[Macuyiko]

Hopefully we'll come to a point soon where AI and deep learning can help out in this space. I remember models being trained that could predict the country based on a street view photo with a high degree of accuracy, which would be a good first step. The next one would be to either match a top-down satellite photo with a picture, or figure out a way to match street view photos based on similarity.

Given that Google actually has the data as well as the skills and computing power to help out Europol, I'd have hoped they'd be jumping at the opportunity to do so. At the least it would make for great PR, but perhaps they're afraid of the possible privacy issues and discussions as well (I can already see the headlines: "Google knows the location of any yard appearing in a photo!").

In any case, it's telling how humans are so good at this stuff and picking up on clues and patterns. It'll still take a while before machines are as good.

[zenexer]

That’s not really how any of this works—you can’t simply throw a bunch of unrelated data at some algorithms and expect usable output. (There’s also no such thing as “deep learning”.)

Law enforcement and related organizations already use machine learning quite a bit, particularly for image enhancement. Yes, Google does have a lot of images of various locations from a top-down perspective, but that isn’t helpful for accurately determining a location from the images that Europol collects.

Also, keep in mind that Europol is only posting images here when all other means of determining location and identity have been exhausted. The images are usually indoors and don’t contain enough information for ML to be of any significant use. You might be able to narrow it down to a probably country based colors and design patterns, but that’s hardly sufficient and not solid enough evidence to actually do anything.

[areyousure]

> There’s also no such thing as “deep learning”.

What do you mean by that?

https://en.wikipedia.org/wiki/Deep_learning

[Macuyiko]

> That’s not really how any of this works—you can’t simply throw a bunch of unrelated data at some algorithms and expect usable output.

That is never what I claimed. First, note that I took a pretty (IMO) balanced view and indicated that this is still a hard setting. Second, note that I did indicate that sufficient training (i.e. labeled) data would be required.

This is what was possible in 2016: https://www.theverge.com/2016/2/25/11112594/google-new-deep-...: "The new deep-learning program churns through millions of photos to determine the best match."

Also see project of a fast.ai participant: "Which of the 110 countries a satellite image belongs to?" (point 13 here: https://forums.fast.ai/t/deep-learning-lesson-2-notes/28772)

> (There’s also no such thing as “deep learning”.)

- https://www.deeplearningbook.org/

- https://www.coursera.org/courses?query=deep%20learning

- https://eu.udacity.com/course/intro-to-tensorflow-for-deep-l...

- https://www.edx.org/professional-certificate/ibm-deep-learni...

- https://www.deeplearning.ai

- https://www.fast.ai/

> Yes, Google does have a lot of images of various locations from a top-down perspective, but that isn’t helpful for accurately determining a location from the images that Europol collects. You might be able to narrow it down to a probably country based colors and design patterns, but that’s hardly sufficient and not solid enough evidence to actually do anything.

Maybe not completely, but again: being able to narrow it down would already be an incredible help, especially for outdoor pictures (which were also shown in the article's video). I never claimed that a model would completely replace the human process.

Also, I find the downvotes (not saying you) on my initial comment to be in pretty bad form. I'm not Jeremy Howard or Andrew Ng, but don't think I was blowing smoke, and work in the area of data science and ML.

[tomohawk]

Maybew Europol could do a Kaggle competition?