Sketchy AI

Have you ever been somewhere abroad where you didn’t understand the language at all?

Yes, sure, I’ve been to Japan last year to cover a robotics conference and demo

Cool! So imagine someone there pointing at a robot prototype and saying something to you in Japanese

OK, that’s easy to imagine … since it did happen! A lot!

Great, so they point at the robot and say “shiro” …                             

… yes? So?

So what do you get from that? Is the name of the robot, of one of its parts, its color, its function?                                          

Right, maybe I see what you’re getting at “ostensive definitions”, pointing and saying a word can be quite vague …

Exactly. Tagging an image is vage and can be ambiguous. It’s not like the AI knows from the outset what is what                      

In “man riding a bike” anything could be the bike or the man. It needs to learn by combining a lot of those images                      

Generalizing from the data, but somewhat … blindly

Generalizing from very messy data. We’re not feeding it a textbook on photography or drawing portraits, but random tagged images from the internet                        

Ouch, yes, I can see why that might trip it up … but what about the scribbling thing?

Kids learn to draw in stages and there’s a specific stage, somewhere when they’re around 3 or 4 when they start drawing people and faces

Right, with the legs directly attached to the head, like an octopus!

Yep, but more importantly, they have started to realize that bodies and faces have parts: eyes, noses, fingers, etc.

But then they draw them all over the place … ah, I think I get it …

The AI doesn’t know exactly what’s what, just that it needs to be somewhere in the picture            

And then you get bikes with too many pedals, just like scribbles of people with too many fingers or a nose next to an eye

I don’t want to sound too negative here, it is definitely an interesting development and has attracted a lot of interest