Needed for Learning or Just Toxin Removal?
Does sleep help problem solving in an informational way aside from refreshing you so you can focus? In 2018, a hypothesis was proposed1Lewis, P.A., Knoblich, G., & Poe, G.R. (2018). How Memory Replay in Sleep Boosts Creative Problem-Solving. Trends in Cognitive Sciences, 22, 491-503.:
Memory replay mechanisms in non-REM can abstract rules from corpuses of learned information, while replay in REM may promote novel associations.
Recently, deep learning blogger Carlos E. Perez took another look at the replay hypothesis via Twitter. He pointed to a recent paper which proposes this extension to replay2Kurth-Nelson, Z., Behrens, T.E., Wayne, G., Miller, K.J., Luettgau, L., Dolan, R., Liu, Y., & Schwartenbeck, P. (2022). Replay and compositional computation. Neuron. arxiv:
replay is able to implement a form of compositional computation where entities are assembled into relationally-bound structures to derive qualitatively new knowledge
This goes beyond prior replay conjectures in that it allows not just repeating the same spatial pathways, but non-spatial pathways that have not occured yet while awake. And then you wake up and have new general concepts about the world.
As Perez tweeted:
I like making the analogy of “defragmentating memory” to that of sleep. Perhaps we aren’t truly learning while we are away; we are just collecting experiences and generating new deductions. It’s only when we are asleep that we conjure up new abstractions and, thus learn.
This particular version of compositional replay depends on certain types of mental representations which are a contested area of cognitive science and philosophy of mind. Even though they propose the hippocampus is a sequence generator to “construct generic content-free sequences into which any content can be slotted” it still seems to depend on traditional notions of content and discrete entities stored in memory.
That doesn’t mean it’s wrong, but it could involve neural structures of some sort that are not necessarily explicit representations or symbols or “content” as thought of in mainstream AI and cognitive science. Representations as a core part of theories of memory has been questioned a lot3Hutto, D.D., & Peeters, A. (2018). The roots of remembering: Radically enactive recollecting. New Directions in the Philosophy of Memory.:
Going the radically enactivist way is to conceive of remembering as a matter of ‘on-the-fly’ construction that can be grounded in structural synaptic changes in the brain as well as other structural changes in the environment without assuming the existence of stored and retrieved contents.
But let’s say we’re ok with mainstream cogsci mental representations. Some would still argue that for neurocognitive systems to have and learn representations requires embodiment, embeddedness and enaction4Piccinini, G. (2022). Situated Neural Representations: Solving the Problems of Content. Frontiers in Neurorobotics, 16.. If so, can semantic construction actually occur at all during sleep?
Why Do We Need Sleep?
Perez is working off the premise that biological entities need sleep and/or dreams:
We know our memories of dreams are not real…Sleep replays our memories in a way that is decoupled from our senses. It is disembodied learning. It is within a context that we have abandoned agency and passively explore memories of the past. But it is more than being a couch potato consuming information.
And:
general intelligence is always involved in sleeping. Aquatic mammals like dolphins and whales need to be always awake. They would drown otherwise. But they achieve this feat by having one hemisphere sleep and the other awake. My conjecture is “Any general intelligence will have sleep as a necessary ingredient of its cognitive function.”
But the main function of sleep might simply be cleaning toxins from brain cells.5Xie, L., Kang, H., Xu, Q., Chen, M., Liao, Y., Thiyagarajan, M., O’Donnell, J., Christensen, D.J., Nicholson, C., Iliff, J.J., Takano, T., Deane, R., & Nedergaard, M. (2013). Sleep Drives Metabolite Clearance from the Adult Brain. Science, 342, 373 – 377. If so, dreaming is the side effect of the decoupling from senses, which in turn has to happen to protect the body hurting itself zombie-style.
That whales evolved to sleep with one half of a brain at a time to prevent drowning (or perhaps more accurately, prevent suffocation) could just be a result of their embodied situation not that “offline” replay is required for intelligence.
Perez replied to me on Twitter: “It could be that the side effect of this cleaning of toxins has no cognitive value. However, I am arguing from the evidence in deep learning that shows replay functionality has value.”
Replay—and the extended form of generating new abstractions—in networks could be the thing…or a thing. But deep learning is based on artificial neural nets and other computer science methods that have very little to do with real biological brain networks. But still maybe this is a general phenomena that applies to both.
I realize memory consolidation theories are popular for the main roles of sleep. But I don’t see why they might not be side effects either. For instance: consolidation / generation functions always running but sleep simply affords them to run faster / bigger.
It would be interesting if a human could develop without the rate and/or type of memory consolidation / generation that supposedly happens in sleep. However, to test that would be difficult: it would require a baby who does not sleep and somehow (perhaps with technology) still cleans the brain cell toxins out.
My subtitle of this post is a bit misleading in that of course there’s a third option: Perhaps in evolutionary history a side effect (at first) of sleep enabled the advancement of certain species that could develop / learn faster.
On Dreams
Another note on this discussion: although we do have memories evoked in dreams, not everything we experience in dreams is from memory.
Dreams are a mishmash of abstractness and concreteness and arbitrary transitions. And importantly—new compositions. Certainly I have dreams with people I’ve never seen in real life before and completely fake environments (although inspired by real environments).
Now you might think: Wait, generation? Doesn’t that support the extended replay hypothesis (composition)?
Not necessarily. Those could be at totally different levels. Dreaming doesn’t mean you’re experiencing the cognitive architectural changes in the “backend.” Decoupling the senses in sleep presumably means there’s a bunch of informational noise or lack of information, leaving the system to essentially hallucinate experience.
In summary, even if compositional replay during sleep is correct, it’s not necessarily dependent on or even reflected in dreams. And it might not use the typical notion of representations.