Funny and interesting, and it was voted a finalist so I have to post it - but I don't believe this at all.
Main concern: how do you encode memories in a chemical? That is, suppose you want to encode the fact "My name is Tim". It's easy to imagine how this works as neural connections: during infancy, you hear language, your brain forms a predictive model of the language, you get neurons representing the sounds T, I, and M, and the word "name", and you get links between them. Eventually there's some complicated network connecting name, T + I + M, the sight of a nametag, the motor program of shaking hands, and all the other concepts relating to your name being Tim. We know something like this has to work because LLMs can have this type of memory and they're nothing but connections.
How would you implement this in a protein? All I can think of is a binary-like code with literal text - using one amino acid to represent 0, another to represent 1, and coding the ASCII string "M-Y-N-A-M-E-I-S-T-I-M", and then having a decoder somewhere in the brain. But this doesn't match the phenomenology of human memory, which is sensory (eg associated with pictures, smells, etc) and works even without language (eg in preverbal children). CF https://slatestarcodex.com/2017/09/07/how-do-we-get-breasts-out-of-bayes-theorem/ , which asks a similar question about genetic memory.
But also, even supposing you could do this - in order for the organ transplants to make sense, you need for the heart to be producing or storing these chemicals (why?) in such large quantities that they get into the bloodstream, cross the BBB (why would you have a transporter for these?), make it into the brain, and perfectly interface with the host's own memory-decoder system (how?). Then either the heart has to keep making these constantly (why?) or there has to be some mechanism for a single low-dose exposure to get the host brain to start making them itself (why?) And if the cannibalism claim is to make sense, you have to be able to absorb these extremely complex chemicals through the digestive tract, which is usually doing a pretty good job preventing you from absorbing complex chemicals. Why do this, rather than just keep the chemicals in the relevant part of the brain?
Also, if we had these chemicals, a natural application of them would be genetic memory, eg in the womb you get a big dose of your mother's memory chemicals and now you understand the world as well as she does. As far as I know there are far fewer stories of people mysteriously having their mother's memories compared to their organ donors' memories, adjusted for the ratio of people who have mothers vs. organ transplants.
Also, the writer asks whether losing your arm might make you forget something, or eating an animal liver might give you some of the animal's memories. But many people lose limbs and eat animals, and as far as I know nobody reports these phenomena.
Agreed, I think this review goes way too far. Sure there are other mechanisms for memory in the brain besides synapses. No, I don't believe anything about heart transplants transferring memory.
I think the way you would encode "TIM" in protein is in a protein network. You could conceivably do this inside a single cell, AFAIK.
The problem is cells don't have mouths to say "my name is Tim" with. Within-cell memory is absolutely real and important, but it's just not the right scale for organism-level behavior.
You could have a network of lots of different proteins, but I don't understand how the symbols would be grounded. For example, what would it mean to say that arginine = M instead of N? Unless it bottoms out in some kind of behavior like saying an M or hearing an M, it's just arginine.
...maybe this is just the same point you're making.
Not a protein expert, but I think many proteins are very weird and have many different possible configurations, and those configurations can be influenced by all sorts of features in the environment. In theory you could have a "name" protein that has a different configuration for each of the 1k most common names in the English language or something. This would be a weird way to do things, though!
Have you read "How Life Works" by Phillip Ball? As someone who had done their best to avoid modern molecular biology in grad school, I found it very useful for making sense of the snippets I had heard. It made molecular biology sound a frightening amount like social science.
I still don't understand how this would be grounded. That is, I 100% agree a protein could have 1000 different states. But what makes those states correspond to 1000 different names, as opposed to 1000 different shades of blue?
Suppose you hear the name Tim, and it equals state #489. You need the brain (which is doing some kind of auditory processing and has decided that "Tim" is important enough to devote a protein state slot to) to encode that by modifying the protein to state #489. Then the protein goes to the heart, and the heart is transplanted into a different person. Then the protein goes to the second person's brain. Unless there is a natural mapping between names and protein states, how does the second person get the same protein state code where it knows that a protein in state #489 equals the name "Tim"?
(for that matter, how does the first person establish a protein-state-to-name code which is consistent across all relevant brain cells?)
I agree! These are all very good reasons why you would not want to encode people's names in a protein. It is just too far away from the physical stimulus, there's no easy way for the encoder and decoder to coordinate on a shared meaning.
But many of these problems are also true for synaptic memory. I don't think cortex or hippocampus transplants would transfer memories either.
The vegetarianism thing might work though? The transplant cells are not used to getting meat bits, and when they get meat bits, they tell the brain, somehow, "hey, knock it off with the meat bits".
re the mother part: even if the heart transplant version worked, it would make sense that it doesn't work at birth, because you would need the model developed by being alive for a while to actually translate those memories into anything useful. it doesn't help to have a signal that resonates with the name "tim" if you don't know how to identify discrete sounds out of the signals from your ears yet.
(on the other hand, hell, maybe this what people thinking they remember their past lives is--oh you thought you drowned and were reincarnated? well maybe your mom had a traumatic experience in a swimming pool once)
I thought the review was fairly clear that a complex memory like "my name is Tim" would be out of scope for the proposed non-synaptic memory, and nobody is saying that the ceullular system would completely replace SPM. The definition from the review is clear in that non-synaptic process would be an addition to, not replacement of, the synaptic system, because the synaptic system clearly exists and works to a great degree:
>The formation, consolidation, and retrieval of learning and memory in biological systems often involves stimulus-dependent, non-synaptic molecular and intracellular processes. These processes do not just serve synaptic-weight-based mechanisms, but provide complementary mechanisms. They are necessary for making and keeping long-term memories, but not sufficient, and interact with synaptic-weight-based mechanisms in nontrivial ways.
A single cell organism doesn't have or need a concept of names, so its memory system has no reason to evolve that kind of capability. As organisms grow more complex, they need more complex memories, and thus more capable memory systems. As long as the old system isn't actively harmful, it sticks around because evolution is a hoarder. It keeps doing whatever it's been doing forever and, biochemistry being as baroque as it is, the old systems interacting in useful and significant ways with the new wouldn't surprise me at all.
As you say, there are some fairly basic objections to forming complex memories through transplants, let alone food, and the review might be overplaying that angle in the introduction. But the basic idea of non-synaptic memory has been well argued, I suppose (speaking as an interested reader).
I agree but think the even bigger point is that the reviewer is defining "memory" too loosely to be of interest to me. If I can't recall it then it's not what I call a 'memory'. The only evidence presented towards recallable memory is the transplant evidence, which the reviewer themselves admits is no weirder than many other things we don't believe in.
Still, I love it when people embrace odd theories like this and the gut microbiome one you 'contra'd earlier this year. Aesthetically and emotionally, I want people to go down oddball science interest trajectories and stand in the face of conventional wisdom or even common sense. (Just don't put them in charge of our health system next time, please!)
Neurons themselves are nothing but chemical/biological processes. Is it really weird that some alternative combination of chemical/biological interactions could serve the same purpose ?
At any given point in time, a neural network embodies a particular mathematical algorithm, and any given algorithm can presumably be instantiated in many different ways.
I don't think it's any more intuitively obvious a priori that a memory can be stored in the form of a neural network than in some other form. The obviousness that storing memories in the form of a neural network is possible mostly comes from the confirmation we got from modern AI, and it seems to me that dismissing other mechanisms as much more unlikely a priori is hindsight bias.
Re: "we don't get our mother's memories", uhm, yes we do ? How did you learn to breath ? It was thanks to your mother's and father's DNA, which is not a neural network.
Separately, I think much of the presented evidence in the review is rather weak. The experiments about grinding worms into dust and the like would be big if true, but they seem rather untrustworthy at first glance. But a weak form of the proposed theory is obviously true (re:DNA, individual cells), and a somewhat stronger version does not seem a priori very implausible to me.
The review is very nicely written and raises interesting points. But I want to severely push back on one point. The neuroscientific community is quite misrepresented here.
The author claims that the consensus is the strong SPM hypothesis. But no one believes this. No one. At least not among the neuroscientists who are actually biologists who work with individuals neurons. And I have talked to some real experts in the field, and (as a bit of an outsider) I have done academic research in the field. Not working with actual neurons, but building models of spiking neural systems that were supposed to be pretty close to biology.
Let me try a more accurate description. The human brain is pretty complicated. In order to understand it, we need to look at it at several levels of abstraction.
A) At the lowest levels, we try to understand the chemistry of how a single spike is formed.
B) On slightly higher levels, we may investigate how the input form other neuron excites another, without modelling each single spike on a chemical level.
C) Even higher, how patterns of spiking look like in an ensemble of neurons.
And so on, until we are at regions of the brains like the prefrontal cortex, or even higher up at behavioral level.
By the way, all this is not even covering yet learning, it is just describing a snapshot of the system.
If you are at level B, how a neuron excites another: you can't work with weights on that level. Weights will not explain what's going on. Everything is totally non-linear. When two incoming spikes arrive at the same time at the same dendritic branch, then the effect is much larger than the sum of the two. Especially if the one further from the soma (the blob at the center of the neuron) comes a little before the other. Timing matters a lot. Some people say that neurons should not be described as "integrate-and-fire" (adding up the inputs and fire when they are above a threshold), but rather as "coincidence detectors" (they fire when they get two simultaneous input, screw their weights). Everything is awfully complicated.
But then you want to abstract this whole mess. Because the higher level of abstractions are also important, and you need them.
What is the best abstraction of a neuron and its synapses? Perhaps first you only keep like 10 parameters per neuron and 2-3 per synapse. If you want to make it even simpler, then you keep 3-4 per neuron, and perhaps 1 per synapse. But what if you want to go even more abstract, what do you keep?
And then the consensus probably is: if you really force me to reduce everything to a single number, then probably a "weight" per synapse is our best bet. But it's a simplification.
And now, when it comes to learning: the strong SPM hypothesis claims that only weights change. On higher levels, this is a useful abstraction. But on abstraction level B it is so obviously wrong. Of course the neuron is plastic and changes! All of the time, and all parts of it, not just the synapses. The soma changes, the ion pumps change, the genetic expressions change, the timing of the incoming spikes change. Those are things that we have abstracted away at the higher levels, but they are there. And this does not even touch the formation of new synapses, the growth of the dendrites into new directions, the formation of new neurons. For what we know, even the non-neurons in the brain, like glia cells, may change and play a role. This is why I say that no one believes the strong SPM hypothesis.
Although the author did describe the strong SPM hypothesis as "it's all in the weights", I sorta took their meaning to be more like "it's all in the neurons", where perhaps weights was an (intentional or not) simplification of what that looks like. Would it change your stance if they had de-emphasized the "weights" part?
I could subscribe a lot more to "it's all in the neurons", yes. I would add some caveats that there are other surrounding cells that may play a role, but yes, "it's all in the neurons" is a standard hypothesis. But I don't think the author of the review means "it's all in the neurons" with the strong SPM hypothesis. Right after introducing the strong SPM hypothesis, they are pretty explicit about it:
"The crux of my negative review of the SPM hypothesis is this: cells are extraordinarily complex molecular machines, and there’s a lot going on inside of them that the SPM hypothesis implicitly neglects. We often abstract away most of the biophysical complexity of neurons, which are cells. As cells, they take up physical space, and can have weird, complicated shapes. They talk to other (not necessarily neural) cells. Each individual neuron has a complicated (gene regulatory) network inside it, whose complexity parallels that of many of our models of entire neural circuits. Do we really think that none of this complexity is involved in processes as complicated and multiscale as learning and memory?"
I mean, I agree, they clearly literally meant weights. But if I had to guess they probably didn't know that much about the details and were using a gloss they had heard elsewhere. Probably they now wish they had been more accurate now, because it wasn't that important to the point anyway.
> How can we remember things for years with synapses that turn over on a time scale of weeks or less?
We don't understand it very well, but this question misses an important point: SOME synapses change very fast, in the order of weeks or less. Especially in regions like the hippocampus where long-term memories are NOT stored.
I don't think we know very much about the synapses that don't change. This is because they are hard to measure. You can manipulate animals so that changing synapses light up. But it's harder to make non-changing synapses light up. Also, it's easy to look into a mouse (uhm, if you are a top-notch neuroscientist) and then look into the same mouse a week later and look for changes. But it's much more expensive, and probably also harder, to make the same experiment with a year time lag instead of a week.
I think the standard model is that indeed some synapses change very fast. Fastest is hippocampus, where many things (not everything) changes within days, in some other regions it is more a matter of weeks or months, and probably some regions like the prefrontal cortex contain subsets of neurons and synapses that don't change in years or decades. This is where our long-term memory sits. Now, this model has not much experimental backup, because it's hard to test experimentally, but it doesn't have a lot of evidence against it either.
One nitpick: Connectionism is stronger than the claim that "all of the things that make us human—our ability to talk, think, reason, remember, and so on—follow from networks of interacting neurons, and changes in the strengths of connections between those neurons." Classical / symbolic cognitive theories (typically viewed as the competitors to connectionist theories) also claim that all cognitive functions are implemented via neurons in the brain. The difference is that symbolic cognitive theories view those neurons as implementing symbolic computation (like a computer, where, e.g., logic gates are implemented via transistors), while connectionist theories view cognitive functions as directly resulting from sub-symbolic, distributed, associative patterns across those neurons (like an LLM). See https://plato.stanford.edu/entries/connectionism/.
I know this is not central to the post, but seemed worth pointing out. (Source: I'm a PhD in cognitive science.)
If anyone wants a quick rating on the kookiness scale from a licensed neuroscientist*: the basic biological facts reported are not at all kooky, but the speculations and extrapolations are extremely kooky.
It is not at all kooky to say that single cells have "memories" of some kind, and that these are important, but not well-understood. This is universally recognized in neuroscience and not remotely controversial.
But these are cell-memories of cellular things, not people-memories of people things. It is extremely kooky to suggest that that single-cell memories could be responsible for the reported personality changes in transplant patients.
Kooky things are not always wrong. But they are kooky!
*Not a memory or cellular specialist; also not actually licensed.
There's actually an interesting recent paper on the topic of cellular memory: Kukushkin et al., "The massed-spaced learning effect in non-neural human cells" (not sure if I can link it here), which basically showed that “spacing effect” in learning (training works better when it's broken into several sessions over time) emerges in non-neural human cells, like kidney cells. If I understand correctly, they used human cell lines and found that four brief, 10-minute-spaced "pulses" of a chemical signal (forskolin in this case) drove stronger and longer-lasting gene transcription than a single massed pulse, i.e., a cellular analogue of better memory with spacing. I've see some interviews with the lead author of this paper, and he argues, on this basis, that memory fundamentally depends on cellular machinery common to all cells.
I love this line of inquiry if only for all the weird hypotheses it generates
* does eating meat from free-range animals cause you to absorb some of their... wildness? is that why some people like it a lot? does our mass-produced meat industry manage to produce meat that is missing some vital nutrient of "experience"? does eating factory-farmed meat lead to something like anxiety because you vaguely remember the horror of it?
* do blood transfusions confer some of someones's vitality to you? is this why some people are weirdly into them?
* is this why spiders eat their mates or their parents? to learn how to spin webs better?
* do babies absorb some of their mothers' memories through breast milk?
* does thinking that your memories live in your brain cause us to "ignore" our bodily memories, leading to some dysfunction? is this why bodywork-type therapies help, because they remind you to pay attention to the rest of your "mind"?
* is there some concrete experiential benefit to swallowing, um, semen?
no opinion on the veracity tho. just fun to speculate
Kooky hypotheses are fun! The problem is there's a massive "motte and bailey" element to this, and non-specialists might not pick up on the distinction.
Cells definitely have a form of "memory", but these hypotheses require a massive leap from "having memory of some sort" to "transmission of arbitrary information via transplantation or ingestion". These are super different things and the former is not evidence of the latter!
Context: I'm college-educated and work on software engineering that uses LLMs, so I'm vaguely familiar with concepts of synaptic weights from cultural absorption.
I got to section 4 of this review and realized ... "I hadn't thought much about it, but I had always thought that neuron strength, neuron pruning, and other qualities of the neuron itself might be important, alongside just the strength and existence of the synapse." ... Is that the distinction it's fighting against, the literal interpretation of SPM as only being about synapses? Colloquially I had understood it to be more then just synapses, even though I suppose it's in the name. Just providing a data point that some colloquial understanding never abandoned the ideas this review seeks to resurrect.
> Do we really think that none of this complexity is involved in processes as complicated and multiscale as learning and memory?
Natural selection would tend to eliminate brains that were too easily changed by so many processes. Consider how genes of large effect are disfavored relative to lots of genes of small effect, which better enable a target to be hit rather than overshot.
> People think that epigenetic changes like DNA methylation are responsible, since these changes can be fairly stable, and can be heritable
> Relatedly, it seems unlikely that no organism could learn or remember anything before synapses existed. There had to be some other, simpler mechanism that worked before large-scale, synapse-based neural networks were common.
Why must that have been so? Why is it impossible for there to have been organisms without memory, until the capacity for memory evolved later?
> Maybe your memory of your fifth birthday party is in your arm, or leg, or heart
If that were the case then people would LOSE memories on losing a limb (or getting a heart transplant).
> Can you forget something about your fifth birthday party if someone chops your arm off?
There are many people who have lost limbs. Is there any evidence of them losing memories as a result?
Like a few others here (demost_, maybe others), I am a neuroscientist, I teach classes in learning and memory, and do research in learning and memory. I have been in this field for over 40 years (yes I am old) With that background I believe I can speak to how the synaptic plasticity (in general) and synaptic weight (specifically) hypothesis is viewed and taught in the field. It is a strawman to claim that there is a strict doctrine that synaptic weights are both necessary and sufficient to account for learning and memory that is widely believed and taught.
How does the field treat synaptic plasticity? It is indeed typical to teach about LTP and LTD, but only as one mechanism. It took many years for the field to become persuaded that LTP really was a major mechanism underlying memory, and most courses that have time to do more than just mention it do go into the evidence for LTP being a memory mechanism, as well as what we still don't know about it. In addition most courses acknowledge other interesting examples of plasticity, including white matter plasticity, epigenetics, network dynamics, and even molecular mechanisms! Researchers presenting talks in which they have focused on synaptic plasticity mechanisms in my experience typically acknowledge that there are alternative possible mechanisms as well. I haven't heard anyone in the field claiming that synaptic weights are necessary and sufficient for learning.
I also have not had the experience that AI researchers starting really pushing the idea of synaptic weights underlying learning and memory in the 2010s. The idea was around from much longer, including in the 1980s when I was an undergraduate and connectionism was becoming stronger. At the time there was always an acknowledgement that calling these connections synapses was just by analogy and was never intended to be biologically realistic. I even wrote a paper in undergrad about what we knew then about biological synaptic plasticity and implications of that research for connectionism. I still find that AI researchers who I interact with are quick to acknowledge the lack of biological grounding and the limitations of the synaptic weights view.
With that in mind, the review falls apart for me. The second half argues against a view that may be held in some non-scientific circles but is not an accurate characterization of the field, yet claims to be arguing against a dogma held in the field. The first half is way out on a limb -- many fascinating anecdotes that I did enjoy reading, but probably NOT the best cases to make for a possible molecular basis for memory given the complexity of some of the declarative memories described.
Funny and interesting, and it was voted a finalist so I have to post it - but I don't believe this at all.
Main concern: how do you encode memories in a chemical? That is, suppose you want to encode the fact "My name is Tim". It's easy to imagine how this works as neural connections: during infancy, you hear language, your brain forms a predictive model of the language, you get neurons representing the sounds T, I, and M, and the word "name", and you get links between them. Eventually there's some complicated network connecting name, T + I + M, the sight of a nametag, the motor program of shaking hands, and all the other concepts relating to your name being Tim. We know something like this has to work because LLMs can have this type of memory and they're nothing but connections.
How would you implement this in a protein? All I can think of is a binary-like code with literal text - using one amino acid to represent 0, another to represent 1, and coding the ASCII string "M-Y-N-A-M-E-I-S-T-I-M", and then having a decoder somewhere in the brain. But this doesn't match the phenomenology of human memory, which is sensory (eg associated with pictures, smells, etc) and works even without language (eg in preverbal children). CF https://slatestarcodex.com/2017/09/07/how-do-we-get-breasts-out-of-bayes-theorem/ , which asks a similar question about genetic memory.
But also, even supposing you could do this - in order for the organ transplants to make sense, you need for the heart to be producing or storing these chemicals (why?) in such large quantities that they get into the bloodstream, cross the BBB (why would you have a transporter for these?), make it into the brain, and perfectly interface with the host's own memory-decoder system (how?). Then either the heart has to keep making these constantly (why?) or there has to be some mechanism for a single low-dose exposure to get the host brain to start making them itself (why?) And if the cannibalism claim is to make sense, you have to be able to absorb these extremely complex chemicals through the digestive tract, which is usually doing a pretty good job preventing you from absorbing complex chemicals. Why do this, rather than just keep the chemicals in the relevant part of the brain?
Also, if we had these chemicals, a natural application of them would be genetic memory, eg in the womb you get a big dose of your mother's memory chemicals and now you understand the world as well as she does. As far as I know there are far fewer stories of people mysteriously having their mother's memories compared to their organ donors' memories, adjusted for the ratio of people who have mothers vs. organ transplants.
Also, the writer asks whether losing your arm might make you forget something, or eating an animal liver might give you some of the animal's memories. But many people lose limbs and eat animals, and as far as I know nobody reports these phenomena.
Agreed, I think this review goes way too far. Sure there are other mechanisms for memory in the brain besides synapses. No, I don't believe anything about heart transplants transferring memory.
I think the way you would encode "TIM" in protein is in a protein network. You could conceivably do this inside a single cell, AFAIK.
The problem is cells don't have mouths to say "my name is Tim" with. Within-cell memory is absolutely real and important, but it's just not the right scale for organism-level behavior.
You could have a network of lots of different proteins, but I don't understand how the symbols would be grounded. For example, what would it mean to say that arginine = M instead of N? Unless it bottoms out in some kind of behavior like saying an M or hearing an M, it's just arginine.
...maybe this is just the same point you're making.
Not a protein expert, but I think many proteins are very weird and have many different possible configurations, and those configurations can be influenced by all sorts of features in the environment. In theory you could have a "name" protein that has a different configuration for each of the 1k most common names in the English language or something. This would be a weird way to do things, though!
Have you read "How Life Works" by Phillip Ball? As someone who had done their best to avoid modern molecular biology in grad school, I found it very useful for making sense of the snippets I had heard. It made molecular biology sound a frightening amount like social science.
I still don't understand how this would be grounded. That is, I 100% agree a protein could have 1000 different states. But what makes those states correspond to 1000 different names, as opposed to 1000 different shades of blue?
Suppose you hear the name Tim, and it equals state #489. You need the brain (which is doing some kind of auditory processing and has decided that "Tim" is important enough to devote a protein state slot to) to encode that by modifying the protein to state #489. Then the protein goes to the heart, and the heart is transplanted into a different person. Then the protein goes to the second person's brain. Unless there is a natural mapping between names and protein states, how does the second person get the same protein state code where it knows that a protein in state #489 equals the name "Tim"?
(for that matter, how does the first person establish a protein-state-to-name code which is consistent across all relevant brain cells?)
I agree! These are all very good reasons why you would not want to encode people's names in a protein. It is just too far away from the physical stimulus, there's no easy way for the encoder and decoder to coordinate on a shared meaning.
But many of these problems are also true for synaptic memory. I don't think cortex or hippocampus transplants would transfer memories either.
The vegetarianism thing might work though? The transplant cells are not used to getting meat bits, and when they get meat bits, they tell the brain, somehow, "hey, knock it off with the meat bits".
re the mother part: even if the heart transplant version worked, it would make sense that it doesn't work at birth, because you would need the model developed by being alive for a while to actually translate those memories into anything useful. it doesn't help to have a signal that resonates with the name "tim" if you don't know how to identify discrete sounds out of the signals from your ears yet.
(on the other hand, hell, maybe this what people thinking they remember their past lives is--oh you thought you drowned and were reincarnated? well maybe your mom had a traumatic experience in a swimming pool once)
I thought the review was fairly clear that a complex memory like "my name is Tim" would be out of scope for the proposed non-synaptic memory, and nobody is saying that the ceullular system would completely replace SPM. The definition from the review is clear in that non-synaptic process would be an addition to, not replacement of, the synaptic system, because the synaptic system clearly exists and works to a great degree:
>The formation, consolidation, and retrieval of learning and memory in biological systems often involves stimulus-dependent, non-synaptic molecular and intracellular processes. These processes do not just serve synaptic-weight-based mechanisms, but provide complementary mechanisms. They are necessary for making and keeping long-term memories, but not sufficient, and interact with synaptic-weight-based mechanisms in nontrivial ways.
A single cell organism doesn't have or need a concept of names, so its memory system has no reason to evolve that kind of capability. As organisms grow more complex, they need more complex memories, and thus more capable memory systems. As long as the old system isn't actively harmful, it sticks around because evolution is a hoarder. It keeps doing whatever it's been doing forever and, biochemistry being as baroque as it is, the old systems interacting in useful and significant ways with the new wouldn't surprise me at all.
As you say, there are some fairly basic objections to forming complex memories through transplants, let alone food, and the review might be overplaying that angle in the introduction. But the basic idea of non-synaptic memory has been well argued, I suppose (speaking as an interested reader).
I agree but think the even bigger point is that the reviewer is defining "memory" too loosely to be of interest to me. If I can't recall it then it's not what I call a 'memory'. The only evidence presented towards recallable memory is the transplant evidence, which the reviewer themselves admits is no weirder than many other things we don't believe in.
Still, I love it when people embrace odd theories like this and the gut microbiome one you 'contra'd earlier this year. Aesthetically and emotionally, I want people to go down oddball science interest trajectories and stand in the face of conventional wisdom or even common sense. (Just don't put them in charge of our health system next time, please!)
Neurons themselves are nothing but chemical/biological processes. Is it really weird that some alternative combination of chemical/biological interactions could serve the same purpose ?
At any given point in time, a neural network embodies a particular mathematical algorithm, and any given algorithm can presumably be instantiated in many different ways.
I don't think it's any more intuitively obvious a priori that a memory can be stored in the form of a neural network than in some other form. The obviousness that storing memories in the form of a neural network is possible mostly comes from the confirmation we got from modern AI, and it seems to me that dismissing other mechanisms as much more unlikely a priori is hindsight bias.
Re: "we don't get our mother's memories", uhm, yes we do ? How did you learn to breath ? It was thanks to your mother's and father's DNA, which is not a neural network.
Separately, I think much of the presented evidence in the review is rather weak. The experiments about grinding worms into dust and the like would be big if true, but they seem rather untrustworthy at first glance. But a weak form of the proposed theory is obviously true (re:DNA, individual cells), and a somewhat stronger version does not seem a priori very implausible to me.
The review is very nicely written and raises interesting points. But I want to severely push back on one point. The neuroscientific community is quite misrepresented here.
The author claims that the consensus is the strong SPM hypothesis. But no one believes this. No one. At least not among the neuroscientists who are actually biologists who work with individuals neurons. And I have talked to some real experts in the field, and (as a bit of an outsider) I have done academic research in the field. Not working with actual neurons, but building models of spiking neural systems that were supposed to be pretty close to biology.
Let me try a more accurate description. The human brain is pretty complicated. In order to understand it, we need to look at it at several levels of abstraction.
A) At the lowest levels, we try to understand the chemistry of how a single spike is formed.
B) On slightly higher levels, we may investigate how the input form other neuron excites another, without modelling each single spike on a chemical level.
C) Even higher, how patterns of spiking look like in an ensemble of neurons.
And so on, until we are at regions of the brains like the prefrontal cortex, or even higher up at behavioral level.
By the way, all this is not even covering yet learning, it is just describing a snapshot of the system.
If you are at level B, how a neuron excites another: you can't work with weights on that level. Weights will not explain what's going on. Everything is totally non-linear. When two incoming spikes arrive at the same time at the same dendritic branch, then the effect is much larger than the sum of the two. Especially if the one further from the soma (the blob at the center of the neuron) comes a little before the other. Timing matters a lot. Some people say that neurons should not be described as "integrate-and-fire" (adding up the inputs and fire when they are above a threshold), but rather as "coincidence detectors" (they fire when they get two simultaneous input, screw their weights). Everything is awfully complicated.
But then you want to abstract this whole mess. Because the higher level of abstractions are also important, and you need them.
What is the best abstraction of a neuron and its synapses? Perhaps first you only keep like 10 parameters per neuron and 2-3 per synapse. If you want to make it even simpler, then you keep 3-4 per neuron, and perhaps 1 per synapse. But what if you want to go even more abstract, what do you keep?
And then the consensus probably is: if you really force me to reduce everything to a single number, then probably a "weight" per synapse is our best bet. But it's a simplification.
And now, when it comes to learning: the strong SPM hypothesis claims that only weights change. On higher levels, this is a useful abstraction. But on abstraction level B it is so obviously wrong. Of course the neuron is plastic and changes! All of the time, and all parts of it, not just the synapses. The soma changes, the ion pumps change, the genetic expressions change, the timing of the incoming spikes change. Those are things that we have abstracted away at the higher levels, but they are there. And this does not even touch the formation of new synapses, the growth of the dendrites into new directions, the formation of new neurons. For what we know, even the non-neurons in the brain, like glia cells, may change and play a role. This is why I say that no one believes the strong SPM hypothesis.
Although the author did describe the strong SPM hypothesis as "it's all in the weights", I sorta took their meaning to be more like "it's all in the neurons", where perhaps weights was an (intentional or not) simplification of what that looks like. Would it change your stance if they had de-emphasized the "weights" part?
I could subscribe a lot more to "it's all in the neurons", yes. I would add some caveats that there are other surrounding cells that may play a role, but yes, "it's all in the neurons" is a standard hypothesis. But I don't think the author of the review means "it's all in the neurons" with the strong SPM hypothesis. Right after introducing the strong SPM hypothesis, they are pretty explicit about it:
"The crux of my negative review of the SPM hypothesis is this: cells are extraordinarily complex molecular machines, and there’s a lot going on inside of them that the SPM hypothesis implicitly neglects. We often abstract away most of the biophysical complexity of neurons, which are cells. As cells, they take up physical space, and can have weird, complicated shapes. They talk to other (not necessarily neural) cells. Each individual neuron has a complicated (gene regulatory) network inside it, whose complexity parallels that of many of our models of entire neural circuits. Do we really think that none of this complexity is involved in processes as complicated and multiscale as learning and memory?"
I mean, I agree, they clearly literally meant weights. But if I had to guess they probably didn't know that much about the details and were using a gloss they had heard elsewhere. Probably they now wish they had been more accurate now, because it wasn't that important to the point anyway.
> How can we remember things for years with synapses that turn over on a time scale of weeks or less?
We don't understand it very well, but this question misses an important point: SOME synapses change very fast, in the order of weeks or less. Especially in regions like the hippocampus where long-term memories are NOT stored.
I don't think we know very much about the synapses that don't change. This is because they are hard to measure. You can manipulate animals so that changing synapses light up. But it's harder to make non-changing synapses light up. Also, it's easy to look into a mouse (uhm, if you are a top-notch neuroscientist) and then look into the same mouse a week later and look for changes. But it's much more expensive, and probably also harder, to make the same experiment with a year time lag instead of a week.
I think the standard model is that indeed some synapses change very fast. Fastest is hippocampus, where many things (not everything) changes within days, in some other regions it is more a matter of weeks or months, and probably some regions like the prefrontal cortex contain subsets of neurons and synapses that don't change in years or decades. This is where our long-term memory sits. Now, this model has not much experimental backup, because it's hard to test experimentally, but it doesn't have a lot of evidence against it either.
One nitpick: Connectionism is stronger than the claim that "all of the things that make us human—our ability to talk, think, reason, remember, and so on—follow from networks of interacting neurons, and changes in the strengths of connections between those neurons." Classical / symbolic cognitive theories (typically viewed as the competitors to connectionist theories) also claim that all cognitive functions are implemented via neurons in the brain. The difference is that symbolic cognitive theories view those neurons as implementing symbolic computation (like a computer, where, e.g., logic gates are implemented via transistors), while connectionist theories view cognitive functions as directly resulting from sub-symbolic, distributed, associative patterns across those neurons (like an LLM). See https://plato.stanford.edu/entries/connectionism/.
I know this is not central to the post, but seemed worth pointing out. (Source: I'm a PhD in cognitive science.)
If anyone wants a quick rating on the kookiness scale from a licensed neuroscientist*: the basic biological facts reported are not at all kooky, but the speculations and extrapolations are extremely kooky.
It is not at all kooky to say that single cells have "memories" of some kind, and that these are important, but not well-understood. This is universally recognized in neuroscience and not remotely controversial.
But these are cell-memories of cellular things, not people-memories of people things. It is extremely kooky to suggest that that single-cell memories could be responsible for the reported personality changes in transplant patients.
Kooky things are not always wrong. But they are kooky!
*Not a memory or cellular specialist; also not actually licensed.
There's actually an interesting recent paper on the topic of cellular memory: Kukushkin et al., "The massed-spaced learning effect in non-neural human cells" (not sure if I can link it here), which basically showed that “spacing effect” in learning (training works better when it's broken into several sessions over time) emerges in non-neural human cells, like kidney cells. If I understand correctly, they used human cell lines and found that four brief, 10-minute-spaced "pulses" of a chemical signal (forskolin in this case) drove stronger and longer-lasting gene transcription than a single massed pulse, i.e., a cellular analogue of better memory with spacing. I've see some interviews with the lead author of this paper, and he argues, on this basis, that memory fundamentally depends on cellular machinery common to all cells.
The How the Hippies Saved Physics link goes to a "critical error on this website" page. I suggest replacing it with an Amazon (https://a.co/d/cSo6tgX) or Goodreads link (https://www.goodreads.com/book/show/11048785-how-the-hippies-saved-physics).
(Yes, I had to click the link for a book with such a name)
I love this line of inquiry if only for all the weird hypotheses it generates
* does eating meat from free-range animals cause you to absorb some of their... wildness? is that why some people like it a lot? does our mass-produced meat industry manage to produce meat that is missing some vital nutrient of "experience"? does eating factory-farmed meat lead to something like anxiety because you vaguely remember the horror of it?
* do blood transfusions confer some of someones's vitality to you? is this why some people are weirdly into them?
* is this why spiders eat their mates or their parents? to learn how to spin webs better?
* do babies absorb some of their mothers' memories through breast milk?
* does thinking that your memories live in your brain cause us to "ignore" our bodily memories, leading to some dysfunction? is this why bodywork-type therapies help, because they remind you to pay attention to the rest of your "mind"?
* is there some concrete experiential benefit to swallowing, um, semen?
no opinion on the veracity tho. just fun to speculate
Kooky hypotheses are fun! The problem is there's a massive "motte and bailey" element to this, and non-specialists might not pick up on the distinction.
Cells definitely have a form of "memory", but these hypotheses require a massive leap from "having memory of some sort" to "transmission of arbitrary information via transplantation or ingestion". These are super different things and the former is not evidence of the latter!
Context: I'm college-educated and work on software engineering that uses LLMs, so I'm vaguely familiar with concepts of synaptic weights from cultural absorption.
I got to section 4 of this review and realized ... "I hadn't thought much about it, but I had always thought that neuron strength, neuron pruning, and other qualities of the neuron itself might be important, alongside just the strength and existence of the synapse." ... Is that the distinction it's fighting against, the literal interpretation of SPM as only being about synapses? Colloquially I had understood it to be more then just synapses, even though I suppose it's in the name. Just providing a data point that some colloquial understanding never abandoned the ideas this review seeks to resurrect.
> Do we really think that none of this complexity is involved in processes as complicated and multiscale as learning and memory?
Natural selection would tend to eliminate brains that were too easily changed by so many processes. Consider how genes of large effect are disfavored relative to lots of genes of small effect, which better enable a target to be hit rather than overshot.
> People think that epigenetic changes like DNA methylation are responsible, since these changes can be fairly stable, and can be heritable
Those tend to be erased in mammals https://www.razibkhan.com/i/82002047/if-only-it-werent-for-that-blank-slate Again, there are reasons for natural selection to favor that.
> Relatedly, it seems unlikely that no organism could learn or remember anything before synapses existed. There had to be some other, simpler mechanism that worked before large-scale, synapse-based neural networks were common.
Why must that have been so? Why is it impossible for there to have been organisms without memory, until the capacity for memory evolved later?
> Maybe your memory of your fifth birthday party is in your arm, or leg, or heart
If that were the case then people would LOSE memories on losing a limb (or getting a heart transplant).
> Can you forget something about your fifth birthday party if someone chops your arm off?
There are many people who have lost limbs. Is there any evidence of them losing memories as a result?
Like a few others here (demost_, maybe others), I am a neuroscientist, I teach classes in learning and memory, and do research in learning and memory. I have been in this field for over 40 years (yes I am old) With that background I believe I can speak to how the synaptic plasticity (in general) and synaptic weight (specifically) hypothesis is viewed and taught in the field. It is a strawman to claim that there is a strict doctrine that synaptic weights are both necessary and sufficient to account for learning and memory that is widely believed and taught.
How does the field treat synaptic plasticity? It is indeed typical to teach about LTP and LTD, but only as one mechanism. It took many years for the field to become persuaded that LTP really was a major mechanism underlying memory, and most courses that have time to do more than just mention it do go into the evidence for LTP being a memory mechanism, as well as what we still don't know about it. In addition most courses acknowledge other interesting examples of plasticity, including white matter plasticity, epigenetics, network dynamics, and even molecular mechanisms! Researchers presenting talks in which they have focused on synaptic plasticity mechanisms in my experience typically acknowledge that there are alternative possible mechanisms as well. I haven't heard anyone in the field claiming that synaptic weights are necessary and sufficient for learning.
I also have not had the experience that AI researchers starting really pushing the idea of synaptic weights underlying learning and memory in the 2010s. The idea was around from much longer, including in the 1980s when I was an undergraduate and connectionism was becoming stronger. At the time there was always an acknowledgement that calling these connections synapses was just by analogy and was never intended to be biologically realistic. I even wrote a paper in undergrad about what we knew then about biological synaptic plasticity and implications of that research for connectionism. I still find that AI researchers who I interact with are quick to acknowledge the lack of biological grounding and the limitations of the synaptic weights view.
With that in mind, the review falls apart for me. The second half argues against a view that may be held in some non-scientific circles but is not an accurate characterization of the field, yet claims to be arguing against a dogma held in the field. The first half is way out on a limb -- many fascinating anecdotes that I did enjoy reading, but probably NOT the best cases to make for a possible molecular basis for memory given the complexity of some of the declarative memories described.