Fig. 5: Predictive Coding and FGL.
From: A predictive approach to enhance time-series forecasting
A Illustration of predictive coding in the brain. Information is received by the primary visual cortex (V1), which then propagates to different areas of the brain with more complex levels of abstraction. This propagation takes the form as a predictive estimator: higher level areas aid in prediction of lower level areas. The difference between the prediction and true output is the predictive error. Areas with lighter colors represent low-level abstractions, where darker colors represent the increasing representational demand (thoughts, movement, etc). B Hierarchical FGL propagates information via uncertainty; in other words, FGL substitutes the predictive estimator with the uncertainty of each layer. This uncertainty is conveyed via the KL divergence of between each layer’s probability distribution with the successive layer. The difference between each layer’s prediction and true output is the predictive error. In both models, areas lower in the hierarchy process information in a delayed manner, as they are the last to receive it. As the demand for complexity increases, so does the predictive error.
