There was an article over on Nature.com a few days ago entitled "Intellect Thrives on Sleep". It looks into how sleep is very important for mammals, as they tend to go over the things they have just experienced. This supposedly allows them to perform better the next day!

Dreams are also crucial, and I've always personally believed this; the dream allows you to explore otherwise improbable scenarios (falling off a tall building ;) and unconsciously probe your brain's reaction. If one though of the dreams as a search-space, dreams would be right at the edges! :P

I've never thought that dreams were particularly meaningful in a Freudian sense, but I've often wondered just what it is that happens in the brain during sleep.

Presumably some sort of "offline processing" takes place where damaged calls are repaired and perhaps information from the day is either selected as important and remembered or considered trivial and erased.

One of the difficulties in reinforcement learning is being able to propogate the reinforcement signal back in time to associate it with causal events. Perhaps sleep plays a role in this, with events from the previous day being associated together with their likely outcomes.

- Bob

motters,
That's a really good thought. It may be useful in the long run to have a reinforcement learner periodically "sleep" to make those necessary connections that animals readily do. The tricky part would be how to propagate that info, but embedding the "sleeping" aspect into a reinforcement algorithm would be trivial. Again, good idea.

-Rob

Humans and animals don't want to think about everything. They like to sleep; this is surely why evolution has developped the ability to do this while we sleep!!

For computers, this means nothing. They can be made pro-active, ALWAYS computing rewards, understanding scenarios. Then its just a matter of doing more of it, and more efficiently...

My thoughts about sleep were provoked mainly by playing with my Rodney robot. Eventually I'd like to leave the robot running continuously and this necessarily means that during the night (or periods of prolonged darkness) the robot has nothing to do.

The constraints of realtime performance mean that some types of operation which would be computationally expensive aren't possible. This would include operations similar to the "data mining" that companies increasingly to in order to learn more about their customers. The robot could potentially record a lot of visual and other information during the day, but processing this large amount of data in realtime to extract useful patterns would be difficult. Instead it might be better to carry out this sort of analysis while the robot is dormant in its "sleeping" mode.

In animals sleep probably has several purposes, the most obvious one being purely physiological repair of cells and removal of waste products. Sleep also helps to prolong lifespan by slowing metabolism and reducing general wear and tear. Whether sleep has any function in terms of information processing is unknown, but the fact that we (and other animals) dream would suggest that some sort of operation is being carried out. The rapid eye movements which are characteristic of a dreaming state suggests to me that recall of visual memories and physical eye movements are closely interelated.

- Bob

Alex,
I must agree with motters. While computers can be set up to continually re-evaluate their rewards, I feel that it would be useful during down times (ours not the bots) to re-evaluate previously gained info. Or, even as motters suggested, evaluate some of the info for the first time. This could be implemented per a "sleep" mode that is used to reprocess older data or process computationally difficult aspects of the problem.

Yes, I agree there's benefit in thinking about everything later. But lets not get carried away; it's still just the same computation -- albeit delayed.

In humans, the brain works in a different fashion while we sleep; there's very little knowledge about it, but the fundamental difference is conscious and unconscious 'thinking'.

Maybe consciousness is just a trick with time.

- Bob

Hello alex,

Sleeping is essential, why? I don't know. Im sure of one thing, we don't sleep becose we like it, but, becose we need it.

By exemple, a really rare genetic desease cause a person to stop sleep. Consequenses:

1) after a cupple of days, apparition of nevroses.
2) 1 or 2 day after, psychotic pathologies ( symthoms of schizophrenia ).
3) some days after, really hard pathologies, near to die.
4) after 1 week or 2, he die.

The last time I ear of this desease is a person died in Paris. One person on a million or more have this genetic desorder ( and there isn't any cure for this desease ).

So, this demonstrate that we dont sleep for fun, but, for survival.

Salutations,

Fred

Who claimed otherwise? I think too much sleep is bad, but there is a fine line to find!

Though normal people can't die of not sleeping. You will pass out after a while, though it depends how long you can resist! But if you do something psychotic in the later stages, then chances are you might kill yourself :)

We sleep to let our body repair and replenish it's cells. This includes but is not exclusive to the mind.

Take, for example, the cat-nap. You can sleep 15 minutes on the couch and feel refreshed afterwords. Did you dream? Most likely not since most dreaming is started after many hours of sleep. Most of us dream in the morning, because our brain cells have finished repairing and replenishing and have nothing else to do while waiting for the body to finish. (This is just an observation, so please comment.)

I've noticed that after I've had hard runs (ie physically running for exersize) that I need to sleep but start dreaming right away. Another case that the brain isn't tired yet but the body needs rest.

Thoughts?

The links and optimisations between the physical and mental?

Interesting article...
When it comes to AI I definitely would think that 'sleep' would be very important. Why? For the restructuring of memories and the pathways for the 'brain'.

It's rather hard to explain in a forum like this, but I'll try...
Say that the 'mind' is used for memory storage and is separate from the 'brain' which is used for the relaying of signals from sensory nodes to the mind. The structure of the 'mind' itself is not so important for this example, yet the structure of the brain is extremely important and would be the purpose of sleep.

We animals have had eons of evolution to create a brain/mind relation that allows us to control certain functions without thought. A newly created AI doesn't have this advantage and will need to evolve (although it's time for evolution should be minuscule compared to animals). When will it create the 'structure' of a brain that allows for automated (yet controllable) responses without the down time sleep provides?

While sleeping, the mind has the opportunity to 'relive' / 'restructure' the memories accumulated and to restructure the brain itself allowing for higher functionality, effectiveness and speed.

Just like people who take the time to learn advanced reading techniques and visualise what they read rather than seeing the words; or people who study advanced memory techniques and can recall almost any audible/visual memory at will. Many of these people don't rely on sleep to take get the most of these techniques; they rely on meditation/quiet moments of reflection and sleep for them is mostly a physical need. Some also teach themselves to remain mentally alert while sleeping.

The concept of dreaming being used for memory reflections is also arguable by studying and practising lucid dreaming techniques. Rarely do I not dream, most of the time I go directly into a lucid dream when falling asleep and honestly there are allot more things to dream about than the memories I have from recent activities. Those people who remain mentally alert while sleeping sometimes create lucid dreams that are interactive with there current sensory inputs.
then there is the whole field of OOB studies.

The point is that the restructuring of the mind/brain connection is trainable to such an extent that it doesn't need to be a conscious effort and that creates an environment where sleep is purely a physical thing, and dreams/meditation are purely mental and the two only join because it makes sense to use the time your body is 'down' for mental activities.

I think this will hold true for the AI engines of the future where sensory input and reactions need to be effective for real-time response as well as reflective analysis.

/Gonjin