a question: can neural networks be made to do this?
mlevin at presto.cs.tufts.edu
Tue Jul 14 13:17:07 EST 1992
I have a question about applying neural nets to a certain problem.
Suppose I have a certain generative algorithm, that takes a small
string as input, and produces a two-dimensional image (array of color
indeces) as output. The L-systems are an example of such an algorithm,
and so are fractal functions when plotted. Suppose also that
producing an image from such a seed is simple (though computationally
intensive), but finding a seed to produce a given image is very
difficult to do (for a human being). Is it possible to train up a
neural network bu giving it an image, having it guess at a seed, and
then giving it the correct seed to let it adjust its weights. The hope
then is that after a while the net's guesses will become good, and it
can be used to derive seeds for arbitrary images. There is an infinite
supply of training material, because there is an infinite amount of
randomly-chosedn seeds to produce images.
So, does anyone have any suggestions about this? Is it feasable to
feed a network an image as input? What do I need to do to make it
return legal seeds as guesses? Are there any public domain neural net
packages that may be used to try this? How does the size (number of
nodes, layers, etc.) depend on the difficulty of the problem? Please
send me any info on this you may have at mlevin at presto.cs.tufts.edu.
More information about the Comp-bio