Hacker News Reader

Comment by canada_dry | Hacker News Reader

> The second step in classification requires analysis. It takes 30 to 45 seconds to get a look at a new detection and discover what kind of engine she’s running, what kind of hydraulic pumps are online, and what screw blade configuration she is using.

One can imagine ML making this job somewhat obsolete in the not-too-distant future.

View on HN

Not sure why this was downvoted - it almost certainly is heavily automated already, and the article hints at it. Tom Clancy described a primitive version in The Hunt For Red October in the 80s. You'd expect the Navy to be very interested.

I was at a day of research presentations in January before everything shut down. One of the presentations was on using ML to identify individual dolphin speech signatures--with some degree of accuracy. One of my thoughts was that the more sophisticated navies probably already have this sort of thing pretty well developed.

Is it really something for ML? Shazam isn't ML, and it's the same basic thing. Take the current sound and then compare it to a huge database of known sounds. Report the result.

Agreed. ML and AI are (currently) useful in situations where there is a lot of data, of different types, usually from disparate sources, often sparse, with no known pattern or causal relationship. Shazam and sonar analysis don't really fall into any of those areas.

Can confirm that USN and sonar systems have been using "ML" techniques since the 90s. Y'all were late to the game!

That's if they aren't already.

Without having the data I can't try but I'd guess that Fourier analysis (or something a bit more clever to deal with noise e.g. Welch) would get you a lot of the way there even without invoking AI

Even if you're also using ML, Fourier transforms can be pretty useful in this sort of analysis.

This is not based on any classified information and sonar signal processing is not something i know much about, but cyclostationarity (and other higher-order statistical signal processing) does tend to be useful when processing signals that are generated by processes where there's multiple periodicities at work: https://en.wikipedia.org/wiki/Cyclostationary_process#Angle-...

Signal separation techniques for characterizing cyclostationary signals are also more robust to noise. For example, analysis of the second order FFT can resolve the signals of incipient faults in rolling bearings because it can resolve the signals of the bearing components rotating (at a much lower frequency than the shaft) whereas simple FFT analysis would generally only measure the magnitude of the structural resonance of the machine, with spikes at the mains, rotational, and slip frequency.

From TFA: "Patterns are matched and measured with onscreen tools and cursors. Automatic artificial intelligence algorithms flag potential targets for the sonar operators to verify. "

If you ever watched the hunt for red october, they already had computers doing this (doubt it was ML'd tho).

It is ML depending on what we're calling ML. Is gradient descent/MLE ML? If yes, then the Dallas (the US sub) in Red October was using ML.

Clancy did embellish a few things and speculated rather a bit to make an entertaining scenario.

All told, a good solid read, and accurate in a bunch of not-quite critical details, but don't lean too hard on the edges of the story.

Same with Red Storm Rising. Very well told, close to a solid "what if", but lots of little touches to make it entertaining as well.

https://xkcd.com/1897/

Reply on news.ycombinator.com