Scaling Neural Networks to Dynamic Topologies

50 concepts51 papers3 proposals

Traditional deep learning architectures are built on a rigid assumption: the dimensionality of the input and output spaces must remain fixed. If you train a Multi-Layer Perceptron (MLP) or a standard Convolutional Neural Network (CNN) to optimize a system with ten variables, it cannot natively accept an eleventh.

In real-world infrastructure—such as wireless communication networks, sensor arrays, and routing topologies—this fixed-dimension constraint is a critical bottleneck. Wireless networks are inherently fluid; users connect and disconnect, base stations switch on and off, and hardware configurations change continuously. As Sharma et al. and Sanguinetti et al. observe, conventional neural networks designed for resource allocation lack the structural flexibility to handle these fluctuations. When the network topology shifts, the underlying architecture becomes invalid, requiring either a multi-model approach or complete retraining from scratch.

To overcome this limitation, a prominent research direction focuses on topological scalability and generalization: designing neural networks that can be trained on small, static configurations but deploy seamlessly onto larger, dynamically changing network structures without retraining.

The Mechanism: Permutation Equivariance and Graph Neural Networks

The reason standard models fail at topological scaling lies in how they process data. An MLP maps specific input indices to specific output indices via a fixed weight matrix. If the order of users in a wireless cell is shuffled, or if a new user is introduced, the spatial relationships break down.

To build intuition for why topological generalization is possible, consider the physics of a wireless network. The optimal power allocation for a user depends on their local channel conditions and the interference caused by neighboring devices. This physical reality possesses two mathematical properties:

  1. Size Invariance: The fundamental rules governing signal interference do not change whether there are 5 users or 500 users.
  2. Permutation Equivariance: Shuffling the arbitrary indexing of the users should shuffle the resulting optimal allocations in the exact same manner, without changing the underlying values.

Graph Neural Networks (GNNs) naturally inherit these properties. By modeling a communication network as a graph—where nodes represent devices (users, base stations, or antennas) and edges represent physical channels or interference links—GNNs replace global weight matrices with localized message-passing operations.

In a message-passing layer, a node updates its state by aggregating features from its immediate neighbors using a parameterized function (such as a small MLP) that is shared across the entire graph. Because these parameters are applied locally and uniformly, the network's weights are completely decoupled from the total number of nodes or edges.

Shen mathematically demonstrates that leveraging this permutation equivariance allows GNNs to achieve unprecedented sample efficiency and scalability, solving radio resource management problems for thousands of users in milliseconds. Similarly, Ribeiro et al. introduce Random Edge Graph Neural Networks (REGNNs), proving that treating interference patterns as dynamic graphs allows a model trained on small topologies to transfer its optimization capabilities directly to larger, unseen networks.

Deep Unfolding: Embedding Algorithmic Flexibility

While pure black-box GNNs offer size invariance, ensuring they generalize robustly to complex mathematical constraints requires a hybrid approach known as deep unfolding (or algorithm unrolling).

Classical optimization algorithms (such as the Weighted Minimum Mean Square Error, or WMMSE) handle variable network sizes effortlessly because they are defined by iterative mathematical operations rather than fixed data structures. However, they are computationally slow. Deep unfolding bridges this gap by mapping the sequential iterations of an optimization algorithm directly into the layers of a neural network.

Instead of forcing a network to learn an optimization policy from scratch, the architecture is structured to mirror the mathematical steps of the classical algorithm, using GNN layers to learn and accelerate the tuning of the algorithm's internal parameters.

[Input Topology (Variable Size)] 
               │
               ▼
┌────────────────────────────────────────┐
│ Unfolded GNN Layer 1 (Iteration 1)     │ ──► Shares local weights
└────────────────────────────────────────┘
               │
               ▼
┌────────────────────────────────────────┐
│ Unfolded GNN Layer 2 (Iteration 2)     │ ──► Inherits algorithmic structure
└────────────────────────────────────────┘
               │
               ▼
[Near-Optimal Allocation (Scaled Output)]

The authors of several foundational papers demonstrate the power of this paradigm:

  • Chowdhury et al. unfold the WMMSE algorithm using GNNs for power allocation, preserving the structural flexibility of the original mathematical framework while achieving massive execution speedups.
  • Segarra et al. apply deep unfolding to energy-aware power allocation, showing that the resulting model maintains high energy efficiency even when evaluated on network sizes completely different from the training data.
  • Tang et al. extend this to multi-cluster wireless networks, demonstrating that unfolded deep graph learning generalizes across physical layouts without requiring structural redesigns.

Generalization Across Diverse Network Domains

By shifting from fixed-dimension vectors to flexible graph structures, researchers have successfully brought topological scalability to a wide array of complex networking problems:

Beamforming and Transceiver Design

In Multi-User MIMO systems, beamforming vectors must adapt as users move and cell densities change. He et al. developed the Interference Channel GNN (ICGNN) to scale beamforming to a variable number of users without retraining. To handle hardware constraints alongside user fluctuations, Zhu et al. combined graph attention networks with variational autoencoders, allowing transceivers to scale efficiently across varying cell densities and feedback bit budgets.

Dynamic Infrastructure (UAVs and RIS)

When networks incorporate mobile elements like Unmanned Aerial Vehicles (UAVs) or Reconfigurable Intelligent Surfaces (RIS), the topology changes spatially as well as numerically. Wang et al. and Tang et al. leverage GNNs to jointly optimize UAV positioning and beamforming for fluctuating ground populations. For next-generation surfaces, Chen utilizes graph architectures to maximize energy efficiency in Rate-Splitting Multiple Access (RSMA) networks, outperforming rigid deep learning baselines in structural adaptability.

Routing and Sensor Networks

The problem of topological rigidity extends beyond wireless signals to data packet routing and audio processing. Suárez‐Varela et al. demonstrate that GNNs can learn routing optimization policies on a single network topology and accurately predict performance on entirely unseen internetwork configurations. In the domain of sensor arrays, Essid addresses ad-hoc microphone arrays where nodes frequently join or leave, using attention mechanisms to handle variable input dimensions dynamically.

Ultimately, this research direction shifts the deep learning paradigm for infrastructure from memorizing static configurations to learning the underlying physics of interaction. By embedding permutation equivariance and algorithmic structure directly into the network architecture, these models achieve the fluid scalability required for real-world deployment.

Key Foundational Work

Deep Learning Power Allocation in Massive MIMO (2019) — contributed 2 concept elements to this direction.
Deep Learning Based Online Power Control for Large Energy Harvesting Networks (2019) — contributed 1 concept elements to this direction.
Towards Optimal Power Control via Ensembling Deep Neural Networks (2018) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_topological_scalability_c3161b7c,
  title = {topological scalability},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/topological-scalability/rdr_c3161b7c949e4db6bb43c84f158e03e8},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_topological_scalability_c3161b7c_v1,
  title = {topological scalability},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/topological-scalability/rdr_c3161b7c949e4db6bb43c84f158e03e8/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake