SCHOOL OF ENGINEERING AND COMPUTER SCIENCE

The (random) talk relates to GPUs:

A system I have been wanting to build involves the storing of the rules of a GP system in a texture. Think of storing the weights of a neural network in the data of the texture image. A GPU (on most video cards) has pixel shaders that can run (parallel x 256) an arbitrary program using the texture as input, thus applying a genetic program, or neuron model, on the data it is given.

-> Hence rule-based

The program would write the new state of the neuron/whatever back to texture (via a temporary buffer), and would iterate as usual. Thus the negative feedback would tend towards some set goal

-> Hence feedback

Now, the pixel shaders are not the only shaders; there are also vertex shaders, for use in texturing. They are the thing that makes a 3D video card '3D'. These have various methods built in (and highly optimized) for interpolating coordinates for use by th (2D) pixel shaders. One, of course, is bilinear-interpolation.

-> Hence bilinear

So I figure what the random thing means is:

A system that encodes a captured image on one texture, and a set of rules and weights on another.

A polygon is formed in world space, almost matching up with the screen, but not quite.

The captured image, and the rules, are sent to the GPU, with the vertex shader using bilinear or similar interpolation to determine coordinates for the pixel shader, which applies a genetic program on each pixel, both image and rules are output as textures.

The image is displayed, while the input rules, and possibly input image, are subject to feedback. This is based on the delta rule using the output image and a goal (say, with a fitness function between "wow, that looks fantastic" and "it's just a blank screen")

The intention is to use the system for fancy graphics.

(The use of the vertex shader would be new, and I imagine similar to the video camera trick: point a video camera at a monitor connected to its output. The camera "sees" a polygon (the monitor), not quite exactly on the "screen" of its projection, but close. Hence feedback results in a simple fractal. This would advance that significantly)

Or, it could mean something else entirely..

-- smartwill - 04 Mar 2009