RPN 2: On Interdependence Function Learning Towards Unifying and Advancing CNN, RNN, GNN, and Transformer

Jiawei ZhangarXiv 2024 · DOI · arXiv

Introduction

In modern deep learning, data types are often treated as distinct domains requiring specialized architectures: Convolutional Neural Networks (CNNs) for images, Recurrent Neural Networks (RNNs) or Transformers for text, and Graph Neural Networks (GNNs) for relational data. This fragmentation stems from the traditional machine learning assumption of independent and identically distributed (i.i.d.) data. However, real-world data is inherently connected—pixels depend on neighbors, and words depend on context.

Building on the Reconciled Polynomial Network (RPN), RPN 2 introduces a unified mathematical framework that explicitly models these relationships through "interdependence functions." By shifting the focus from independent attributes to the links between them, RPN 2 provides a theoretical and practical bridge between the "Big Four" architectures.

The RPN 2 Framework

The core innovation of RPN 2 is the Interdependence Function ($\xi$). Unlike standard models that treat input features in isolation, RPN 2 maps inputs into an interdependence space to generate an Interdependence Matrix. This matrix captures the structural, numerical, or statistical relationships between data instances and attributes.

The RPN 2 framework maps diverse data types (image, language, time-series, graphs) into a unified interdependence and transformation space.

As illustrated in the figure above, the process involves:

  1. Data Batching (e): Organizing raw inputs ($x_1, x_2, ...$).
  2. Interdependence Mapping (h → i): Using $\xi(\cdot)$ to project data into an interdependence space, resulting in the Interdependence Matrix (f).
  3. Data Transformation (i → j): Applying a transformation function $\kappa_\xi(\cdot)$ to produce the final reconciled representation (g).

Unifying the "Big Four"

RPN 2 demonstrates that the primary difference between major architectures lies in their specific definition of interdependence:

  • CNNs: Defined by a grid-based structural interdependence, assuming local geometric relationships.
  • RNNs: Defined by a chain-based structural interdependence, modeling sequential order.
  • GNNs: Defined by a graph adjacency matrix, capturing explicit relational links.
  • Transformers: Defined by bilinear interdependence, utilizing learnable forms to replicate the attention mechanism.

By adjusting the interdependence function, a single RPN 2 architecture can morph into any of these backbones, offering a "Grand Unified Theory" of neural design.

Empirical Performance and Theoretical Bounds

RPN 2 was evaluated across a "multimodal decathlon," including vision (CIFAR-10), language (AGNews), time-series (stock and traffic data), and graphs (Cora citation networks).

Key results include:

  • Vision: Achieved 85.2% accuracy on CIFAR-10, outperforming standard CNN baselines (84.7%).
  • Language: Reached 92.2% on AGNews, surpassing LSTM models.
  • Graphs: Hit 83.0% on the Cora dataset, beating traditional Graph Convolutional Networks (81.5%).

Beyond performance, the authors utilize VC-dimension and Rademacher complexity to show that explicitly modeling interdependence tightens generalization error bounds. This suggests that the model is inherently more robust against overfitting because it respects the underlying structure of the data.

Conclusion and Future Directions

RPN 2 represents a significant advancement in function-oriented learning. While it successfully unifies existing architectures, future work will focus on true temporal dynamics (moving beyond spatial chains for time-series) and scaling the framework to the level of Large Language Models (LLMs). The framework is currently implemented in the tinyBIG library, providing a toolkit for researchers to explore interdependence-based learning.

Summary updated

BibTeX
@misc{zhang2024rpn,
  author = {Jiawei Zhang},
  title = {{RPN 2: On Interdependence Function Learning Towards Unifying and Advancing CNN, RNN, GNN, and Transformer}},
  year = {2024},
  eprint = {2411.11162},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2411.11162},
}
Figure