[Intro to brain-like-AGI safety] 5. The “long-term predictor”, and TD learning

post by Steven Byrnes (steve2152) · 2022-02-23T14:44:53.094Z · LW · GW · 27 comments

Contents

  5.1 Post summary / Table of contents
  5.2 Toy model of a “long-term predictor” circuit
    5.2.1 Toy model walkthrough part 1: static context
      5.2.1.1 David-Burns-style exposure therapy—a possible real-life example of the toy model with static context?
    5.2.2 Toy model walkthrough, assuming changing context
  5.3 Value function calculation (TD learning) as a special case of long-term prediction
    5.3.1 Switch (i.e., value = expected next reward) vs summation (i.e., value = expected sum of future rewards)?
    5.3.2 Example
  5.4 An array of long-term predictors involving the extended striatum & Steering Subsystem
  5.5 Five reasons I like this “array of long-term predictors” picture
    5.5.1 It’s a sensible way to implement a biologically-useful capability
    5.5.2 It’s introspectively plausible
    5.5.3 It’s evolutionary plausible
    5.5.4 It’s neuroscientifically plausible
    5.5.5 It offers a nice way to make sense of a wide variety of animal psychology phenomena, including Pavlovian conditioning, devaluation, and more
  5.6 Conclusion
  Changelog
None
27 comments

(Last revised: July 2024. See changelog at the bottom.)

5.1 Post summary / Table of contents

Part of the “Intro to brain-like-AGI safety” post series [? · GW].

In the previous post [AF · GW], I discussed the “short-term predictor”—a circuit which, thanks to a learning algorithm, emits an output that predicts a ground-truth supervisory signal arriving a short time (e.g. a fraction of a second) later.

In this post, I propose that we can take a short-term predictor, wrap it up into a closed loop involving a bit more circuitry, and we wind up with a new module that I call a “long-term predictor”. Just like it sounds, this circuit can make longer-term predictions, e.g. “I’m likely to eat in the next 10 minutes”. This circuit is closely related to Temporal Difference (TD) learning, as we’ll see.

I will argue that there are a large collection of side-by-side long-term predictors in the brain, each comprising a short-term predictor in the Learning Subsystem [LW · GW] (more specifically, in the extended striatum—putamen, nucleus accumbens, lateral septum, part of the amygdala, etc.) that loops down to the Steering Subsystem [LW · GW] (hypothalamus and brainstem) and then back up via a dopamine neuron. These long-term predictors make predictions about biologically-relevant inputs and outputs—for example, one long-term predictor might predict whether I’ll feel pain in my arm, another whether I’ll get goosebumps, another whether I’ll release cortisol, another whether I’ll eat, and so on. Moreover, one of these long-term predictors is essentially a value function for reinforcement learning.

All these predictors will play a major role in motivation—a story which I will finish in the next post [AF · GW].

Table of contents:

• Section 5.2 starts with a toy model of a “long-term predictor” circuit, consisting of the “short-term predictor” of the previous post [AF · GW], plus some extra components, wrapped into a closed loop. Getting a good intuitive understanding of this model will be important going forward, and I will walk through how that model would behave under different circumstances.
• Section 5.3 relates that model to Temporal Difference (TD) learning, which is closely related to a “long-term predictor”. I’ll show two variants of the long-term predictor circuit, a “summation” version (which leads to a value function that approximates the sum of future rewards), and a “switch” version (which leads to a value function that approximates the next reward, whenever it should arrive, which may not be for a long time). The “summation” version is universal in AI literature, but I’ll suggest that the “switch” version is probably closer to what happens in the brain. Incidentally, these two models are equivalent in cases like AlphaGo, wherein reward arrives in a lump sum right at the end of each episode (= each game of Go).
• Section 5.4 will relate long-term predictors to the neuroanatomy of the striatum and brainstem.
• Section 5.5 will offer five lines of evidence that lead me to believe this story: (1) It’s a sensible way to implement a biologically-useful capability; (2) It’s introspectively plausible; (3) It’s evolutionarily plausible; (4) It’s neuroscientifically plausible; (5) It’s compatible with the psychology literature on “Pavlovian conditioning”, “revaluation”, “incentive learning”, and more.

5.2 Toy model of a “long-term predictor” circuit

A “long-term predictor” is ultimately nothing more than a short-term predictor whose output signal helps determine its own supervisory signal. Here’s a toy model of what that can look like:

• The blue box is the short-term predictor of the previous post [AF · GW]. It optimizes its output signal such that it approximates what the supervisor signal will be in 0.3 seconds (as an example).
• The purple box is a 2-way switch. The toggle on the switch is controlled by genetically-hardwired circuitry (gray oval), according to the following rules:
  1. By and large, the switch is in the bottom setting (“defer-to-predictor mode”). This setting is akin to the genetically-hardwired circuitry “trusting” that the short-term predictor’s output is sensible, and in particular producing the suggested amount of digestive enzymes.
  2. If the genetically-hardwired circuitry gets a signal that I’m eating something right now, and that I don’t have adequate digestive enzymes, it flips the switch to “override mode”. Regardless of what the short-term predictor says, it sends the signal to manufacture digestive enzymes.
  3. If the genetically-hardwired circuitry has been asking for digestive enzyme production for an extended period, and there’s still no food being eaten, then it again flips the switch to “override mode”. Regardless of what the short-term predictor says, it sends the signal to stop manufacturing digestive enzymes.

Note: You can assume that all the signals in the diagram can vary continuously across a range of values (as opposed to being discrete on/off signals), with the exception of the signal that toggles the 2-way switch.^[1] In the brain, smoothly-adjustable signals might be created by, for example, rate-coding—i.e., encoding information as the frequency with which a neuron is firing.

5.2.1 Toy model walkthrough part 1: static context

Let’s walk through what would happen in this toy model.^[2] To start with, assume that the “context” is static for some extended period of time. For example, imagine a situation where some ancient worm-like creature is digging in the sandy ocean bed for many consecutive minutes. Plausibly, its sensory environment would stay pretty much constant as long as it keeps digging, as would its thoughts and plans (insofar as this ancient worm-like creature has “thoughts and plans” in the first place). Or if you want another example of (approximately) static context—this one involving a human rather than a worm—hang on until the next subsection.

In the static-context case, let’s first consider what happens when the switch is sitting in “defer-to-predictor mode”: Since the output is looping right back to the supervisor, there is no error in the supervised learning module. The predictions are correct. The synapses aren’t changing. Even if this situation is very common, it has no bearing on how the short-term predictor eventually winds up behaving.

The times that do matter for the eventual behavior of the short-term predictor are those rare times that we go into “override mode”. Think of the overrides as like a sporadic “injection of ground truth”. They produce an error signal in the short-term predictor’s learning algorithm, changing its adjustable parameters (e.g. synapse strengths).

After enough life experience (a.k.a. “training” in ML terminology), the short-term predictor should have the property that the overrides balance out. There may still be occasional overrides that increase digestive-enzyme production, and there may still be occasional overrides that decrease digestive-enzyme production, but those two types of overrides should happen with similar frequency. After all, if they didn’t balance out, the short-term predictor’s internal learning algorithm would gradually change its parameters so that they did balance out.

And that’s just what we want! We’ll wind up with appropriate digestive enzyme production at appropriate times, in a way that properly accounts for any information available in the context data—what the animal is doing right now, what it’s planning to do in the future, what its current sensory inputs are, etc.

5.2.1.1 David-Burns-style exposure therapy—a possible real-life example of the toy model with static context?

As it happens, I recently read David Burns’s book Feeling Great (my review [LW · GW]). David Burns has a very interesting approach to exposure therapy—an approach that happens to serve as an excellent example of how my toy model works in the static-context situation!

Here’s the short version. (Warning: If you’re thinking of doing exposure therapy on yourself at home, at least read the whole book first!) Excerpt from the book:

For example, when I was in high school, I wanted to be on the stage crew of Brigadoon, a play my school was putting on, but it required overcoming my fear of heights since the stage crew had to climb ladders and work near the ceiling to adjust the lights and curtains. My drama teacher, Mr. Krishak, helped me overcome this fear with the very type of exposure techniques I’m talking about. He led me to the theater and put a tall ladder in the middle of the stage, where there was nothing nearby to grab or hold on to. He told me all I had to do was stand on the top of the ladder until my fear disappeared. He reassured me that he’d stand on the floor next to me and wait.

I began climbing the ladder, step by step, and became more and more frightened. When I got to the top, I was terrified. My eyes were almost 18 feet from the floor, since the ladder was 12 feet tall, and I was just over 6 feet tall. I told Mr. Krishak I was in a panic and asked what I should do. Was there something I should say, do, or think about to make my anxiety go away? He shook his head and told me to just stand there until I was cured.

I continued to stand there in terror for about ten more minutes. When I told Mr. Krishak I was still in a panic, he assured me that I was doing great and that I should just stand there a few more minutes until my anxiety went away. A few minutes later, my anxiety suddenly disappeared. I couldn’t believe it!

I told him, “Hey, Mr. Krishak, I’m cured now!”

He said, “Great, you can come on down from the ladder now, and you can be on the stage crew of Brigadoon!”

I had a blast working on the stage crew. I absolutely loved climbing ladders and adjusting the lights and curtains near the ceiling, and I couldn’t even remember why or how I’d been so afraid of heights.

This story seems to be beautifully consistent with my toy model here. David started the day in a state where his short-term-predictors output “extremely strong fear reactions” when he was up high. As long as David stayed up on the ladder, those fear-reaction short-term-predictors kept on getting the same context data, and therefore they kept on firing their outputs at full strength. And David just kept feeling terrified.

Then, after 15 boring-yet-terrifying minutes on the ladder, some innate circuit in David’s brainstem issued an override—as if to say, “C’mon, nothing is changing, nothing is happening, we can’t just keep burning all these calories all day. It’s time to calm down now.” The short-term-predictors continued sending the same outputs as before, but the brainstem exercised its veto power, and forcibly reset David’s cortisol, heart-rate, etc., back to baseline. This “override” state immediately created error signals in the relevant short-term-predictors in David’s amygdala! And the error signals, in turn, led to model updates! The short-term predictors were all edited, and from then on, David was no longer afraid of heights.

This story kinda feels like speculation piled on top of speculation, but whatever, I happen to think it’s right. If nothing else, it’s good pedagogy! Here’s the diagram for this situation; make sure you can follow all the steps.

5.2.2 Toy model walkthrough, assuming changing context

The previous subsections assumed static context lines (constant sensory environment, constant behaviors, constant thoughts and plans, etc.). What happens if the context is not static?

If the context lines are changing, then it’s no longer true that learning happens only at “overrides”. If context changes in the absence of “overrides”, it will result in changing of the output, and the new output will be treated as ground truth for what the old output should have been. Again, this seems to be just what we want: if we learned something new and relevant in the last second, then our current expectation should be more accurate than our previous expectation, and thus we have a sound basis for updating our models.

5.3 Value function calculation (TD learning) as a special case of long-term prediction

At this point, ML experts may recognize a resemblance to Temporal Difference (TD) learning. It’s not quite the same, though. The differences are:

First, TD learning is usually used in reinforcement learning (RL) as a method for going from a reward function to a value function. By contrast, I was talking about things like “digestive enzyme production”, which are neither rewards nor values.

In other words, there is a generally-useful motif that involves going from some immediate quantity X to “long term expectation of X”. The calculation of a value function from a reward function is an example of that motif, but it’s not the only useful example.

(As a matter of terminology, it seems to be generally accepted that the term “TD learning” can in fact apply to things that are not RL value functions.^[3] However, empirically in my own experience, as soon as I mention “TD learning”, the people I’m talking to immediately assume I must be talking about RL value functions. So I want to be clear here.)

Second, to get something closer to traditional TD learning, we’d need to replace the 2-way switch with a 2-way summation—and then the “overrides” would be analogous to rewards. Much more on “switch vs summation” in the next subsection.

 

Third, there are many additional ways to tweak the circuit which are frequently used in AI textbooks, and some of those may be involved in the brain circuits too. For example, we can put in time-discounting, or different emphases on false-positives vs false-negatives (see my discussion of distributional learning in a footnote of the previous post [LW · GW]), etc.

To keep things simple, I will be ignoring all these possibilities (including time-discounting) in the discussion below.

5.3.1 Switch (i.e., value = expected next reward) vs summation (i.e., value = expected sum of future rewards)?

The figures above show two variants of our toy model. In one, the purple box is a two-way switch between “defer to the short-term predictor” and some independent “ground truth”. In the other, the purple box is a two-way summation instead.

The switch version trains the short-term-predictor to predict the next ground truth, whenever it should arrive.

The summation version trains the short-term-predictor to predict the sum of future ground truth signals.

The correct answer could also be “something in between switch and summation”. Or it could even be “none of the above”.

RL papers universally use the summation version—i.e., “value is the expected sum of future rewards”. What about biology? And which is actually better?

It doesn’t always matter! Consider AlphaGo. Like every RL paper today, AlphaGo was originally formulated in the summation paradigm. But it happens to have one and only one nonzero reward signal per game, namely +1 at the end of the game if it wins, or -1 if it loses. In that case, switch vs summation makes no difference. The only difference is one of terminology:

• In the summation case, we would say “each non-terminal move in the Go game has reward=0”.
• In the switch case, we would say “each non-terminal move in the Go game has a reward of (null)”.

(Do you see why?)

But in other cases, it does matter. So back to the question: should it be switch or summation?

Let’s step back. What are we trying to do here?

One thing that a brain needs to do is make decisions that weigh cross-domain tradeoffs. If you’re a human, you need to decide whether to watch TV or go to the gym. If you’re some ancient worm-like creature, you need to “decide” whether to dig or to swim. Either way, this “decision” impacts energy balance, salt balance, probability of injury, probability of mating—you name it. The design goal in the decision-making algorithm is that you make the decision that maximizes inclusive genetic fitness. How might that goal be best realized?

One method involves building a value function that estimates the organism’s inclusive genetic fitness (compared to some arbitrary—indeed, possibly time-varying—baseline), conditional on continuing to execute a given course of action. Of course it won’t be a perfect estimate—real inclusive genetic fitness can only be calculated in hindsight, many generations after the fact. But once we have such a value function, however imperfect, we can plug it into an algorithm that makes decisions to maximize value (more on this in the next post [AF · GW]), and thus we get approximately-fitness-maximizing behavior.

So having a value function is key for making good decisions that weigh cross-domain tradeoffs. But nowhere in this story is the claim “value is the expectation of a sum of future rewards”! That’s a particular way of setting up the value-approximating algorithm, a method which might or might not be well suited to the situation at hand.

I happen to think that brains use something closer to the switch circuit, not the summation circuit, not only for homeostatic-type predictions (like the digestive enzymes example above), but also for value functions, contrary to mainstream RL papers. Again, I think it’s really “neither of the above” in all cases; just that it’s closer to switch.

Why do I favor “switch” over “summation”?

An example: sometimes I stub my toe and it hurts for 20 seconds; other times I stub my toe and it hurts for 40 seconds. But I don’t think of the latter as twice as bad as the former. In fact, even five minutes later, I wouldn’t remember which is which. (See the peak-end rule.) This is the kind of thing I would naturally expect from switch, but is an awkward fit for summation. It’s not strictly incompatible with summation; it just requires a more complicated, value-dependent reward function. As a matter of fact, if we allow the reward function to depend on value, then switch and summation can imitate each other.

Anyway, in upcoming posts, I’ll be assuming switch, not summation. I don’t think it matters very much for the big picture. I definitely don’t think it’s part of the “secret sauce” of animal intelligence, or anything like that. But it does affect some of the detailed descriptions.

The next post [AF · GW] will include more details of reinforcement learning in the brain, including how “reward prediction error” works and so on. I am bracing for lots of confused readers, who will be disoriented by the fact that I’m assuming a different relationship between value and reward than what everyone is used to. For example, in my picture, “reward” is a synonym for “ground truth for what the value function should be right now”—both should account for not only the organism’s current circumstances but also its future prospects. Sorry in advance for any confusion! I will do my best to be clear.

5.3.2 Example

After all that abstract discussion, I should probably go through a more concrete example illustrating how I think the long-term predictor works in the domain of reinforcement learning.

Suppose I’m downstairs and feeling cold. An idea pops into my head: “I’ll go put on my cozy sweater”. So I walk upstairs and put it on.

What just happened? Here’s a diagram:

This is the same kind of diagram as Section 5.2 above, but instead of predicting a digestive enzyme signal, the blue box is predicting a special signal I call valence. This circuit is closely related to the value function and reward function of actor-critic reinforcement learning, as above (and more in the next post, along with my five-post Valence series [LW · GW]).

Here the “override mode” is the pleasure from putting on the cozy sweater, which can be grounded in innate brainstem circuitry related to warmth, tactile feelings, and so on. At that point, when the sweater is on, the brainstem “knows” that something good has happened.

But the more interesting question is: what was happening during the thirty seconds that it took me to walk upstairs? I evidently had motivation to continue walking, or I would have stopped and turned around. But my brainstem hadn’t gotten any ground truth yet that there were good things happening. That’s where “defer-to-predictor mode” comes in! The brainstem, lacking strong evidence about what’s happening, sees a positive valence guess coming out of the striatum and says, in effect, “OK, sure, whatever, I’ll take your word for it.”

And now we’re in the static-context situation discussed above (Section 5.2.1), and therefore the short-term predictor will (after some life experience) output a good guess for the next override valence, whenever that may occur. And since the cozy sweater will be a positive valence override, the valence guesses preceding it will likewise be positive. And therefore I will be motivated to get the sweater, including spending 30 seconds walking up the stairs.

“…Hang on,” you might be thinking. “How was that an example of static context? First you were on the ground floor, and then you were walking up the stairs, then opening the drawer, and so on. That’s changing context, right?? Section 5.2.2, not 5.2.1!”

My response: Yes, part of the context is changing, such as my immediate sensory inputs. But I claim the most important part of the context data is not in fact changing, and that’s the idea, in my head, that I’m on my way to put on a sweater. That idea is in my head the whole time, from conceiving the idea, through walking up the stairs, right through the moment when I’m pulling on my cozy sweater. And that idea is represented by some set of cortical neurons, just like everything else in my conscious awareness, and thus it’s included among the context signals.

This “static-ish context” is critical for how the learning algorithm can actually work in practice. TD learning doesn’t necessarily need to laboriously walk back the credit assignment 0.1 seconds at a time, all the way back from the pleasure of the sweater to the bottom of the stairs. It can just do one step, from the pleasure of the sweater to the idea in my head that I’m putting on a sweater.

5.4 An array of long-term predictors involving the extended striatum & Steering Subsystem

Here’s the long-term-predictor circuit from above:

I can lump together the 2-way switch with the rest of the genetically-hardwired circuitry, and then rearrange the boxes a bit, and I get the following:

Now, obviously digestive enzymes are just one example. Let’s draw in some more examples, add some hypothesized neuroanatomy, and include other terminology. Here’s the result:

Excellent! We’re halfway to my big picture of decision-making and motivation. The rest of the picture—including the “actor” part of actor-critic reinforcement learning—will come in the next post [AF · GW], and will fill in the hole in the top-left side of that diagram. (The term “Steering Subsystem” comes from Post #3 [AF · GW].)

Here’s one more diagram and caption for pedagogical purposes.

In the next two subsections, I will elaborate on the neuroanatomy which I’m hinting at in this diagram, and then I’ll talk about why you should believe me.

5.5 Five reasons I like this “array of long-term predictors” picture

5.5.1 It’s a sensible way to implement a biologically-useful capability

If you start producing digestive enzymes before eating, you’ll digest faster. If your heart starts racing before you see the lion, then your muscles will be primed and ready to go when you do see the lion. Etc.

So these kinds of predictors seem obviously useful.

Moreover, as discussed in the previous post (Section 4.5.2) [AF · GW], the technique I’m proposing here (based on supervised learning) seems either superior to or complementary with other ways to meet these needs.

5.5.2 It’s introspectively plausible

For one thing, we do in fact start salivating before we eat the cracker, start feeling nervous before we see the lion, etc.

For another thing, consider the fact that all the actions I’m talking about in this post are involuntary: you cannot salivate on command, or dilate your pupils on command, etc., at least not in quite the same way that you can wiggle your thumb on command.

(More on voluntary actions in the next post [AF · GW], when I talk more about the cortex.)

I’m glossing over a bunch of complications here, but the involuntary nature of these things seems pleasingly consistent with the idea that they are being trained by their own dedicated supervisory signals, straight from the brainstem. They’re slaves to a different master, so to speak. We can kinda trick them into behaving in certain ways, but our control is limited and indirect (see Section 6.3.3 of the next post [LW · GW]).

5.5.3 It’s evolutionary plausible

As discussed in Section 4.4 of the previous post [AF · GW], the simplest short-term predictor is extraordinarily simple, and the simplest long-term predictor is only a bit more complicated than that. And these very simple versions are already plausibly fitness-enhancing, even in very simple animals.

Moreover, as I discussed a while back (Dopamine-supervised learning in mammals & fruit flies [LW · GW]), there is an array of little learning modules in the fruit fly, playing a seemingly-similar role to what I’m talking about here. Those modules also use dopamine as a supervisory signal, and there is some genomic evidence of a homology between those circuits and the mammalian telencephalon.

5.5.4 It’s neuroscientifically plausible

The details here are out-of-scope, and somewhat more complicated than I’m making it out to be, but for (a major subset of) Thought Assessors, I believe the correspondence between my toy models and real neuroanatomy is as follows:

• The “short-term predictor” is a set of medium spiny neurons (a.k.a. spiny projection neurons) of the extended striatum (caudate, nucleus accumbens, lateral septum, part of the amygdala, etc.)
  • Most of the extended striatum is dedicated to just a few Thought Assessors, namely the “valence guess” Thought Assessor (part of the brain’s “main” / “success-in-life” RL system), or the analogous Thought Assessors for any of the brain’s narrow-RL systems (those are off-topic for this series but see Section 1.5.6 of my valence series [LW · GW]).
  • A smaller portion of the extended striatum—specifically, parts of the amygdala, lateral septum, and nucleus accumbens shell—comprises all the dozens-to-hundreds of “visceral” Thought Assessors, related to things like cortisol, freezing, immune system activation, and so on.
• The “Output” signal from the short-term predictor (a.k.a. the “guess”, or “scorecard” entry) goes to one of various Steering Subsystem areas (e.g. GPi, GPe, SNr, SNc, various cell groups in the hypothalamus).
• The “Supervisor” signal going into the short-term predictor is one or more dopamine neurons coming back (directly or indirectly) from the Steering Subsystem.^[4]

(Caveat: I say “a major subset of Thought Assessors” are medium spiny neurons in the extended striatum, rather than “all Thought Assessors”, because I think the cortex learning algorithm has different loss functions in its different subregions, and thus I think that certain bits of the cortex are in some respects more like Thought Assessors than like parts of the cortex “Thought Generator” which I’ll be talking about at length in the next post. I’m leaving this caveat out of my diagrams in order to (over)simplify the discussion.)

Thus, for example, my claim is that if you zoom into (say) the central amygdala, you’ll find dozens of little genetically-determined subregions, each projecting to a specific set of genetically-determined Steering Subsystem targets, and with a genetically-determined narrowly-targeted supervisory signal coming back (directly or indirectly) from often the same Steering Subsystem target. Thus, for example, maybe thus-and-such little group of amygdala neurons is universally associated with getting goosebumps. But each individual brain would learn within its lifetime a different set of patterns that trigger this group of goosebumps-related neurons.

Is that hypothesis consistent with the experimental evidence? My impression so far, from everything I’ve read, is “yes”, although I admit I don’t have super-solid evidence either way. More specifically:

For the valence-guess Thought Assessor, the gory details are out-of-scope, but I think my story is elegantly compatible with textbook basal ganglia circuitry, particularly the famous motif of “cortico-basal ganglia-thalamo-cortical loops”.

For the “visceral” Thought Assessors: On general priors, there’s certainly no question that the genome is capable of wiring up dozens-to-hundreds of little cell groups in dozens-to-hundreds of specific and innate ways—this is ubiquitous in the Steering Subsystem. What about more specifically? I’ve found a few things. As one small-scale example, Lischinsky et al. 2023 studied two little embryonically-defined amygdala subpopulations, and finds that these two subpopulations wind up triggering in very different circumstances and connecting to very different downstream cells. As another example, at a larger scale, Heimer et al. 2008 says that the bundle of outputs of (what I’m calling) the “visceral” Thought Assessors are just an horrific mess, going every which way, as would be expected from hundreds of cell groups projecting to hundreds of different Steering Subsystem targets:

In the case of some macrosystem outputs, such as, for example, those from the accumbens core, the fibers course among these structures in tightly fasciculated fashion, giving off robust “bursts” of terminations only within or in the vicinity of particular structures. In other cases, good examples being the accumbens shell and extended amygdaloid outputs, the fibers are more loosely, if at all, fasciculated and have countless varicosities and short collaterals presumed to be sites of synaptic or parasynaptic transmission that involve not only defined structures but also points of indeterminate neural organization all along the course of the medial forebrain bundle. … As a consequence of their descent into this poorly defined organization, it is exceedingly difficult to conceive of precise, definitive mechanisms that outputs from the forebrain, including from the macrosystems, might engage to orchestrate purposeful, adaptive behavior from the extensive and sophisticated repertoire of autonomous, albeit restrictively programmed, hindbrain-spinal motor routines. But it is equivalently hard to dispel the notion that this is precisely what happens, and it seems advisable to presume that we are at present unable to perceive the relevant functional-anatomical relationships rather than to think they don’t exist.

As another piece of suggestive evidence, Lammel et al. 2014 mentions so-called “‘non-conventional’ VTA [dopamine] neurons” in “medial posterior VTA (PN and medial PBP)”. These seem to project to roughly the “visceral” Thought Assessor areas that I mentioned above, and it’s claimed that they have different firing patterns from other dopamine neurons. That seems intriguing, although I don’t have a more specific story than that right now. Relatedly, a recent paper discusses certain dopamine neurons which burst when aversive things happen (Verharen et al. 2020), and four of the five regions these neurons seem to "supervise" are associated with visceral Thought Assessors, in my opinion.^[5] This makes sense because if the mouse freezes in terror, then that’s a positive error signal for the freeze-in-terror visceral Thought Assessor, as well as the raise-your-cortisol-level visceral Thought Assessor, and so on.

5.5.5 It offers a nice way to make sense of a wide variety of animal psychology phenomena, including Pavlovian conditioning, devaluation, and more

See my later post Incentive Learning vs Dead Sea Salt Experiment [LW · GW] for details on that. See also the last section of that post [LW · GW] for some contrasts between my model and a couple others in the psychology and neuro-AI literature.

5.6 Conclusion

Anyway, as usual I don’t pretend to have smoking-gun proof of my hypothesis (i.e. that the brain has an array of long-term predictors involving extended striatum-Steering Subsystem loops), and there are some bits that I’m still confused about. But considering the evidence in the previous subsection, and everything else I’ve read, I wind up feeling strongly that I’m broadly on the right track. I’m happy to discuss more in the comments. Otherwise, onward to the next post [AF · GW], where we will finally put everything together into a big picture of how I think motivation and decision-making work in the brain!

Changelog

July 2024: Since the initial version, I’ve made a bunch of changes.

The biggest changes are related to neuroanatomy. I was just really confused about lots of things in 2022, but I hope I’m converging towards correct answers! In particular:

I’ve gone through several iterations of the diagrams in this post, particularly including which neuroanatomy corresponds to which boxes.

There used to be a much longer discussion of neuroanatomy in Section 5.4, but much of that was incorrect, and the rest got moved into the newly-rewritten Section 5.5.4.

I originally had a discussion of dopamine diversity. I mostly deleted that, apart from a couple sentences, having been convinced that at least one of the pieces of evidence I discussed (Engelhart et al. 2019) was not related to “visceral Thought Assessors”, but rather had a different explanation. (Thanks Nathaniel Daw for the correction.) I also moved the discussion of distributional RL into a footnote of the previous post.

More minor things:

I added a note that all the diagrams in this post should be understood as the kinds of diagrams you see in FPGA design—i.e. at every instant, every block is running simultaneously, and every line (wire) is carrying a numerical value. It’s not like the kinds of diagrams you see in descriptions of serial program control flow, where we follow a path through the diagram, executing one thing at a time.

I’ve also changed the wording on the labels of the various diagrams a few times since the initial version. For example, where it used to say “Will lead to reward?” and “It was in fact leading to reward”, it now says “Valence guess” and “Actual valence”, respectively. The previous wording was OK, but I think the connotations are a bit better here, and also more consistent with my later Valence series [LW · GW].

I added a second concrete example story (involving a cozy sweater) as Section 5.3.2.

I replaced a brief discussion of the “dead sea salt experiment” with a link to my more recent post dedicated to that topic [LW · GW], in Section 5.5.5.

1. ^{^}

   To be clear, in reality, there probably isn’t a discrete all-or-nothing 2-way switch here. There could be a “weighted average” setting, for example. Remember, this whole discussion is just a pedagogical “toy model”; I expect that reality is more complicated in various respects.

2. ^{^}

   I note that I’m just running through this algorithm in my head; I haven’t simulated it. I’m optimistic that I didn’t majorly screw up, i.e. that everything I’m saying about the algorithm is qualitatively true, or at least can be qualitatively true with appropriate parameter settings and perhaps other minor tweaks. 

3. ^{^}

   Examples of using the terminology “TD learning” for something which is not related to RL reward functions include “TD networks”, and the Successor Representations literature (example), or this paper, etc.

4. ^{^}

   As in the previous post, when I say that “dopamine carries the supervisory signal”, I’m open to the possibility that dopamine is actually a closely-related signal like the error signal, or the negative error signal, or the negative supervisory signal. It really doesn’t matter for present purposes.

5. ^{^}

   The fifth area where that paper found dopamine neurons bursting under aversive circumstances, namely the tail of the striatum, has a different explanation I think—see here [AF · GW].

27 comments

Comments sorted by top scores.