I have been playing a lot of a particular game the past few days and I just had a dream about it, and it got me thinking. It was clearly a dream about the game, but the details were all wrong. I’m having a hard time remembering the particulars but in addition to remixing the game elements, I think other ideas were introduced that had nothing to do with the game, but still had the same visual aesthetic as the game.
Which kind of reminds me of stable diffusion. I wonder if dreaming is just the processing of raw sense memories into concepts and relationships. It’s interesting to me that (according to a brief skim of the Wikipedia page on Dreams which I’m finding a delightful read) we only remember dreams that get interrupted, which seems to imply that it’s not important that we remember our dreams as long as the processing gets done. I find that strange considering how often people credit dreams with flashes of insight or sudden clarity on some topic. Maybe those creative breakthroughs would have occurred to us even if we didn’t remember the dream that inspired them. My experience is that the vast majority of my dreams are complete nonsense, devoid of interesting ideas. But some of my dreams have made their way into my “Unsorted D&D concepts”.
What do you think? How do you think about dreams? Do you think this is nonsense? I’m a lil baked. It’s the holidays y’all and they are not going well. I just want to chat.
Reinforcement learning involves a reward signal (e.g. a score in a game) which I don’t think is present here.
Diffusion models, as you’ve also mentioned, seem a better metaphor. These try to generate a structured image (e.g. matching a prompt) from noise. Perhaps your visual cortex is just trying to make sense of random sensory input while your eyes are closed.
It’s also interesting to think about dreaming in terms of the more general set of representation learning techniques. As I understand it you’re trying to process the day’s experiences and reflect on past memories - sifting through and deciding what to retain and what to forget - essentially mental filing.
You may be interested in Deep dream. This is a program that runs an e.g. convolutional neural network in reverse. Instead of adjusting it’s belief about whether an image should be classified as a dog or not it adjusts a given input image so that it looks more and more like a dog. The results are pretty psychedelic!
Yeah I think I used it incorrectly. I had in mind the notion of adding new data to existing ideas to form more sophisticated versions of them. And I guess dreaming might do that if you have the same type of experience all the time. I think that’s probably why dreams about being at work (at least, my dreams) tend not to include some of the trippier aspects of my other dreams. They usually don’t make sense, but they still feel very mundane and specific. Usually I’m just doing my job the whole time, with varying levels of anxiety about how badly things are going for no clear reason. All that similar input makes the concept very specific.
But you’re right, it’s not really driving toward anything, it’s just semi-random jumbles of ideas and associations.
That image is Trypophobia lol