Dreaming as neural “annealing” (turning up the temperature)

11:45>> okay. >> Exactly. So >> Okay. Okay. >> Maybe dreaming could fix something like this, right? Yes. And the analogy that I want to make is actually a physics analogy. Something called >> annealing. >> Okay. >> Whenever we have a material that is in a certain state and we want to get it to a new thermodynamic equilibrium, what you can do is you can raise the temperature. >> Okay. >> When you raise the temperature, you're raising the noise, right? >> And so in a physics thing, suppose I've got a potential landscape, right? Which is [clears throat] hills and valleys. >> My state is stuck in a local valley, but it's not like the global minimum. Yes. Right? It's stuck in some local solution and I want to get to the global minimum

12:25which is the actual solution. >> Yes. >> If the temperature is too small, my state is only going to really jiggle around that local solution. But if I raise the temperature, >> then my state is able to explore a lot of different >> avenues because the jiggle >> has increased. >> That's the idea of a kneeling, right? And what people have started to figure out is perhaps the biological function of sleep and specifically dreaming is to raise the parameter of your neural network >> such that you start exploring different >> spaces in your brain. Okay, this happens

13:07a lot of times in um like artificial neural networks. Yes. >> Like when you're when you're trying to code up an LLM or like some kind of AI, a lot of times it's going to get stuck in a local minima. >> Mhm. >> And what you what you kind of do is you like increase there there's like various tricks that you can do to make it not get stuck in that local minima. One of the ways is to increase the temperature. So you increase the jump size in your gradient descent. The other thing you can do is introduce momentum so that if it's going in one direction, even if that's local minima, it's going to have enough to like jump over that energy barrier. And so these the same things is sort of a theory that's happening in brains. Now, is there like a

13:48neurological idea like is there a mechanistic philosophy in our brain that we can actually point to here? Right. Yep. >> It turns out there is. There's a lot of research on how this annealing phenomenon, this raising of temperature is actually happening in our brain. So the REM sleep initiation >> and just briefly and just because I want to make sure I'm tracking the way you described it earlier was when you're in the waking state, there are these tight >> parameters around where to explore. >> Exactly. The temperature is low. >> The temperature is low. And so we're basically trying to ide we're trying to understand if you know is does dreaming loosen >> those parameters

14:29>> the process we made we made the analogy that's akin to a kneeling in physics >> but the simple way of saying it is is turning off the guard rails a little bit to expand the ability to explore into other areas. >> Yes, that's exactly right. Okay. Yeah. [clears throat] And that's that's effectively what's happening. And there is actually neurological studies that show that this is happening at a cellular level, okay, at like the biochemistry level. So REM sleep initiation is driven by um the activation of these REM on cells in our brain. Okay? And they're located in like the LDT and the PPT. That's the lateral dorsal and tegmental nuclei in the brain stem. Okay? Okay, the brain stem is the

15:09part that connects our brain to the spinal cord and make sure our brain can talk to the rest of the body. Now, these cells, what they do is they flood the cortex and the hippocampus with massive amounts of acetylcholine and the norepinephrine goes to zero. Norepinephrine is kind of an inhibition. Acetyloline is an activation. >> Okay. During wakefulness, there's high amounts of norepinephrine. So, you have a lot of inhibition. That's those guardrails that you were talking about. But during dreaming >> that's gone. >> So now without that constraint now these cortical networks have a really high temperature. They have a lot of noise. They're able to explore all the different possibilities, right? And then

15:51when we think about like dream phenomenology, right? All those like bizarre hallucinatory >> yes >> nature of the dreams like when you're like flying or like you're like random crap happens during dreaming. Yes. that might be a subjective experience of this