Predictive Coding Graphs are a Superset of Feedforward Neural Networks

A new theoretical paper by researcher Björn van Zwol provides a crucial mathematical bridge between neuroscience-inspired AI and mainstream deep learning. The work, titled "Predictive Coding Graphs are a Superset of Feedforward Neural Networks," was accepted at the prestigious NeuroAI Workshop at NeurIPS 2024. In its 11 pages, the paper formally proves that Predictive Coding Graphs (PCGs)—a generalization of predictive coding networks—define a mathematical superset that encompasses standard feedforward artificial neural networks, also known as multilayer perceptrons (MLPs). This establishes a rigorous foundation connecting these biologically plausible models to the tools powering today's AI applications.

This proof has significant implications for the future of AI architecture design. By positioning PCGs as a superset, the research reinforces earlier proposals to study non-hierarchical and more general network topologies for machine learning tasks. It moves the discussion beyond simply mimicking the brain's predictive processing and shows how its computational principles can formally subsume and extend current engineering approaches. The work underscores the broader notion that network topology itself is a critical, underexplored variable in AI systems, potentially paving the way for more efficient, robust, and flexible models that aren't constrained by strictly layered, feedforward designs.