Foggy PCN explores predictive-coding networks that update internal states to reduce reconstruction error. The experiments vary attention, injected noise, and the way noise changes during inference and training.
The work includes denoising tasks and reconstruction variants. One question is whether a model can use noise to represent uncertainty in a useful way rather than treating every perturbation as something to remove immediately.
The project contains several implementations and analysis scripts. Reported results are specific to their tasks and configurations. The next useful step is to compare those variants under consistent conditions and distinguish better reconstruction from better downstream decisions.