Hacker News Reader

Comment by tantalor | Hacker News Reader

"machine learning algorithms decode the patterns created when polarized light shines through the glass"

Unclear why ML is required versus old-school methods.

"As fans of The Mote in God's Eye already know, one remaining question must be answered for any data storage method expected to last for millennia—what happens when the technological and cultural context surrounding a storage medium collapses? Silica addresses this problem also, by using initial "ground truth" tracks. The team is using machine learning algorithms to re-read Silica's data, and in the event of the loss of those trained algorithms, fresh algorithms can train very rapidly on the "ground truth" tracks, which teach them how to interpret the rest of the data." https://arstechnica.com/gadgets/2019/11/microsofts-project-s...

You just need to get your civilisation back the point where you can write ML algorithms first.

Are our brains so inferior? Wouldn't someone who studies those ground-truth records long enough learn to see the data?

Yes, they would - and it actually says that in the article. But you'd then need some non-ml way to read & play it back, which would still require readily available computers & lasers, etc...

ML is the new way to refer to old-school methods such as regression, statistical inference, etc.

Ohh, ML is the new term for AI now.

We've finally accepted the computer doesn't have any information in it yet, and that we have to (somehow) feed some in.

Heh.

That was my thought. If the data is written in some specific format, shouldn’t it be able to be read just as easily, without resorting to ML?

I think the idea is to make it readable even if civilisation ends and we don’t have any record of the format when the tech is re-invented afterward — A built-in Rosetta Stone.

But you'd still need to be able to run the classifier, right? What's the difference between that and a piece of code that just knows the format? There's still a format here, regardless of how it's being decoded.

I’m not sure I follow. The idea (if it works is another matter) is that a future civilisation can say “we don’t know what this format is, let’s throw machine learning at it” and even though that machine learning system doesn’t know anything about the format when it starts, it can work it out on its own.

That would work regardless of the format, so why use ML today? Today you just write a decoder that assumes the format, in the future it can be reversed engineered using whatever means they have available to them.

Would it? I have not heard anyone suggest that JPEG could be decoded using only a self-training ML system.

Why wouldn't it? What makes a format more amenable to being decoded via ML?

This format comes with its own metaphorical Rosetta Stone.

Reply on news.ycombinator.com