Raffi Hotter (@raffi

Raffi Hotter

516 posts

Raffi Hotter

@raffi_hotter

co-founder @alephneuro

San Francisco, CA

Joined November 2018

Raffi Hotter
@raffi_hotter
May 29, 2024
200x compression is mathematically impossible for this data. Given the specs of Neuralink's chip, you cannot do better than 5.3x compression. Here's why 🧵 (1/10)
Felix
@felix_red_panda
May 24, 2024
Neuralink started a compression challenge. They're asking for people to find methods to losslessly (!!) compress files to 1/200th (!) the size but the files are extremely noisy so thats certainly not possible 🙁 (spectrogram of one randomly chosen file for illustration)
519K
Raffi Hotter
@raffi_hotter
May 29, 2024
Replying to @raffi_hotter
Signals are not infinitely compressible. You cannot compress a signal to less bits than its entropy. This is formalized by Shannon's noisy-channel coding theorem. (2/10)
en.wikipedia.org
Noisy-channel coding theorem - Wikipedia
19K
Raffi Hotter
@raffi_hotter
Jan 6, 2025
Why are MRIs so damn heavy? They weigh over 10,000 pounds, or about 3 cars' worth. MRI design knowledge is hidden inside large companies, so I tried to work it out from first principles. 🧵 (1/12)
20K
Raffi Hotter
@raffi_hotter
May 30, 2024
Sorry, I should have clarified that this thread is about lossless compression. The challenge asks for a lossless solution. My point is that if 200x compression is desired, it has to be lossy.
6.3K
Raffi Hotter
@raffi_hotter
May 29, 2024
Replying to @raffi_hotter
Thank you @santiaranguri ,@selenazhxng and Simon Tartakovsky for thinking through this with me! Here's the code for the entropy calculations: github.com/rhotter/neural… (10/10)
GitHub - rhotter/neuralink-bounds
From github.com
13K
Raffi Hotter
@raffi_hotter
May 29, 2024
Replying to @raffi_hotter
It's hard to compute the entropy of a signal. But part of the Neuralink signal is noise. We can lower bound the entropy of the signal by computing the entropy of the noise. We look at the spec sheet of Neuralink's chip to compute the entropy of the noise. (3/10)
19K
Raffi Hotter
@raffi_hotter
May 29, 2024
Replying to @raffi_hotter
Neuralink's chip introduces noise when it amplifies and digitizes the signal. From their spec sheet: - standard dev of chip noise is 5.9 µV - signal is quantized in 7 µV intervals (4/10)
19K
Raffi Hotter
@raffi_hotter
May 29, 2024
Replying to @raffi_hotter
So you cannot compress the signal to more than 1.9 bits per time point. Since the signal was originally 10 bits, that's a max of 10 bits / 1.9 bits = 5.3x compression. (7/10)
18K
Raffi Hotter
@raffi_hotter
Jan 16, 2025
I've been trying to think of new ways of imaging the brain (non-invasively). So I started making a table of all the physical properties in the brain (the rows) and all the ways to modulate them (the columns). Lmk if you have things to add to the table!
3.7K
Raffi Hotter
@raffi_hotter
May 29, 2024
Replying to @raffi_hotter
These two numbers are close to each other. So noise will often cause the signal to hop to the next quantization level. E.g. a signal of 3.12 µV digitizes to 0 µV, but adding 5 µV of noise causes the signal to digitize to 7 uV instead. (5/10)
16K
Raffi Hotter
@raffi_hotter
Feb 7, 2025
animation for upcoming blog post on why EEG sucks
00:00
5.3K
Raffi Hotter
@raffi_hotter
May 29, 2024
Replying to @raffi_hotter
2) My simulations assume the chip noise is iid (independent and identically distributed). The chip noise could have correlations in time, which would reduce the entropy. But it seems unlikely at this high frequency, and it likely isn't enough to make 200x feasible. (9/10)
14K
Raffi Hotter
@raffi_hotter
Oct 15, 2024
Check out our results for new ways of brain-computer interfacing! Blog post: brainhack.vercel.app
marley 📐
@_marleyx
Oct 15, 2024
Can we invent new brain-computer interface modalities? @raffi_hotter and I got 9 friends together and built a lab at home to test two totally new imaging methods: acoustoelectric imaging & functional ultrasound through the skull 🧵 story that involves nV measurements, pretty
GIF
00:00
16K
Raffi Hotter
@raffi_hotter
May 29, 2024
Replying to @raffi_hotter
Some caveats: 1) The specs from the chip are from Neuralink's 2019 paper. Perhaps the specs have improved since then. But I also ran noise simulations at 1/2 and 1/4 the noise levels, and the max compression is still only 9.5x and 63x, respectively. Far from 200x. (8/10)
14K