SoylentNews is people
SoylentNews
The US government has placed software designed to train neural networks to analyse satellite images under new export controls in a bid to prevent foreign adversaries using said code.
The decision, made by Uncle Sam's Bureau of Industry and Security (BIS), is effective today. Vendors shipping software subject to the controls – in that the applications help machine learning systems annotate satellite images in a particular way – will have to apply for a license to sell their products to customers outside of the US and Canada.
"Items warrant control for export because the items may provide a significant military or intelligence advantage to the United States or because foreign policy reasons justify control," the BIS said.
Hah!
This discussion has been archived.
No new comments can be posted.
US to Restrict Export of AI Technology for Image Analysis
|
Log In/Create an Account
| Top
| 21 comments
| Search Discussion
The Fine Print: The following comments are owned by whoever posted them. We are not responsible for them in any way.
I don't want a pickle,
I just wanna ride on my motorsickle.
And I don't want to die,
I just want to ride on my motorcy.
Cle.
-- Arlo Guthrie
(Score: 2) by ikanreed on Wednesday January 08 2020, @07:06PM (3 children)
There's huge fucking hardware requirements when you start talking about the 50 layers used in ResNet-50 operating in reasonable time, but the overall code structure is remarkably compact.
(Score: 2) by HiThere on Wednesday January 08 2020, @07:35PM (2 children)
There's certainly a lot of generalized structural similarity, but there are also a large number of specialized functions. Most of the code will use identical modules, but some will require specialized modules, and those will need to be integrated into the rest of the code.
Perhaps I'm overestimating the abilities of the current AIs, but a real AI will need to focus the object recognition on what it's looking for. Skim over most of the image without processing deeply, but when it sees something that looks plausible, evaluate it more closely. Then, if the image is partially obscured, or rotated in an uncommon way, it will need to pass that up to a higher level which will evaluate it again in a larger context.
Even simple object recognition isn't that simple, and when you're looking something in particular the processing gets extra complex (though better at producing the right answer). Part of recognition is deciding whether something that you've possibly seen is reasonable to see in that context. (This is why the students see the demonstrator being stabbed with a knife rather than a banana. So it can lead to errors.)
Javascript is what you use to allow unknown third parties to run software you have no idea about on your computer.
(Score: 2) by ikanreed on Wednesday January 08 2020, @07:44PM (1 child)
Well, now you're starting to get unfair. Those t-shirt encryption perl scripts aren't dumping all of TLS with its socket management, streams, error resiliency, key validation, certificate lookup, and the host of complex infrastructure around encryption. They're shifting some bits and doing some math. Of course there's application specific contextualization that needs to happen.
The actual neural net guts are real damn simple, at least in terms of "where the magic happens".
(Score: 2) by HiThere on Wednesday January 08 2020, @11:14PM
Well, if you restrict it to actual visual image processing, and don't consider object recognition as significant, then yeah, you could do it, but you everyone else already has that capability, and has had it for decades.
Javascript is what you use to allow unknown third parties to run software you have no idea about on your computer.