Reproduce “Hybrid Predictive Coding: Inferring, Fast and Slow” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Alexander Tschantz, Beren Millidge, Anil K Seth, Christopher L Buckley

Tschantz et al. introduce Hybrid Predictive Coding (HPC), an architecture that integrates rapid amortized inference with slower iterative inference to offer a unified explanation for both fast and slow perceptual processes. The model achieves classification accuracy comparable to standard predictive coding while requiring significantly fewer inference steps for familiar stimuli, demonstrating adaptive computational efficiency and data robustness.

Key insights

  • Perception can be effectively modeled as a dynamic trade-off between fast, approximate amortized inference and slower, precise iterative inference, with the system adaptively allocating computational resources based on internal uncertainty.
  • The amortized pathway can learn to directly predict the outcomes of the more computationally intensive iterative inference process, enabling rapid and accurate initial estimates for familiar stimuli.
  • A unified framework combining generative (top-down) and discriminative (bottom-up) processing provides a comprehensive and biologically plausible account for the diverse temporal characteristics observed in biological perception and learning.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.