back to article Facebook uses one billion Instagram photos to build massive object-recognition AI that partly trained itself

Facebook has trained its most advanced semi-supervised computer vision system yet on a dataset of a billion public images taken from Instagram, its other social network. Known as SEER, short for SElf-supERvised, this massive convolutional neural network contains over a billion parameters. If you show it images of things, it …

  1. IGotOut Silver badge

    And yet...

    ...its too hard for Facebook to stop ads for land in protected areas in the Amazon rainforest.

    1. Charlie Clark Silver badge

      Re: And yet...

      It's an advertiser and will normally sell any kind of to anyone and worry about the morals or legality if anyone complains…

  2. Efer Brick


    Or not?

  3. TeeCee Gold badge

    "one billion Instagram photos"

    Ok, one question. Just what use do Zuckerbitch and his minions have for an insane AI that hates humanity with a vengeance?

    1. Chris G

      Re: "one billion Instagram photos"

      If you think it hates humanity now, just wait until after it has watched a million tik tok videos!

  4. Muscleguy

    Basically this is how human infants learn. They have eyes which see the world and objects in it. They then begin to learn language and names for things. Except in humans the process is ongoing. As our infants progress we feed their brains more complex stuff with names again and again and again.

    Just doing it once is not the way to go. Maybe they need to get a better AI to better label the images to then better correct this one.

    1. Ben Tasker

      > Maybe they need to get a better AI to better label the images to then better correct this one.

      That's basically the definition of an Adversarial Network isn't it? One AI generates/calculates, whilst the other tries to pick holes in it. Stuff that passes scrutiny is the output

    2. Doctor Syntax Silver badge

      The Treachery of Images.

      You're missing out a stage. Vision is only one aspect of this. A baby learns about the world by correlating inputs from all the senses. It will learn that a banana has an inside which is edible. It will also learn that it is much smaller than a golfing umbrella, something which is not necessarily obvious from an image. And it will be making these correlations before gaining language.

      An object and a picture of the object are two different objects.

    3. Charlie Clark Silver badge

      That's probably how things will go. You can create a network that learns where the first network got it wrong and create a feedback loop, with a view to improving various steps later on.

      The important thing is the speed of the training given the volume of data because once the model is considered "good enough" it can be deployed on kinds of devices.

  5. Spiz


    But is it intelligent enough to realise that 90% of the images on instagram have been photoshopped or modified in some way, or will it always think that human females have perfect skin, a 10 inch waist and an arse the size of Manchester? i.e. Does it have a Thot filter?

  6. TimMaher Silver badge


    I refer everyone to the comments in this issue’s “Something for the weekend sir?”.

    ‘Nuff said.

  7. Anonymous Coward
    Anonymous Coward

    ... and is then trained how to caption those pictures

    match found:SARAH CONNER

  8. Anonymous Coward
    Anonymous Coward


    How long before the AI "common sense" decides that humans are a wasteful resource?

  9. WolfFan

    HAL 209

    I’m sorry, Dave, but you have 20 seconds to comply.

    1. Steve Davies 3 Silver badge

      Re: HAL 209

      It really is time for Zuck to be cast off into space (fired into the sun would be good)

      Hey Elon, got any space on one of your exploding rockets?

    2. Lance 3

      Re: HAL 209

      Pretty sure it will be more like ED-209; so what humanity does in that 20 seconds to comply won't actually matter.

  10. Anonymous Coward
    Anonymous Coward

    1 billion images

    i hope that they have signed releases for each and every image. If not then all it takes is one DMCA takedown to make this an exercise in futility.

    What? You mean to say that anyone who posts on Instagram gives Zuck full rights to use the image for anything they want?

    Good job that I never have felt the urge to visit any of the sites of crap that your company runs.

    1. Anonymous Coward
      Anonymous Coward

      Re: 1 billion images

      Doesn't require a DMCA....the owner just needs to remove it or delete their account.

      "This concerns many advertisers, with as much as 52.8% of Digital Marketers considering government regulation or the threat of regulation as an obstacle in their ability to leverage user data.

      On the other hand, according to a Cisco Customer Privacy Survey taken in 2019, 84% of their respondents expressed their concern for data privacy. They believed that they should have more control over how their data is processed, and 80% of those people even went to far as to mention that they would be willing to take action to defend it."

      Yet these same people just put stuff out there for the world to see without a care in the world.

      The bigger thing though, "Facebook told us SEER remains a proof-of-concept idea and won’t be used to power any of the web giant's features or products for the moment." The moment can end whenever they decide. Coming tagging, then facial recognition, etc.

      1. Freezus

        Re: 1 billion images

        Facebook has been using facial recognition to auto-tag photos since at least 2011, nothing new here

    2. John Brown (no body) Silver badge

      Re: 1 billion images

      "i hope that they have signed releases for each and every image. "

      Back in the day, both AOL and Geocities (and most likely many others) laid claim to a perpetual licence to use anything you put on their servers for any use they could think of. I would imagine most hosting services still do that. Although in this day and age, that's a little more risky since they probably don't want to be laying claim to and using any illegal material their users may post.

  11. sgp

    Common sense

    Always sounds good until you realize it tells you more about your own views and beliefs than anything else.

  12. Tree

    They cannot be trusted.

    Glad I did not join Pinterest or faceBUTT. If Suckerburg cannot recite his Terms of Service, they are not for my benefit.

    1. Anonymous Coward
      Anonymous Coward

      Re: They cannot be trusted.

      Most presidents of companies or CEO's cannot recite the ToS. They didn't write it, the legal team of the company wrote it.

      Could you recite the T&C's for this site?

      1. Lusty

        Re: They cannot be trusted.

        The point is that the terms should be simple enough to recite if they are genuine.

        We’ve fallen into a situation where businesses think they need to trick customers into buying things. The idea that you can just be good at something is almost forgotten.

  13. Boris the Cockroach Silver badge

    one billion

    cat pics

    Need I say more?

  14. seven of five








  15. Pascal Monett Silver badge

    But of course

    "areas including medical imaging where it’s difficult to amass large labelled datasets from private clinical data "

    Yes, great idea, let's hand over private clinical data to El Zuck.

    What could possibly go wrong ?

POST COMMENT House rules

Not a member of The Register? Create a new account here.

  • Enter your comment

  • Add an icon

Anonymous cowards cannot choose their icon

Other stories you might like