Hacker News new | ask | show | jobs
by Kylekramer 2703 days ago
It tries and fails to address it. I could barely count the number of zeros I would have to put in front of the 1 in the percentage of Facebook's photo data this meme covers. And a good portion of their dataset will have EXIF timestamps. Training an algorithm on the meme only would be insane waste of their data set.
3 comments
bbrian 2703 days ago
> their dataset will have EXIF timestamps

It used to be the case (still is?) that the dates on uploaded photos weren't applied to the photo album. I remember having to go through holiday snaps and change the date on each from the upload date to the actual date. The images were also resized down from what was uploaded.

So, if they have the originals with the full EXIF data, I'd like to be able to use that for my old photos!

link
bigbugbag 2703 days ago
Chances are they have and you won't be able to use them.

Why would it be so ? Because it profits facebook which is the only reason facebook exists. It profits them to have original with EXIF for data mining and you gave them permission to do so while also giving them the data, and it profits them to not make them available to you to save on bandwidth and processing costs.

link
Setepenre 2703 days ago
eh, I think not. There is a reason why they resize the photos. To save space. Even at facebook scale the amount of space they save by doing this must be enormous.

Additionally, images used in AI are usually scaled down a lot more, 224x224 for something like resnet50. which means that they do not need your high quality original and the smaller one they generated are fine.

link
cmmartin 2703 days ago
They don't need everyone to train their AI to detect aging. They don't even need a substantial sample size. All humans age in relatively similar ways.
link
baroffoos 2703 days ago
Why would they even bother with this when they already have much better data though.
link
DoctorOetker 2703 days ago
perhaps they forgot to split a test set :)

while I'm sure they already have a great dataset going back far, most of that dataset will be a small fragment of the world that already had internet 10 years ago, in the mean time a lot of people in other parts of the world have come online and may have old pictures of themselves

link