First at all, sorry ! It will be a very long post !
Most of the information can be found at :
which is an excerpt of Stephen LaBerge’s “Lucid Dreaming” (1985).
In this post, i’ll submit two hypotheses about dream function. They are based on the following remarks :
=> dream imagery is produced from random impulses ;
=> neural networks(1) are effective brain simulators.
From these remarks, I will try to find out why random inputs can be useful in a neural network and therefore in brain.
i Neural networks are AI computer programs which attempt to imitate the way a human brain works. A neural network works by creating connections between artificial neurons, the computer equivalent of biological neurons. The organization and weights of the connections determine the output.[/i]
Dream imagery and random impulses :
In 1977, Drs. Allan Hobson and Robert McCarley of Harvard University found that in REM sleep,
Considering the dream functions, we will retain that Hobson and McCarley thought it was the maintaining process of the neurons, activating and testing at regular intervals the brain circuits that underlie our behavior-including cognition and meaning attribution ; and that this test program is essential to normal brain-mind functioning.
In 1983, Nobel Laureate Francis Crick and Graeme Mitchison proposed another function of dream sleep. Their theory was derived from the hypothesis that the cerebral cortex, as a completely interconnected network of neurons,
Thus, the dream function
In brief, it seems that random impulses in REM could be useful in maintaining, testing and correcting the brain neural network.
Comparison with artificial neural networks :
Moreover, we can wonder if we’ll find an equivalent in computer neural networks.
And we’ll find it in NN learning indeed. Learning data must be proposed to the NN randomly. If not, the network can be “traumatized” by the first data and learning fails.
That’s my first hypothesis, and it’s not so far from Hobson’s. We can compare a NN with a river system. Imagine a landscape with hills and valleys. Random water drops on the landscape will follow different ways and irrigate different parts of the landscape. In case of NN trauma, it’s like all the drops always go down in the same river : the output of the NN will be the same, disregarding of the inputs. It’s the “parasitic mode of behaviour”, whose Hobson is talking about.
Let’s suppose a dream where a first random input is elephant, and the brain makes the following associations : elephant, trunk, sex. And for the second random input, cliff : cliff, grotto, vagina, sex. It must be a problem somewhere !
In a dream, while stimuled by random inputs, the brain could be able to recognize such problems and correct them.
Hobson also stresses the problems which could be caused by the modifications produced by experience.
These problems find their equivalent in artificial NN too, and are called catastrophic interferences :
In conclusion, we can formulate the hypothesis that random inputs are useful in artificial neural networks in correcting learning problems due to excessively organized input data causing traumas, and enabling progressive learning.
It could be se same in human brain.
Congrats to all those who read this long post.