you that are parents, I can almost guarantee to you that no parent on this planet, no matter how loving and caring, has ever pointed out 10,000 unique coffee mugs to their children, to try to teach the children what is a coffee mug. So, AI systems today require much more labeled data to learn than when a human child or than would most animals."
Sure you can show just one image of an object on a table to a human and tell him the name and the kid can already tell you if the object is present or not in a picture or in the room, but humans have trained on huge amounts of data since birth (if you count 24 images per secyearsnumber of color detection cells in the eye, just to mention vision). Yes, a very small part is ‘labeled’ (if this means having the image presented with the corresponding to a text name or a sound name), but we still train on huge quantities of such non completely unrelated ‘labeled’ data in the first place. When you show a cup to your kid, as mentioned in the video as example, you don’t just show it one image for 1/24th of a second, and it’s not the first time he has seen such an object in the past. You don’t show him 10000 different cups, but he will still see in a matter of seconds way more data than 10000 different low-resolution images, even just from the cup you showed him. You are likely going to let him play with it or with a toy representation of a cup, and he might forget the name of the object if you mentioned it only once, but get a huge understanding of the concept of this object that no AI given current technology/processing power can ever hope to achieve.
That is to say, if you give a large enough network enough data for long enough to the degree of what a human is having/exposed to, for sure you could get at least the same result: pretty good recognition of an object by the algorithm even with a single labeled example (and understanding of the concept by letting it play a bit with it.) and not need to show 10000 examples…