On the necessity of sleep

Community Lounge
1

I just watched Jeff Hawkins - Human Brain Project Keynote Screencast and near the end, Jeff said,

“We should be able to build brains that are really tuned for certain types of problems, they never get tired and they’re really good at it.”

I’ve heard this before, that AI never sleeps. But that idea has always bothered me, and now, I’m making a fuss about it :slight_smile:

I think a general AI build like our brains must sleep. I think there’s something about sleep that must be required by the informational structure of our intelligence. In other words, I don’t think we sleep merely because we need to do maintenance on the hardware of the body or brain, rather I think we do maintenance on our software as well - maintenance that changes it’s operation and therefore necessitates the loss of consciousness.

My intuitive theory is that there’s something about sleep that’s informationally fundamental; that it’s not only a result of the particular cyclical circumstance that biological life evolved in (with light and dark cycles), but is somehow an information theoretic requirement of complex, distributed models.

It seems strange that virtually all animals sleep if it were not required to facilitate the maintenance of their minds. Surely it would be an evolutionary advantage to adapt the ability not to sleep, at least for prey animals. But everybody sleeps.

I guess what I’m trying to say is that it seems to me like not every single reason we sleep comes directly from a physical reason; that there’s some kind of informational maintenance reason we have this cycle.

I don’t know what that reason could be, but I’ll try to give a flavor of what I’m talking about:

  • Our minds spend all day making connections that make sense in the moment. We create meaning by forming informational structures contained within a very narrow subset of all possible informational structures: only those structures that make coherent sense with the rest of our mind. Only those structures that fit our narrative, that connect our present to our past and allow us to exist in time.

  • Perhaps as an informational (mental, not physical) homeostatic mechanism, our millions of models inside our minds need to spend some time much less constrained by their neighboring models; they need to stretch their legs, as it were, and explore and thereby recalibrate themselves by acting independently from the group of other models that they are typically associated with during waking life. Perhaps by doing so, they can reconnect with the group as a slightly, by their own metric, improved model. Of course, while all our models are acting independently we lose our overall coherence and therefore our consciousness and must, therefore, sleep.

  • The models inside our head can be viewed as individuals in a group, where each individual is merely a group of smaller model-individuals. If those models act independently the group ceases to exist. Perhaps during sleep, all structures and all boundaries dissolve.

  • Another way to think of it is anthropomorphically; every model that can exist wants to exist, wants a seat at the table, wants a vote, but that means a time must be provided that contradictory models are allowed to express themselves; better to let those arise all at once while the body is shut down than allow them to trickle in during conscious focused attention.

Another reason I think sleep must have a purely informational-management component to it is that people who don’t sleep lose their minds - they don’t simply just get dumber or slower as the buildup of junk proteins in the brain might suggest they should; instead, they lose grasp on reality. It seems as though, far from merely being a hardware cleanup mechanism, sleep safeguards our particular architecture of distributed model fuzzy consensus intelligence from, for lack of a better term, some kind of ‘divergent over-fitting’ tendency that it naturally has: that it’s an informational-homeostatic mechanism for distributed model consensus.

Anyway, I’d be interested in hearing your thoughts on the topic? Does anyone know of any research for or against this kind of idea? Does this idea make sense?

Do you believe, as I do that the more complicated AGI we create the more we will find it effective and efficient to incorporate a time for disillusionment and sleep? Or is our requirement of sleep merely hardware derived? Is it required only by our particular biological hardware, and silicon-based intelligence will not need to experience it?

8 Likes
Thought Experiment / Consciousness & Awareness
2

I have a different take on the need for sleep.

The hippocampus/EC structures are tuned for one-shot episodic memory; the cortex is tuned for slower Hebbian learning.

The difference between these two methods is capacity; the hippocampus can hold about one days worth of memories and the rest of the cortex can hold vastly more. Obviously - evolution has tuned the hippocampus capacity to our daily cycle.

When the hippocampus is full it causes negative effects like hallucinations and is painful to the point where sleep deprivation is used as a form of torture.

One of the optimizations of the hippocampus is that the contents are specific to the most recent events. This means that these contents can be “colored” by the limbic system with the outcomes of the events. I believe that this is where experienced events are given personal relevance, both good and bad. The known connections between the hippocampus and limbic system support this narrative.

During sleep the contents of the hippocampus are used to reinforce the experience of the day. Since the connections are 1:1 with the rest of the cortex a delta in contents would be the the same as highlighting information that has to be “transferred” to the cortex. Multiple presentations of the newly learned contents (with emotional learning mixed in) of the hippocampus would be imposed on the cortex to develop and reinforce long term memory. The well known “spindle” waveform is a form of readout of this difference and are the driving force in this transfer. When the two systems have “equalized” transfer of learning is complete.

12 Likes
Cortical grid cells and Abstract ideas
Thought Experiment / Consciousness & Awareness
Grids into Maps!
Why do members here think DL-based methods can't achieve AGI?
3

What you’ve described is pretty cool!

You see the necessity of sleep (as far as information requirements are concerned) as less of an exploration / calibration process, and more of a data transfer process. Emptying the ram onto the disk, if you will?

I guess, technically you see our requirement of sleep as a function of our biology then? Because were we built on hardware other than the cortex and the hippocampus we’d probably be able to engineer a faster or parallel way to transfer the information? Or is it truly an information-theoretic requirement in your estimation?

3 Likes
4

I would say that you have done a good job of summarizing my ideas.

It seems to be the simplest way to connect up all of what I have read on the topic and related structures.

2 Likes
5

I recently watched and read a few things on the importance of sleep as I was curious about dreams. It seems like, just like many areas of neuroscience these days, that there is a lot of new findings in this area. To me it seems like the sleep and dream are a time of shutdown of consciousness for the purpose of organizing/associating the days experiences in the context of other long term memories.

The researcher Mathew Walker really scared me about not getting enough sleep!

4 Likes
6

I haven’t watched the video yet, but your comment here is exactly the kind of thing I’m talking about. It’s like we associate our daily experiences with whatever information is immediately pertinent to the goals of our conscious attention (usually shorter-term memories). But we need a time and a way to integrate the lessons learned by so doing into a larger context (memories gained further in the past or over a longer time-period). During that time of remote associative processes any narrative about how these associations make sense is tenuous and ever changing, thus, perhaps a loss of consciousness and dreams are required during this time.

Anyway, thanks for your comment, I think it help clarify things for me.

7

Having been messing around with event sourcing as an architectural construct for the recent while, I would think that every day you’re filling up the neurological equivalent of RAM and every night while you are shut down, your RAM image is added to the event stream… when you awake, you are a new you because the event stream that make you what you are has been changed. Now it’s morning and time to make more RAM.

3 Likes
8

This is the essence of what @Bitking suggested I think.

2 Likes
9

To add a little more background on the nature of what is being encoded:

I believe that everything that you learn is framed and parsed in terms of what you have learned before; you learn delta coding based on your prior life experiences. If you extend this all the way back you build on your innate built-in instinctive and emotional wiring as dictated by your genetics. It is nature first, then nurture.

To flesh this out - as you experience the world prior memories are being recalled as part of your predictive process; this is exactly what HTM proposes and there is nothing new in what I am saying. Perception is a process of active recall of prior memories. To the degree that you recognize what you are seeing - there is nothing novel so no learning is being triggered. But nothing is ever really exactly this same - there are always differences. Some examples might be different viewing angles or different contexts.

These differences are triggering your novelty detectors (Surprise!) - the most basic feature of the HTM canon. This novelty triggers learning throughout the hierarchy of connected maps everywhere that novelty is detected. We are always refining our inner models with these new perceptions.

This novelty/surprise based learning seems to be at different rates in the cortex vs the hippocampus. There is considerable research that shows that the learning rate can also be modified by chemicals released by the limbic system. It is very likely that different emotions generate more rapid learning in the hippocampus; perhaps differently based on the type of recognition in the amygdala.

Here is where I will strike out on my own with a proposal based on various hints I have seen in the literature; as the learning occurs I propose that the metabolites accumulate in the cell body, more than likely in the dendrite. This is what I call the ROE (residue of experience) - or a chemical memory independent of direct modifications of the synaptic connection nodes. This is not essential to the proposed process but a possible enhancement. More on this later.

When the brain switches to sleep mode it is time to take the more rapidly/deeply learned new engrams in the hippocampus which are themselves changes (delta) generated when recalling older prior memories in the cortex and push them onto the cortex. This clears these new memories from the hippocampus and prepares the system to start over with a fresh slate when you wake up.

One of the slippery parts of this (to me anyway) is that we remember patterns WHICH INCLUDE FRAGMENTS OF SEQUENCES as part of our memories; we learn transitions. Part of what will be added onto our prior learned patterns will be collections of transitions in the context of older recalled patterns. (I can easily see how patterns can be held and added up because my thinking is a form of 2D pattern manipulation - I don’t inherently see sequences of transitions so this is very hard for me to visualize but it must be happening!)

I have been looking for research that shows that some sorts of memory enhancement chemical are circulated during the dreaming process to enhance the transfer of memories. I do think that I will find them as there are several lines I have seen that indicate that the learning rate is being modified during spindle waveforms. I think that this interacts with the aforementioned “residue of experience” to help modify the generation of new connections.

Even if the ROE theory has no foundation “spindle” brain wave are known to excite the hippocampus in dreaming - driving a recall in both the cortex and hippocampus at the same time. If the hippocampus has leaned a response that makes it fire faster from new learning (using the “standard” spike timing learning ) this could trigger learning in the related cortex. This excitation sweeps through again and again as long as there are significant phase differences in response between the two areas.

Dreams? What to make of them. They are always personal and often filled with all sorts of deep meaning and connections. I am not surprised. As I indicated - your daily learning is based on personal recall of prior memories. Some of this may require that pools of previously un-related material be recalled to match what you are perceiving. As closely related parts of the prior learning are modified in the dreaming state you may be making new connections that “finally” connect partial patterns that have been building up. Part of very old memories may end up being reinforced to the point where they go from almost forgotten to being closer to the surface.

What is the episodic part of our neural mechanism supposed to make of our mental furniture being shifted around and modified while these little bits of new learning is being consolidated? Perceiving these recalled patterns as they are being recalled and reinforced is what we call “dreaming.”

Dreams are made of the stuff of our prior experience being modified on the fly so it is likely to pull up the emotional freighting that is associated with the recall of the original, somewhat unrelated, old memories that are being are pressed into service. As I said - I am not surprised that these are perceived as having deeply personal meanings that you may not be able to recognize as they really are novel synthesis of your prior experiences.

I may come back to this post and tune it up but this is a good start on getting out some ideas that I have been harboring for a long time.

6 Likes
Determinism
Dream's Duration
Cortical grid cells and Abstract ideas
Reinforcement Learning and HTM Algorithm
About memory
ML and Deep Learning to automatically create AGI?
How does short term memory work?
Grids into Maps!
Basic system level diagram of the brain
Efficient dendritic learning as an alternative to synaptic plasticity hypothesis
10

What happens if you use time of day as a physical marker the way you encode space relative to you? Not sure if ToD would be more tethered to time after waking up or time since sunrise… or time before sleep…

Every day you would build up a timed array of mental states with attached emotional measures…

At the end of the day you would store states of mind as desirable, as causal, as consequential, etc.

It’s a funny thought, but when you are thinking, you are loading in partial momentary states of mind and using them to override the system that’s virtualizing input data that combined state of mind then becomes a part of the record of things that happened to you during that part of the day.

You are constantly crafting potential future states of mind that might be useful in already encountered situations.

1 Like
11

I can’t comment on the circadian rhythm part - that seems to me more of a computer thing and less of a biological thing. Computers use clocks and timestamps where biology does not place as much emphasis on strict timing. Sure there are day and night cycles but I don’t think it gets a lot more specific than that.

I am pretty sure that you encode “events” as clusters of related perceptions. The roughly 3 minute window of “here and now” seems to give some framing of human perception for events. These can be chained together with some sort of “elongation mechanism” that I really don’t understand.

Longer windows seem to be coded by some different mechanism - more of a background environment state thing. This is more strongly tied to the same orienting mechanism when you enter a room (remapping) and it becomes your current environment.

1 Like
12

Whatever we need sleeping for, with the right architecture and fast enough hardware it can be implemented in an artificial system without freezing the main runtime.

Dolphins partially (for deep sleep only) solved the problem even using regular biological hardware by letting one half of their brain sleep at a time.

4 Likes
13

Since the perceptions itself each store information about the (likely) next transition(s), events could be encoded as the first memory only, and perhaps its context.

Thinking about the first memory of an event and about the right context would automatically trigger the right transitions which in turn would make us think about the successive events.

Therefore, successions of “scenes” could be encoded as “first scene + its context”. In other words, memories do not need to store a succession of transitions and/or as a succession of perceptions, but only the initial state. The contents of each scene (which are each encoded in its own part of the cortex) will each store the transition to the next one.

2 Likes
14

I’ve thought a bit more about this, and I think sleep may be necessary to bring intelligence back into alignment with itself.

It seems (at least biological) intelligence is necessarily distributed. Thus intelligence is also necessarily a distributed consensus phenomenon. Therefore it seems to me that as we are awake, the different aspects of our brains have a natural tendency to misalign and decay their coherence with one another. This seems like it because different areas of the brain are paying attention to different things.

It seems that sleep, where the whole brain participates in one cycle, is a way of re-aligning the frequencies and cycles of intelligence on every scale such that we don’t go insane.

Anyway, that’s not very specific or scientific, but I hope this flowery language helps someone’s intuition on this topic develop.

3 Likes
15

I don’t see how such a transfer could occur for a few reasons, primarily LTP and LTD timing would not be transferred due to the different frequency of sleep spindles. This difference in frequency would ater the synaptic positioning along the dendrite [1] as a preliminary issue.

[1] https://www.researchgate.net/publication/51714617_Frequency-Dependent_Changes_in_NMDAR-Dependent_Synaptic_Plasticity

Secondly, consider the SDR model whereby the patterning requires a carry over of state of prior inputs for the new learn element to be added. On a singular SDR area/column this may well work, however spindles are out of temporal sequence (in parallel, hence why we halucinate/dream) the issue then becomes a problem with distant dendritic connections which are completely out of sequence and would create incoherent synaptic connections ? (between columns / areas)

Also, when considering a delta only, I can see how LTP could be transferred easily but how would you then transfer the effect of LTD as sleep spindle timing is likely faster (and slower) than the original sensory input ? Too fast and it becomes proximal, too slow it becomes LTD.

In order to perform a transfer as such the delta’s would have to both map the dendritic connection information and relative timing ?

That said I still have no idea as to why we can’t stay awake 24x7, although I really like the idea of it’s due to coherence as that makes sense.

Just a hunch and some ideas… and still learning…

16

There-s mounting evidence that sleep evolved before brains
and it has a much lower level function than “neurons”

We may ask whether brains evolved to adapt their functioning to the general metabolic destruction/reconstruction cycle.

An hypothetical example - short vs long term memory.

Maybe short term memory works by “exhausting” inhibitory synapses with strong localized signals, which allows target neurons to become responsive to the same signal received through normal synapses.

The hypothesis being that it might be faster to “fry” a synapse then let sleep’s restorative physiology to repair it while the neurons selected for more permanent memories are “trained” to store the recent acquired patterns.

PS or even a general process by which synapses are permanently grown then destroyed either by frying via overload or by starving/forgetting

4 Likes
17

This is not a theory looking for a problem to solve, this is an attempt to explain a known process.
Note that the work in this field centers on the details at the cell level. I have been looking at this from the higher level of the nature of the engram patterns and what is special about these patterns to trigger the consolidation mechanisms described in these papers.

There are several high-level constraints that any useful theory has to explain.

  1. I start from the position that the index theory matchs much of the material I have read:
    Max Bennett: An Attempt to Model The Neocortical Microcircuit in Sensory Neocortex - Sept 9, 2020 - #11 by dmac
  2. I refine this with the known fact that you can recall episodes during the day before they are consolidated into the cortex. This tells me that the EC/HC index communication to the cortex is always active.
  3. Add to this that HM and other patients can remember episodes before the loss of the EC/HC complex so episodes can be transferred to the cortex and function there.
  4. The various engrams share the same computing fabric (A palimpsest) so the “readout” of the engrams must be ether sequential (seems unlikely) or a parallel process that transfers the “new” contents from all the engrams at the same time. How do they stay separated at they move from the EC/HC or map-2-map in the cortex? What is special about the newly trained columns that flags those bits to signal to the map on the other end of the fiber bundle?
  5. During the training period (experience of the episode) the same bundles convey experience to the EC/HC to form the index in the first place. These same bundles are bidirectional. It is entirely possible that this is the key to pushing the “new” memories back for consolidation, aligned with where they were experienced from in the first place, with the HC time cells mediating the high-level sequence part of this playback as opposed to the strict sequential replay of the original experience at the lower hierarchical levels. Do remember that the HC has a bunch of specialist cells that do object vectors, direction, distance, orientation, and time - and possibly much more that has yet to be discovered. This all had to be developed in the hierarchy (terminating in the the temporal lobe) to be presented to the the EC/HC as an experience.

Sorry in advance - this one is good but paywalled:
https://www.nature.com/articles/s41593-019-0467-3

3 Likes
18

One thing I don’t get is why consolidation has to be during sleep only. Brain areas go into very similar alpha-wave micro-sleep / spacing-out state every few minutes, I would guess that’s for consolidation too. Although it may not be as effective as spindles, which seem to be unique to actual sleep. Also, HM couldn’t maintain memory for longer than a minute or so, that means HC must be consolidating it all day? Sleep maybe more specific to pruning-out memories.

19

Capturing it all day.

20

I thought you agree that capture happens in the cortex too? Yes, there is that indexing thing in HC, but cortical columns probably also have reference frames. It’s hard to believe that we can’t use those fresh cortical memories until the next day, relying only on HC during the day.