The overfitted brain
November 16, 2020 5:56 AM   Subscribe

How Artificial Neural Networks Paved the Way For A Dramatic New Theory of Dreams “ The goal of this paper is to argue that the brain faces a similar challenge of overfitting, and that nightly dreams evolved to combat the brain's overfitting during its daily learning. That is, dreams are a biological mechanism for increasing generalizability via the creation of corrupted sensory inputs from stochastic activity across the hierarchy of neural structures.”
posted by dhruva (64 comments total) 41 users marked this as a favorite
 
Sounds a lot like psychedelics.
posted by swr at 6:11 AM on November 16 [4 favorites]


dreams are a biological mechanism for increasing generalizability via the creation of corrupted sensory inputs from stochastic activity across the hierarchy of neural structures.

Man, if I had a nickel for every time I've said exactly that...
posted by gusottertrout at 6:11 AM on November 16 [38 favorites]


I feel like I am 85% of the way to understanding this? The abstract mostly makes sense but I need, like, an analogy or something. Also reading the phrase "in vivo or in silico" makes me feel like I live in a cyberpunk future (which I guess we all do).

So are they basically saying overfitting is when you learn one thing really well but can't generalize it, and that this is a problem with artificial neural networks which suggests maybe it's also a problem with our actual brains? And that since one way to fix it in computers is to just inject some "noisy" messy input maybe dreams serve that purpose for us? Cool.
posted by Wretch729 at 6:37 AM on November 16 [8 favorites]


The discussion of superstition misses an important point: the effect on the mental state of the practitioner. There is no way that wearing special underwear can affect the outcome of a basketball game that I'm watching on TV. But if I'm playing in that game and it affects my mental state, say it makes me more relaxed, then it definitely can have an effect. Why do researchers studying artificial intelligence suddenly turn around and act as if natural intelligence does not exist?
posted by hypnogogue at 6:53 AM on November 16 [6 favorites]


Wretch729: That's basically it.

Though it's seems a little off that this is called a "new theory". I'm glad the paper cites The "Wake-sleep algorithm" paper (Hinton, Dayan, et al., 1995), because... the idea is already there? I mean, they didn't make a bit deal out of the neuroscience connections, but... it is called the wake-sleep algorithm.

On review: the linked article is a bit too breathless, but the paper is nice, and summarizes some history on potentially convergent results. I don't have the feeling it's saying anything groundbreaking, but it would be cool if someone could verify these ideas with experiments on animals or humans.
posted by Alex404 at 7:04 AM on November 16 [1 favorite]


Isn't dreaming just the result of the brain stem occasionally kicking the tires of the of intellect while we are asleep for whatever biological reason it does? I mean, it's clear the intellect while unconscious is trying to interpret this "false data" within the framework of everyday impressions, as well as subconscious fears, and not really doing that great a job of it. Maybe if we have to struggle so hard to find a purpose for dreaming, it really doesn't have a "purpose" at all, but is just a mostly harmless side effect of the brain stem's sleep cycle.
posted by jabah at 7:08 AM on November 16


Athlete superstitions aren't a perfect example, but it's an accessible one. Humanity provides no shortage of examples of hasty generalization to choose from.

it would be cool if someone could verify these ideas with experiments on animals or humans.

Yeah, it was a bit disappointing to see that the researchers had proposed some experiments but not actually done them yet. As is it's just a hypothesis. It would be ironic if experiments did not bear out the hypothesis, suggesting that it was itself an example of overfitting by analogizing too simplistically from artificial neural networks to biological ones.
posted by jedicus at 7:09 AM on November 16 [6 favorites]


Given that neural networks are very tenuously related to actual brains, I'm pretty skeptical of generalizations that go in this direction. It's a cool analogy they've drawn, but I wonder if it's over-reliant on the assumption that we're successfully simulating the brain in silico.
posted by little onion at 7:15 AM on November 16 [16 favorites]


If anyone wants to watch their brain overfit, just stare at a random textured surface like a popcorn ceiling or a big slab of wood or a piece of granite and wait until you see a face. You will see one eventually but I'm pretty sure it's not a real face.
posted by GuyZero at 7:21 AM on November 16 [8 favorites]


Wretch729: you basically have it, but need to add an element of "too often the training data had people wearing purple socks performing the activity we want recognized, so the resulting network over-emphasizes purple socks". Injecting noise into the training data or dropping out nodes at random as part of training helps to prevent weird one-off false positives like that from getting too deeply embedded in the underlying structure of the network.

My overall reaction to the paper is that it's a neat idea, but I take massive issue with section 2.2 ("Dreams are for Memory Consolidation"), in particular this paragraph:
"First, there is a lacking strong theoretical argument as to why offline replay of episodic memories would assist memorization, rather than introducing errors, since ground truth is absent offline. Indeed, neuroscience has shown that re-accessing memories generally changes them, rather than enforces them (Duvarci and Nader, 2004)"
When cognitive scientists with a strong computer science background talk about "memory consolidation" what they're actually concerned with is memory fragmentation, which is a fundamental problem faced by *all* Turing machines/information processing systems of any kind, including the meat we self-important neural patterns inhabit. No matter what kind of substrate you're storing memory in, arbitrary additions/removal leads to gaps. In an organic system this is going to be particularly problematic because it doesn't just produce inefficient usage of available memory substrate but will also drastically ratchet up access times the longer it goes uncorrected.

The brain must have some means of addressing this, and much like defrag.exe works by shuffling fragmented blocks off to one side and then shuffling them back in a more compact format, an analogous operation in the human brain would produce semi-accurate replay of existing memories that begin to bleed over into completely unrelated events/phenomenon as it is reseated into a new neurotopological "neighborhood," because of signal bleeding (we are chemical soup, not ones and zeros).

My gut reaction is that they're on to something here, but that it isn't at odds with defragmentation or simulation training. There is absolutely room for dreams to serve all three goals simultaneously, and I suspect that's the case.
posted by Ryvar at 7:36 AM on November 16 [8 favorites]


As someone neuroatypical, I'm a bit curious about some aspects of the idea, for example, on a different site I frequent, some of the participants regularly post dreams they've had, and those dreams are way, way, more involved and elaborate in unusual detail than almost any I've had, save for a handful of examples throughout my life in periods of excessive stress. Their dreams may involve celebrities or strangely unique environments, where mine basically never do.

They cover roughly a small range of repeating spaces and "characters" that I can easily map to real life echoes, even with the distortions the dreams produce. While dreaming I remember the locales, even if, as is often the case, I can't find something within them. They are, for lack of a better word, familiar. Which leads me to wonder if that's associated to why I seem to process and generalize information and connections so differently than many others seem to as well, and maybe why I look to the arts in a different fashion than others seem to, suggested by the last section of the article.
posted by gusottertrout at 7:41 AM on November 16 [2 favorites]


"in silico" sounds cool but it should be "in silicio", right?
posted by 7segment at 8:08 AM on November 16 [1 favorite]


"in silico" sounds cool but it should be "in silicio", right?

In silico has been around for ~30 years, so we're probably stuck with it. If it makes you feel better, the Latin term for silicon was only invented in the early 19th century, and as a wise man once said, "all words are made up".
posted by jedicus at 8:22 AM on November 16 [5 favorites]


It's just pseudo-Latin made up to sound like the terms in vitro and in vivo. It has become the common way to refer to findings from computer simulations as opposed to biological experiments.
posted by little onion at 8:24 AM on November 16 [1 favorite]


If it makes you feel better, the Latin term for silicon was only invented in the early 19th century,

I'm no prescriptivist, but it does actually make me feel better that the same community who made up the term also modified it. Words will change, but I find there are cases where the (often diglossic, or otherwise unbalanced) power dynamics involved can certainly justify some purist caution.
posted by ipsative at 9:05 AM on November 16


Not that I think that's the case here. Just, generally.
posted by ipsative at 9:05 AM on November 16


I did not read the paper, just the article, so bear with me. What I got is that, as humans, we tend to relate things that aren´t really related, in a correlation equals causation sort of way. Apparently artificial intelligences do, too. Dreaming may be a way to break those false links by filling our heads with unrelated activity and information.
posted by olykate at 9:12 AM on November 16 [2 favorites]


Exploring the latent space of a GAN is remarkably similar to what dreams feel like in retrospect. Colors, shapes, and forms seamlessly bleeding into each other, constantly morphing and changing. I think dreams are a spatial/conceptual version of what that does for 2D images.
posted by Rhaomi at 9:14 AM on November 16 [1 favorite]


It does seem like dreams are a function of the imagination, which is something that runs all the time, day and night. Imagination + perception = wakefulness, and thus the role of sleep is simply to turn off perception, not to "turn on" dreaming.

This article attempts to address the question "Why do we need to turn perception off at all?" I think people have been onto the idea that "dream states integrate daily experience thru the quietude of senseless imagination" for a while now. I'm not sure if the statistical learning metaphor actually adds any insight here, it's just the modern metaphor of the day.
posted by grog at 9:24 AM on November 16 [4 favorites]


Interesting, seems it has merit as a first order approximation. To march the idea to death (and dress up the corpse in Lederhosen), perhaps the tentative comparisons & predictions involved in this noisy simulation process goes some way towards explaining why we have the sense that dreams are predictive.
posted by dmh at 9:30 AM on November 16 [1 favorite]


Another argument in favor of that afternoon nap... most excellent!
posted by JoeXIII007 at 10:05 AM on November 16


Now, here I go again, I see
The crystal vision
I keep my visions to myself
It's only me, who wants to
Wrap around your dreams and
Have you any dreams you'd like to sell
Dreams are a biological mechanism
For increasing generalizability
Via the creation of corrupted sensory inputs
From stochastic activity...
posted by Cardinal Fang at 10:23 AM on November 16 [9 favorites]


Any time dreams come up, it seems that the only criteria is that we dream. No one seems to deal with the actual content of dreams. Given that remembering dreams is hit or miss, though you can train yourself to remember dreams with increasing skill, the problem is in reporting dreams. Until such time that these researchers actually can relate the dream content in a meaningful way to their reasoning of why we dream, then I might pay attention. And generalizing from dream content across populations seems less likely given the immense variety of dreams that are reported. There appears to be a lot of individual character to dreams per person. Oh, I will confess that I find models of brain behavior reduced to models of computation (defrag memory?) a little sad and reductionist.
posted by njohnson23 at 11:49 AM on November 16 [3 favorites]


Isn't dreaming just the result of the brain stem occasionally kicking the tires of the of intellect while we are asleep for whatever biological reason it does?

This "kicking the tires" comes at a cost: The brain needs to be able to shut off movement so the dreamer doesn't act things out. This is a complex mechanism and doesn't always work properly, resulting in problems like sleep paralysis. Why go to all that evolutionary trouble instead of simply not "kicking the tires", which it already does during non-REM sleep anyway? It stands to reason that REM sleep is just too important to not do.
posted by swr at 12:08 PM on November 16


Sleepwalkers with sharp sticks and knives reworking the practice of hunting sabertooth kitties is probably an evolutionary dead end.
posted by sammyo at 12:14 PM on November 16


Thank you, njohnson23.

As someone who has dreamed all his life, figured out a way to game a consistent nightmare when I was a kid, and has kept dream journals since 1988, it is incredibly frustrating that I do feel able to contribute to this thread mostly out of the fear that my experience is not data and therefore worthless.
posted by goalyeehah at 12:16 PM on November 16 [1 favorite]


And it just may turn out when we finally invent an actual android it may be necessary for it to dream of electric sheep!
posted by sammyo at 12:24 PM on November 16 [3 favorites]


years ago when I dug deep into such stuff, I grew very frustrated of science fiction that sought to explain God. I didn't mind the stories that explored the stuff of God and gods, but the ones that actually delivered an explanation -- they left me cold. It just felt, as njohnson23 just put it, a little sad and reductionist. To take a realm of such complexity and mystery and just sort of shrug it off with a hypothetical plausiblity -- it felt like traveling in a strange and exotic land with somebody who kept looking for a hamburger joint.

My dreams remain impossible. And I prefer them that way.
posted by philip-random at 12:41 PM on November 16 [5 favorites]


So the dreams we remember are just the tip of the iceberg of this un-overfitting process that is happening to your thoughts and experiences from the previous day? Like, if I read a chapter on quantum mechanics I might not then dream about it, but the cells and connections that were involved need to undergo this same process?
posted by polymodus at 12:47 PM on November 16


polymodus: if this paper is correct, then yes. Even if the paper isn't correct, ALL complex systems that perform computation and memory storage/retrieval and can overwrite or reuse that memory will experience fragmentation and eventually need some sort of cleanup sweep. Whether dreaming is when that happens or where overfitting is corrected or both is an open question but how we experience dreams - from the most realistic to the completely surreal and outlandish - makes it a strong candidate for required neurostructural maintenance work.

To objections of reductionism: it doesn't matter whether the system in question is a gerbil burrowing into the Mongolian steppe, a 16-bit microprocessor from 1990, a human mind or a pan-galactic super-intelligent swarm of planet-eaters: it will need to deal with memory fragmentation. This is fundamental laws-of-thermodynamics-grade stuff, here: nobody and nothing gets a pass.

Apologies for the anthropic principle, but: even posting a comment to Metafilter is proof that the poster has some capacity for memory defragmentation and it is working.
posted by Ryvar at 1:03 PM on November 16 [1 favorite]


The superstition angle seems misleading, particularly if you're going to claim this is analogous to machine learning. I think it's more like if you learned to play a version of Tetris where each shape was always the same color, you might end up using color as part of the way you play. But then if you try a version of Tetris with a different color scheme, you wouldn't be able to play as well.

So your dreams try to shake things up. "A square that's yellow? That's crazy! But, I guess it would still fit, wouldn't it?"

So it's more about the brain trying to disregard irrelevant detail. And when you put it that way, this explanation seems related to one of the other prominent theories, that dreams are part of the mechanism for remembering and forgetting things.
posted by straight at 1:29 PM on November 16 [1 favorite]


Even if the paper isn't correct, ALL complex systems that perform computation and memory storage/retrieval and can overwrite or reuse that memory will experience fragmentation and eventually need some sort of cleanup sweep.

The brain is not a von Neumann machine.
posted by logicpunk at 1:48 PM on November 16 [6 favorites]


Wouldn't dreaming be helpful in correcting overfitting only if the brain made no distinction between dream experiences and waking experiences? Wouldn't the dream experiences have to be added to the same dataset as the real life ones for this to work? I don't feel like that's what actually happens in my brain. It seems like dream experiences get separated out as unreliable.

For instance, in my dreams, it's common for pushing on the brakes in my car to have only a very slow, gradual effect on my forward momentum, so I overshoot where I want to stop by dozens or even hundreds of feet. In my dreams, it's a familiar experience. But the many times I've had that experience seem to have had no effect at all on my waking brain's expectation of what will happen when I hit the brakes. And of course I wouldn't want them to.
posted by Redstart at 1:57 PM on November 16


Seems related to simulated annealing to me.
posted by sjswitzer at 2:41 PM on November 16


The brain is not a von Neumann machine.
posted by logicpunk


Correct. Any system with non-deterministic memory writes will eventually run into the brick wall of fragmentation. Doesn't need to strictly adhere to von Neumann architecture, have an instruction set or encode in binary, or even be a machine. Over two decades ago this was covered mid-semester in freshman Intro to Cognitive Science 101, and portrayed as our equivalent to the second law of thermodynamics or Godel. If that's changed I'd appreciate a link.
posted by Ryvar at 2:43 PM on November 16


just stare at a random textured surface like a popcorn ceiling or a big slab of wood or a piece of granite and wait until you see a face.

I’ve “heard” that on mushrooms you can see a face in practically every cloud.
posted by sjswitzer at 2:51 PM on November 16


some people need mushrooms?
posted by sammyo at 2:52 PM on November 16 [2 favorites]


just stare at a random textured surface like a popcorn ceiling or a big slab of wood or a piece of granite and wait until you see a face.

That overfit's not necessarily all coming from your brain. Rabbits, eg, have fairly specific hawk feature detectors right on the retina.
posted by rhamphorhynchus at 2:57 PM on November 16 [1 favorite]


None of this explains bathroom dreams, you know the ones, where you have to go but can't find an acceptable place to eliminate, especially when I do not actually have to get up in the middle of the night.
posted by seanmpuckett at 3:44 PM on November 16 [1 favorite]


I’ve “heard” that on mushrooms you can see a face in practically every cloud.

Well then the iPhoto face detector must have been on 'shrooms.
posted by RobotVoodooPower at 4:01 PM on November 16


late to the party, but here's a few things...

There are a bunch of different kinds of 'simulation' that happen when training machine learning models. Two different types that might be interesting in this context:

a) Augmentation: Taking a particular input and doing any of a thousand things to deform it without changing it's underlying meaning. Eg, for audio: changing the volume, changing the EQ, adding reverb, mixing in different kinds of background noise at different levels, etc. This helps avoid the tank problem by giving the model a wider range of experiences than is immediately available in the training set. Typically this kind of thing is more a technique for classification and perception algorithms.

b) Simulations: Reinforcement learning, OTOH, tries to model sequences of actions and reactions. (for example, playing a board game or driving a car.) To extend datasets, it then helps to simulate different scenarios, perhaps with other augmentations. This is probably much closer in spirit to dreaming, which tends to have some narrative structure.

These of course aren't mutually exclusive! So you can apply a lot of augmentations while creating a simulated situation. So it wouldn't be suuuuper surprising to me if the brain did lots of these things all at once, in the time available.

(Meanwhile, I completely agree with everyone else who's said that using what's currently popular to explain how dreams work is probably wrong. The history of people using bees as a model of human society - and projecting their own values onto the hive - is similarly fraught.)
posted by kaibutsu at 5:28 PM on November 16 [1 favorite]


This is my subscribed theory of humor. The brain goes round and round doing life stuff while awake and when it get off track somehow and hits an incongruous state... we laugh and flood our brain "... via the creation of corrupted sensory inputs from stochastic activity across the hierarchy of neural structures." and laugh our asses off to reset the brain back into functional survival mode. Most if not all humor is a bit of a cultural taboo, something you're not supposed to think about and when you find yourself there thinking about it, you laugh. Go ahead, try to find something that's funny that also isn't a bit mean. The fully routine day to day isn't funny. The funny is always a gaffe of some sort if you look hard enough. That's why we laugh and flood our brain with random, it's all to stop that wrong thinking and bring us back to right thinking. Somehow we enjoy this. We seek out being a bit miffed and laughing so hard that you're crying. Dreams are just unconscious humor.
posted by zengargoyle at 5:29 PM on November 16


Redstart: Wouldn't dreaming be helpful in correcting overfitting only if the brain made no distinction between dream experiences and waking experiences?

I was reminded of this very cool video sequence suggesting that octopus 'dreams' have real physical consequences.
posted by dhruva at 6:33 PM on November 16 [1 favorite]


"Will I dream?"
posted by DevilsAdvocate at 7:20 PM on November 16


Any system with non-deterministic memory writes will eventually run into the brick wall of fragmentation.

it's just weird to call what the brain does a memory 'write' at all. assuming you agree with the idea that a memory is stored in the network of weighted connections between neurons (a 'neural network' if you will), it's not like there's a discrete network that encodes any specific memory.

if you could identify all the connections that constitute a particular memory, you could probably go in with an electron knife and excise that memory, but you are going to also degrade other information in the network that is partially stored in those connections as well. encoding new information into the brain doesn't overwrite existing information if there's not enough 'space': it gets integrated with what's already there.

which isn't to say that there aren't issues with trying to encode too much information in a limited capacity network; hell, some of them might even look like the increased 'access time' you refer to. two similar memories (e.g., your 10th vs 11th birthday) may interfere with each other, making it more difficult (i.e., take longer with more errors ) to recall specific events from one or the other. Furthermore -

Over two decades ago this was covered mid-semester in freshman Intro to Cognitive Science 101...

oh, well. i guess i'll just take my ph.d. in cognitive science and go pound sand.
posted by logicpunk at 8:56 PM on November 16 [7 favorites]


I’ve “heard” that on mushrooms you can see a face in practically every cloud.

I've never tried psychedelics, but every night I stand in front of my toilet and urinate before going to bed. Depending on how well-hydrated I am, this can take some time...and I pee standing up so my gaze is usually fixed upon the textured painted wall behind my toilet.

I always manage to find at least one --usually several-- humanoid skulls in the paint, staring woefully back at me as I void my bladder.

Kind of sounds like "overfitting" is another word for pareidolia.
posted by Doleful Creature at 9:50 PM on November 16


If "overfitting" is the brain's way of dealing with information overload, or generalizing it, one would think that it might be in sorting, prioritizing, categorizing and creating a kind of hierarchy of importance for the information one takes in.

As an example, imagine you went to a cabin for a weekend with some friends. This is something you do every year with mostly the same group, but with some variation. You go boating with your buddies. The day is hot and sunny and during the excursion the propeller comes off the motor and you have to paddle back to shore.

In memory, certain things will stand out more than others. That specific boating excursion might be a main prompt because of the unusual circumstance of loosing the propeller, making for a unique memory of some added emotional weight so it will be easy to recall.

That going to the cabin with some of those friends was a regular occurrence might make remembering which time your buddy Joe said or did X harder to recall exactly because it is tied to a number of other similar memories, where remembering Joe's partner Y was along and only went with you once might help place or label other occurrences during that specific trip by associating it to the unique element of Y's accompaniment, even as that might not help you remember which specific year that happened because your association with Y doesn't have a strong memory trigger aside from that trip.

All the different elements of the various trips are essentially compressed, along with everything else you experience and feel, into a sort of hierarchy of importance or categories. If you are asked to think about hot and sunny days, you will have a storehouse of experiences that you will refer to, but most of them will be "untagged" with significance, the ones that retain a stronger pull having something else of import that makes that day stand out becoming a key experience that stands in as a central reference point with a select group of other days where sunshine and heat matter to the "story", other general instances may still exist in a vague connection with sunshine and heat, but lack specificity.

Another generalized slot will be all the experiences you've had with Joe, where some will be linked to specific days or situations, while others might be more vague in their time frame and location. There will be likewise generalized and specific slots for cabin visits, boating, and everything else, which may or may not hold prominence of memory but might be able to be recalled with secondary prompts tying events to times or people to places or whatnot. The significance of a memory might be in the feeling associated with it or the singularity of the event, holding it more "ready" for access than other information, which is how people can be surprised when they hear thing A happened 20 years ago while thing B was only 5 years back because thing A feels far more present in the mind.

Sorting all this information into a kind of hierarchy or into more generalized informational nodes obviously takes some effort and isn't something we do entirely consciously, even as our feelings about events will dictate some sense of how we rank significance. This work then is done some other way, where our memories are being matched to feelings and slotted into points of reference, immediate need or long term association easier and harder to access depending on how the experience is or is not generalized with others.

This seems like something that dreaming might be assisting in and might be influenced by as the shape of the dreams may be driven by the information being sorted, even if only in a more symbolic way, associating connections like creating analogies, and how we add different casts of "characters" and places to our dream world as we grow, while retaining some strong attachments to old "characters" and situations, like the late for a test at school dreams people report long into adulthood. But, hey, what do I know? It's not my area of expertise, but it does sort of fit with some other things I do study as a possibility.
posted by gusottertrout at 12:04 AM on November 17 [2 favorites]


Hmm, sorry to go on, but I guess my assumption would be that dreaming is might be better described as a weighting of experience/memory, in both the statistical sense of assigning all the experiences increased or decreased significance by the significance of our emotional/intellectual response and in the sense of the gravity of the experiences warping our generalized concepts by dint of certain experiences having that greater importance or weight to them, which are so often reflected in dreams, directly or by seeming analogic "likeness" or symbolic association.
posted by gusottertrout at 1:25 AM on November 17 [1 favorite]


The body as software is a seductive pseudoscience, as was comparing the body to a machine in the age of industrialization. It can be a useful metaphor to teach children about the human body, for example.

But: the brain is not ‘overfitting’ — it is not using someone’s machine learning algorithm.

One could just as easily make the claim that some brains are stuck in an infinite loop of comparing anything they come across to machine learning. Why not? The brain is a loop function!
posted by romanb at 1:35 AM on November 17 [6 favorites]


But: the brain is not ‘overfitting’ — it is not using someone’s machine learning algorithm.

Right, the difference is that machine learning, as I understand it anyway, is trying to fit information into already determined patterns of "correctness", while humans constantly "rewrite" their conceptual pattern understanding based on new information. If, for example, I like Joe my immediate associations with him are likely to be warm memories, but if Joe would become abusive, the recollections might shift to those that reinforce a negative perspective instead. My way of categorizing my relationship with Joe or Joe himself is not a constant, but adaptive and that will inform the recall of my history and memories.
posted by gusottertrout at 2:12 AM on November 17


As a metaphor, sure, the brain has a CPU and some storage space and it uses an ML algorithm in some sort of buggy artistic screensaver application, and so on.

You could say that the authors are over-fitting a metaphor. It would be equally valid to write a paper on how a human leg functions like the wheel of a car. Sure, I guess?
posted by romanb at 5:23 AM on November 17


A few points for logicpunk:
1) Dropping a no-shit-Sherlock one-liner like "the brain is not a von Neumann machine" into the thread and then suddenly about-facing with "oh btw check out my PhD" is kind of a party foul, and I think you know that.

2) My point was not to compare your level of experience to my own (I had a serious mental health crisis one semester before finishing the usual Cog Sci kiddie's Comp Sci/Psych dual-major - which is not nothing but far less than your level of expertise), but rather to say that even in the Dark Ages of AI when it was considered a dead field this was about as fundamental as it gets.

I have no idea if you were in the field then, but we still had to fight our professors tooth and nail to get any acknowledgement of neural-based approaches. Every last person senior to the grad students was still beating the dead horse of formal semantic logic and responses to "these fuckin' kids and their bottom-up neural-based fetishes" were reactionary to say the least.

And yet: even back then they were teaching memory fragmentation as fundamental regardless of whether it was RAM modules or meat. ANNs were considered a dangerously irresponsible hook to hang your thesis on and still...

3) My "If that's changed I'd appreciate a link" was very intentionally BOTH snarky/dismissive AND a genuine request. Like, if you've got something, I'm legit all ears.

That bullshit out of the way:
assuming you agree with the idea that a memory is stored in the network of weighted connections between neurons (a 'neural network' if you will), it's not like there's a discrete network that encodes any specific memory.

Right. My assumption is memory encoded across many, many unevenly distributed synaptic connections that only vaguely correspond to any kind of semantic element (eg this one's 0.17% "truck", that one's 0.93% "red" and they're both used in a lot of other things but they both definitely get involved when you see a red truck). This does not change the fundamental fact that when memory is written - whether by electrons flipping discrete bits or a few hundred synaptic connections forming - the new memory may not be a good fit for its otherwise-ideal location within the substrate.

if you could identify all the connections that constitute a particular memory, you could probably go in with an electron knife and excise that memory, but you are going to also degrade other information in the network that is partially stored in those connections as well. encoding new information into the brain doesn't overwrite existing information if there's not enough 'space': it gets integrated with what's already there.

We're mostly on the same page. I'm assuming all brains have barely-correlated neurotopology and semantic topology and that where you'd like to place new information because of how neatly it slots into the semantic topology is not necessarily a physically available location at the neural level because corresponding synapses at any one site are finite. All of this is made vastly more complex and messy by the lack of clear distinction between recalling a memory, writing a memory, and re-executing a mental model used to problem-solve or forward-predict at the time of the memory.

which isn't to say that there aren't issues with trying to encode too much information in a limited capacity network; hell, some of them might even look like the increased 'access time' you refer to. two similar memories (e.g., your 10th vs 11th birthday) may interfere with each other, making it more difficult (i.e., take longer with more errors ) to recall specific events from one or the other

Bingo. Because the new memory has to get sited sub-optimally at the neural level this ultimately reduces to something akin to a storage indexing problem if you squint really, really hard. Horribly sub-optimal bridging connections become necessary for there to be any sort of coherence. Repurposing local dead branches presumably has a physical cost and is inefficient for a single memory. And all of this is due to writes being both non-deterministic and non-uniform in size. Hence: fragmentation is a serious problem any large-scale neural network is going to have to contend with, same as an x86 CPU writing to DRAM modules. Nobody and nothing gets a free pass.

Where and when does the reseating take place? Well, since dreams feature recent elements being combined haphazardly with unrelated elements and sometimes really weird shit completely out of left field - even anxieties from forward-prediction discards - that maps pretty neatly with what you'd expect if neurally-encoded memories were being shuffled from one topological neighborhood to the next. There's going to be bleeding from heavy synaptic activity at the new site spilling over into neighbors (ie if the new connection is 0.17% "truck" then a lot of things tied to the other 99.83% are now potentially in play)...this is suspiciously close to our subjective experience of dreams. Far moreso than any other human experience I'm aware of, and my ultimate point is that I don't know how the authors could honestly write anything as broadly dismissive as the memory consolidation section in the linked paper. If we had this as a theory in the late 90s (okay: speculative hypothesis but a very strong and popular one), saying there's no theoretical basis twenty years later is a bit of a head-scratcher. Worse still, solving broader memory fragmentation issues seems like a problem that is very efficient (low energy cost and thus a good survival adaptation) to be solving while addressing overfitting. They're rather similar cleanup activities that need to take place while the brain isn't occupied with survival, because they're going to be reseating connections and spilling over into semantically-unrelated areas while in progress. There's no good reason for the paper's suggested exclusivity that I can see, here, which was my original objection and remains so.
posted by Ryvar at 5:31 AM on November 17 [1 favorite]


not necessarily a physically available location at the neural level because corresponding synapses at any one site are finite
...and over-purposing any individual synaptic connection or local group of connections will lead to a critical loss in fidelity, especially in regions with heavy recent activity.

Sorry, important additional nuance that will, hopefully, stave off obvious elementary objections.
posted by Ryvar at 6:16 AM on November 17


>I’ve “heard” that on mushrooms you can see a face in practically every cloud.

> some people need mushrooms?



with mushrooms you can watch a music visualizer and periodically close your eyes so that you capture the rhythmically shifting patterns of lines behind your eyelids, seeing the dancing lines with your eyes closed that you don't need to actually look at the visualizer at all. once you've gotten the pattern caught in your mind you can then play with it — see if you can make it spin faster, or pulse along with the beat more subtly or more dramatically, or reverse its direction of spin.

once you've done that you can gradually realize that although what you consider "you" can give the spinning spirals suggestions on how they should move, ultimately the pattern is sustained by some other part of your brain, some other thing that is not "you," not exactly. moreover, you can discover that if you are too forceful in your suggestions to the pattern, if you believe you can honestly straightforwardly command its movement, that the pattern can then decide to stop listening to you and then the pattern, now a righteously angry spinning purple pink and blue spider, can tell you that it will "show you what it really is."

there can be a moment of disorientation as the pattern rushes forward into what you consider your consciousness, and then absorb it, altering it in nameless ways, showing you in great detail what it is like to be commanded by something outside yourself.

you can then find yourself — no, not yourself, a perspective with the absence of self — flying very quickly over the face (this is a story about seeing faces, after all) of a small planetoid, not quite as small as the little prince's world but still small enough that you know that this swiftly flying perspective can circumnavigate it in a few minutes. you can then encounter a tall narrow many-faceted mountain, a beautiful crooked shard of stained glass rising from the small planet, and your perspective — remember, this is not you, this is just a perspective that you, insofar as there is a you, is in some way experiencing — this perspective will then gracefully slide up up up the shining stained glass mountain. you (insofar as there is a you) can experience a deep sense of what you can in retrospect describe as nostalgia while this smooth slow ascent occurs.

you can see a face toward the top of the mountain — i remind you once more that this story about seeing faces. the face in/of the mountain is not quite at the top, maybe two-thirds of the way up, and you can think that it looks something like the old man of the mountain that used to stand in new hampshire. except of course it's made of stained glass, because the entirety of the mountain is made of stained glass, and you are looking it full-on rather than in profile.

you can behold that stained glass face, and as you do so you can be suffused with the sense that what you are seeing is the lack of anything resembling a god, the total and inescapable universal impossibility of any god or of any sense or order that a god might provide, the sense that there is no noumenon whatsoever behind what we call the phenomena that we claim to perceive. and then you can, you must touch that face. the face of the absence of god.

you can take a deep breath, a breath that feels better than any breath you have ever taken, and your perspective, moving impossibly fast, can ascend through the blue atmosphere to the blackness of space. and then after that you can open your eyes. you can if you wish also become wholly comfortable with abandoning the delusion that "yourself" is a word that means anything at all. you can look at your own face in the mirror and decide whether you want to see it as a face or just as a carefully arranged, carefully curated collection of planes and facets. you can feel a little bit like you're nothing at all — a true nothing, a good nothing — for days after.

later you can wish you could go back there, that you could see that empty stained-glass face once more and then feel your own true emptiness again. but sadly you can never return.
posted by Reclusive Novelist Thomas Pynchon at 7:06 AM on November 17 [3 favorites]


and then you can, you must touch that face. the face of the absence of god

Oftentimes, when I am struggling with particularly acute psychiatric distress (on the order of once a month or so), my wife - kind soul that she is - will ask me, "hey, is there anything I can get you that would help?" My answer is always the same, somewhere between a favorite in-joke and a genuine cry for help:

"The death of God in the human heart."

I have never taken any drugs that were not prescribed by a doctor, nor had a sip of alcohol in my forty years on this planet. Given the scope and scale of what I'm up against, that is probably for the best.
posted by Ryvar at 7:24 AM on November 17 [1 favorite]


i've never done drugs either. they sound neat, though.
posted by Reclusive Novelist Thomas Pynchon at 7:33 AM on November 17


"with mushrooms you can watch a music visualizer and periodically close your eyes so that you capture the rhythmically shifting patterns of lines behind your eyelids, seeing the dancing lines with your eyes closed ..."

this is the first time I have started reading a comment, said "oh right - that's loquacious" and been wrong
posted by thelonius at 7:44 AM on November 17 [5 favorites]


I've done plenty of psychedelics and trust me, there's a mushroom in every cloud.
posted by philip-random at 8:43 AM on November 17


Does anyone remember who was doing some of this at CogSci in the early '00's (before the Deep Learning era)? A sleep/dream phase to avoid overfitting? Maybe Robert French?
posted by spbmp at 1:18 PM on November 17


Hinton's wake-sleep algorithm for unsupervised neural nets was in '95. On the sleep-memory neuro side you can find a pretty exhaustive review from 2013 here. Ringing faint bells for me from back then are Giuditta's sequential hypothesis and Crick & Mitchison (both '95 the latter specifically wrt ANNs).

While I'm poking around: The Deep Learning paper from 2015, for those curious. 32K cites and climbing, Christ.
posted by Ryvar at 3:13 PM on November 17 [1 favorite]


haha, the 'real' deep learning paper for 2015 was the ResNet paper, which apparently has 61k citations. Apparently I'm behind on my reading.
posted by kaibutsu at 4:23 PM on November 17 [2 favorites]


Your second link has ImageNet classification with deep convolutional neural networks clocking in at just under 74K citations, but that's cheating since it's from 2012 (...okay, yes, fair cop, your paper is the one I should've linked to).

More helpful for anyone else reading this far down the thread, I think, would be this Deep Learning Papers Roadmap, which hits all the landmark papers both in general and task-specific.
posted by Ryvar at 4:59 PM on November 17


Nah, the Nature paper's a totally fine one to link to. I find it kinda amazing the number of citations on the Big neural net papers, though. (Batch normalization is at 22k, "Attention is All You Need" is at 14.5k...)
posted by kaibutsu at 6:32 PM on November 17 [1 favorite]


A few points for logicpunk:
1) Dropping a no-shit-Sherlock one-liner like "the brain is not a von Neumann machine" into the thread and then suddenly about-facing with "oh btw check out my PhD" is kind of a party foul, and I think you know that.


Coming in hot with a pretty basic mix-up regarding 'consolidation' as optimizing read-write access times vs memory stabilization, and following it up with a 'this-is-so-fundamental-i-can't-believe-i-have-to-explain-it' is not exactly a rhetorical power move either. talking about defragging a brain doesn't warrant a particularly nuanced rebuttal. they don't work that way.

cognitive scientists, especially the ones fixated on the computer metaphor of the brain, bother me in ways similar to hardcore bayesian brainers in that they have a very particular view of how the brain should work that skews their interpretation of how the brain actually works.

'defragging' doesn't work as a metaphor for a few reasons. one, it assumes there's a single substrate on which a memory is written, and that the initial 'write' of the memory is necessarily sub-optimal because there's already stuff there taking up space. maybe it's the case that the initial 'write' isn't optimal, but that doesn't really matter because initial storage in hippocampus is only temporary anyway. eventual consolidation in cortex isn't automatic - you can probably remember what you ate for breakfast today, but probably not 10 years ago. transferring the memory to cortex can involve modifications of existing synapses, but it doesn't happen in a single step, and the memory is likely altered in the process - salient details are preserved, minor details are not. consolidation can also involve the formation or removal of new synapses.

maybe all this can be interpreted in terms of the computer metaphor - growing new synapses is like allocating additional diskspace or whatever, or that the loss of irrelevant information is just cleaning up bits that got orphaned during the defrag, but in order to accommodate the defragging metaphor, you have to give the defragger additional scope. it has to be able to select the 'important' information, it has to do space allocation, as well as the ultimate goal of ensuring that every stored memory is addressable/recoverable in a timely fashion. at that point we're talking about a larger disk management system instead of just a humble defragger.

even in your reply you were backtracking on a pure defrag story with synaptic bleedover effects from shuffling data around, which is pretty much not what a defragger should do - like I would be pretty unhappy if I defragged my hard drive and now every time i tried to start chrome i got internet explorer because they're pretty similar.

a fairly standard reply to this (and i'm not saying you've made it) would be that the brain is attempting something that amounts to defragging, but because it's just meat, obviously it's not going to do it perfectly. which is where i started this screed - trying to impose an imperfect metaphor on how the brain works can be useful in a limited fashion, but if you start taking it too seriously you spend most of your time generating excuses for your metaphor instead of accepting that the brain doesn't work like that.
posted by logicpunk at 8:36 PM on November 18


« Older Ancient Clippy from the Deep   |   Fiddle+dance Newer »


You are not currently logged in. Log in or create a new account to post comments.