A brief introduction to neural chemistry and a simple cppn (compositional Pattern producing Networks) DEMO

Recently fascinated by the direction of Neuro-evolution (neuroevolution), the feeling is a very good research field after deep learning. One of the leading factors in this field is the evolution of network parameters and structures, modeled on human genetic mechanisms. Note that even the network structure can evolve, that is, unlike traditional neural networks, structures are defined in advance.

The most recent research in this area is Professor Stanley at the University of Florida, and the main contributions are neat,hyperneat and novelty Search, who are interested to go to Google.

Neat and hyperneat are a kind of indirect coding algorithm, which aims at updating the topological structure and parameters of neural network, and novelty search is an algorithm to guide neat and hyperneat how to update. The relationship between neat,hyperneat and neural networks is this:

The parameters of our neural network (ANN) for decision-making and classification are determined by a network called compositional Pattern producing Networks (CPPN), and the cppn topology is evolved by neat and hyperneat algorithms, Novelty Search Guides The optimization of neat and hyperneat algorithms.

CPPN Network is similar to the traditional neural network, but the activation function is more special, richer, and the network structure does not have the concept of a layer, the input is the coordinate value x, Y, can refer to the 2009 proposed hyperneat the article on page 5th of Figure 2.

Here is a very simple CPPN demo, written in Matlab

Clear;clcwid = 32;hei = 32;map = Zeros (hei, WID); for h = 1:hei for    w = 1:wid        %map (h, w) = sin ((h^2 + w^2) *pi/1 %+ exp (((H-HEI/2) ^2 + (W-WID/2) ^2) + H + W;        %map (h, W) = exp (((H-HEI/2) ^2 + (W-WID/2) ^2);        Map (h, w) = sin ((h^2 + w^2) *pi/100) + exp (((H-HEI/2) ^2 + (W-WID/2) ^2)/800);    Endendmap = map-min (map (:)); map = 255 * Map/max (map (:)); Imshow (uint8 (map))

As you can see from the code, the input is the coordinate value, and the network structure is simple: it is a Gaussian function + sine function. The final map map is

More complex cppn generated images can be found in the http://picbreeder.org/

A brief introduction to neural chemistry and a simple cppn (compositional Pattern producing Networks) DEMO