Predictive coding theory holds that the brain perceives by minimizing prediction errors through cycles of neural activity. However, some visual perception, including complex object recognition, happens too quickly for such cycles. This paper proposes that the initial fast 'feedforward sweep' performs amortized inference, using a learned function to map data directly to beliefs, while slower recurrent processing performs iterative inference, sequentially updating beliefs for greater accuracy. A hybrid predictive coding network combining both methods is introduced, implemented in a biologically plausible neural architecture using local Hebbian rules. The hybrid model achieves rapid perception for familiar data while retaining context-sensitivity and sample efficiency for novel situations, and adaptively balances both inference modes based on uncertainty.
Cognitive systems are often modeled as input/output devices with separate perceptual and motor modules, a view that resonates with the separation principle of control theory. This paper presents a minimal sensorimotor model based on that principle and shows its limitations when external forces—such as environmental perturbations or interference from other agents—are not accounted for. As an alternative, the authors propose a nonmodular architecture grounded in active inference, which demonstrates robustness to unknown external inputs. In linear models, this robustness is achieved through a mechanism equivalent to integral control, offering a principled way to handle disturbances that the agent cannot directly control.