Hacker News Reader

Comment by ericsoderstrom | Hacker News Reader

The author says that training their own model would have been too hard due to lack of training data, but evidently Rekognition had sufficient training data to make it work? Why can't NYT use the same training set Rekognition uses? Does Amazon somehow have a secret non-public collection of celebrity photos?

View on HN

It shouldn't take an intern too long to collect a representative set of Congress people and other high officials for training. Maintaining it would not be an undue burden. That would eliminate the false positive matches for all the unwanted celebs. Clearly Amazon's models aren't that great to begin with so there's little reason to stick with them.

Wrap it up into a simple native app and you can bypass the MMS BS. Even better, a sufficiently capable dev could integrate an opensource recognition library [1] to have it entirely implemented on the device.

[1] https://github.com/rudybrian/tuFace

Hi! I'm Jeremy, one of the developers.

We'll probably work on something like this for the next version. One reason it's harder than you think: We would have to buy / own rights to the photographs before we could use them to train -- most of those photos are owned by Getty or the AP. And our own photographs are perfectly lit and square, which made them awful for training face recognition.

The other hangup (which I didn't get to in the article) is having to add / remove people. New members are constantly being added and that's a maintenance burden for us. Amazon usually has the new member within a day or two. (Our team is very small and we have a lot of other responsibilities!)

But good points, definitely.

"We would have to buy / own rights to the photographs before we could use them to train..."

Is this actually true?

In USA I don't think it is because the end use is transformative [1].

In UK it would be tortuous because it relies on Fair Use to temporarily store the images in order to extract the facial structure data. Fair Dealing is really draconian in comparison.

[1] https://www.lib.umn.edu/copyright/fairuse

Really doesn’t matter - the legal team at the NYT thinks it might be, and lawyers exist to tell people “you’d better not”.

And it's our job, as someone who knows what a computer is, to move forward with common sense if they're overreaching which, is their job.

They have every incentive to be as conservative in their advice as possible, and no incentives to "allow" risks. Doesn't increase their compensation any.

> We would have to buy / own rights to the photographs before we could use them to train -- most of those photos are owned by Getty or the AP.

I think your model would be covered by derivative art... unless you started selling the model itself.

> We would have to buy / own rights to the photographs before we could use them to train

Please do elaborate on who's enforcing this mindset on you/your team.

But...doesn't congress maintain some sort of api for available bio data - possibly including photos - to avoid the maintenance issue brought up in the piece? A quick google search shows propublica has sych an api, and it seems to have originally been developed by NYT... https://projects.propublica.org/api-docs/congress-api/

Rekognition crawled and annotated millions of images of different celebrities to train their face recognition model. Once you have an accurate model for a lot of classes it's much easier to add new ones with just a few samples.

> Once you have an accurate model for a lot of classes it's much easier to add new ones with just a few samples.

This is pretty cool. Do you know of any good references for stuff like this? Not sure what the right topic name would be: online learning? streaming?

https://blog.keras.io/building-powerful-image-classification...

This is a good start

Transfer learning

This is known as transfer learning, see [0] for an approachable example.

[0] https://www.mathworks.com/help/nnet/examples/transfer-learni...

Thanks for the replies!

Reply on news.ycombinator.com