Introduction
Predicting wireless signal (RF) behavior in complex 3D environments is a cornerstone of modern telecommunications, yet it remains a significant challenge. Signals undergo reflection, diffraction, and scattering, creating a multipath environment that traditional models struggle to capture efficiently. While Neural Radiance Fields (NeRF) have been adapted for RF modeling, they typically rely on a "point-in-point-out" approach, processing voxels along a ray independently. This ignores the physical reality that a signal's state at any point depends on the medium it has already traversed. Furthermore, the reliance on standard Multi-Layer Perceptrons (MLPs) often results in "black-box" models that lack explainability.
To address these issues, this paper introduces KANNA-NeRF (KAN with liNear Attention NeRF), a framework that shifts the paradigm to a "ray-in-ray-out" approach. By integrating Kolmogorov-Arnold Networks (KAN) and Linear Transformers, the model achieves superior accuracy, efficiency, and physical consistency.
Methodology
Ray-in-Ray-Out Architecture
The core innovation of KANNA-NeRF is treating the entire signal ray as a sequence rather than a set of isolated points. This allows the network to learn spatial dependencies and interactions along the propagation path, mimicking actual physical wave behavior.
Efficient KAN with PowerMLP
Kolmogorov-Arnold Networks (KANs) are mathematically well-suited for RF propagation because they learn activation functions on the edges rather than just weights at nodes. However, standard KANs using B-splines are computationally expensive. KANNA-NeRF utilizes "PowerMLP," which replaces splines with power-of-ReLU functions. This maintains the high non-linear mapping capabilities of KANs while keeping the computational footprint comparable to standard MLPs.
Linear Transformers and Causal Masking
To process long sequences of points along a ray efficiently, the model employs Linear Transformers (Performers). These use random feature maps to achieve linear time complexity, avoiding the quadratic memory costs of standard attention mechanisms. Crucially, a causal mask is applied to the attention mechanism. This ensures that the signal representation at any point only "attends" to points it has already passed through, strictly adhering to the physical principle of causality in wave propagation.
Dual-Branch Design
The architecture is split into two specialized branches:
- Attenuation Network: Models energy loss as the signal traverses the medium.
- Radiance Network: Models secondary emissions and reflections, contributing to the final signal representation.
Experimental Results
KANNA-NeRF was validated against several baselines, including MLP, VAE, DCGAN, and the original NeRF², across both synthetic and realistic datasets.
- Synthetic Performance: In a simulated "Bedroom" environment, KANNA-NeRF achieved a Signal-to-Noise Ratio (SNR) of 28.15 dB, significantly outperforming NeRF² (21.65 dB) and standard MLPs (17.67 dB).
- Real-World Robustness: On the MIMO-CSI (Argos) dataset, the model reached 25.00 dB SNR, demonstrating its ability to handle real-world noise and hardware distortions.
- Efficiency: By utilizing PowerMLP and Linear Attention, the model maintains parameter storage scaling at $O(L \cdot d^2)$, making it viable for large-scale deployment in complex urban or indoor environments.
Discussion and Explainability
Beyond performance, KANNA-NeRF offers enhanced explainability. By analyzing the Jacobian matrix of the model, researchers can visualize which 3D coordinates most heavily influence signal predictions. This allows engineers to identify specific environmental obstacles causing signal degradation, transforming the model from a black box into a diagnostic tool.
While the model currently requires pre-processing to handle hardware-specific offsets, future work aims to integrate these corrections into an end-to-end learning process. The framework also holds potential for expansion into other wave-based sensing domains, such as acoustic modeling and 6G integrated sensing and communication (ISAC).
Summary updated