Digital Digital Burgess Conference Follow-up:
Conference was held August 29-September 1 1997, Banff Alberta, Canada.
On Neural Nets and Aesthetic Values
I'd be curious to know if you could train a neural net on the *input* expression for the image, not just the output image. In other words, if you could train a net to find expressions that produce intersting images. The question to be answered here is whether there is any correlation between the input expression and the aesthetic value of the resulting image.
An advantage of this technique is that, since the input is an expression, rather than an image with close to a million pixels, each with one of a possible millions of color values, the neural net would be far simpler, and consequently much faster.
Second, rather than training the net once, and then letting it run freely, I would like to see the neural net combined with Sim's installation, where someone steps on a foot switch to indicate their favorite image. Each time someone votes for an image, the net would be trained. When nobody is there to vote, the net would breed images. Because the network would be sporatically trained, this would avoid the problem of having the same kinds of images win all the time.
Now, let's take this one step further. The result of the above exercise is not just some (supposedly) good images, but also a trained neural network. Could you have different trained networks compete, and have people vote on which network produces better images? Then combine the "genes" of the two networks...