The overfitted brain: Dreams evolved to assist generalization (2021)

"The Genius Neuroscientist Who Might Hold the Key to True AI"

Karl Friston's free energy principle might be the most all-encompassing idea since the theory of natural selection. But to understand it, you need to peer inside the mind of Friston himself.

The gist of the problem is Karl Friston didn't come up with the idea in question, yet somehow he's often promoted publicly much as if he did.

Variational free energy methods generally are pretty amazing. Thanks to mathematicians, physicist, chemists and so on over the course of the last three centuries or so for giving us that. Free energy methods in perception and learning are also very interesting, and that goes back at least to Geoff Hinton, Richard Zemel and Radford Neal in the mid 90s (arguably there is relevant prior work in psychology and energy based ANN before that too). It's hard to believe Friston's application of the theory to brain function (as opposed to earlier in brain imaging) would be independent of having arrived at University College London, where people at Gatsby Unit (Hinton, Peter Dayan, etc) had been working with the methods in innovative and and effective ways for machine learning/artificial intelligence/theoretical neuroscience since the 90s. Yet if you entered the subfield through Friston's citations, which many people now do, due in no small part to glowing personality portraits like the one presented by WIRED, you might easily come away with the impression that the whole field suddenly manifested out of whole cloth between the ears of this Genius Neuroscientist, some time in the 00s.

Friston does have one real distinguishing contribution to his name though, in being the first person to reach wide attention making vocal and bold claims that the "free energy principle" amounts to a largely comprehensive explanation for how the human brain works. Good for him if that turns out to be the case I guess, but personally I'm not aware of any demonstrations that it actually is such a powerful explanation for any mechanisms of brain function.

I have no personal knowledge of Karl Friston, but I've heard he's often well liked in personal contacts. I don't really know enough about the medical imaging methods era of his work to comment on it.

Basically, it's a very badly reviewed article.

Most of the times I don't remember having dreamt are those when I've been more or less living on autopilot, with nothing novel in work or leisure time. Recently I've had an explosion of vivid dreams that lines up almost perfectly with doing things I'd not had prior experience with when programming in both work and hobby projects.

I've been on a pretty consistent one cup of coffee and occasional second cup over the course of the pandemic. The only way caffeine impacts my sleep is if I have it past 1-2PM, in which case I find it harder to fall asleep and may wake up periodically but will still dream.

The other thing I've noticed is people who smoke a lot of weed seem to report a lack of dreams.

As someone who smokes weed I heard this as well. But from experience I still dream, just that its harder for me to remember them if at all.

I stopped drinking caffeine for a few years (maybe 3?) and have restarted also for a few years (probably 5 now) and I dream fairly rarely both when drinking caffeine and not.

Personally, the amount of sleep I get correlates strongly with the amount that I remember dreams.

I suspect caffeine intake is correlated with sleep deprivation bidirectionally. People who don't have the free time to afford enough sleep tend to drink more caffeine to compensate, and people who drink more caffeine tend to have lower quality of sleep.

So, I suspect that it's the length of sleep that's really the relevant factor here.

I've also heard from several friends that nicotine patches cause very vivid dreams almost every night. Caffeine certainly isn't nicotine, but the case for stimulants in general causing a lack of dreams. Maybe caffeine is an outlier in this regard, but, like I said, I suspect caffeine intake is primarily a confounder for sleep deprivation.

First day light headache. Second day onward totally fine.

Dreams are back!! Not kidding at all. Over the last few days I’ve had much more vivid and sticking dreams. Some of them bad, but dreams nonetheless!

Thanks for putting forth this theory. I will be reducing my caffeine intake permanently. I beleive dreaming is a healthy part of being alive, and I’ve been missing it.

I have confirmed this with experiments - I have woken up in the night, written down my dream and went back to sleep.

When I woke up in the morning I couldn't remmeber a thing and the notes seemed like they were written by someone else. There was no 'yeah, i can sort of remmember it" - memory was totally blank. The experiment is fairly repeatable - not every night, but couple times a week.

Psychedelics can give a taste of this with the experience of separation of self, where it feels like you're an observer behind the one way mirror watching a focus group deliberate on various topics.

Intelligence is a great example. It's obviously beneficial for both survival and reproduction, so should naturally be highly selected for. But in a society where higher intelligence correlates with lower reproduction, evolution will "select" for lower intelligence. Of course it's not evolution selecting for anything, but our own traits paired against our own outcomes. If it turns out to be beneficial for breeding, we could gradually even "devolve." But the notion of devolution is again somewhat nonsensical, because evolution has no direction.

So for many things I think looking for an evolutionary cause is not really meaningful. If evolution had a direction, we'd all be tardigrades.

Because it is learning? If you had this supersmart function that could evaluate situations, why not ask it things when you are training another algorithm that tries to evaluate things cheaper?

Edit: Or just to categorise things in general.

You don't believe animals have a conscious? I thought all humans believes that a dog for example had a consciousness.

> What you're proposing is a dialog between a conscious and subconscious self, while in a dream state, where one or the other (or both) forms of being are fully aware of what questions to ask

You say this is complex, but there is definitely a subroutine that determines when you should sleep, when to wake you up etc. That routine calling you to do some bookkeeping while you are sleeping doesn't add any complexity at all. I don't see why that routine would be smart enough to do everything else, but not smart enough to ask some questions about things when you sleep...

But we do know that our subconsciousness can learn to drive or to walk home etc, since we can do that without thinking about it. It has to learn those things somehow, and asking our consciousness about how all the things works, how to react or think about things when we sleep, that is the obvious ways for it to learn and test things from its control. I don't believe you can run the consciousness functions without us remembering it in some way, and that becomes the dreams.

Edit: Maybe the hard part to understand is that people don't want to imagine themselves as a small part of a larger system. We aren't the main thing in control of ourselves, instead there is something that can shut us down, turn us up, invoke parts of us at will etc, that is in control. It just reuses some of the code that runs our consciousness to run other evaluations.

Any assumption that such a process doesn't exist is therefore disproven immediately, we know that none of those can be true. So, since we know that such a process exists, it asking us questions just like we ask it questions seems like a very useful property to have to handle all sorts of bookkeeping and learning. It is a very obvious mechanic in a brain containing two intelligent systems interacting so tightly as these two does.

Since hearing about this theory, I've been able to pin point my dreams 90% of the time to themes that were discussed or experienced during the day I fell asleep. Including topics I had only worried about but never mentioned to anyone.