MJD 59,436

This entry is part 18 of 21 in the series Captain's Log

A week ago, for the first time in decades, I spent several days in a row doing many hours of hands-on engineering work in a CAD tool (for my rover project), and noticed that I had particularly vivid dreams on those nights. My sleep data from my Whoop strap confirmed that I’d had higher REM sleep on those nights.

These were dreams that lacked narrative and emotional texture, but involved a lot of logistics. For example, one dream I took note of involved coordinating a trip with multiple people, with flights randomly canceled. When I shared this on Twitter, several people replied to say that they too had similar dreams after days of intense but relatively routine information work. A couple of people mentioned dreaming of Tetris after playing a lot, something I too have experienced. High-REM dreaming sleep seems to be involved in integrating cognitive skill memories. This paper by Erik Hoel, The Overfitted Brain: Dreams evolved to assist generalization, pointed out by @nosilver, argues that dreaming is about mitigating overfitting of learning experiences, a problem also encountered in deep learning. This tracks for me. It sounds reasonable that my “logistics” dreams were attempts to generalize some CAD skills to other problems with similar reasoning needs. REM sleep is like the model training phase of deep learning.

Dreams 2×2

This got me interested in the idea of tapping into the unconscious towards pragmatic ends. For example, using the type of dreams you have to do feedback regulation of the work you do during the day. I made up a half-assed hypothesis: the type of dream you have at night depends on 2 variables relating to the corresponding daytime activity — the level of conflict in the activity (low/high) and whether or not the learning loop length is longer or shorter than 24 hours. If it is shorter, you can complete multiple loops in a day and the night-time dream will need to generalize from multiple examples. If it’s less than a complete loop, there is no immediate resolution, so instead you get more dreams that integrate experiences in an open-loop way, using narrative-based explanations. If there’s no conflict and no closed loops, you get low dreaming. There’s nothing to integrate, and you didn’t do much thinking that day (which for me tends to translate to poor sleep). I made up the 2×2 above for this idea.

I have no idea whether this particular 2×2 makes any sense, but it is interesting that such phenomenology lends itself to apprehension in such frameworks at all. I rarely remember dreams, but I think even I could maintain a dream journal based on this scheme, and try to modulate my days based on my nights.

This also helps explain why people in similar situations might have similar dreams (such as all the “corona dreams” that many of us were having during early lockdown months). It also lends substance to the narrative conceits of stories like H. P. Lovecraft’s Call of Cthulhu, which begins with people having widespread similar dreams (where it turns into science fiction is in the specificity of the shared motifs and symbols that appear).

You don’t need to buy into dubious Jungian or Freudian theories of individual and collective dreaming to think about this stuff in robust ways. The development of deep learning, in particular, offers us a much more robust handle on this phenomenology. Dreams are perhaps our journeys into our private latent spaces, undertaken for entirely practical purposes like cleaning up our messy daytime learning (there’s other theories of dreaming too of course, like David Eagleman’s theory that we dream to prevent the visual centers from getting colonized by other parts of the brain at night, but we’re only hypothesizing contributing causes, not determinative theories).

Series Navigation<< MJD 59,396MJD 59,459 >>

Get Ribbonfarm in your inbox

Get new post updates by email

New post updates are sent out once a week

About Venkatesh Rao

Venkat is the founder and editor-in-chief of ribbonfarm. Follow him on Twitter

Comments

  1. During my student years I sometimes discussed the paranormal with friends and we shared the opinion that it had dreamlike qualities – dreams let loose – which blended into objective reality without prevailing. “The world is deep, and deeper than day had thought” as Nietzsche said.

    As it seems, Deep Learning suggests yet another epistemological framework which allows for an accommodation to weird realism. Sure, Lovecraft warned his readers that it won’t be fun and there are modes of madness which are worse than death. We just don’t believe enough in the old ones and their projections and poisons to be too scared. Not moving too far into this direction and prohibiting dreams from infesting the real might serve as a protective layer – for the time being.

  2. Mark Solms’ book “The Hidden Spring: A Journey to the source of consciousness” posits that dreams are inherently exploratory. If you pick up a skill during the day, for instance, you want to keep exploring outcomes of using that skill in different ways and different situations. This is slightly different from avoiding overfitting, but ties in well with your theory of learning loops.

  3. I did not understand the idea of ‘learning loop length’. Can someone explain this with the help of an example or two. Thanks.

    • How long it takes you to do a trial-and-error cycle. Eg print a design to see if it fits, then redesign and reprint if necessary.

    • Same with me.
      “whether or not the learning loop length is longer or shorter than 24 hours. If it is shorter, you can complete multiple loops in a day and the night-time dream will need to generalize from multiple examples. If it’s less than a complete loop, there is no immediate resolution, so instead you get more dreams that integrate experiences in an open-loop way” –> If I understand this right, the loop should happen within 1 day (and before one goes to sleep). If so, it cannot be 24 hours. Something like 12-16 hours?
      May be I am missing the whole idea.

  4. Once in undergrad, I was working on a math proof for a couple hours, took a dream laden nap and the solution came to me.

  5. I am sorry for being a psychotic liar. Being psychotic is pretty much dreaming while you’re awake. As if that will solve the problem. Cringe. The bad scary things you say about people are often true about me. You’re better than tv.