Last night I had the most horrifying dream I can remember in a long time.  (Won’t recount the content here, since it was very disturbing, and not very interesting apart from that.)

I’ve become more resentful of my dreams lately.  When I was younger, I used to think of the dream world as a space of open possibility, where I could explore ideas and aspects of my psyche too wild or too deeply buried to be available to my waking consciousness.  But the older I get, the less my dreams seem like deep, creative art and the more they seem like tropey, superficial art.

Considered as stories, they’re almost all thrillers, horror stories, or some mixture of the two genres.  They’re plot-driven, almost entirely free of characterization, with no “room to breathe,” no free exploration.  Everything is always focused on some big Problem I am confronting, what a writer would think of as the plot hook.  How can I avert the impending disaster?  How can I complete my task in the face of an implausibly extensive series of obstacles?  What, if anything, should I do in a scary situation I don’t understand?  How can I live with myself, knowing I’ve committed some great sin?

I think my dreams have grown more unpleasant in the last five years or so, and I’ve usually attributed any resentment I feel to this trend.  But I recorded some of my dreams in high school, and looking back at that text file, I notice that although the emotional tone was less aggressively negative back then, the plot devices and structures were the same.  I was trying to do something (often, transport myself to some particular place) and was repeatedly thwarted; I was placed in a baffling situation and mocked for not understanding it; there was always a Problem, and everything else was secondary to it.  Dream logic was an enemy, as it is now, making it impossible to know how to behave appropriately, morally, or effectively.  (In one dream from high school, I was horrified when I accidentally killed a friend’s bird while pet-sitting, then horrified in a different way when, inexplicably, the friend didn’t seem to care.)

I would resent my dreams less if I were more convinced that they were scaring me every night for a good cause.  But from what I have read of the science, dreams seem oddly disposable at best and actively harmful at worst.  Depression is associated with spending more time in REM sleep, and acute sleep deprivation has an antidepressant effect which can be traced specifically to the resulting REM deprivation.  Most antidepressants decrease REM, and MAOIs obliterate it almost completely – which would seem to rule out any theory of REM that makes it vital to human life, since people on MAOIs manage to do just fine (or as fine as they can given that they needed a MAOI to begin with), and never, like, spontaneously combust from a lack of dream-supplied spiritual sustenance or something.

Warning: here is where this post veers off into speculative bullshit, and speculative bullshit that probably over-weights my own dream experiences over others’.

My current pet theory of the function of REM is that it is doing the same thing as experience replay in DeepMind’s reinforcement learning algorithms.  (I make no claim that this is unique to me; I think a lot of people have independently had this idea.)  In experience replay, instead of learning immediately from each experience as it happens, an AI will build up a buffer of past experiences, and at some regular interval will learn from a random selection of those experiences, presented in no particular order.  This is meant to deal with the problem that successive experiences tend to be correlated, which can cause the AI’s parameters to spiral out of control unless the experiences are shuffled (“decorrelated”) first.

“REM sleep is experience replay” fits with a number of notable features of dreaming.  It fits with the way that dreams often use environments and situations from one’s past, but with different time periods jumbled together.  It fits with the tendency toward difficult and confusing situations, particularly those that are frustratingly unpredictable.  (Generally the AIs I’m talking about are either trying to predict what will happen, or just what “reward”/”punishment” value it will have for them, and it is popular to use an “attention” mechanism that preferentially replays frames on which the AI’s predictions were unusually bad.)

It fits with what I’ve heard about REM and memory.  REM seems to be important for certain kinds of memory, which sound like the kinds of memories tied to performance in difficult situations.  From Wikipedia (my emphasis):

REM sleep may favor the preservation of certain types of memories: specifically, procedural memory, spatial memory, and emotional memory. In rats, REM sleep increases following intensive learning, especially several hours after, and sometimes for multiple nights. Experimental REM sleep deprivation has sometimes inhibited memory consolidation, especially regarding complex processes (e.g., how to escape from an elaborate maze).^[92]In humans, the best evidence for REM’s improvement of memory pertains to learning of procedures—new ways of moving the body (such as trampoline jumping), and new techniques of problem solving. REM deprivation seemed to impair declarative (i.e., factual) memory only in more complex cases, such as memories of longer stories.^[93]

“New ways of moving the body” sounds like a case where being able to do simulated replays of your own failures would help.  Less so with declarative memory, where (outside of certain school tests) there is rarely a specific past experience you can replay for practice.  (If you try to learn to juggle, you will accumulate many memories of failing to execute a particular maneuver, but for most of the facts you know, you do not have any memories of failing to remember that specific fact.)

This interpretation puts my own dream frustrations in a disheartening light.  If I were a more ordinary mammal or bird, I would benefit from simulating past challenges.  In my dreams, I would replay memories of (say) chasing prey or being chased by predators, especially those where I made inaccurate predictions; I would learn to do better at these tasks, to beat my own high score.  But in my life as a modern human, most of the challenging and unpredictable situations I face are caused by complex social systems I cannot control or predict (even in principle), or complex interpersonal dramas where the least harmful course of action may be to do nothing at all.  Yet my brain subjects itself to endless remixes of these situations, in a Sisyphean search for a better high score, for a solution that does not exist.