PrediRep: Modeling hierarchical predictive coding with an unsupervised deep learning network.

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Cognitive Science

Background:

• Hierarchical predictive coding (hPC) explains cortical function via prediction error minimization.
• Existing deep learning models deviate from hPC principles, limiting neuroscientific applications.

Purpose of the Study:

• Introduce PrediRep, a novel deep learning network adhering to hPC architectural principles.
• Validate PrediRep's functional alignment with hPC compared to existing models.
• Provide a tool for in silico exploration of cortical predictive coding.

Main Methods:

• Trained PrediRep and existing hPC-inspired models on a next-frame prediction task.
• Compared functional alignment with hPC using an all-level loss function (PrediRepAll).
• Evaluated information processing, representation activity, and prediction accuracy across hierarchical levels.

Main Results:

• PrediRepAll demonstrated high functional alignment with hPC.
• PrediRep processed input-relevant information at higher hierarchical levels.
• PrediRep maintained active representations and accurate predictions across all levels with fewer parameters.

Conclusions:

• Architectural adherence to hPC principles is crucial for functional accuracy.
• PrediRep offers a lightweight, biologically plausible model for neuroscience research.
• PrediRep facilitates in silico investigation of predictive coding and empirically testable predictions.