The point is that the networks use an entirely different set of criteria than humans for recognizing images. Where humans are able to recognize the "whole", these networks demonstrably become trained to recognize small, specific patterns that distinguished the object from everything else in the training set.
@Kenshin I'm not going to discuss this with you unless you demonstrate you've read what I linked. It clearly shows that contemporary image recognition networks do not recognize the whole.
the human brain is thought to work in a similar ay
@ACuriousMind I've read what you linked
I've also read that the researchers in the future may try to encorporate a "none of the above" classification but so far it has been too difficult for them
"Thus, they effectively have a way to communicate “none of the above”. It would be interesting to try training them with a specific “none of the above” class, but that would require assembling a set of pictures to put in that class that we are sure do not belong in one of the other 1000 classes. We’d like to do that, but assembling these images will take some time. Stay tuned."
@ACuriousMind From the article:Does this optical illusion phenomenon happen to animals or humans? Yes. Humans are susceptible to optical illusions. Such illusions are designed to hack the way our brains see the world. Similarly, these images hack the way neural networks see the world.
@Slereah Yeah, I'm pretty sure that if that was a viable fix it would long have been deployed
@Kenshin Yeah. But optical illusions that fool us don't fool the networks, and vice versa. So they're not "good at what humans do". We both recognize images, but in entirely different ways.
@ACuriousMind yes but we were trained differently to the networks. The networks are trained with pictures and trained to identify specific classes of objects. Humans learn to idnetify millions and millions of object classes and learn from a wide vareity of stimuli, not carefully selected images.
So I'm not convinced it is the structure of the neural network itself that makes it behave differently to humans, but rather how these particular neural networks were created
As Slereash states, the "nodes" in the brain are evolved to already have concepts for light, colour etc.
@Slereah There are the wonderful images where you think the upside-down face is smiling, but it's actually horrifying when viewed in the correct orientation
I think the big thing that differentiates us from current neural networks for image recognition is we can tell "That looks like a face but it's not a face"
Actually, there is another building which collapsed in 911. It is the No.7 building, with 47 floors, not hit by an airplane.
All the three buildings collapsed in a very regular way.
Frankly speaking, I find this weird. Look at the remains of the old buildings around the world. Generally, some...
What are some easy to perform physics experiments that can be done at home (with not too much special equipment) and that allow to actually measure/plot data and draw conclusions from it?
My son is pretty interested in everything physics and we've done just about all possible (qualitative) exper...
Why the short wavelength limit of x-ray produced is independent of nature of target but the characteristic spectra depend on its nature?
Can I extend the concept of Bragg's diffraction condition to nanoscale level?
Actually, there is another building which collapsed in 911. It is the No.7 building, with 47 floors, not hit by an airplane.
All the three buildings collapsed in a very regular way.
Frankly speaking, I find this weird. Look at the remains of the old buildings around the world. Generally, some...
Why the short wavelength limit of x-ray produced is independent of nature of target but the characteristic spectra depend on its nature?
Can I extend the concept of Bragg's diffraction condition to nanoscale level?
