Topological Neural Networks over the Air

Simone Fiorellino, Claudio Battiloro, Paolo Di Lorenzo

Abstract

Topological Neural Networks (TNNs) offer a sophisticated framework for processing data on complex structures like cell complexes, enabling the modeling of multi-way interactions beyond simple pairwise connections. However, existing TNN research often assumes ideal communication channels, neglecting the impact of wireless impairments such as fading and noise in decentralized sensing. This paper presents AirTNN, a novel framework that incorporates wireless channel characteristics directly into the TNN architecture. By utilizing the wireless medium as a noisy "shift operator" and implementing a "noise-in-the-loop" training strategy, AirTNN demonstrates exceptional robustness and accuracy in decentralized tasks compared to traditional models.

Introduction

In decentralized AI systems, such as IoT sensor networks, agents must share information with neighbors to perform collective learning. While Graph Neural Networks (GNNs) are widely used, they are limited to pairwise interactions. TNNs address this by operating on higher-order structures like cell complexes, where interactions can occur across shared vertices (lower neighborhoods) or shared polygons (upper neighborhoods).

Despite their potential, TNNs face significant challenges when deployed over wireless networks. Real-world signals suffer from Rayleigh fading and Additive White Gaussian Noise (AWGN), which can cause performance to collapse if the model is trained on idealized data. While "AirGNNs" have been proposed for simple graphs, they lack the capacity to handle the higher-order topological relationships essential for modeling physical systems like power grids or traffic flows.

Methodology

The primary innovation of this work is the Topological Filter Over-the-Air (AirTF). This approach integrates the physical layer directly into the neural network's filtering operations.

Cell Complexes and Neighborhoods

Unlike standard GNNs, AirTNN operates on cell complexes. This allows the network to aggregate information from both lower and upper neighborhoods, capturing complex structural dependencies.

Representation of lower neighborhood networks, cell complexes, and upper neighborhood networks used in AirTNN.

Over-the-Air Computation

Signal aggregation is performed using uncoded, analog transmission, where the wireless channel itself serves as the "shift operator." The authors mathematically model Rayleigh fading and AWGN within the filtering process. To ensure robustness, the model is trained with a "noise-in-the-loop" approach, sampling random channel coefficients during backpropagation. This forces the learnable weights to adapt to the inherent volatility of the wireless medium.

Experimental Setup

The model was evaluated using a synthetic dataset based on a Stochastic Block Model (SBM) with 70 nodes. This graph was extended into a cell complex using a cycle basis. The task involved source localization: identifying the community of origin for "edge signals" (flows) diffused across the topology and observed under noisy conditions.

Results and Discussion

AirTNN was compared against standard GNNs, TNNs, and the graph-based AirGNN. The results highlight the advantages of incorporating topological information into wireless AI.

Accuracy of AirTNN compared to other models across varying Signal-to-Noise Ratios (SNR).

Robustness to Noise and Fading

As shown in the performance analysis, AirTNN maintains high accuracy even as the Signal-to-Noise Ratio (SNR) decreases. At 20 dB SNR, AirTNN achieves approximately 80% accuracy, whereas standard TNNs and GNNs—which do not account for channel effects—drop to near-random performance levels (10-20%). Furthermore, AirTNN consistently outperforms AirGNN across various fading intensities, demonstrating that higher-order topological redundancy provides a significant buffer against communication errors.

Topological Redundancy

A key insight is that TNNs are inherently more resilient to noise than GNNs. By processing signals through both upper and lower neighborhoods, the model benefits from structural redundancy, allowing it to recover information even when specific links are degraded by fading.

Conclusion

AirTNN bridges the gap between topological deep learning and wireless communication. By treating the channel as an integral part of the network architecture, it enables robust decentralized sensing on complex infrastructures. Future research will explore more complex non-stationary channel models, higher-order signals on 3D meshes, and the hardware challenges of precise sensor synchronization required for over-the-air computation.

Summary updated

Abstract

Topological neural networks (TNNs) are information processing architectures that model representations from data lying over topological spaces (e.g., simplicial or cell complexes) and allow for decentralized implementation through localized communications over different neighborhoods. Existing TNN architectures have not yet been considered in realistic communication scenarios, where channel effects typically introduce disturbances such as fading and noise. This paper aims to propose a novel TNN design, operating on regular cell complexes, that performs over-the-air computation, incorporating the wireless communication model into its architecture. Specifically, during training and inference, the proposed method considers channel impairments such as fading and noise in the topological convolutional filtering operation, which takes place over different signal orders and neighborhoods. Numerical results illustrate the architecture's robustness to channel impairments during testing and the superior performance with respect to existing architectures, which are either communication-agnostic or graph-based.
BibTeX
@inproceedings{fiorellino2024topological,
  author = {Simone Fiorellino and Claudio Battiloro and Paolo Di Lorenzo},
  title = {{Topological Neural Networks over the Air}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2024},
}
Figure