I spy with my little eye….A-I?
This is Sandra Tsing Loh with the Loh Down on Science.
Human intelligence can easily tell elbows from oboes. But ARTIFICIAL intelligence networks are TAUGHT using thousands of pictures. This deep learning technique has achieved extraordinary success at teaching A-I to see and identify objects!
But Philip Kellman at UCLA wondered what A-I’s REALLY see.
He asked two well-taught computers to do what they do best..
The networks were great with silhouettes and scrambled images, correctly identifying fifty percent or better! But they were TERRIBLE when it came to altered or undefined images.
Example? A teapot with a golf ball pattern was called a golf ball. And a glass polar bear apparently resembles a...can opener?
Researchers think the A-I networks are focusing on small details, not the bigger picture. Textures and patterns are favored over shape. Adding or subtracting key details easily bamboozles the A-I’s.
This technology has come a long way. But Kellman cautions that improvements are still needed. Especially before we trust A-I to say, drive our cars!
So Watson may beat you at Jeopardy - but you’ll still rule at I Spy!