Jonathan Whitaker posts neural cellular automata image classifier demo
——0——
Jonathan Whitaker posted a neural cellular automata demonstration in which each pixel functions as an independent cell running a shared lightweight neural network. The system classifies hand-drawn images by building a consensus prediction over successive time steps, as shown when an elephant sketch reached 45.2 percent probability for the correct class. A live web interface titled Live Moodle NCA lets users draw on a canvas, select targets, and adjust steps while viewing updating bar charts and heatmaps. Alex Mordvintsev and Joscha Bach shared the project.
