a1cst/genreg-radial

GENREG-RADIAL is a small research demo that shows off a system for teaching a computer to recognize motion, without using the kind of training method most AI models rely on today. Most machine learning systems learn by gradually adjusting internal numbers through a process called backpropagation. This project skips that entirely. Instead, it uses an approach called neuroevolution, where many candidate solutions are tried and the better ones are kept and combined, similar to natural selection. The demo itself needs nothing more than Flask, a lightweight Python web server, to run. All of the results shown on the page were computed ahead of time by the real evolution process and saved as data files, so visitors do not need a graphics card or any machine learning libraries installed to see it work. The page walks through three experiments. The first asks the system to name which of ten animated motions it is watching, based on a short six frame clip, even though the shape performing the motion and its position keep changing. The second experiment tests how well the trained system holds up when conditions change, such as new colors, inverted backgrounds, or different image sizes, and includes a method for patching weaknesses without starting over. The third experiment builds a simple form of visual attention: a model learns to track a moving red cursor on screen, then uses its own tracking guess to figure out what shape the cursor is pointing at. There is also an interactive version where a visitor can move their mouse over shapes and watch the system respond live. The full source code used to produce these results is included in the project for anyone who wants to inspect or reproduce the work, though redoing the training requires a graphics card and the complete GENREG codebase. This is one part of a larger research project called GENREG.