99%+ energy-based model demo
I trained energy-based models for nonogram puzzles which are NP-complete.
This model was trained on a laptop GPU. The public demo uses CPU inference on a small cloud host so is slow — local performance is ~30ms; the first solve may take a second while the worker spins up.
The model does not backtrack: it descends through an energy landscape shaped by the clues and its weights, converging toward a valid grid when the puzzle is within its capability.
More on energy-based models: NYU Deep Learning SP20 — week 7.1