Compression could be machine learning's "killer app"


Originally published at:


Unless I’m mistaken the link to the article seems to be borked


It seems to be broken, which is better then being Bjork-ed.


“ML systems could convert your speech to text, then back into speech using a high-fidelity facsimile of your voice at the other end”

Eh - give me a facsimile of another voice - maybe Eartha Kitt.


I imagine that would be a huge boon for communications when we finally go to Mars.


Fixed! Sorry, dropped a quote.


fun fact, Bjork’s voice is already run through an ML algorithm to convert it from her native multi-dimensional quantum wave-forms into regular human languages


Is it a Middle-Out Compression Algorithm with a Weissman score of less than 2.9 and optimal tip-to-tip efficiency?



Machines deciding what is important and what isn’t? WCPGW?


Not borked, compressed!


It might be fun! Like running your text through Google Translate into Urdu and then back to English, except machine learning is much, much more alien than Urdu.

Also, I’m surprised nobody has pointed out that bandwidth is cheap and getting cheaper, while confused communication can be very very expensive.


Like if it transforms your elephant to a sofa.


This idea appears in Lucifer Rising, a Doctor Who novel (don’t judge me) from the nineties. I remember thinking it was interesting although silly. Basically, in the distant future, instead of recording video, computers log events in some kind of script-like narrative form (it’s a little vague) and use that to play back reconstructed 3D scenes. Aspects of the plot revolve around a murder investigation, and how evidence in this format is highly dependent on how you reconstruct facial expressions, emphasis etc.

In terms of compression, I could sort of buy the idea, because I don’t think storage space will ever become so abundant that people can’t find ways to run out of it, and even in the distant future it might not be feasible to constantly record everything without some kind of drastic compression measure like this.

But as a medium, this idea takes you into much deeper epistemological waters than you might think, if you hadn’t read McLuhan and/or paid attention to the continual setbacks of 20th-century AI research. The default post-Enlightenment assumption is that impressions of the world can be broken down in a single, canonical, hierarchical way, as in, a photo can be filed under “still life” / “images of flowers” / “in vases” / etc. If that were true, AI would be straightforward, if not easy. But it’s not even close to true – there are a trillion different ways to summarise even a simple photo – and cognitive science has been (is being) very grudgingly forced to admit that these complexities are not peripheral; they strike right at the heart of the single-giant-encyclopedia model of knowledge.

When you see a movie from the 1930s, it feels more alien than a novel from the 1830s, because movies record all the now-strange details and mannerisms that the filmmaker didn’t mean to call attention to. Where a novel would say “he walked down the street,” a movie can’t help but say “he walked down a street where there were all these funky-looking cars and a mother hitting her child on the face and ads for radioactive soap and people smoking in a deli”. If you used a machine to convert the latter into the former, that could be very interesting but it’s not “compression,” it’s an apples-to-oranges conversion.


Being Bjork-ed isn’t bad. It’s just weird.


This topic was automatically closed after 5 days. New replies are no longer allowed.