Predictive coding networks for temporal prediction.

Area of Science:

• Computational neuroscience
• Neural networks
• Perception

Background:

• The brain infers dynamic world states from sensory input, a process not fully understood.
• Predictive coding theory explains perception but often focuses on static stimuli.
• Key questions about temporal predictive coding's neural implementation and properties remain unanswered.

Purpose of the Study:

• To formulate a temporal predictive coding model suitable for biological neural networks.
• To investigate the computational properties and neural implementation of temporal prediction.
• To explore how the brain might predict future stimuli using biologically plausible mechanisms.

Main Methods:

• Developed a temporal predictive coding model implemented in recurrent neural networks.
• Utilized local neuronal inputs for activity dynamics and local Hebbian plasticity for learning.
• Compared model performance to the Kalman filter for linear and nonlinear systems.

Main Results:

• The model approximates Kalman filter performance in predicting linear system behavior.
• Networks exhibit biologically plausible Gabor-like, motion-sensitive receptive fields when trained on natural dynamic inputs.
• The model generalizes effectively to nonlinear systems.

Conclusions:

• Biologically plausible recurrent networks can perform temporal predictive coding.
• The model offers a framework for understanding neural computation in temporal prediction.
• This work bridges computational theory with potential neural mechanisms for sensory prediction.