What the world looks like to an algorithm

Artificial intelligence is governing more and more of our lives, but the way it sees and understands the world is completely different from you or me. For this video, we found a way to look around inside AI’s “brain.” First, we asked fellow humans to guess paintings made by a computer program. And then, the opposite: we asked AI to guess our doodles. It’s a game of Pictionary that explains a lot about our future alongside AI.

[VergeScience]

Did you guess any of these AI-generated images correctly?

[phatdookie]

*So that is why videos get demonetized when they shouldn't be.*

[halo_origins]

I'm drawing a spider!
*Draws 10 legs

[vidoma]

Wtf.. this video was sponsored by the US Army?!

[XSpImmaLion]

It's an interesting exercise and thought experiment, but just so people know, the algorithms and processes that are actually being employed in recent big AI projects are far more complex and way more developed than the examples given.

Like, leaps and bounds further than simply using TensorFlow for some experiments or other types of simple neural networks with very basic training procedures for fun. So I'd advise against conflating something like this admitedly fun and kinda absurd idea of reverse engineering what an AI sees, and the tech that automated cars are actually using.
The distance between those is like an electric scooter to a car in complexity. Sure, they might have some few commonalities, but they are nowhere near in complexity.

To be clear, this experiment doesn't come close even to examples of AI being used in non life threatening examples, like what smartphone cameras use to recognize objects in a scene, or stuff like Google Lens, or face ID, or AI assistants and whatnot. It is related since it's one interesting technique among multiple others, but it's far from being only that. For automated cars it's several degrees removed from the actual thing.

For starters, automated cars don't use only machine learning and AI for their detection system. They are already using, right now before even going into widespread production, an array of several different techniques that goes far beyond even what regular people can see. I posit that it would be kinda hard to even represent that for a normal human to see, because it's looking at a whole ton of data that would be hard to represent in a single drawing/shot. It's a mish mash of multiple cameras, radars, laser scanners, GPS other types of sensors, and a bunch of other stuff.

As for the first fatal Tesla crash, for starters, it didn't even have a system with a whole lot of AI in the first place. Tesla autopilot is not an automated driving system (only a driving assist system), and it's not very representative to the stuff that's being developed for autonomous cars. It's a small piece of the puzzle at best, and it should have never been sold with a deceiving name like "autopilot". It's just driving assist, which lots of other cars also have.

So there. It's an interesting video, but it slips a bit on information and unnecessary fearmongering.

[varagor]

Presented by US Army? That's kinda creepy.

[TommoCarroll]

New verge science + AI!? Yup, I’m game (pun well and truly intended)

[douira]

the google quick draw AI might be better at Pictionary for hand-drawn shapes than this one

[ahamedFuad]

Hey that's mind blowin 😱, I still wonder how *The Verge* team manages to find these kinda cool stuffs 😕

[moth.monster]

You know what they say.... a bird in the hand gets crushed by the fist

[MouseGoat]

6:38 a yess, because ruining over stop signs is for sure a great move LOL

[mickmoon]

Would like to see the source code for that inversion or de-engineering image? Surely if you fed the inverted image to deep learning it should be able to guess what the image is?

[TheJohnRowley]

This was great in every way - Interesting subject told in a concise inventive visually entertaining way with an engaging narrator. Wonderful, more from the team who made this please

[eli_leprosy]

"Electric Fan"

No that's an elephant.

[suprotikdey]

This is the main cause why variety of data is required during training. Also, its true that a classifier will always try to classify an unknown object to any of its known classes. It was built like that only.

[lynwood]

Why is the US army sponsoring you guys, this is the second video that I've noticed

[cig_in_mouth]

Thanks verge science and us army for doing this but I need more keep it up

[BitchItsJules]

Lol it's a bit like a little kid who doesn't know much yet.
"Look, Daddy, it's a polar bear bird!"

[xCRiPPLESx]

Why was this "presented by the US army"? wtf?

[jonashubotter]

This comparison is really unfair. How the algorithm see the world depends on the data it was trained with. And that data is certainly not the same data your human mind was trained with. Every neural network serves a purpose that is more specific than your mind. So there'll always be the edge cases where they'll see something humans cannot. I mean Google's and Amazon's classifiers weren't taught to distinguish abstract art and pictures.