Optical Illusions and Machine (Vision) Learning

The Coffer Illusion

The ABC had an interesting article on optical illusions and the workings of the brain today.
The Coffer Illusion (above) is really spooky!   Looks like rectangles right? Look longer enough and another shape appears (circles). Took some time with me originally, page but once I could see them I could come back even hours later and they would be almost immediately apparent.  It's like your brain has been "programmed" with the circles so you see them again.  Can I ever un-see them??? TODO come back in a week and check :-)

So the question I had was given that these sort of optical illusions tell you something about how the brain works, do they tell you anything about Machine Vision? For example, could ML algorithms be trained to recognise either (or both) rectangles or circles in this image? How about optical illusions that require switching between recognising one object or another?  E.g .this one which has a single face, or 2 faces kissing?




This article on "Computers learn to fall for optical illusions" is relevant. 

It points out that it's possible to "debug" or understand how computer machine learning vision is identifying images by introducing optical illusions, as the problem is that the algorithms are often not identifying the features that humans recognise. I suspect this approach is similar to the GAN algorithms. None of these look like the object to me, but they do (at the pixel level) to the ML vision algorithm:





Meaningless to humans, these images are recognized by a computer with great certainty as common objects. The white noise versions on top and pattern versions below were created by slightly different methods. (Credit: Cornell)


Yes, I was right about the GAN connection, see this article, which also points out that animals are also tricked (but human's not so easily?) by similar illusions.

And research from 10 years ago (UCL, where I was in 2004) on similar topic.