The overfitted brain: Dreams evolved to assist generalization | Hacker News Reader

The overfitted brain: Dreams evolved to assist generalization

cell.com

178 pointshheikinh76 comments

> Notably, the techniques that researchers employ to rescue overfitted artificial neural networks generally involve sampling from an out-of-distribution or randomized dataset. The overfitted brain hypothesis is that the brains of organisms similarly face the challenge of fitting too well to their daily distribution of stimuli, causing overfitting and poor generalization. By hallucinating out-of-distribution sensory stimulation every night, the brain is able to rescue the generalizability of its perceptual and cognitive abilities and increase task performance.

Bit of jump there from randomized/out of sample data to dreams are for generalization.

And if the brain is overfitted, how can it generate out-of-distribution samples? Seems like wildly cyclical reasoning.

How can computers calculate more accurately than the human brain but also have PRNGs or even true hardware RNGs?

But that’s the thing, it’s NOT random. It’s just unlikely scenarios. Dreams still follows the general model of the world.

The brain is a lot of different components working together. Perhaps the bits which do statistical learning are functionally distinguishable from the bits which generate permutations on the data.

(This is in fact how modern neural network training often works: start with a set of 50,000 images, run them through a bunch of permutations — translation, rotation, lighting changes, etc. — and produce a training set of 1 million images which yield better generalization when trained against.)

You'd have to posit multiple models in the brain. For example, one which simulates the world (call it the "world model"), and one which determines the best actions to take in the world (call it the "policy"). The policy might be overfit even while the world model is not. In dreaming, you are drawing samples from the world model and using them to train the policy.

I can't remember what it was, but I've read something within the past 3 months, possibly on these pages, which very much suggested human cognition being derived from a series of overlapping partial models of the world, similar to what you described (but with more simulation models in that layer)

Maybe something like turning up the exploration vs exploitation parameter for tree search AI algorithms.

Of course it is. But that wouldn't make a very interesting hyperbolic headline.

Also a bit of a jump from neural networks to brains.

Read something recently theorizing dreams are to prevent the visual processing part of the brain from being repurposed when we sleep...

That's a very interesting hypothesis actually.

This is what I want to read an article about

Can you explain what that means? Being repurposed for what and why would that be something we’ve evolved to avoid?

Bitcoin mining

Damn cryptojackers

It's like how the visual neurons of blind people get repurposed for other senses. Turns out this happens so quickly we can start to lose visual function in just the 8 hours our eyes are closed while we sleep.

Can't remember source but heard this theory on a podcast.

Yeah, I think that's David Eagleman's idea.

All I know is my dog sometimes yips in his sleep, and it can be fairly loud-- I will forever wonder what he's yipping at, or how that behavior could possibly be selected for. Brains are weird.

I noticed mine seems to have vivid dreams with barking after being woken up at night. Interestingly, the paper claims:

> It is notable that approximately 50% to 70% of the time subjects report having a dream upon a sudden waking during sleep.

That's interesting. I keep a dream journal and most of my entries are on days where I had interrupted sleep.

Waking up and realizing you were dreaming is not the same as waking up and then dreaming

Dogs and their ancestor's wolves are pack animals, maybe the sleep barking wakes the other pack mates making them more aware of their surroundings and better able to defend the pack from threats during the night? Maybe even The barking itself helps to ward off threats that would otherwise attack a sleeping animal.

Is there even anything that preys enough on wolves where enough of them are being killed by it that the sleep barking would be selected against?

I thought dreams only happen durring rem and you are paralyzed durring rem stages. So you wouldn't be physically reacting to anything. I think those movements only happen in other stages of sleep

What's your source for the paralysis being complete during REM and movements happening only during other stages?

That seems suspicious to me given that a) REM is named for movement, and b) these mechanisms definitely aren't perfect, as things like night terrors and sleepwalking demonstrate.

For what it's worth, the American Kennel Club claims that dog motions do happen during REM: "We can’t know exactly what dogs dream about, but researchers have observed that certain breeds of dogs tend to perform breed-specific behavior in their sleep when the pons is inactivated. Pointers, for instance, point, and English Springer Spaniels exhibit flushing behavior during REM sleep." https://www.akc.org/expert-advice/lifestyle/why-dog-twitch-i...

This is fairly easy to google its pretty basic sleep info...

https://www.sciencedaily.com/releases/2012/07/120711131030.h...

sleepwalking isn't related to rem stage, most movement doesn't occur during rem stages

https://www.mayoclinic.org/diseases-conditions/sleepwalking/....

https://www.uofmhealth.org/health-library/hw48331

I don't think its farfetched to guess this applies to other mammals

I'm not saying it isn't the general case. I'm saying it isn't absolute. I also am not saying sleepwalking is part of REM; I'm saying it is an example of how mechanisms for quiescence are not perfect.

And since we're talking specifically about your claim about dogs, I'd like the dog reference specifically, please.

https://www.petmd.com/dog/general-health/do-dogs-dream

This suggests that all mammals brain work simmiliar and are paralyzed durring rem.

I'm not sure why your bringing up sleepwalking and night terrors, those don't occur durring rem, so it had nothing to do with paralysis disfunctioning

And I'm not even sure what point your getting at with your comments? Just demanding sources. I feel like you can Google these things your self

I know someone who sleepwalks their dreams. People are demanding sources because you are claiming "always" when the reality is almost certainly just "usually."

I'm talking about general cases, not disorders. Sleeping walking doesnt occur durring rem, Google it. there's other disorders that allow for movement durring rem. What does disorders have to do with anything?

I did Google and found a source contradicting your claim, as well as nothing supporting it.

As I've explained, I bring those up as examples that the sleep system is not 100% reliable, so I think it's a mistake to go from the general case of "most motor control is shut down during REM sleep" to "nobody in any species ever twitches during a dream".

My understanding is that dogs and some other mammals don't experience the same extent of paralysis as humans during REM sleep. I don't know the reason for it, but dogs flail in REM sleep. Scared the crap out of me when we first adopted our dog. I've only ever owned cats and I thought she was having a seizure.

I had a cat crawl under the covers and sleep next to me. Once I woke up because it was.. dream running or something?!?

The feet twitched back and forth along my ribs and I am quite ticklish!

I was amused and glad it wasn't claws out dream fighting!

So I think cats can't be ruled out regards moving while asleep too. :-)

My pet theory is a bit similar to the article but in a more computer-architecture way. I tend to think that sleep is our brain performing "GC" our experiences, i.e. choosing what information to retain and what to discard. Since we don't have a separate circuit to do that, we basically reuse the same one (processor and its "bus") for a different purpose, by redirecting its output to /dev/null, and that's a dream. This kinda explains why we don't remember much of it because the result is meant to be discarded. There's no scientific evidence to this but it's quite fun to think this way.

edit: the GC theory also explains why we have a better memory after sleep!

Oh wow. Cell. Please also see previous dicussion: https://news.ycombinator.com/item?id=23956715

I vaguely remember an Indian Paper (or an author with an Indian name) that made the same claims and it got destroyed on HN. But my memory may fail me.

It's Cell Patterns, a new open access journal from Cell Press that started in April 2021.

> Sleep loss, specifically dream loss, leads to an overfitted brain that can still memorize and learn but fails to generalize appropriately

Curious where this claim came from. Obviously a lack of sleep will degrade your day-to-day performance, but I almost never have dreams.

Or you just don’t remember them.

It's a very distinct feeling waking up thinking "I did not dream last night" vs "I had a dream last night but don't remember it" IMO

Do you have any evidence for that. It's very likely you had dreams with both experiences but just didn't realise it.

It seems more likely that the brain has some kind of retention filter on dreams to try to sort the random useless signals from, for example when you deduce the structure of the benzene molecule. Probably all humans dream every night, unless there's some major issue with sleep, but only dreams passing the filter are remembered. Those on the threshold are "remembered as existing" but not the content.

I think that probably has more to do with how recently you were in REM sleep (which is when most dreams occur). If you have the sensation of waking up from a dream, whether you remember it or not, you likely woke up from REM sleep. If you don't think you dreamt at all, then perhaps you just were in REM sleep much earlier during your sleep.

From what I recall, everybody dreams an approximately equal amount of time every night, the only major difference is how much of it is forgotten when you wake up.

I wish I recalled (..or saved) the paper, but waking up releases a chemical in the brain that makes you actively and quickly forget what you just dreamed about.

That’s interesting. I’ve noticed that if I forget a dream and then meditate later in the day it often comes back or at least bits and pieces do some 5-15 minutes into the meditation.

Without very, very strong evidence, that's BS, or more politely put: a retroactive fit of the concept to the operationalization, in this case probably the REM observation which vaguely correlates with dreaming.

For an alternative theory see https://psyarxiv.com/k6trz

This seems true to me. Correlation doesn't imply causation, but I've had my fair share of bizarre dreams where two completely separate concepts try to wedge themselves together in interlocking dream logic. No examples to speak of directly, as even holding them together after awakening is almost impossible.

Also here

https://news.ycombinator.com/item?id=23956715

https://news.ycombinator.com/item?id=27166536

https://news.ycombinator.com/item?id=23922449

It strikes me as rather unintuitive that the brain should be generating its own “out-of-distribution” data. It’s training itself on itself? Compare with an adversarial network, which can be composed of two entirely separate entities, whereas the biological brain is know to have a lot of “bleed-through”, e.g. our memories influence our perception and vice versa.

I looked and was gratified it cited Francis Crick and Graeme Mitchinson's paper from 1983 that had a very similar idea.

A theory of dreams should imo mention a few things:

* Lucid dreaming. In particular does lucid dreaming prevent the generalization benefits?

* Sensory deprivation hallucinations. It would seem this is connected to dreams, but is it?

* Tetris effect. It seems that seeing a lot of a given pattern primes the brain the brain to look for it everywhere.

Seems like it isn't the randomization (high entropy) but rather the excessively low entropy. Sleeping brains are colder and more predictable (due to massive synchronization). REM sleep might be described as when our cortex turns on and tries sensemaking all the intrinsic oscillations.

These theories are fun, but has it been shown that dreaming is an essential part of the function of sleep? If you could suppress dreaming but maintain sleep, would organisms still get the same benefit? Maybe what is required is simply a period of inactivity, and dreams are the effect of white noise cascading through the brain during that period.

From what I remember, all of the phases of sleep are essential — if you suppress any of them, the person becomes unwell (possibly even dies). That includes the REM phase in which we sleep.

I don't know if you can arrange to have healthy REM with zero dreaming, though.

I used to dream a lot; in the past year since I started taking THC as medication I haven't had any. I put this down to the THC basically letting you dream while you're awake, which seems to be more effective than doing it while asleep. As far as I can tell the purpose of dreams seems to be a way for the brain to visualize its own structure and repair itself (i.e. retraining your pattern matching on other inputs, not just visual).

I have too many dreams. It actually sucks since you don't get great sleep.

I have a sleeping disorder and if I wake up and remember having a dream I almost always also feel well rested.

I remember 1-3 dreams every night. All my dreams are vivid dreams, where I know they are dreams while they are happening and I even have some degree of control over the content.

Those are "lucid dreams". My experience is that lucid dreams do lead to you be tired, probably because we tends to wake up after lucid dreams, though we wake up momentarily in normal life too, we just don't remember it.

Maybe lucid dreaming overrides the benefits of sleep since the mind is actively engaged again.

It's all fascinating, and there are no real answers whatsoever.

Also, side questions:

1. Is your working memory pretty good? What about your IQ? 2. How vivid are these dreams and how logically-sequential are they? I.e. can you remember a clear chain of events from the dream?

That's the term I was looking for.

Memory is good. It used to be even better (I could recite important parts of meetings verbatim when I first joined my company 9 years ago). I think I'm still above average, but I'm not sure how to rule out bias on that.

I've never taken an actual IQ test. When I was 5 my verbal/vocabulary skills tested on the level of a normal 10 year old. I was curious about IQ in college and took 2 unofficial tests which turned out very consistent with scores of 121 and 124. I don't really see a lot of value in IQ tests, even the official ones.

The dreams vary. Some are very detailed. Some might not be, or it's just that I don't remember as much detail except in the final dream of the night (I sometimes remember up to 3 dreams per night). Usually they are logical/realistic, but sometimes include unrealistic elements (that I know during the dream are unreal). I can usually remember the chain of events in the final dream of the night very well. I can usually remember the order that I had the dreams in too. The earlier dreams I can usually remember the high level chain of events for the other dreams, but maybe not the smaller details.

Any reason for the questions?

Yea, I asked because ability to lucid dream (and remember those dreams) is closely tied to a high working memory. A good test for working memory is called "dual n-back". Basically, its a modified version of remembering a list of numbers / letters while the list is adding and dropping items, except with a spatial aspect. More info here if interested:

https://www.gwern.net/DNB-FAQ

Thanks for replying without context. Oh, and the reason for the chain of events question is personal curiosity. I can usually remember maybe 3 coherent actions, but there is rarely a clear start and end to the dream that I can remember.

Have you had a sleep study? I used to have tons of vivid lucid dreams during sleep apnea. Now I have much less detailed dreams and they're less frequent, after fixing my sleep apnea. It's a bit sad because the dreams were fun, but I do feel way more rested these days.

No sleep study. I do snore sometimes, but there are times, like on vacation, when I do sleep well. Maybe I'll look onto that. Thanks!

I find it fascinating how a strong atheistic perspective and premise still tries to ascribe meaning to things.

“Evolved to assist” and also most of the comments here that aren’t discussion of noticing animals dreaming.

There's nothing contradictory about using teleological language to describe evolution. Evolution is an algorithm that maximizes fitness in a changing environment over time. It does things "in order to" maximize fitness in the same way that you or a robot might do things "in order to" maximize utility.

Evolution is an algorithm that carries on traits that happened to be carried along when someone successfully mated. It speaks nothing of ‘fitness’ in any useful sense beyond ‘was successful at mating’.

It _tends_ to yield things that correlate well with survivability to time of mating; but to imply anything of intent to it is inaccurate and ascribing purpose where there is none in this model.

> It speaks nothing of ‘fitness’ in any useful sense beyond ‘was successful at mating’.

That's the definition of fitness.

https://en.wikipedia.org/wiki/Price_equation

And yet this still speaks nothing to the 'purpose' implied by so many. Even the article says it's a statistical reality and nothing more.

Similarly your brain acts in ways that are correlated with things like dopamine release. If you analyze it at the level of interacting atoms, you won’t see anything like “purpose” or intentionality there.

The idea of intention is something that only emerges after some abstraction. It’s just as much a metaphor when applied to humans as when applied to evolution, and just as valid.

That's like calling love exclusively a response to a chemical reactions, and I don't think the broad audience believes in that way when using the word 'love'.

Ctrl-F "mania" says something. But on the level of barroom talk this sounds like a plausible explanation of the prodromal effects of sleep derivation.

DNN is a nice hammer, but brain is not a nail. Hopefully this hype will disappear soon.

They talked about this on an episode of Very Bad Wizards

Or sleep is needed to clean the toxins in the brain and dreams are just a funny consequence of disconnecting you from the world

Reply on news.ycombinator.com