Károly Zsolnai-Fehér describes research experiments in which a model analyzes another network’s weightsWeight space learning treats the trained numerical parameters of neural networks as data a system can analyze, predict or transform. to identify learned game behaviors and makes targeted changes to model behavior.
He discusses a reported math benchmark improvementEvaluation is the systematic assessment of a system's behaviour or results against explicit criteria. and emphasizes that the intervention addressed a repetition problem rather than simply teaching the model mathematics.
Károly Zsolnai-Fehér stresses that the reading, repair and improvement examples used separate setups, mostly on small systems, and that an attempted improvement of another model did not succeed. He also describes generating a game-playing network directly.
Watch on YouTube
