AI Still Sucks at Optical Illusions, So at Least we Have That Going for us

By Jennings Brown on at

Optical illusions take advantage of shortcomings in the visual system. Certain special designs can trick our brain as it’s trying to process the information that’s coming in. Learning more about what can trick our minds would help us learn more about the human mind itself.

That’s part of the reason why Robert Williams and Roman Yampolskiy, two researchers at the University of Louisville in Kentucky, want to develop machine learning systems that can create new optical illusions—they hope to learn more about the “very specific tricks that cause us to misjudge color, size, alignment and movement of what we are looking at,” they wrote in a recent paper on their study. “It is also important to consider whether making a perceptual mistake similar to humans constitutes having a visual experience similar to humans.”

Williams and Yampolskiy wanted to build this generative adversarial network using the same method as a recent machine learning system that was trained to generate new images of human faces, using a neural network that was fed thousands of photos of faces.

Alas, they couldn’t pull it off.

Neural networks are dependent on the volume of material that they can “learn” from. According to the research paper, there are only a few thousand static optical illusions that exist. And the researchers estimate that there are probably only a few dozen different types of optical illusions, like Fraser’s Spiral illusion, Hermann Grid illusion, and Zöllner illusion.

The researchers were able to compile a dataset of more than 6,000 optical illusion images that they gave to a neural network. But ultimately, that wasn’t enough for the machine to figure out how to make new optical illusions.

So for now, you can rest easy that the robots aren’t able to find new ways to fuck with our minds.

[MIT Technology Review]

Featured image: Michael McCauslin (Flickr)