These experiments examine attention mechanisms built around predictive-coding updates. The models include attention modules, transformer-style encoders, and classifiers, alongside image-reconstruction and next-token experiments.
One text-classification path uses AG News. Other scripts explore reconstruction and several attention variants. This gives the work a few concrete tasks, rather than treating attention behavior as a result on its own.
A useful comparison needs a fixed task, comparable model capacity, and the same training budget, with attention behavior and task performance examined separately.