Comment by seesawtron | Hacker News Reader

Comment by seesawtron | Hacker News Reader

There are two schools of thoughts as to where "memory" is stored in neuronal networks. The larger group of neuroscientists believe it is at the synaptic level, as huge amount of research has shown how synapses change when they undergo Long Term Potentiation (LTP) or Long Term Differntiation (LTD) which relate to increase and decrease in synapse size while undergoing learning. The former correlates to strengething of a synaptic connection and the latter to the opposite.

Gallistel, Hesslow (PI of ferret study, [0]) and colleagues constitute the second, relatively smaller, group of neuroscientists who believe synapses are only an "effect" that one sees as a result of learning. The true mechanisms are either hidden in the nucleus, cell membrane or somewhere inside the cell [1]. This group so far has only very few substantially convincing experiments and more hypotheses. The ferret study [0] is one such experiment in this direction which was published in 2015. I am not aware of any more data to prove any of the hypothesis.

But of course even the inherent mechanisms that guide synapse formation and alteration are in the end guided by proteins "inside" the neuron. To me it seems these two groups are looking at the same idea at different steps of the memory learning pipeline.

[0] https://www.pnas.org/content/112/45/14060 [1] https://arxiv.org/ftp/arxiv/papers/2101/2101.09774.pdf

In general in biology, when there is long-standing dispute as to whether a certain system relies on Mechanism X or Mechanism Y, my impression is that the answer almost invariably turns out to be that X, Y, and previously-unsuspected Z all play a role.

Good take.

Here:

- X=neural network geometric configuration,

- Y=individual synapses due to the various neurotransmitters,

- Z=cytoskeleton (already suspected to play a role)

Yep,

It seems reasonable to suspect that there's a lot of duplication in the human brain - as demonstrated by the periodic articles about people who function normally with brains compressed/shrunk to a tiny percentage of normal size. And I guess that duplication comes from the multiple systems all able to provide a similar function.

For example, man with 10% normal brain size: https://www.newscientist.com/article/dn12301-man-with-tiny-b...

I think scientists should consider a third possibility. That brains overall are actually far more computationally complex (or has far more state-based, CPU-like-behavior) than it needs to be for the intelligence involved in human actions.

It seems plausible to me that the smaller the detail space one goes in, the more overall computation is limited by data-movement bandwidth rather than discreet computation operations. Intel has a couple extra CPUs in just sitting on top of their regular CPUs (manage engine and all-that). GPU computation is often bandwidth limited as well. So basically, perhaps it's possible to do storage and computation in multiple fashions in the brain and nature being nature, does all of them 'cause nothing prevents it. And that ultimately results in computation that might band width efficient but which only a bit of the CPU-like-processes because there isn't a way do things other ways.

I think it's a bit more than looking at different steps, it's about what's fundamental architecturally vs. not. The synaptic side is saying the internals aren't fundamental, in the same sense that you can have ANN's that are nothing but weights and connections. Gallistel is saying the weights and connections aren't fundamental, or at least trivial compared to a state-storing/state-processing machinery inside. Maybe both exist, but either one being more fundamental or important than the other is a salient conceptional difference.

I wonder why both mechanisms could not be in place simultaneously.

E.g. brain can run on glucose or on ketones; muscles can run on oxygen producing CO₂ or without it producing lactic acid, etc. The body has a number of alternative mechanisms, this may be another such pair.

Synapses are formed outside the neuron in the extracellular space (ECS), at the end of axon terminals called "boutons" which are essentially storehouses for vesciles which are tiny packages containing neurotransmitters. The internal mechanisms of the neurons as well as ongoing biochemistry at the location of a synapses "guides" the transfer of proteins necessary for strenghening of removal of these synapses.

So its possible that both mechanisms occur simultaneously, there's just not enough evidence to clearly understand these (yet).

What’s incredible to me is how long the human brain remains a trade secret from nature. Have we ever made a product so well concealed that its IP would withstand being sent for more than 5 years to every “spy” country in the world? Let alone be sitting there for a millenium on top of every human... If that’s not nature’s way to taunt us, its capacity to remain a secret is at least admirable.

Why not both? These ideas seem compatible.

as for 'storage of durations' mentioned in the interview there is a well known paper on "time cells" in the hippocampus [0] by Howard Eichenbaum et al. which doesn't seem to refer to Purkinje cells (which only exist in the cerebellum [1]).

There is significant evidence for temporal memory maintained in the hippocampus and entorhinal cortex [2], just like grid cells[3] in the hippocampus used as coordinate system for spatial and abstract navigation [4] while time cells facilitate "navigation" in temporal dimension.

[0] Hippocampal "time cells" bridge the gap in memory for discontiguous events https://pubmed.ncbi.nlm.nih.gov/21867888/

[1] Basic anatomy of human memory https://courses.lumenlearning.com/wsu-sandbox/chapter/parts-...

[2] Time cells in the human hippocampus and entorhinal cortex support episodic memory https://www.pnas.org/content/117/45/28463

[3] Grid cells: http://www.scholarpedia.org/article/Grid_cells

[4] Time (and space) in the hippocampus https://www.sciencedirect.com/science/article/pii/S235215461...

Truly, I believe there are multiple architectures for information processing in the body, and not only in the brain/neurons. Think of the computing landscape where you have a salad bowl of ANN's implemented on TPU's, some GPU's, some specialized ASIC's, DSP's, some CPU's. There is no reason to believe efficient information encoding through evolution ends up with one architecture. It's going to turn out to be as varied as the differentiated cells in the body, though there may be some unification in foundational units at the equivalent level of transistors of something, maybe some molecular machinery that stores a bit or activates a switch. Presumably that's what Gallistel is looking for. It does seem wasteful for a whole neuron to be the basic unit of information processing, so I agree there should be something more atomic inside.

That is an interesting line of thought. Cortex and non-cortical regions are somewhat different in terms of their cell types composition which could inherently support different computation and learning mechanisms.

I always saw the hormones that regulate our emotions as primitive forms of memory, older than neurons. So an organism can remain in a certain state for a while.

There is also the possibility that RNA is used for memory:

https://www.bbc.com/news/science-environment-44111476

Nice story, but actually it is superconductors that lurk inside, with absurdly high critical temperatures.

https://arxiv.org/abs/1812.05602

Reply on news.ycombinator.com