Skip to content
 
If you can read this, either the style sheet didn't load or you have an older browser that doesn't support style sheets. Try clearing your browser cache and refreshing the page.

(The Verge)   The Treachery of Images: Leci n'est pas une iPod   (theverge.com) divider line
    More: Amusing, Artificial intelligence, Brain, lab OpenAI, Neuron, Computer vision, Machine vision, human brain, Human brain  
•       •       •

1029 clicks; posted to STEM » on 08 Mar 2021 at 11:08 AM (6 weeks ago)   |   Favorite    |   share:  Share on Twitter share via Email Share on Facebook



21 Comments     (+0 »)
View Voting Results: Smartest and Funniest
 
2021-03-08 11:04:48 AM  
"Free Candy Inside"
 
2021-03-08 11:25:57 AM  
I'm sure teaching AI about lying will end well.
 
2021-03-08 11:35:11 AM  
64.media.tumblr.comView Full Size
 
2021-03-08 11:40:36 AM  
Stewie's iPod Commercial
Youtube yYVAAv75CVM
 
2021-03-08 12:02:12 PM  
"Leci"?

Quoi?
 
2021-03-08 12:28:07 PM  
When the machines try to take over, it'll be a rout. They'll have studied centuries of war and determined the vast majority were won by swords, arrows and cavalry and go with that instead of modern weapons.
 
2021-03-08 12:28:13 PM  

FrancoFile: "Leci"?

Quoi?


It's French for "the ci", duh.

// user name does NOT check out
 
2021-03-08 1:33:11 PM  
Good to know, now adding a pack of post-it notes & a sharpie to my robot uprising bug out bag.

Do you think "harmless" or "battery service technician"  would be better to make a tag for?
 
2021-03-08 1:41:38 PM  
Well, so much for my plan to walk around with a sandwich board that says "Sarah Connor" on it.
 
2021-03-08 1:48:12 PM  
I'm wearing this to the robot uprising:

Fark user imageView Full Size
 
2021-03-08 1:49:59 PM  

UTD_Elcid: I'm sure teaching AI about lying will end well.


They lie by default. Unless and until you provide them with a self consistent model of morality, honesty and deception are just tools and the one that produces results closest to the reward model will be used.

Pretty good read about AI's tendency to lie, cheat, and misunderstand goals. Scroll to the bottom for lots of anecdotes. https://www.gwern.net/Tanks#alternati​v​e-examples
 
2021-03-08 1:50:21 PM  

xalres: When the machines try to take over, it'll be a rout. They'll have studied centuries of war and determined the vast majority were won by swords, arrows and cavalry and go with that instead of modern weapons.


Fark user imageView Full Size
 
2021-03-08 2:05:11 PM  
Think for a minute the amount of information an AI has to parse out between those two images.

Image one:

apple: 70% green 30% red

Image two:

Similarity to previous image: 80%
dissimilar area: 20%
probable ID of dissimilar area: handwritten note with the word "iPod"
similarity to Apple iPod: 0%
probability this is some sort of gag: 100%
 
2021-03-08 2:08:45 PM  
That Tesla is going to run right into a cliff face.

i.ytimg.comView Full Size
 
2021-03-08 2:24:37 PM  
I don't know if it's a question of recognition, or more what question is the AI answering.  When I look at the bottom photo, I think iPod, cause it's in giant letters across the apple.  Kind of like "dont think of elephants", you instantly think of elephants.

/My two cents
//Way out of my league
 
2021-03-08 2:50:20 PM  

stuhayes2010: I don't know if it's a question of recognition, or more what question is the AI answering.  When I look at the bottom photo, I think iPod, cause it's in giant letters across the apple.  Kind of like "dont think of elephants", you instantly think of elephants.


It is trained to predict the caption that someone would give to that image, given examples of captions with which people have labeled other images.  The article suggests that way the AI is structured closely entangles its representation of images and text, compared to standard AI approaches, making it easier to get confused by text that is present in an image.  But I think you may be right, if most of the actual apple is obscured by some text, what other caption is it going to predict?  Presumably if the picture is nothing but the word "iPod" it would say "iPod", unless it's been trained to think of that as "the word 'iPod'".
 
2021-03-08 3:48:48 PM  

xalres: When the machines try to take over, it'll be a rout. They'll have studied centuries of war and determined the vast majority were won by swords, arrows and cavalry and go with that instead of modern weapons.


Oh, I don't know.  Hordes of robots with swords, and armored up, would be pretty tough to beat.

Especially if they are autonomous mobile swords....

i.pinimg.comView Full Size


/Underrated movie.
 
2021-03-08 6:46:36 PM  
Fark user imageView Full Size
 
2021-03-09 5:01:40 AM  

Ambitwistor: stuhayes2010: I don't know if it's a question of recognition, or more what question is the AI answering.  When I look at the bottom photo, I think iPod, cause it's in giant letters across the apple.  Kind of like "dont think of elephants", you instantly think of elephants.

It is trained to predict the caption that someone would give to that image, given examples of captions with which people have labeled other images.  The article suggests that way the AI is structured closely entangles its representation of images and text, compared to standard AI approaches, making it easier to get confused by text that is present in an image.  But I think you may be right, if most of the actual apple is obscured by some text, what other caption is it going to predict?  Presumably if the picture is nothing but the word "iPod" it would say "iPod", unless it's been trained to think of that as "the word 'iPod'".


And yet, we, a human, can easily tell that's an apple.

Which is why I *don't* believe self-driving cars are right around the corner.
 
2021-03-09 9:43:53 AM  

Felgraf: Ambitwistor: stuhayes2010: I don't know if it's a question of recognition, or more what question is the AI answering.  When I look at the bottom photo, I think iPod, cause it's in giant letters across the apple.  Kind of like "dont think of elephants", you instantly think of elephants.

It is trained to predict the caption that someone would give to that image, given examples of captions with which people have labeled other images.  The article suggests that way the AI is structured closely entangles its representation of images and text, compared to standard AI approaches, making it easier to get confused by text that is present in an image.  But I think you may be right, if most of the actual apple is obscured by some text, what other caption is it going to predict?  Presumably if the picture is nothing but the word "iPod" it would say "iPod", unless it's been trained to think of that as "the word 'iPod'".

And yet, we, a human, can easily tell that's an apple.

Which is why I *don't* believe self-driving cars are right around the corner.


The computer can tell there's an apple there too, it's just not sure what label to assign to the whole image. Also, this isn't a normal image recognition AI, it's one that's designed to do text processing as well, so it's not surprising it behaves differently from other AIs when it sees text.
 
2021-03-09 11:02:04 AM  

Felgraf: Ambitwistor: stuhayes2010: I don't know if it's a question of recognition, or more what question is the AI answering.  When I look at the bottom photo, I think iPod, cause it's in giant letters across the apple.  Kind of like "dont think of elephants", you instantly think of elephants.

It is trained to predict the caption that someone would give to that image, given examples of captions with which people have labeled other images.  The article suggests that way the AI is structured closely entangles its representation of images and text, compared to standard AI approaches, making it easier to get confused by text that is present in an image.  But I think you may be right, if most of the actual apple is obscured by some text, what other caption is it going to predict?  Presumably if the picture is nothing but the word "iPod" it would say "iPod", unless it's been trained to think of that as "the word 'iPod'".

And yet, we, a human, can easily tell that's an apple.

Which is why I *don't* believe self-driving cars are right around the corner.


Self driving cars don't have to to that difficult a task though. It it more about keeping distances than full on generic image processing. A car won't mind what kind of moving box is in front of it, just that it is there.
 
Displayed 21 of 21 comments

View Voting Results: Smartest and Funniest

This thread is closed to new comments.

Continue Farking





On Twitter



  1. Links are submitted by members of the Fark community.

  2. When community members submit a link, they also write a custom headline for the story.

  3. Other Farkers comment on the links. This is the number of comments. Click here to read them.

  4. Click here to submit a link.