Creating both artistic and scientific, suppose that the creator is endowed with sufficient autonomy to enable it to overcome its cultural achievements,
the artificial creation of autonomous systems therefore requires acting independently of their designers.
The word is a neologism connecvolution built on the two terms of connectionism and evolutionism, for the first, thanks to neural networks,
an ontogenetic adaptation phenotypes, and the second producing, using genetic algorithms, the conditions of 'phylogenetic evolution of a population of genotypes.
By evolving populations genetically (neural networks or robots) it is possible to select those best that verifie conditions.
More generally, the theory of neural Darwinism reflects the organization and evolution of the neural networks themselves.
By relying on these considerations, I propose to develop a system for artificial creation which should memorize and categorize perceptions
of the world and then recognize its own actions.
More specifically, it would be to build a system that many entry artistic accomplishments and categorize
would be capable of acting on the world by creating itself from other works by genetic evolution categories
and that it would perceive.
Reentrant structures neural networks could adequately model such a process loop, while Darwinian
selection process should be insure a no supervised evolution potentially innovative.
Concepts as important as artificial creation, artificial emotion and consciousness are thus raised,
leading to requestionner their natural counterparts.
I propose a method of constructing an artificial creation system that observes its environment by
categorizing perceptions using unsupervised learning.
Thus every perception is a certain state of an associative memory.
If we code these states as a population of genomes, it is possible to apply Darwinian evolution
(by genetic algorithms) to promote those best suited to a certain evaluation function.
One way would be to note the image (which does not exist) corresponding thereto, which implies
that a method is found to go upstream the network.
For this you can use an inverse supervised learning network(obtained by swapping the inputs and outputs),
couples are learning the inverse couples (state, experience) of the first learning.
To any new state obtained by crossing two states corresponds a new image that can be evaluated.
I made a prototype providing very many artistic images to neural networks inputs, which generate
populations of genomes whose children give, by go back upstream in the network, an infinity of new images.
A similar experiment could be attempted with music.
This is explained in detail in
Création artificielle, and