59 comments

[ 3.6 ms ] story [ 22.0 ms ] thread
What an impressive, and extremely creepy, use of state of the art neural net technology.
ah good, this means the data can be "poisoned"

They go further and use the machine to locate images that do not have location cues, such as those taken indoors or of specific items. This is possible when images are part of albums that have all been taken at the same place.

I'm going to start including pictures of the moon surface in all my albums.

> I'm going to start including pictures of the moon surface in all my albums.

Sorry to say it but they're going to be able to exclude whatever you throw at it.

And even if you found some way to meaningfully degrade the quality, they'll analyse what you do and find a solution for that.

They are not going to suddenly throw their hands up in the air and decide to give up.

It's a matter of how many resources can you throw at the problem and how many can they to counteract you.

They're going to be able to exclude whatever you throw at it.

How, exactly? You're talking about something that's one of the most poorly understand areas in AI at present, and making it sound like it's a solved problem.

> How, exactly? You're talking about something that's one of the most poorly understand areas in AI at present, and making it sound like it's a solved problem.

With manpower. I said that they are going to analyse whatever users do to wilfully degrade the service and will find a way to exclude that.

Just like they do with search for years. Somehow they manage to keep their search relevant even though so many people constantly work on coming up with new SEO tricks that would seriously degrade this service if left unchecked.

Of course I cannot know how _exactly_ they will do that, even the persons working on that service cannot know at this moment how exactly they will keep this working in the future since they do not know yet what people will try to do.

I like your enthusiasm but look at the number of images again, and remember that Google is notorious for trying hard not to use humans. I think the statistical significance will probably overwhelm moon pictures, and that's what they'd use long before some human starts poking around for weird data.
Or they could train a network to recognize persons likely to try and poison their data sets and use it as an exclusion. You could go on to argue that those persons could also poison that data, but then they could handle it in a similar way. This goes process could theoretically go on ad infinitum.
Exactly. Machine learning is an inequality-expanding lever just like many other forms of capital. This is one of the reasons Google et al are so dangerous.
This has about as much reasonable intellectual bearing as suggesting that language is an inequality expanding lever just like many other forms of capital and one of the reasons Google and others are so dangerous.

This is just the Luddite fallacy repeated over and over. By this logic, we should never develop new informatic methodologies, we should never develop new forms of analysis, we should never develop new software – because of economies of scale exist.

I'll the state the obvious: that's stupid. There's no way to soften that blow.

This will not be of statistical significance and thus will not affect the performance of the system.
An impressive, if not very technically interesting, result. I wonder if it can get cooler with hierarchical sets of neural nets with ever smaller grids...

An aside - I was really annoyed by "small-scale experiment shows that PlaNet reaches superhuman performance at the task of geolocating Street View scenes". It's absurd to say computers are superhuman at searching the web for instance, and though this is a much more impressive perception-based result it is still fundamentally very data intensive. Would be nice if 'superhuman' was used a bit less now that Deep Learning has already been proven to be capable of really impressive performance in perception tasks.

It would be even more impressive if they could compress the images, given that the backgrounds are already known.
Years ago I was like Richard Stallman is totally crazy!, saying stuff like don't post my pictures on Facebook and such. Today his fear is really really justifiable...
It's not anything a human can't do in theory, identifying plants and common architecture features. If you are uploading photos on facebook, it's far more likely you'll leak geographic data through the photo's metadata, your ip, and gps.
Metadata is basically giving your data...

Stallman was asking to not put any kind of his pictures on Facebook, which means your face is recognized now by Facebook, and there's plenty of things to do with that...

I'm all in favor of credit where credit is due, but that's like saying "the Unabomber was right" because examples exist of modern technology being misused in the past 20 years. If anything, his advocacy has damaged the credibility of legitimate privacy and civil rights advocates.

Surely, SURELY RMS is not one and only person you have ever heard in your entire life warn about the dangers of putting personal information online, or the dangers of government surveillance. That's all just common sense and being relatively informed and skeptical about the world and the people around you.

RMS says a heck of a lot more than "don't post your pictures on Facebook", much like PETA says a lot more than just, "treat animals ethically", or Earth First says a lot more than just, "don't destroy the environment". If sea levels rise much more in my lifetime, one lesson I will not draw from it is, "Earth First was right!"

If you care about this stuff, donate to the ACLU, not the FSF.

Really? Let's not forget the NSA revelations that occurred so recently. Thanks for reminding me to renew my fsf membership
The ACLU would make much better use of your money.
I'm also a member of the ACLU.
Everything that man says about privacy and digital freedom is ultimately proven true given enough time. Unfortunately he's still considered a crackpot.
Today his fear is really really justifiable...

Why, exactly?

Usually I like people to know where the photos I share publicly are taken. Indeed, that's why I go to the trouble of captioning or manually geotagging them.

It's so good, it guesses the correct continent less than 50% of the time. I bet it even gets the correct planet 100% of the time :)
The median error is 1100 km. Which is better than that. But I agree with you that this is nothing to worry about. I skimmed through the paper and thought "wow that's really interesting it can beat humans and get results that good", but it's not so good that it's a significant leak of privacy. Like the top comments seem to be scared about.

The dataset might be a little biased though. I wonder how many pictures are just of people's faces, or food, or random indoor scenes. It's much more difficult to detect a location from such an image. How good are the results if you exclude the images that don't have any clues at all?

As I understand it, the algorithm can identify common types of plants unique to different regions, or architectural features, or popularity of types of cars. So an outdoor photo is more likely to be recognized. But I'm not 100% sure what features it's using. In fact the creators don't even know!

Mark my words--this tool is going to be quickly hijacked by meme groupies who will submit dick pics and post the location results:

"Check it out, I'm Mogadishu!"

You could improve the accuracy by using the timestamp of when the photo was taken & looking at the amount of daylight in the photo.
Will this really work? Cameras are by design able to accurately control the rate (and duration) of light they let in to create an image.
For known cameras like iPhone, that shouldn't hurt since you can read the exposure time out of the EXIF.
If you can also get accurate white balance, you may be able to get useful information from the colour of the sky as an alternative.
> the amount of daylight

wouldn't that be much effected by weather conditions and air quality ?

That would actually improve the accuracy since the image can be compared to historical weather records.
Possibly also length of shadows? (sounds tricky though)
You'd let the neural network do the heavy lifting there
This could be a dumb question, but could it be possible that the Neural Network would have already identified this correlation?
It's a great question, but the only information about the photos they used was the visual data from the pixels and what albums they were in.
You would probably be interested in the computer vision work of professor Nathan Jacobs over the last decade: http://cs.uky.edu/~jacobs/

He has done some amazing work extracting geo-location and scene structure based on daylight patterns, shadows cast by clouds, rainbows, etc in images and image sequences.

This is only about using the pixels, not anything from the exif data. The only additional information was used in a later iteration of the model, taking advantage of knowing that photos in the same album were more likely to be taken in the same places. Also, there was a section in their paper about comparing PlaNet to human players at the GeoGuessr.com game, specifying that both had the exact same information available; there's no time information with GeoGuessr.

I'm sure any serious organization actually putting together a product to locate images that have time—but not location—data would try to take advantage of the outdoor color temperature, lighting, global weather records, text recognition, and all that kind of stuff. This was a more focused research project.

Ah, the luxury of having so much labeled data. Average machine learners like us will never have access to such low hanging fruit...
There are probably average machine learners working at companies like Google
(comment deleted)
Any idea who would have "126 million images along with their accompanying Exif location data."?

That seems an insane amount to me.

Android phones that backup all the photos to Google Photos? Seems like a good place to start.
Sounds like 'all public Google+ photo albums'. (I say public because given what Facebook reports about daily photo uploads, a cumulative 126m seems like it would be way too small for G+ if it covered all Google Docs/Photos uploads including private ones.)
Google Street View
Yup. And we've already been tagging locations with adresses for years through that captcha.
That isn't exactly needed for this. The GPS location will give you a street address close enough for the metrics used by this.
They mentioned Flickr as a source of images.
We're closer to my nirvana: I go on vacation and take NO pictures, just enjoy myself, but photobomb other people constantly. When I get home, Google makes a vacation album for me, by identifying my face in all those other pictures! With knowledge of my itinerary it should be easy now!
Do you guys have an url where we can try it out?

Or you think it's like every "new" Google product from a few years where there is a lot of talking before anything.

Blame the headline. It's an interesting AI research paper by a team with two Google employees on it, not a product.
in the article there is a game to identify locations by photos. its pretty neat.

https://www.geoguessr.com/

It's a great game, but a little misleading to use as a proof of concept since presumably they trained it with google streetview data.
In 2008, when not all of us are using smartphones, we always remind the smartphone users to protect their privacy. We told them not to upload their photos. But now, who cares? Everyone should know where I am!
“Our model uses only 377 MB, which even fits into the memory of a smartphone”

Okay. But wouldn't this be more useful as a web service?

-Realtime feedback -Better privacy
Bad guys posting internet selfies are now going to get hunted by law enforcement which is a good thing.