sleep

It's weird that we lie down every day to cease our consciousness, and sometimes to hallucinate. There are physiological benefits to sleep --- for example, it allows the body to repair itself in various ways --- but it's also clearly cognitively necessary. A person deprived of sleep for several days becomes dysfunctional and hallucinatory, and can eventually suffer long-term psychiatric damage. So sleep is clearly doing something.

Possibly/probably it's doing several things - there's no reason to expect a single 'right answer'. Sleep itself consists of multiple states, with different effects, and of course evolution will get as much juice out of it as possible.

Some hypotheses:

Learning / practicing in simulation: a dream is a simulated world where we can safely practice a skill, allowing us to fine-tune our policy for handling a given situation.

Amortized inference: as in the wake-sleep algorithm, we can train ourselves to make inferences by simulating experience for which we know the 'true' underlying parameters.

Negative-phase contrastive learning: each of our neurons is learning a generative model of its context, and jointly they are learning a generative model of our experience. There is a danger that individual neurons overfit to properties of the network that don't reflect the world: for example, a network might be wired so that a particular neuron always sees images with some sort of banding as its input, but the banding exists regardless of what's going on in the world. So we need a process to feed in 'negatives' and unlearn any signals associated with them. This is roughly the theory that Geoff Hinton espouses on the Robot Brains podcast. It would explain why not doing this eventually leads to hallucinations as our perceptual system does only positive updates.

Phenomena to be explained: