CPPNs are similar to artificial neural networks (ANNs). A CPPN takes in the (x, y) position of a pixel as input, and outputs a grayscale value (MNIST) or tuple of HSV color values (ImageNet) for that pixel. Like a neural network, the function the CPPN computes depends on the number of neurons in the CPPN, how they are connected, and the weights between neurons. Each CPPN node can be one of a set of activation functions (here: sine, sigmoid, Gaussian and linear), which can provide geometric regularities to the image. For example, passing the x input into a Gaussian function will provide left-right symmetry, and passing the y input into a sine function provides top-bottom repetition. Evolution determines the topology, weights, and activation functions of each CPPN network in the population. arxiv.org/abs/1412.1897
Posted on: Mon, 29 Dec 2014 17:06:35 +0000
Trending Topics
Recently Viewed Topics
© 2015