Raffi Hotter
@raffi_hotter
Followers
1K
Following
2K
Media
40
Statuses
351
neurotech | prev: biophysics @harvard, neurotech intern @theteamatx, math/cs @mcgillu @ucberkeley
San Francisco, CA
Joined November 2018
Check out our results for new ways of brain-computer interfacing!. Blog post:
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
1
1
78
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).
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)
41
53
839
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).
3
5
194
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)
7
13
123
@ElmoTheHokage 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.
4
0
111
Thank you @santiaranguri ,@selenazhxng and Simon.Tartakovsky for thinking through this with me!. Here's the code for the entropy calculations:.(10/10).
4
1
97
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).
3
0
94
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)
2
2
93
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).
2
1
87
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!
11
4
92
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)
2
0
85
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).
4
1
83
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).
1
0
80
What is the entropy of this noise? I simulated the noise and computed its entropy. It's 1.9 bits per timepoint. (6/10)
2
0
71
.@selenazhxng and I are hosting a learning “potluck” in Cambridge on Sunday!. Instead of bringing a dish, you bring a topic to teach another person 1-on-1.
6
0
63
.@_marleyx and I are super excited to unveil our first product. The world's first telekinetic citrus
4
2
53
This is incredibly cool. It's a way to reduce optical scattering in tissue. Scattering limits how deep you can see inside tissue at high resolution. For example, a mouse brain is about 1 cm in diameter, but you can only see the first few hundred microns optically. (1/5).
Congrats to my friend and colleague Guosong Hong for his stunning and original discovery, published today in Science, on clearing tissues *in living animals* with a common food dye!. The dye is tartrazine, used in Doritos!.
1
4
39
How do you make friends who inspire you? Joining Interact is the best hack I know.
Applications are now open for the 2022 Interact Fellowship!. Learn more and apply here by February 15th:
2
2
39
Ultrasound propagation through the head. [Creds: Lluís Guasch et al., 2020]
1
2
32
Disclaimer: I don't work at GE/Siemens/Philips, so this is just my best guess. Please tell me if I'm wrong!. Here's a spreadsheet with all my calculations:. A big thank you to Simon Tartakovsky for thinking through this with me!. (12/12).
1
0
26
Did you know EEG was invented to see if humans could telepathically communicate via electric fields?
3
1
26
2
1
22
I read the ASME standard and it seems you'd need ~1 cm of stainless steel casing to prevent buckling. This comes out to 1,200 pounds of stainless steel!. So the heavy part in an MRI is actually the metal needed to contain the vacuum! (8/12)
1
1
21
I wrote a list of questions a few years ago related to brain-computer interfacing. Sharing them, in case other people would find them interesting:. (link in next post)
1
0
20
MRIs are so heavy that you can't even transport them using a freight elevator. Instead, a common way of installing an MRI is ripping out a wall and bringing it in with a crane. Crazy! (2/12)
1
0
19
Why do humans love music so much? Are there any good evolutionary reasons?.
6
0
18
But what about low-field MRI scanners like @Hyperfine?. The magnetic field used in those scanners is ~50x lower than a typical scanner. So you don't need superconductors. But the signal-to-noise ratio scales with field strength, so the low field scanner is noisier. (10/12)
1
0
18
Some questions:.- Could you use methods other than a vacuum to keep the wire cold? Like a refrigerator?.- What if you used higher temp superconductors?.- Why not use lighter metals?.- What if you used stiffening rings to reduce buckling?.- Where's the rest of the weight?.(11/12).
1
0
18
This post is really interesting. The idea is that when you add noise in the diffusion process, the noise first gets rid of the high spatial frequencies. As the image gets noisier and noisier, the lowest spatial frequencies are the last to go.
Diffusion is the rising tide that eventually submerges all frequencies, high and low 🌊. Diffusion is the gradual decomposition into feature scales, fine and coarse 🗼. Diffusion is just spectral autoregression 🤷🌈
1
0
17
Is it just me or has the field of neuroscience learned very little over the past ~50 years?.
3
0
18
The magnetic field in an MRI is generated by a bunch of wire wrapped around a tube (i.e. a massive solenoid). At first, I thought the weight was mostly from the wire. It turns out you need ~50 km of wire (!!) to generate a 3 Tesla field. But this only weighs ~100 pounds. (3/12)
1
0
17
Made a little animation of spherical harmonics (link below)
1
0
15
There's another somewhat heavy part. The large magnetic field generates a big (Lorentz) force on the wire. Without a strong enough material to contain these forces, the wire would rip apart. I calculated that you need ~400 pounds of aluminum to contain the forces. (9/12)
1
0
14
Nope. It turns out ~10 pounds of liquid He is sufficient. But we're getting closer!. The problem is, if you don't insulate the He, it will boil off (and liquid He is very rare + expensive). To insulate liquid helium, MRIs use a vacuum layer (like a Thermos!) (6/12)
2
0
14
Vacuums don't weigh anything, so that's not the problem either. But, large vacuums exert tremendous force on their container (think about the pressure differential). So the vacuum container needs to be extremely strong and thick. (7/12)
1
0
14
@thenabanana You might like Chapter 9 “n-Dimensional space” from The Art of Doing Science and Engineering. Lots of cool related results presented there.
2
0
13
A neat signal processing result is that derivatives are high-pass filters and integrals are low-pass filters.
0
0
13
Wires with 0 resistance are called superconductors. We haven't yet figured out how to make room-temperature superconductors. So, we cool the wire to -269 ºC (4 K) to make it superconducting. People use liquid helium for this. Could the weight be from the helium? (5/12).
1
0
12
So where is the rest of the weight?. When you send large currents through a wire, it heats up a lot (heat = current^2 x resistance). But if you decrease the wire resistance, the heat goes down. The trick in MRI is to make the wire have 0 resistance. (4/12)
1
0
12
This is one of the best articulations of the goals of neurotech I've read. "From a PR standpoint, neurotech doesn’t have a single, intuitive goal like climate tech, longevity research, or to a lesser extent AI. Instead it’s like the early days of personal computing".
1
1
12
So awesome to be part of this amazing group of people.
A $1,000 brain scanner. A crypto key you can’t lose. Settling new cities in Africa. This is The Frontier. Meet the winners of our latest tournament:
1
0
10
Superlenses are so cool! They overcome the diffraction limit by using a material with a negative index of refraction
1
0
10
It turns out that the refractive index and the absorption coefficient are linked, through what's known as the Kramers–Kronig relations. If you know the absorption at every single wavelength, you can determine the refractive index, and vice-versa. (4/5).
3
0
8
The diffraction limit is a limit on bandwidth, not resolution. You can surpass the resolution limit by assuming your signal is sparse (super-resolution techniques).
1
1
8
Health insurance has this really weird property: the richer you are, the cheaper the health insurance you should buy.
4
0
8
Asking friends "do you have feedback for me?" has been so valuable.
1
1
7
@ansonyuu @_ivyzhang @_marleyx we first heard about AE from @thomas_rribeiro, who has this magical talent for finding fascinating research that nobody knows about. we learned about functional ultrasound first from a paper by @SumnerLN++. They used it for a bci in macaques
0
0
8
@LauraDeming I think you'd like @a__bra's page that has some very pretty math animations!. . For example, here's one on the matrix lie group orbits
1
0
8
found this gem of a blog about MRI physics. it answers questions like:.- how does SNR scale with field strength?.- why are T1 and T2 of water what they are?.
3
1
7
Collaborative web browsing is awesome. I've been beta testing Curius and have discovered so many interesting reads because of it.
0
0
7
is this agi?.
We're releasing a preview of OpenAI o1—a new series of AI models designed to spend more time thinking before they respond. These models can reason through complex tasks and solve harder problems than previous models in science, coding, and math.
3
0
6
@_marleyx and I are running an experiment!. We're reading a neurotech paper with a small group of people who have expertise in math/physics/electrical eng/mech eng/material science/chemistry. dm either of us if you're interested!.
1
1
7
If you’re curious about the latest BCI progress, check out this thread 👇.
Brain-computer interface (BCI) 2021 year in review 🧵. 2021 was a big year! We saw an intracortical BCI decode handwriting, the 1st country to pass a bill on neuro-rights, & much more. Below we cover the most exciting BCI updates across academia, industry, & public policy.
1
1
6
The reason for this is that the frequency spectrum of images decays with frequency, but the noise added is the same across all frequencies. So diffusion is like auto-regression but in the frequency domain.
1
0
6
We’ve hosted this in the past and it’s been excellent!. Some past topics:.- Quines (self-replicating code).- Feminism in the works of Mary Wollenstonecraft.- Neuroscience learning vs AI learning. DM me or @selenazhang if you want to come!.
1
0
6
Depression will be the largest medical BCI market, according to the Morgan Stanley BCI report
1
0
6
@algekalipso My understanding is that amplifiers have two main types of noise:.1. Flicker noise, or 1/f, which is dominant at lower frequencies.2. Thermal noise, whose frequency spectrum is flat. As you go to higher frequencies, thermal noise dominates.
1
1
5
Congrats to my friend and colleague Guosong Hong for his stunning and original discovery, published today in Science, on clearing tissues *in living animals* with a common food dye!. The dye is tartrazine, used in Doritos!.
0
0
6
I made this while trying to understand why EEG has poor resolution (it turns out it's connected to spherical harmonics!).
0
0
5
The common explanation for why EEG is bad is that each electrode listens to a large population of neurons. But that alone doesn't explain the poor resolution. In a CT scan, each measurement also listens to a population of spots, but the resolution is still very high!.
0
0
6
BCI is focused on reading/writing to neurons, but what about synapses? E.g. downloading knowledge to your brain.
1
0
4
Scattering happens when 2 materials have different refractive indices. In tissue, lipids and water have different refractive indices, so you get scattering. The idea here is to increase the refractive index of water so that it more closely matches that of lipids. (3/5).
1
0
4
Less than a week left! Message me if you have questions!.
There is one week left to apply to Interact's 10th Fellowship cohort. Learn more about Interact and apply now at
0
0
5
@sabrinasngh Hot take: school (if done properly) is pretty good for learning physics! You can walk into a class at any uni and just start taking it.
1
0
5
Congrats Alex on the launch!! Looks awesome.
We are excited to finally have our Product Hunt launch!.Check out our post here:.
0
0
4
If the only cars allowed on the road were self-driving cars, would self driving become an easy problem?.
4
0
5
The conscious part of the brain is the CPU and the unconscious part is the GPU.
0
0
5
Could the aliens be already here but hiding with invisibility cloaks?.
0
0
6
@algekalipso If the noise was mostly just present at low frequencies, then it would be highly correlated in time.
0
0
4
We're incredibly fortunate to be partnering with @guillefix @selenazhxng @yush_g @SebastienZany @turboblitzzz.
0
0
4
does anyone else picture proteins in their minds as kind of like transformers? (from the movies).
0
0
4
Magnetofluorescence is one of my favourite recent discoveries in science!. We’re hosting @AndrewGYork for a whiteboard talk tomorrow night in sf. Details in the next tweet.
Meet MagLOV: an engineered protein that responds STRONGLY to magnetic fields. This is a fluorescence timelapse of MagLOV in E. coli. We're waving a (small) magnet under the plate. Can you tell where the magnet is?. Want some? It's on Addgene now!
1
0
4
An underrated feature of ChatGPT is that it works well in places with low internet speeds (traditional web pages are much slower).
0
0
3
@jasminewsun This post+comments were very interesting:.
Why does privacy matter? What are the best principled arguments for it? Where to set the boundary [i.e., when is it best to require that certain actions be more broadly known]?. My own answer to this bundle of questions isn't very good. I've read a few classic books, papers,.
1
0
4
The way they do this is by adding a dye that absorbs in the ultraviolet range. Why does adding an absorber reduce the scattering? (2/5).
1
0
3
@yasmeenbrain For one cool application of information theory in the brain, Chapter 3 of Principles of Neural Design is excellent (in particular, the section "A neuron’s information capacity")
1
0
4
"If people are sufficiently worried, then there is a lot less to worry about. But if no one is worried, that's when you should worry.".
With recorded COVID-19 cases (outside china) so eerily matching an exponential, I couldn't resist making a primer on exponential/logistic growth. At least 3 counterintuitive things about this kind of growth seem worth putting into the discussion.
0
0
2
@uzpg_ There’s a neat info theory argument for why neurons are silent most of the time (they spike only ~0.1% of the time). The argument is that low spike rates maximize the information per unit of energy.
@yasmeenbrain For one cool application of information theory in the brain, Chapter 3 of Principles of Neural Design is excellent (in particular, the section "A neuron’s information capacity")
0
0
4
Why is YouTube search so much better than Google search for finding introductory technical material?.
4
0
3