🏛 3D Progressive Neural Network: Frozen Columns, Lateral Adapters and Catastrophic Forgetting
Three stone pedestals built in Blender hold the columns of a progressive network that learns XOR, a ring and a sine stripe one after another: earlier columns are frozen behind a padlock and feed the new one through lateral adapter cables, while a single fine-tuned network in the front row forgets, with gradient, zero-forgetting and parameter-count checks.
Three stone pedestals built in Blender hold the columns of a progressive network that learns XOR, a ring and a sine stripe one after another: earlier columns are frozen behind a padlock and feed the new one through lateral adapter cables, while a single fine-tuned network in the front row forgets, with gradient, zero-forgetting and parameter-count checks.
3D · Three.js / WebGL renderer · 60 FPS target · runs fully client-side, no install