RPN 2: A Unified Framework for Neural Network Architectures

This paper introduces RPN 2, an enhanced version of the Reconciled Polynomial Network (RPN), designed to model data interdependence in complex domains like language, images, time series, and graphs. Unlike its predecessor, which assumed independence among data instances and attributes, RPN 2 incorporates structural and data interdependence functions into its architecture. This advancement unifies key neural network architectures—CNNs, RNNs, GNNs, and Transformers—under a shared framework, revealing that their distinctions primarily lie in how they define interdependence functions. Beyond unification, RPN 2 demonstrates improved learning performance and paves the way for novel architectures that could outperform existing backbones.