The Cognitive Diffusion Probabilistic Models (CogDPM) is aCogDPM: Diffusion Probabilistic Models via Cognitive Predictive Coding”, showing the connection between diffusion probability model and PC theory.

CogDPM has an accuracy estimation method based on the hierarchical sampling capability of the diffusion model, and uses the accuracy weights estimated from the inherent properties of the diffusion model to weight the guidance signal. The research team showed through experiments that the accuracy weights can effectively estimate the predictability of the data. In the paper, CogDPM is applied to real-world forecasting tasks using UK precipitation and ERA surface wind datasets. The results show that CogDPM surpasses existing domain-specific operational models and general deep forecasting models, and can provide more skillful forecasting results.