Asking machines to identify images

  • Thread starter Thread starter Ryan_m_b
  • Start date Start date
  • Tags Tags
    Images Machines
AI Thread Summary
Researchers have utilized image recognition neural networks to identify and modify images within random noise, creating a feedback loop that results in bizarre and beautiful images after multiple iterations. This process, while fascinating, raises questions about the nature of pattern recognition, as some argue that the neural networks are detecting non-existent patterns initially. Traditional pattern recognition methods, such as Fourier and Hough transforms, are contrasted with this neural approach, highlighting the differences in how patterns are identified and processed. The discussion also touches on the psychological implications of these generated images, suggesting parallels to human cognitive biases and phenomena like pareidolia and hallucinations. The potential for these images to illustrate psychological flaws adds another layer of intrigue to the technology's capabilities.
Ryan_m_b
Staff Emeritus
Science Advisor
Homework Helper
Insights Author
Messages
5,963
Reaction score
726
I can't get over how cool/creepy this is. Researchers took image recognition neural networks and asked them to look for certain images in random noise and once identified modify the image to highlight the pattern. That new image is fed back into the machine and it's asked to do the same again. After multiple repetitions the pictures are bizarre, beautiful or both.

http://www.theguardian.com/technolo...twork-androids-dream-electric-sheep?CMP=fb_gu
 
  • Like
Likes Czcibor, Silicon Waffle and ShayanJ
Physics news on Phys.org
Ryan_m_b said:
Researchers took image recognition neural networks and asked them to look for certain images in random noise and once identified modify the image to highlight the pattern.
I see what you mean ( by most of the images ). I would not call it "pattern recognition" but rather a LSD-trip.

But I think you know, that "real" pattern recognition uses mathematical rigid methods like Fourier-transforms, Hough-transforms, lens correcting functions, amongst other methods, to measure what is going on in an industrial production ?
 
What a cool idea!
 
Hesch said:
I see what you mean ( by most of the images ). I would not call it "pattern recognition" but rather a LSD-trip.

What makes it not pattern recognition? It's not necessarily good pattern recognition initially given that the machine is detecting patterns that aren't there but after it has edited them in slightly subsequent tests detect that patter.

Hesch said:
But I think you know, that "real" pattern recognition uses mathematical rigid methods like Fourier-transforms, Hough-transforms, lens correcting functions, amongst other methods, to measure what is going on in an industrial production ?

Why would you think I know that and what makes any of that "real"?
 
  • Like
Likes Silicon Waffle
That is cool. They can arrange an exhibition.
 
Ryan_m_b said:
Why would you think I know that and what makes any of that "real"?

Sorry, I thought you meant it as a joke.

I think that this example is real:
dba52883-3a65-4e46-bb08-8c15678642b1-bestSizeAvailable.png

The original photo is "filled" with noise which has some distinct feature. You Fourier transform (FT) the noisy picture and will find this feature in the Fourier transform. Remove it from the transform and make the inverse Fourier transform (IFT). Then you'll get the picture to the right.

Say you have a photo of a car driving by. Due to the speed of the car ( crossing the photo with a shutter time = 1/100 sec. ) the car will be blurred on the photo. Now you take two sheets of paper, draw a dot on one of them and a line on the other ( blurred dot ). The line must exactly be as long as the car has been moving on the photo. Also their moving angle must be the same. FT the dot-picture to D, FT the line-picture to L, FT the photo of the car to C. Then:

IFT( C * ( D / L ) ) and you will have a photo of the car, where you can read its registration number.

Hough transforms are used to recognize lines, circles, parabolas and other mathematical shapes. If such a "known" shape occurs in some photo, the Hough transform will find it and will determine its exact size and location within 1/10 of a pixel-distance. Having a "standard-length" as well in the picture, a computer can calculate very accurate dimension in the picture, check "ovality?" as of things meant to be circular, and so on.

Remember that working machines are often moving very fast, thus the human eye sees nothing. A camera needs perhaps 2μs, using a stroboscope, to see everything in the picture ( well, at least after the computer has calculated for another 100ms ).
 
Last edited:
  • Like
Likes Silicon Waffle
we just start with an existing image and give it to our neural net. We ask the network: “Whatever you see there, I want more of it!” This creates a feedback loop: if a cloud looks a little bit like a bird, the network will make it look more like a bird. This in turn will make the network recognize the bird even more strongly on the next pass and so forth, until a highly detailed bird appears, seemingly out of nowhere.
What intrigues me about these images is that they could be used to illustrate similar human psychological and neurological flaws, from confirmation bias, through delusion and pareidolia, all the way to hallucination. In fact, I wonder if the exact same kind of feedback loop isn't at work in all those things.
 

Hot Threads

Recent Insights

Back
Top