I think the point of the article is that there isn't 2^(32x32x3x8) meaningful images that can be encoded - there's far less than that because so many of the possible arrangements of the data are just meaningless noise, or millions of versions of the same thing with slightly altered color values etc.
The article is saying that neural networks are cool, because part of what they do is finding which images actually contain meaningful visual information.
But they don't, they often conflate noise with actual features, as long as the noise has some statistical bias, which given enough random generations, will. Try it yourself, generate random images using a uniform distribution (not gaussian) and run a SOTA classifier on them. Eventually you'll hit some minor false positives.
It errs in the same way we often see objects in clouds or constellations.
I got the point quickly enough (from the title and initial table of snapshots?), but author has lost me as a reader due to saying little with too many words before they could elaborate on the point.
Such nonsense drivel. How much an image contains can be inflated by doing some dishonest BS like creating a unique encoding to optimize for what each image is representing in RGB.
You can clearly see with steganographically encoded images that there is a limit if you want the image to remain and image - and still be able to store whatever arbitrary data in it. Being bound to the arbitrariness of data that can be stored is against the implications of the title, which is why this should be deemed nonsense.
If you want to look at the deeper questions the title invokes, this is the scam of how ThisPersonDoesNotExist and DALL-E work.
They deliver information to the brain, not actual pictures.
What they have created is amazing, but other than fooling other AI's their use is limited. They won't scale, they have gone the direction of information over fidelity.
Many moons ago, I wondered what it would look like if you iteratively generated every possible image. It doesn't sound very useful at all (and it certainly isn't) but I learned C++ and SDL that way.
I've thought about this as well; it's one of those weird things where you realize that with enough pixel count and shading options, every image you can think of exists within the total set of generated images.
"enough" hides a surprising amount of work. Mobilise the entire planet's computers for a decade, assuming a generous Moore's law increase as well, and you might count through seventeen 8-bit grayscale pixels in that decade.
Annihilate the Sun and you won't get enough energy to drive a simple counter through thirty 8-bit grayscale pixels.
I think about it from the perspective of ideas, once we figure out something (how to fly, electricity) we know it. How do you brute force find that info through random/massive computation. Specifically to find the next things we don't know like jumping dimensions or FTL, etc...
Would this be a good way to avoid censorship. Encode an image by mixing it with a private key image, maybe xor the cells or something like that. Post online. Anyone with the key image can decode.
I sort of came at the combinations of small images from a slightly different angle when I was a teenager, I was imagining that if there was a movie of your life (and everyone elses, in fact every possible life you could live) then you could produce every still image of that video if you restricted the bounds of the image to something small but understandable. I was very quickly was disappointed at how big the numbers get :)
wonderful, it looks like people are talking with Zoom, kidding. a image can contain lot more information if it is designed by an professional artist. an image have an message for everyone.
First, article tried to assess information capacity in Shannon style, completely disregarding what signal is (a) two-dimensional and (b) highly redundant, that is, ignoring the traits of being an image.
Second, article taken too much liberties while mixing photorealistic and pixel-art images. The latter is really an art, since there is no formally defined ("machine") transform between these two types. And last but not least these types have significantly different information density profiles.
29 comments
[ 3.2 ms ] story [ 70.6 ms ] threadThe article is saying that neural networks are cool, because part of what they do is finding which images actually contain meaningful visual information.
It errs in the same way we often see objects in clouds or constellations.
You can clearly see with steganographically encoded images that there is a limit if you want the image to remain and image - and still be able to store whatever arbitrary data in it. Being bound to the arbitrariness of data that can be stored is against the implications of the title, which is why this should be deemed nonsense.
Use the whole space [32x32x3 bytes = 3072 bytes] to encode a Busy Beaver.
https://en.m.wikipedia.org/wiki/Information_theory
Specifying an element from a set of 2^n members seems equivalent to encoding n bits of information.
They deliver information to the brain, not actual pictures.
What they have created is amazing, but other than fooling other AI's their use is limited. They won't scale, they have gone the direction of information over fidelity.
This is probably terrible code by anybody's standards but maybe someone wants to take a look: https://github.com/svenstaro/infinerator
Annihilate the Sun and you won't get enough energy to drive a simple counter through thirty 8-bit grayscale pixels.
https://security.stackexchange.com/a/6149/
...except that we also do it by math. It is just so fast and optimized, that we refer to it as instinct.
First, article tried to assess information capacity in Shannon style, completely disregarding what signal is (a) two-dimensional and (b) highly redundant, that is, ignoring the traits of being an image.
Second, article taken too much liberties while mixing photorealistic and pixel-art images. The latter is really an art, since there is no formally defined ("machine") transform between these two types. And last but not least these types have significantly different information density profiles.
No DSP curiosity here.