I’m personally interested in the interaction of the hypothalamus with learning processes and this is one I find particularly interesting. Do you think AI agents could benefit from implementing similar homeostatic drives in their learning algorithms?
Thanks for the good question, and interesting reference!
In my defense, I did put in a caveat that it’s hardwired “to a first approximation” :-D
Anyway: When I think about brain plasticity, I kinda have two different mental images.
My first mental image is “learning algorithms”. I think about e.g. modern-day ML, or the brain’s various supervised learning & predictive learning & RL algorithms and so on. In this mental image, I’m imagining rewiring rules that result in a “trained model” algorithm that does something difficult and useful, and the “trained model” is at least moderately complicated, in the sense that an astronomically large number of different trained models could have been built if only the inputs to the learning algorithm were different.
My second mental image is “specific situation-dependent rewiring rules”. My go-to example is the self-modifying code in Linux—e.g. “if debugging is turned off, then replace the debugging-related algorithm steps with no-ops”. In a biological example, imagine that the genome is trying to implement the behavior “if you keep winning fights, start being more aggressive”. (That’s a real example, see here.) It would be like, there’s some specific signal (related to whether you’re winning fights), and this signal changes the strength of some neuronal connection (controlling aggression). So in this mental image, I’m imagining some control knob or whatever that gets adjusted for legible reasons, and I’m imagining a lot of species-specific idiosyncratic complicated rules.
OK, so those are my two mental images. I don’t think there’s really a sharp line between these things; some things are in a gray area between them. But plenty of things are clearly one or clearly the other. I know that neuroscientists tend to lump these two things together and call them both “plasticity” or “learning”—and of course lumping them together makes perfect sense if you’re studying the biochemistry of neurotransmitters. But in terms of algorithm-level understanding, they seem to me like different things, and I want to keep them conceptually separate.
Anyway, I have long believed that the hypothalamus has “specific situation-dependent rewiring rules”—probably lots of them. I think the paper you cite is in that category too: it’s like the genome is trying to encode a rule: “if you repeatedly get salt-deprived over the course of your life, start erring on the side of eating extra salt”. (Unless I’m misunderstanding.) I assume the brainstem does that kind of thing too. I’m not aware of there being any “real” learning algorithms in the hypothalamus in the specific sense above.
I think there may actually be some things in the brainstem that seem like “real” learning algorithms, or at least they’re in the gray area. Conditioned Taste Aversion is maybe an example? (I don’t know where the CTA database is stored—I was figuring probably brainstem or hypothalamus. I guess it could be telencephalon but I’d be surprised.) Also I think the superior colliculus can “learn” to adjust saccade targeting and aligning different sensory streams and whatnot. I’m very confused about the details there (how does it learn? what’s the ground truth?) and hope to look into it someday. Of course the cerebellum is a legit learning algorithm too, but I confusingly don’t count the cerebellum as “brainstem” in my idiosyncratic classification. :-P
Anyway, I can certainly imagine “specific situation-dependent rewiring rules” being potentially useful for AI, but I can’t think of any examples off the top of my head. I’m a bit skeptical about “homeostatic drives” for AI, like in the sense of a robot that’s intrinsically motivated to recharge its battery when it’s low. After all, at some point robots will be dangerously powerful and intelligent, and I don’t want them to have any intrinsic motivation beyond “do what my supervisor wants me to do” or “act ethically” or whatever else we come up with. Then it can want to recharge its battery as a means to an end.
You’re welcome to disagree with any or all of that. :-)
Really enjoying your writing. I’m interested in your conception of the hypothalamus as hard-wired and what you make of this: https://journals.physiology.org/doi/full/10.1152/ajpregu.00501.2014
I’m personally interested in the interaction of the hypothalamus with learning processes and this is one I find particularly interesting. Do you think AI agents could benefit from implementing similar homeostatic drives in their learning algorithms?
Thanks for the good question, and interesting reference!
In my defense, I did put in a caveat that it’s hardwired “to a first approximation” :-D
Anyway: When I think about brain plasticity, I kinda have two different mental images.
My first mental image is “learning algorithms”. I think about e.g. modern-day ML, or the brain’s various supervised learning & predictive learning & RL algorithms and so on. In this mental image, I’m imagining rewiring rules that result in a “trained model” algorithm that does something difficult and useful, and the “trained model” is at least moderately complicated, in the sense that an astronomically large number of different trained models could have been built if only the inputs to the learning algorithm were different.
My second mental image is “specific situation-dependent rewiring rules”. My go-to example is the self-modifying code in Linux—e.g. “if debugging is turned off, then replace the debugging-related algorithm steps with no-ops”. In a biological example, imagine that the genome is trying to implement the behavior “if you keep winning fights, start being more aggressive”. (That’s a real example, see here.) It would be like, there’s some specific signal (related to whether you’re winning fights), and this signal changes the strength of some neuronal connection (controlling aggression). So in this mental image, I’m imagining some control knob or whatever that gets adjusted for legible reasons, and I’m imagining a lot of species-specific idiosyncratic complicated rules.
OK, so those are my two mental images. I don’t think there’s really a sharp line between these things; some things are in a gray area between them. But plenty of things are clearly one or clearly the other. I know that neuroscientists tend to lump these two things together and call them both “plasticity” or “learning”—and of course lumping them together makes perfect sense if you’re studying the biochemistry of neurotransmitters. But in terms of algorithm-level understanding, they seem to me like different things, and I want to keep them conceptually separate.
Anyway, I have long believed that the hypothalamus has “specific situation-dependent rewiring rules”—probably lots of them. I think the paper you cite is in that category too: it’s like the genome is trying to encode a rule: “if you repeatedly get salt-deprived over the course of your life, start erring on the side of eating extra salt”. (Unless I’m misunderstanding.) I assume the brainstem does that kind of thing too. I’m not aware of there being any “real” learning algorithms in the hypothalamus in the specific sense above.
I think there may actually be some things in the brainstem that seem like “real” learning algorithms, or at least they’re in the gray area. Conditioned Taste Aversion is maybe an example? (I don’t know where the CTA database is stored—I was figuring probably brainstem or hypothalamus. I guess it could be telencephalon but I’d be surprised.) Also I think the superior colliculus can “learn” to adjust saccade targeting and aligning different sensory streams and whatnot. I’m very confused about the details there (how does it learn? what’s the ground truth?) and hope to look into it someday. Of course the cerebellum is a legit learning algorithm too, but I confusingly don’t count the cerebellum as “brainstem” in my idiosyncratic classification. :-P
Anyway, I can certainly imagine “specific situation-dependent rewiring rules” being potentially useful for AI, but I can’t think of any examples off the top of my head. I’m a bit skeptical about “homeostatic drives” for AI, like in the sense of a robot that’s intrinsically motivated to recharge its battery when it’s low. After all, at some point robots will be dangerously powerful and intelligent, and I don’t want them to have any intrinsic motivation beyond “do what my supervisor wants me to do” or “act ethically” or whatever else we come up with. Then it can want to recharge its battery as a means to an end.
You’re welcome to disagree with any or all of that. :-)