Trying to solve a scientific mystery is like starting a startup: timing is everything. Try to solve a problem too soon, and your efforts are wasted. Try to solve a problem too late, and you can’t contribute anything new. But there’s a sweet spot between when a problem becomes solvable in principle and when it becomes obvious, where efforts have the most leverage. A time when the problem you’re trying to solve still looks dreadfully out of reach, but nothing except inertia and muddled thinking is actually standing in your way. I think the mystery of pain and pleasure (i.e., what’s the intrinsic factor that makes things feel good or bad?) is in this sweet spot right now.
After the fold is a partial excerpt from a paper I’m working on. If it piques your interest, please do contact me.
How to tell if your AI is happy: Toward a substrate-independent theory of valence*
Last updated 8-6-14
The mystery: some brain states feel fantastic. Others feel awful. This is a very real difference… but what physical property makes them different?
Similarly- what do all pleasurable brain activity patterns have in common? And why is activity in certain parts of the brain so strongly associated with pleasure?
These are proper questions, which should have proper answers, but at this point we only have correlative observations about ’hedonic hotspots’ in the brain, and no central causal story about what makes them ‘hedonic’. We’ve called this general phenomenon affect, pain/pleasure, psychological valence, hedonic/eudaimonic tone, and a handful of other terms, but none of these labels illuminate the structure of the issue. This seems to be a chicken-and-egg problem: it’s hard to define pain and pleasure if we can’t measure them, and it’s hard to measure them if we can’t define what they are.
The following is an attempt to chart a new approach to this topic, with the ultimate goal of being able to derive valence— in any substrate— as precisely and objectively as we derive gravity.
II. Things we can say about valence:
1). Let’s assume the puzzle of what intrinsically causes pain and pleasure is a legitimate one, and deserves a conceptually-crisp explanation. A good explanation will be an equation, not an essay.
2). Let’s assume the distinction between physical and emotional valence isn’t crisp- that they’re ultimately the same thing, just presented differently- and they should be lumped together.
3). Let’s assume there’s a difference in kind between valence and most kinds of qualia (seeing a green square, for instance). One involves a sort of emotional and behavioral imperative that the other doesn’t.
4). Let’s agree that valence maps onto some normative dimension. I.e., Some things really do just feel better than others.
5). Behavior and valence are connected but ultimately distinct, so let’s avoid defining valence through behavioral metrics. If an organism seeks out a stimulus, chances are that stimulus will correlate with pleasure, but it’s no sure bet (e.g., see OCD and other compulsive behaviors).
6). The question of what intrinsically generates valence is a distinct question from what makes people happy.
In terms of an SAT-style analogy, I’d offer the following:
What physical conditions intrinsically generate valence? : What makes people happy? :: What physical conditions intrinsically generate an electromagnetic field? : How do you maximize the output of an electric motor?
7. Let’s agree that discussion of ‘hedonic brain regions’ is an awfully leaky level of abstraction and is fundamentally unsatisfying. Instead, let’s focus on explaining valence in information-theoretic terms, e.g. patterns within signals and circuits.
8). Though it may be embedded in a complex context, the ‘raw material’ of valence seems like a pretty simple property (a sheen of goodness or badness), so let’s look for something fairly basic.
9). Let’s remember whatever we’re looking for is simple enough that non-humans with simple brains can have plenty of it. As Dawkins noted, “Pain feels primal, like the ability to see colour or hear sounds. It feels like the sort of sensation you don’t need intellect to experience.”
10). The “unreasonable effectiveness” of elegant physical theories should be considered, and we should hold a preference for simpler, more mathematically elegant theories.
11). Let’s look for a property that has computational significance, since it seems we have a brain in order to compute things.
12). The functional purpose of valence seems to be to influence behavior, both in the present and future: pleasure is a reward and strengthens neural circuits, pain is a punishment that disrupts them and propels a change (Cosmides 2011). So, let’s look for something consistent with this syntax.
13). Let’s look for a theory of valence which allows for mixed states (pain and pleasure at the same time) and the fact that some states just seem much more intense than others.
14). Very intense pain and very intense pleasure seem to be informationally/computationally sparse– the closer to either extreme the brain gets, the less room there seems to be for coherent complexity. Let’s look for a signal property which conforms to this inverse relationship.