Comment Permalink

TsviBT8d7-1

Something I'd like WBE researchers to keep in mind: It seems like, by default, the cortex is the easiest part to get a functionally working quasi-emulation of, because it's relatively uniform (and because it's relatively easier to tell whether problem solving works compared to whether you're feeling angry at the right times). But if you get a quasi-cortex working and not all the other stuff, this actually does seem like an alignment issue. One of the main arguments for alignment of uploads would be "it has all the stuff that humans have that produces stuff like caring, love, wisdom, reflection". But if you delete a bunch of stuff including presumably much of the steering systems, this argument would seem to go right out the window.

https://www.lesswrong.com/posts/jTiSWHKAtnyA723LE/overview-of-strong-human-intelligence-amplification-methods#Brain_emulation

See in context

8 Straightforward Steps to Marginally Improve Odds of Whole Brain Emulation

by Dom Polsinelli

24th Mar 2025

7 min read

8

I want to make a small disclaimer that I have no formal training in most of this. I am pretty confident about my understanding of electrophysiology and some robotics but much less so about anything biochem/genetics related. If there are any mistakes I am happy to correct.

Improving Automated Patch Clamping

Patch clamping is the gold standard for electrophysiology and understanding individual cellular properties is just as important as synaptic weights for whole brain emulation. I was personally shocked to find out how sparse the current library of electrophysiology features (hereafter referred to as e-features as keeping with one of the cited papers) was. The Allen Institute has made great strides with their brain atlas. However, there is still quite limited data. Only 107 cells with a full biophysically active model (that is, they tuned a neuron model to recreate observed e-features of these cells)^[1]. I believe there are only ~500 with e-features, morphology, and whole transcriptome but no model fit^[2] . I may not be exactly correct about that and there might be some split for human vs mouse neurons but 100s is the right order of magnitude. I heard a rumor in this PhD dissertation^[3] that there were datasets with >10,000 neurons with at least e-features but when I went into the citations I could not find those. If anyone finds a larger dataset with morphology and e-features please let me know. Doing additional patch clamp experiments would fill gaps in our knowledge of different cell types’ e-features and serve as benchmarks to compare how well we can emulate neurons. It is still unclear how accurate any model has to be to capture all computationally relevant activity in the brain. Even perfect e-features + a model of both electrical and chemical synapses may not be enough.

In any case, there is a seeming lack of neuron models that capture e-features for diverse cell types and specific cells. Even if they can be captured with current models (Hodgkin-Huxley, Izhikevich), this ought to be validated on more than ~500 cells. To do this, we need robust, fast, automated patch clamping. There is work into exactly such a device. Autopatcher is a project in collaboration between researchers at MIT and Georgia Tech that can currently automatically run patch clamp experiments on up to 4 cells simultaneously in vivo. However, it is still far from perfect, successfully attaching to at least 2 cells in less than 50% of cases^[4]^[5].

Measuring mRNA Transcriptome of Cells Before and After Cryopreservation

I believe that measuring the transcriptome of individual cells could help inform recreation of e-features, this will be discussed in the next section. Taking this as given for the time being, can we measure this after cryopreservation? Backing up slightly, the transcriptome is the level of mRNA for certain genes expressed. As an example, if you look at genes coding for a specific ion channel you can infer how dense those ion channels are on that particular cell compared to other neurons which should be helpful for e-features. The Allen Institute has data of transcription levels obtained simultaneously with their patch clamp experiments^[2]. They really knocked it out of the park, they measured transcription, e-features, and filled the cells with dye to recreate the morphology. What more could you ask for if you wanted to make an emulation of a single cell? But what if you want to measure transcription of every cell after cryopreservation? Can mRNA even survive that process? I will readily admit that I did only cursory research into this part as I am not especially well versed in genetics in the first place. However, with relatively little work I was able to find a group investigating just that. Quite frankly, I don’t have the biochem/genetics background to really get into this as much as I should. But looking at their figures I think the way they said it in their conclusion is a fair summary of what I care about: “Using the here-established cryopreservation method, single-cell transcriptome profiles from cells and tissues did not differ from freshly processed material.”^[6]. Now I am not sure how well this would transfer over to preserving a whole brain as their preservation method dissociated the cells and suspended them in their solution before freezing, but I see no reason to believe that it is fundamentally impossible to cryopreserve a whole brain in such a way that the mRNA is still readable. Really I only wanted to include this to confirm to myself that cold in and of itself did not somehow degrade the mRNA.

Predicting Electrophysiology from Morphology and the Transcriptome

So, assuming we have all this we still can’t perform patch clamp experiments on every neuron in the human brain. Maybe someone will object to that but these patch clamp experiments take ~10 minutes each and you have ~80 billion neurons so that just won’t work. I did read a paper about recording electrical activity from the whole brain but I won’t actually be talking about that in this post^[7]. Morphology must be extracted from the cells if you want to do whole brain emulation. You need the connectome already and synaptic weight seems correlated with size/number of dendritic spines so getting the whole morphology of the cell is no extra cost. Ideally you would also have the transcriptome along with this depending on how well preserved the brain is or if the procedure is begun immediately post mortem. Of course, getting the transcriptome does actually incur significant additional costs in terms of the total raw data and how you have to process it. Serial electron microscopy is the gold standard for morphology, but personally I am hopeful that tissue clearing and expansion allows for optical techniques to be successful. Whatever the case, the process of forming serial sections is distinct from one that measures the transcriptome. The brain clearing technique could be some assistance in this arena. To get the mRNA you have to penetrate the cell which would be a good opportunity to also inject a unique dye into each cell which would make them both more distinct when you go to trace the connectome and allow you to keep track of which cell you extracted the nucleus from. This would be a lot of very time consuming work but would be doable on the smallest of scales with current hardware although I may be overconfident.

Fermi estimate for morphology + transcriptome of d. megalaster:

~200,000 neurons
~10 minutes to stick microneedle into cell, fill with dye, extract nucleus
~4 cells can be done in parallel
200,000*10/4 = 500,000 minutes
500,000 minutes is ~1 year

Ok so perhaps that is not viable without some OOM speedup. Let’s say you don’t need to inject each cell individually, you just want to do some nuclear stain that doesn’t damage the mRNA^[8]. You can clear, expand, and dye everything homogeneously. Then imaging on a 2P microscope you can take relatively thick sections and shuttle them off to a lower resolution microscope/micromanipulator that is dedicated to just finding the nucleus and extracting it for transcriptome analysis. You could have as many of these accessory microscopes/micromanipulators as you want until the bottleneck is the speed of imaging the brain in sufficient resolution. I’m not going to do the work to find out how long this would be for d. megalaster but I found a paper that did serial sectioning on a whole mouse brain in a week which gives me hope for a highly parallel method getting both morphology and transcriptome^[9].

Anyway, once you have all this data, is it even sufficient to predict e-features? As far as I can tell that is still somewhat of an open question. I would love nothing more than to have someone confirm or deny this for me. But with my searching I was only able to find one paper from 2020 regarding the prediction of e-features using morphology^[10]. It had some success but did not incorporate the transcriptome data, had a very small dataset of only 104 neurons, and has had no follow up as far as I can tell. I am trying to get in contact with the author but the email listed on the paper looks deactivated. Additionally, the paper from the Allen Institute about measuring e-features, fitting a model to them, and recording morphology and the transcriptome claims that predicting e-features from morphology/transcriptomics is impossible a priori but goes to great lengths showing correlations between the two. I would like to think that I missed something very obvious but it seems to me that nobody has tried to train some machine learning model to integrate morphology and transcriptome data to predict e-features and this really strange to me because it seems like it would be so useful even for non WBE related neuroscience.

Logical Next Steps

I feel like there is a lot of fertile ground here to make marginal progress. The most obvious next steps to me are:

Make automated patch clamp experiments faster and more reliable
Create a large data set of neurons with their morphologies, e-features, and transcriptome
Train a machine learning model to predict e-features from morphology, either derived parameters (soma volume, axon length, etc.) or complete 3D reconstruction of the cell given as planar images
Validate/create new cryopreservation techniques to ensure mRNA is preserved well enough to be sequenced especially with potential tissue clearing and expansion as well as various stains
Combine serial sectioning/imaging acquisition with something that can simultaneously extract mRNA for transcriptomics and label the cell uniquely or otherwise ensure both transcriptome and morphology correspond to the same cell

I am going to start working on 3 because it is the only one I can really try alone without any funding.

I would welcome any related information, especially some that shows people have been working on this problem or convincing arguments on why I shouldn't waste my time chasing this.

^{^}
Nandi A, Chartrand T, Van Geit W, Buchin A, Yao Z, Lee SY, Wei Y, Kalmbach B, Lee B, Lein E, Berg J, Sümbül U, Koch C, Tasic B, Anastassiou CA. Single-neuron models linking electrophysiology, morphology, and transcriptomics across cortical cell types. Cell Rep. 2022 Aug 9;40(6):111176. doi: 10.1016/j.celrep.2022.111176. Erratum in: Cell Rep. 2022 Nov 8;41(6):111659. doi: 10.1016/j.celrep.2022.111659. PMID: 35947954; PMCID: PMC9793758.
^{^}
https://portal.brain-map.org/cell-types/classes/multimodal-characterization/multimodal-characterization-mouse-visual-cortex
^{^}
https://repository.gatech.edu/entities/publication/640bef49-75ce-4675-b4e8-d253a55f2b8b/full
^{^}
Landry CR, Yip MC, Zhou Y, Niu W, Wang Y, Yang B, Wen Z, Forest CR. Electrophysiological and morphological characterization of single neurons in intact human brain organoids. J Neurosci Methods. 2023 Jul 1;394:109898. doi: 10.1016/j.jneumeth.2023.109898. Epub 2023 May 24. PMID: 37236404; PMCID: PMC10483933.
^{^}
https://autopatcher.org/
^{^}
Guillaumet-Adkins, A., Rodríguez-Esteban, G., Mereu, E. et al. Single-cell transcriptome conservation in cryopreserved cells and tissues. Genome Biol 18, 45 (2017). https://doi.org/10.1186/s13059-017-1171-9
^{^}
Kleinfeld D, Luan L, Mitra PP, Robinson JT, Sarpeshkar R, Shepard K, Xie C, Harris TD. Can One Concurrently Record Electrical Spikes from Every Neuron in a Mammalian Brain? Neuron. 2019 Sep 25;103(6):1005-1015. doi: 10.1016/j.neuron.2019.08.011. Epub 2019 Sep 5. PMID: 31495645; PMCID: PMC6763354.
^{^}
Krishnaswami, S., Grindberg, R., Novotny, M. et al. Using single nuclei for RNA-seq to capture the transcriptome of postmortem neurons. Nat Protoc 11, 499–524 (2016). https://doi.org/10.1038/nprot.2016.015
^{^}
Economo MN, Clack NG, Lavis LD, Gerfen CR, Svoboda K, Myers EW, Chandrashekar J. A platform for brain-wide imaging and reconstruction of individual neurons. Elife. 2016 Jan 20;5:e10566. doi: 10.7554/eLife.10566. PMID: 26796534; PMCID: PMC4739768.
^{^}
Michiels, Mario. (2020). Electrophysiology prediction of single neurons based on their morphology. 10.1101/2020.02.04.933697.

CryonicsFuturismTranshumanismWhole Brain EmulationWorld Modeling

Frontpage

8

New Comment

20 comments, sorted by

top scoring

Click to highlight new comments since: Today at 7:32 AM

[-]TsviBT8d7-1

https://www.lesswrong.com/posts/jTiSWHKAtnyA723LE/overview-of-strong-human-intelligence-amplification-methods#Brain_emulation

[-]Steven Byrnes8d40

I’m not an expert myself (this will be obvious), but I was just trying to understand slide-seq—especially this paper which sequenced RNA from 4,000,000 neurons around the mouse brain.

They found low-thousands of neuron types in the mouse, which makes sense on priors given that there are only like 20,000 genes encoding the whole brain design and everything in it, along with the rest of the body. (Humans are similar.)

I’m very mildly skeptical of the importance & necessity of electrophysiology characterization for reasons here, but such a project seems more feasible if you think of it as characterizing the electrophysiology properties of low-thousands of discrete neuron types, each of which (hopefully) can also be related to morphology or location or something else that would be visible in a connectomics dataset, as opposed to characterizing billions of neurons that are each unique.

Sorry if this is stupid or I’m misunderstanding.

[-]halvorz8d30

"They found low-thousands of neuron types in the mouse, which makes sense on priors given that there are only like 20,000 genes encoding the whole brain design and everything in it, along with the rest of the body."

I'm a bit puzzled by this statement; how would the fact that there are ~20,000 genes in the mouse/human genome constrain the number of neuron types to the low thousands? From a naive combinatorics standpoint it seems like 20,000 genes is sufficiently large to place basically zero meaningful constraints on the number of potential cell types. E.g. if you assume that only 15,000 genes vary meaningfully between cell types, and that there are 3000 of those variable genes expressed per cell, chatgpt tells me that the number of potential combinations is too large for it to even estimate its order of magnitude. And that's with a simple, extremely unrealistic binary on/off model of gene expression.

[-]Steven Byrnes8d30

Good question! The idea is, the brain is supposed to do something specific and useful—run a certain algorithm that systematically leads to ecologically-adaptive actions. The size of the genome limits the amount of complexity that can be built into this algorithm. (More discussion here.) For sure, the genome could build a billion different “cell types” by each cell having 30 different flags which are on and off at random in a collection of 100 billion neurons. But … why on earth would the genome do that? And even if you come up with some answer to that question, it would just mean that we have the wrong idea about what’s fundamental; really, the proper reverse-engineering approach in that case would be to figure out 30 things, not a billion things, i.e. what is the function of each of those 30 flags.

A kind of exception to the rule that the genome limits the brain algorithm complexity is that the genome can (and does) build within-lifetime learning algorithms into the brain, and then those algorithms run for a billion seconds, and create a massive quantity of intricate complexity in their “trained models”. To understand why an adult behaves how they behave in any possible situation, there are probably billions of things to be reverse-engineered and understood, rather than low-thousands of things. However, as a rule of thumb, I claim that:

when the evolutionary learning algorithm adds a new feature to the brain algorithm, it does so by making more different idiosyncratic neuron types and synapse types and neuropeptide receptors and so on,
when one of the brain’s within-lifetime learning algorithm adds a new bit of learned content to its trained model, it does so by editing synapses.

Again, I only claim that these are rules-of-thumb, not hard-and-fast rules, but I do think they’re great starting points. Even if there’s a nonzero amount of learned content storage via gene expression, I propose that thinking of it as “changing the neuron type” is not a good way to think about it; it’s still “the same kind of neuron”, and part of the same subproject of the “understanding the brain” megaproject, it’s just that the neuron happens to be storing some adjustable parameter in its nucleus and acting differently in accordance with that.

By contrast, medium spiny neurons versus Purkinje cells versus cortical pyramidal neurons versus magnocellular neurosecretory cells etc. etc. are all just wildly different from each other—they look different, they act different, they play profoundly different roles in the brain algorithm, etc. The genome clearly needs to be dedicating some of its information capacity to specifying how to build each and every of those cell types, individually, such that each of them can play its own particular role in the brain algorithm.

Does that help explain where I’m coming from?

[-]halvorz7d30

Thanks for the reply! I'm familiar with (and am skeptical of) the basic information theoretic argument as to why genome size should constrain the complexity of whatever algorithm the brain is running, but my question here is more specific. What I'm not clear on is how those two numbers (20,000 genes and a few thousand neuron types) specifically relate to each other in your model of brain functioning. Is the idea that each neuron type roughly corresponds to the expression of one or two specific genes, and thus you'd expect <20,000 neuron types?

"For sure, the genome could build a billion different “cell types” by each cell having 30 different flags which are on and off at random in a collection of 100 billion neurons. But … why on earth would the genome do that?"

Interestingly, the genome does do this! Protocadherins in vertebrates and DSCAM1 are expressed in exactly this way, and it's thought to help neurons to distinguish themselves from other neurons, which is essential for neuronal self avoidance: https://en.wikipedia.org/wiki/Neuronal_self-avoidance#Molecular_basis_of_self-avoidance

Of course in an emulation you could probably just tell the neurons to not interact with themselves so this crazy system wouldn't be necessary, but it is a nice example of how biology does things you might a priori think would never happen.

[-]Steven Byrnes6d40

What I'm not clear on is how those two numbers (20,000 genes and a few thousand neuron types) specifically relate to each other in your model of brain functioning.

Start with 25,000 genes, but then reduce it a bunch because they also have to build hair follicles and the Golgi apparatus and on and on. But then increase it a bit too because each gene has more than one design degree of freedom, e.g. a protein can have multiple active sites, and there’s some ability to tweak which molecules can and cannot reach those active sites and how fast etc. Stuff like that.

Putting those two factors together, I dunno, I figure it’s reasonable to guess that the genome can have a recipe for a low-thousands of distinct neuron types each with its own evolutionarily-designed properties and each playing a specific evolutionarily-designed role in the brain algorithm.

And that “low thousands” number is ballpark consistent with the slide-seq thing, and also ballpark consistent with what you get by counting the number of neuron types in a random hypothalamus nucleus and extrapolating. High hundreds, low thousands, I dunno, I’m treating it as a pretty rough estimate.

Hmm, I guess when I think about it, the slide-seq number and the extrapolation number are probably more informative than the genome number. Like, can I really rule out “tens of thousands” just based on the genome size? Umm, not with extreme confidence, I’d have to think about it. But the genome size is at least a good “sanity check” on the other two methods.

Is the idea that each neuron type roughly corresponds to the expression of one or two specific genes, and thus you'd expect <20,000 neuron types?

No, I wouldn’t necessarily expect something so 1-to-1. Just the general information theory argument. If you have N “design degree of freedom” and you’re trying to build >>N specific machines that each does a specific thing, then you get stuck on the issue of crosstalk.

For example, suppose that some SNP changes which molecules can get to the active site of some protein. It makes Purkinje cells more active, but also increases the ratio of striatal matrix cells to striosomes, and also makes auditory cortex neurons more sensitive to oxytocin. Now suppose there’s very strong evolutionary pressure for Purkinje cells to be more active. Then maybe that SNP is going to spread through the population. But it’s going to have detrimental side-effects on the striatum and auditory cortex. Ah, but that’s OK, because there’s a different mutation to a different gene which fixes the now-suboptimal striatum, and yet a third mutation that fixes the auditory cortex. Oops, but those two mutations have yet other side-effects on the medulla and … Etc. etc.

…Anyway, if that’s what’s going on, that can be fine! Evolution can sort out this whole system over time, even with crazy side-effects everywhere. But only as long as there are enough “design degrees of freedom” to actually fix all these problems simultaneously. There do have to be more “design degrees of freedom” in the biology / genome than there are constraints / features in the engineering specification, if you want to build a machine that actually works. There doesn’t have to be a 1-to-1 match between design-degrees-of-freedom and items on your engineering blueprint, but you do need that inequality to hold. See what I mean?

Interestingly, the genome does do this! Protocadherins in vertebrates and DSCAM1 are expressed in exactly this way, and it's thought to help neurons to distinguish themselves from other neurons…
Of course in an emulation you could probably just tell the neurons to not interact with themselves

Cool example, thanks! Yeah, that last part is what I would have said. :)

[-]halvorz6d30

Interesting...I think I vaguely understand what you're talking about, but I'm doubtful that these concepts really apply to biology. Especially since your example is about constraints on evolvability rather than functioning. In practice that is pretty much how everything tends to work, with absolutely wild amounts of pleiotropy and epistasis, but that's not a problem unless you want to evolve a new function. Which is probably why the strong strong evolutionary default is towards stasis, not change.

I guess my priors are pretty different because my background is in virology, where our expectation (after decades of painful lessons) is that the default is for proteins to be wildly multifunctional, with many many many "design degrees of freedom." Granted viruses are a bit of a special case, but I do think they can provide a helpful stress test/simpler model for information theoretic models of genome function.

[-]Dom Polsinelli8d10

Not my claim so I'm not defending this too hard but from my lab experience relatively few genes seem to control bulk properties and then there are a whole bunch of higher order corrections. Literally one or two genes being on/off can determine if a neuron is excitatory or inhibitory. If you subscribe to Izhikevich's classification of bistable/monostable and integrator/resonator you would only need 3 genes with binary expression. After that you get a few more to determine time constants and stuff. I still think whole transcriptome would be helpful, especially as we don't know what each gene does yet, but I am not 100% against the idea that only ~20 really matter with a few thousand template neurons and after that you run into a practical limit of noise being present.

[-]Dom Polsinelli8d30

Thank you for that article, I don't know how it didn't come up when I was researching this. Others finding papers I should have been able to find alone is a continuous frustrations of mine.
I would love to live in a world where we have a few thousand template neurons and can just put them together based on a few easily identifiable factors (~3-10 genes, morphology, brain region) but until I find a paper that convincingly recreates the electrophysiology based on those things I have to entertain the idea that somewhere between 10 and 10^5 are relevant. I would be truly shocked if we need 10^5 but I wouldn't be surprised if we need to measure more expression levels than we can comfortable infer based on some staining method. Having just read your post on pessimism, I am confused as to why you think low thousands of separate neuron models would be sufficient. I agree that characterizing billions of neurons is a very tall order (although I really won't care how long it takes if I'm dead anyway). But when you say '“...information storage in the nucleus doesn’t happen at all, or has such a small effect that we can ignore it and still get the same high-level behavior” (which I don’t believe).' it sounds to me like an argument in favor of looking at the transcriptome of each cell.

Just to be abundantly clear, my main argument in the post is not "Single cell transcriptomics leading to perfect electrophysiology is essential for whole brain emulation and anything less than that is doomed to fail." It is closer to "I have not seen a well developed theory that can predict even a single cell's electrophysiology given things we can measure post mortem, so we should really research that if we care about whole brain emulation. If it already exists, please tell me about it."

I think you make good points when you point out failures of c. elegans uploading and other computational neuroscience failures. To me, it makes a lot of sense to copy single cells as close as possible and then start modeling learning rules and synaptic conductance and what not. If we find out later a certain feature of a neuron model can be abstracted away, that's great. But a lot of what I see right now is people running to study learning rules and they use homogenous leaky integrate and fire neurons. In my mind they are doing machine learning on spiking neural networks, not computational neuroscience. I don't know how relevant that particular critique is but it has been a frustration of mine for a while.

I am still very new to this whole field, I hope that cleared things up. If it did not, I apologize.

[-]Steven Byrnes8d40

Having just read your post on pessimism, I am confused as to why you think low thousands of separate neuron models would be sufficient. I agree that characterizing billions of neurons is a very tall order (although I really won't care how long it takes if I'm dead anyway). But when you say '“...information storage in the nucleus doesn’t happen at all, or has such a small effect that we can ignore it and still get the same high-level behavior” (which I don’t believe).' it sounds to me like an argument in favor of looking at the transcriptome of each cell.

I think the genome builds a brain algorithm, and the brain algorithm (like practically every algorithm in your CS textbook) includes a number of persistent variables that are occasionally updated in such-and-such way under such-and-such circumstance. Those variables correspond to what the neuro people call plasticity—synaptic plasticity, gene expression plasticity, whatever. Some such occasionally-updated variables are parameters in within-lifetime learning algorithms that are part of the brain algorithm (akin to ML weights). Other such variables are not, instead they’re just essentially counter variables or whatever (see §2.3.3 here). The “understanding the brain algorithm” research program would be figuring out what the brain algorithm is, how and why it works, and thus (as a special case) what are the exact set of “persistent variables that are occasionally updated”, and how are they stored in the brain. If you complete this research program, you get brain-like AGI, but you can’t upload any particular adult human. Then a different research program is: take an adult human brain, and go in with your microtome etc. and actually measure all those “persistent variables that are occasionally updated”, which comprise a person’s unique memories, beliefs, desires, etc.

I think the first research program (understanding the brain algorithm) doesn’t require a thorough understanding of neuron electrophysiology. For example (copying from §3.1 here), suppose that I want to model a translator (specifically, a MOSFET). And suppose that my model only needs to be sufficient to emulate the calculations done by a CMOS integrated circuit. Then my model can be extremely simple—it can just treat the transistor as a cartoon switch. Next, again suppose that I want to model a transistor. But this time, I want my model to accurately capture all measurable details of the transistor. Then my model needs to be mind-bogglingly complex, involving many dozens of obscure SPICE modeling parameters. The point is: I’m suggesting an analogy between this transistor and a neuron with synapses, dendritic spikes, etc. The latter system is mind-bogglingly complex when you study it in detail—no doubt about it! But that doesn’t mean that the neuron’s essential algorithmic role is equally complicated. The latter might just amount to a little cartoon diagram with some ANDs and ORs and IF-THENs or whatever. Or maybe not, but we should at least keep that possibility in mind.

In the “understanding the brain algorithm” research program, you’re triangulating between knowledge of algorithms in general, knowledge of what actual brains actually do (including lesion studies, stimulation studies, etc.), knowledge of evolution and ecology, and measurements of neurons. The first three can add so much information that it seems possible to pin down the fourth without all that much measurements, or even with no measurements at all beyond the connectome. Probably gene expression stuff will be involved in the implementations in certain cases, but we don’t really care, and don’t necessarily need to be measuring that. At least, that’s my guess.

In the “take the adult brain and measure all the ‘persistent variables that are occasionally updated’ research program, yes it’s possible that some of those persistent variables are stored in gene expressions, but my guess is very few, and if we know where they are and how they work then we can just measure the exact relevant RNA in the exact relevant cells.

…To be clear, I think working on the “understanding the brain algorithm” research program is very bad and dangerous when it focuses on the cortex and thalamus and basal ganglia, but good when it focuses on the hypothalamus and brainstem, and it’s sad that people in neuroscience, especially AI-adjacent people with a knack for algorithms, are overwhelmingly are working on the exact worst possible thing :( But I think doing it in the right order (cortex last, long after deeply understanding everything about the hypothalamus & brainstem) is probably good, and I think that there’s realistically no way to get WBE without completing the “understanding the brain algorithm” research program somewhere along the way.

[-]Dom Polsinelli8d10

I think I have identified our core disagreement, you believe a neuron or a small group of neurons are fundamentally computationally simple and I don't. I guess technically I'm agnostic about it but my intuition is that a real neuron cannot be abstracted to a LIF neuron the way a transistor can be abstracted to a cartoon switch (not that you were suggesting LIF is sufficient, just an example). One of the big questions I have is how error scales from neuron to overall activity. If a neuron model is 90% accurate wrt electrophysiology and the synapses connecting it are 90% accurate to real synapses, does that recover 90% of brain function? Is the last 10% something that is computationally irrelevant and can just be abstracted away, giving you effectively 100% functionality? Is 90% accuracy for single neurons magnified until the real accuracy is like 0.9^(80 billion)? I think it is unlikely that it is that bad, but I really don't know because of the abject failure to upload anything as you point out. I am bracing myself for a world where we need a lot of data.

Let's assume for the moment though that HH model with suitable electrical and chemical synapses would be sufficient to capture WBE. What I still really want to see is a paper saying "we look at x,y,z properties of neurons that can be measured post mortem and predict a,b,c properties of those neurons by tuning capacitance and conductance and resting potential in the HH model. Our model is P% accurate when looking at patch clamp experiments." In parallel with that there should be a project trying to characterize how error tolerant real neurons and neural networks can be so we can find the lower bound of P. I actually tried something like that for synaptic weight (how does performance degrade when adding noise to the weights of a spiking neural network) but I was so disillusioned with the learning rules that I am not confident in my results. I'm not sure if anyone has the ability to answer these kinds of questions because we are still just so bad at emulating anything.

Edit:

Also, I am not sure if you're proposing we compress multiple neurons down into a simpler computational block, the way a real arrangement of transistors can be abstracted into logic gates or adders or whatever. I am not a fan of that for WBE for philosophical reasons and because I think it is less likely to capture everything we care about especially for individual people.

[-]Steven Byrnes8d20

you believe a neuron or a small group of neurons are fundamentally computationally simple and I don't

I guess I would phrase it as “there’s a useful thing that neurons are doing to contribute to the brain algorithm, and that thing constitutes a tiny fraction of the full complexity of a real-world neuron”.

(I would say the same thing about MOSFETs. Again, here’s how to model a MOSFET, it’s a horrific mess. Is a MOSFET “fundamentally computationally simple”? Maybe?—I’m not sure exactly what that means. I’d say it does a useful thing in the context of an integrated circuit, and that useful thing is pretty simple.

The trick is, “the useful thing that a neuron is doing to contribute to the brain algorithm” is not something you can figure out by studying the neuron, just as “the useful thing that a MOSFET is doing to contribute to IC function” is not something you can figure out by studying the MOSFET. There’s no such thing as “Our model is P% accurate” if you don’t know what phenomenon you’re trying to capture. If you model the MOSFET as a cartoon switch, that model will be extremely inaccurate along all kinds of axes—for example, its thermal coefficients will be wrong by 100%. But that doesn’t matter because the cartoon switch model is accurate along the one axis that matters for IC functioning.

The brain is generally pretty noise-tolerant. Indeed, if one of your neurons dies altogether, “you are still you” in the ways that matter. But a dead neuron is a 0% accurate model of a live neuron. ¯\_(ツ)_/¯

In parallel with that there should be a project trying to characterize how error tolerant real neurons and neural networks can be so we can find the lower bound of P. I actually tried something like that for synaptic weight (how does performance degrade when adding noise to the weights of a spiking neural network) but I was so disillusioned with the learning rules that I am not confident in my results.

Just because every part of the brain has neurons and synapses doesn’t mean every part of the brain is a “spiking neural network” with the connotation that that term has in ML, i.e. a learning algorithm. The brain also needs (what I call) “business logic”—just as every ML github repository has tons of code that is not the learning algorithm itself. I think that the low-thousands of different neuron types are playing quite different roles in quite different parts of the brain algorithm, and that studying “spiking neural networks” is the wrong starting point.

[-]Dom Polsinelli8d30

I apologize for my sloppy language, "computationally simple" was not well defined. You are quite right when you say there is no P% accuracy. I think my offhand remark about spiking neural networks was not helpful to this discussion.

In a practical sense, here is what I mean. Imagine someone makes a brain organoid with ~10 cells. They can directly measure membrane voltage and any other relevant variable they want because this is hypothetical. Then they try and emulate whatever algorithm this organoid has going on, its direct input to output and whatever learning rule changes that it might have. But, to test this they have crappy point neuron models implementing LIF and the synapses are just a constant conductance or something, and then rules on top of that that can adjust parameters (membrane capacitance, resting potential, synaptic conductance, ect.) and it fails to replicate observables. Obviously this is an extreme example, but I just want better neuron models so nothing like this ever has the chance to happen.

Basically, if we can't model an organoid we could

Fix the electrophysiology which either makes it work or proves something else is the problem
Develop theory via reverse engineering to such a point we just understand what is wrong and home in on it
Fix other things and hope it isn't electrophysiology

Three is obviously a bad plan. Two is really really hard. One should be relatively easy provided we have a reasonable threshold of what we consider to be accurate electrophysiology. We could have good biophysical models that recreate it or we could have recurrent neural nets modeling the input current -> membrane voltage relation of each neuron. It just seems like an easy way to cross of a potential cause of failure (famous last words I'm sure).

As for you business logic point, it is valid but I am worried that black boxing that too much would lead to collateral damage. I am not sure if that's what you meant when you said spiking neural networks are the wrong starting point. In any case, I would like higher order thinking to stay as a function of spiking neurons even if things like reflexes and basal behavior can be replaced without loss.

[-]Steven Byrnes7d30

Yeah I think “brain organoids” are a bit like throwing 1000 transistors and batteries and capacitors into a bowl, and shaking the bowl around, and then soldering every point where two leads are touching each other, and then doing electrical characterization on the resulting monstrosity. :)

Would you learn anything whatsoever from this activity? Umm, maybe? Or maybe not. Regardless, even if it’s not completely useless, it’s definitely not a central part of understanding or emulating integrated circuits.

(There was a famous paper where it’s claimed that brain organoids can learn to play Pong, but I think it’s p-hacked / cherry-picked.)

There’s just so much structure in which neurons are connected to which in the brain—e.g. the cortex has 6 layers, with specific cell types connected to each other in specific ways, and then there’s cortex-thalamus-cortex connections and on and on. A big ball of randomly-connected neurons is just a totally different thing.

Also, I am not sure if you're proposing we compress multiple neurons down into a simpler computational block, the way a real arrangement of transistors can be abstracted into logic gates or adders or whatever. I am not a fan of that for WBE for philosophical reasons and because I think it is less likely to capture everything we care about especially for individual people.

Yes and no. My WBE proposal would be to understand the brain algorithm in general, notice that the algorithm has various adjustable parameters (both because of inter-individual variation and within-lifetime learning of memories, desires, etc.), do a brain-scan that records those parameters for a certain individual, and now you can run that algorithm, and it’s a WBE of that individual.

When you run the algorithm, there is no particular reason to expect that the data structures you want to use for that will superficially resemble neurons, like with a 1-to-1 correspondence. Yes you want to run the same algorithm, producing the same output (within tolerance, such that “it’s the same person”), but presumably you’ll be changing the low-level implementation to mesh better with the affordances of the GPU instruction set rather than the affordances of biological neurons.

The “philosophical reasons” are presumably that you think it might not be conscious? If so, I disagree, for reasons briefly summarized in §1.6 here.

“Less likely to capture everything we care about especially for individual people” would be a claim that we didn’t measure the right things or are misunderstanding the algorithm, which is possible, but unrelated to the low-level implementation of the algorithm on our chips.

I definitely am NOT an advocate for things like training a foundation model to match fMRI data and calling it a mediocre WBE. (There do exist people who like that idea, just I’m not one of them.) Whatever the actual information storage is, as used by the brain, e.g. synapses, that’s what we want to be measuring individually and including in the WBE. :)

[-]Dom Polsinelli7d30

First of all, I hate analogies in general but that's a pet peeve, they are useful. But going with your shaken up circuit as an analogy to brain organoids and assuming it is true, I think it is more useful than you give it credit. If you have a good theory of what all those components are individually you would still be able to predict something like voltage between two arbitrary points. If you model resistors as some weird non ohmic entity you'll probably get the wrong answer because you missed the fact that they behave ohmic in many situations. If you never explicitly write down Ohm's law but you empirically measure current at a whole bunch of different voltages (analogous to patch clamps but far far from a perfect analogy) you can probably get the right answer. So yeah an organoid would not be perfect but I would be surprised if being able to fully emulate one would be useless. Personally I think it would be quite useful but I am actively tempering my expectations.

But my meta point of

look at small system
try to emulate
cross off obvious things (electrophysiology should be simple for only a few neurons) that could cause it to not be working
repeat and use data to develop overall theory

stands even if organoids in particular are useless. The theory developed with this kind of research loop might be useless for your very abstract representation of the brain's algorithm but I think it would be just fine, in principle, for the traditional, bottom up approach.

As for the philosophical objections, it is more that whatever wakes up won't be me if we do it your way. It might act like me and know everything I know but it seems like I would be dead and something else would exist. Gallons of ink have been spilled over this so suffice it to say, I think the only thing with any hope of preserving my consciousness (or at least a conscious mind that still holds the belief that it was at one point the person writing this) is gradual replacement of my neurons while my current neurons are still firing. I know that is far and away the least likely path of WBE because it requires solving everything else + nanotechnology but hey I dream big.

To be clear, I think your proposed WBE plan has a lot of merit, but it would still result in me experiencing death and then nothing else so I'm not especially interested. Yes, that probably makes me quite selfish.

[-]Steven Byrnes7d20

As for the philosophical objections, it is more that whatever wakes up won't be me if we do it your way. It might act like me and know everything I know but it seems like I would be dead and something else would exist.

Ah, but how do you know that the person that went to bed last night wasn’t a different person, who died, and you are the “something else” that woke up with all of that person’s memories? And then you’ll die tonight, and tomorrow morning there will be a new person who acts like you and knows everything you know but “you would be dead and something else would exist”?

…It’s fine if you don’t want to keep talking about this. I just couldn’t resist. :-P

If you have a good theory of what all those components are individually you would still be able to predict something like voltage between two arbitrary points.

I agree that, if you have a full SPICE transistor model, you’ll be able to model any arbitrary crazy configuration of transistors. If you treat a transistor as a cartoon switch, you’ll be able to model integrated circuits perfectly, but not to model transistors in very different weird contexts.

By the same token, if you have a perfect model of every aspect of a neuron, then you’ll be able to model it in any possible context, including the unholy mess that constitutes an organoid. I just think that getting a perfect model of every aspect of a neuron is unnecessary, and unrealistic. And in that framework, successfully simulating an organoid is neither necessary nor sufficient to know that your neuron model is OK.

[-]Dom Polsinelli7d30

Yes, I am familiar with the sleep = death argument. I really don't have any counter, at some point though I think we all just kind of arbitrarily draw a line. I could be a solipsist, I could believe in last thursdayism, I could believe some people are p-zombies, I could believe in the multiverse. I don't believe in any of these but I don't have any real arguments for them and I don't think anyone has any knockdown arguments one way or the other. All I know is that I fear soma style brain upload, I fear star trek style teleportation, but I don't fear gradual replacement nor do I fear falling asleep.

As for wrapping up our more scientific disagreement, I don't have much to say other than it was very thought provoking and I'm still going to try what I said in my post. Even if it doesn't come to complete fruition I hope it will be relevant experience for when I apply to grad school.

[-]Adam Shai7d30

why I shouldn't waste my time chasing this.

Some reasons that come to mind very quickly:

- Patch clamp experiments usually take place in slices with artificial cerebrospinal fluid (ACSF). The ephys properties can vary widely based on the experimental prep (angle that slice was taken, the temperature, the specific recipe used for the ACSF, the quality of the patcher, etc. etc.

even if patching worked really well and was robust and reliable, the ephys properties at the soma (where vast majority of patching willt ake place) hardly describe the ephys of the entire dendritic tree, which is very complicated and space dependent, and incredibly nonlinear and variable.

[-]Dom Polsinelli7d10

Based on this and you other comment you seem to be pro GEVI instead of patch clamp, am I correct? Assuming GEVIs were used (or some other, better technology) to find all electrophysiology, why would that be a waste of time? Even if we can get by with a few thousand template neurons and individual tuning is not necessary (which seems to be the view of Steven Byrne and maybe you) how should we go about getting those template neurons without a lot of research into correlating morphology, genetic expression, and electrophysiology? If we don't need them, why would we not? My primary goal is not to defend my plan, I just care about making progress on WBE generally and I would like to hear specific plans if others have them. Studying single cell function just seemed to be the most natural to me. Without that, studying how multiple neurons signal each other or change over time or encode information in spike trains seems like putting the cart before the horse as it were. Again, very glad to be wrong, it just still seems to me that some version of this research has to be done eventually, we haven't done it yet AFAIK, so I should start on what little part I can.

[-]Adam Shai7d20

Under the assumption that capturing the ephys properties of single neurons is important for WBE, it still seems unlikely to me that scaling up patch clamping is a viable path to that. More likely to work would be trying to scale up voltage imaging.

(for the record I don't personally agree with that assumption, for overlapping reasons with what Steven Byrnes thinks).

Moderation Log