I’ll tell you up front that I don’t have the answer to the question posed in the title. It’s a question I’ve wondered about for a while now though.
When humans look up at a cloudy sky they often see shapes in the clouds. Often it’s animals or cartoon characters. This is part of our visual processing system – trying to find objects in the visual field and make some sense of them. The actual broad term for this is apophenia. If you see images and faces it is called Pareidolia.
The term apophenia was coined in 1958 by a German psychiatrist working on Schizophrenia. This has the extra connotation of adding significance to the perception. The more appropriate term pareidolia is actually the one we want – it means seeing shapes as objects instead of what they really are. Like clouds. OR like ink blots in Rorschach tests. OR finding shapes in
jiggly long exposure photos like the one below.
A while back I wondered if an AI could / would do the same thing. Do AIs have imaginations? I’d guess they don’t have one in the original sense of the word “to picture one’s self” because I’m not sure we’ve gotten to self aware sentience just yet.
Not having access to a captive AI to tinker with I can only wonder about it. I guess I could try feeding clouds images to Google’s Lens app and see if it comes up with anything but that somehow feels like a very limited test. I’m more curious what researchers have found out. Or maybe try feeding it scans of ink blots to see what it thinks might be there.
The reason I’m picking this to blog about today is a little article on ScienceDaily. It’s titled: Can an artificial intelligence tell a teapot from a golf ball?
In the article they describe testing AI’s with confusing visual information to find the limits of their abilities. In the test from the title they fed the AI an image of a teapot shape with a golf ball texture. The AI sees golf ball. Teapot is lower on its guess list.
From their results it would appear that the Deep Learning AIs tend to take texture into account far more than they do shape. So outlines and silhouettes confuse them. One would hope that over time shape becomes equally important. But I’m getting the idea that today’s crop of visual AI’s might not exhibit pareidolia just yet.