This is utterly wild... am I a little smooth-brained, or is this totally unrealistic?
- 200 Mbps compressed to under 1 Mbps
- Realtime, i.e: ~1ms latency or better
- Encode and transmit in <10mW (yes, including radio)
- Lossless
- High-entropy input
... if you succeed, just email them your solution - no mention of any reward or Nobel Prize nomination.
https://content.neuralink.com/compression-challenge/README.html
@attie They're really not asking for compression. "one hour of raw electrode recordings from a Neuralink implant" can only be compressed if someone figures out what it means. They want comprehension, not compression
@cliffordheath Given the compute and power budget, that's even more laughable, surely?! And can comprehension really come with a 1ms latency?
@attie yes, it's utterly laughable, more so because they don't seem to realise that this level of comprehension would allow the chimp's brain to be replaced by an electronic one (we could accurately simulate it). If we could figure out how to do that, we wouldn't need Neuralink. If anyone was that smart, the rest of us would be on the menu
@attie Just in case anyone wonders why I think this, think this: "compression is prediction, because it must remove everything predictable"
This is also a reason why their request is likely simply impossible. I expect the signal to have some amount of white noise, and they ask for lossless compression, so they are asking for the noise to be reproduced exactly.
They have 10b samples and want more than 200x compression (presumably over linear PCM, which is the format of the files they published). This can only work if any white noise present itself takes less than 1/200 of the data rate, so if the white noise's amplitude is on the order of 2^-20 of their resolution (encodable in ~1/20 bit/sample).
(I speak of white noise only, because I'm assuming it's uncorrelated in different samples, both across electrodes and across time.)
Yeah, they could simply describe a relaxed comparison that would at least free implementations from having to reproduce quantization noise (e.g. pointwise difference between original and decompressed at most 1 everywhere). But if they wanted to ask people not to faithfully reproduce other kinds of noise, they would need to first mostly solve the problem by describing how to compare two signals for equality modulo all those kinds of noise.
@robryk @attie I imagine they think a comprehensive brain model might permit a coherent signal under the noise (whether white or any other spectral distribution) to be recovered 100%. Clearly they cannot require the original data back, they just want the meaning unchanged, but they don't even know enough information theory to know how to ask for that