Polarisation-Inclusive Spiking Neural Networks for Real-Time RFI Detection in Modern Radio Telescopes

Technical Summary: Polarisation-Inclusive Spiking Neural Networks for Real-Time RFI Detection

Problem Statement
Radio Frequency Interference (RFI) presents an escalating challenge for modern radio astronomy, exacerbated by increasing observatory sensitivity and the proliferation of orbital RFI sources. Current operational constraints require a balance between rising data-processing demands and the need for real-time mitigation. While Artificial Neural Networks (ANNs) have been applied to this task, they rely on continuous, instantaneous values. In contrast, Spiking Neural Networks (SNNs) offer a compelling alternative for real-time detection due to their inherent time-varying nature and potential for extreme energy efficiency when deployed on neuromorphic hardware. However, effectively encoding information for SNNs remains challenging, and the specific utility of including polarisation data within this framework has not been fully explored.

Methodology
This work investigates the integration of polarisation information into SNN-based RFI detection using a supervised time-series segmentation approach. The study utilizes simulated data derived from the Hydrogen Epoch of Reionisation Array (HERA) instrument, retaining all polarisation channels (XX, XY, YX, YY).

• Encoding and Architecture: The authors employ latency encoding, previously shown to be effective for this task. The network utilizes a first-order Leaky Integrate-and-Fire (LiF) neuron model within a multi-layer feed-forward architecture. Two primary strategies for incorporating polarisation are tested:
  1. Full Polarisation: Treating all four polarisation channels as separate input channels.
  2. Degree of Polarisation (DoP): Calculating a DoP metric as the input feature.
     Additionally, a divisive normalisation (DN) pre-processing technique, inspired by neuroscience, is applied to evaluate its impact on feature discrimination.
• Hardware Constraints: Models are trained using the snnTorch framework and evaluated against the constraints of the SynSense Xylo neuromorphic hardware (16 input channels, 1000 internal neurons, 16 output channels). To accommodate these limits, full spectrograms are divided into smaller patches (e.g., 16×16 or 15×16) for training and inference.
• Power Modelling: The study provides three distinct energy usage estimates for deployment on Xylo hardware:
  1. A lower-bound estimate based on the measured spiking activity (FLOPs) of the trained network.
  2. An upper-bound estimate based on the maximum power draw of the Xylo hardware (550µW).
  3. A balanced estimate combining the measured idle power draw (216µW) with the lower-bound activity estimate.

Key Contributions

• Polarisation Integration: This work extends SNN-based RFI detection by explicitly incorporating polarisation data, comparing full-channel inputs against DoP metrics.
• Hardware-Aware Training: The study demonstrates the feasibility of training SNNs specifically sized to fit on edge neuromorphic hardware (Xylo) without significant loss in performance.
• Performance Benchmarking: The paper provides a direct comparison of the proposed SNN against state-of-the-art RFI detection methods, including traditional operational tools (AOFlagger), Auto-Encoders, and other SNN/ANN approaches.
• Energy Estimation: It offers a novel framework for estimating the power consumption of SNNs in radio astronomy, moving beyond theoretical FLOPs to include hardware-specific idle and active power metrics.

Results

• Detection Accuracy: The "Latency + Divisive Normalisation" configuration using full polarisation data on 16-channel patches achieved the highest performance, with an accuracy of 0.997, AUROC of 0.977, and F1-Score of 0.955.
• Comparison to State-of-the-Art: The proposed model outperformed previous methods, including AOFlagger, Auto-Encoders, and ANN-to-SNN conversion techniques, achieving state-of-the-art results on the benchmark dataset.
• Impact of Constraints: Truncating models to fit Xylo hardware resulted in only minor performance degradation. Notably, the "Latency-Xylo + DN" model with full polarisation maintained an F1-Score of 0.936, very close to the best-performing non-constrained models.
• Divisive Normalisation Effects: While DN improved accuracy for full polarisation inputs, it paradoxically reduced performance when applied to DoP inputs, suggesting that the normalisation process amplifies relevant features in one context while suppressing useful signal variations in the other.
• Energy Efficiency: Power estimates indicate that the neuromorphic approach requires only tens of milliwatts per spectrogram. Even under upper-bound estimates, the power consumption is orders of magnitude lower than small GPU-based systems (e.g., Nvidia Jetson Nano, ~5W).

Significance and Claims
The authors claim that this work demonstrates, for the first time, that a directly trained Spiking Neural Network can achieve state-of-the-art performance in RFI detection, surpassing both traditional operational methods and ANN-based approaches. The study argues that SNNs offer a viable, high-performance alternative for radio astronomy, particularly in scenarios where low latency and energy efficiency are critical.

The paper emphasizes that the potential energy improvements are not merely incremental but represent orders-of-magnitude gains. Given the multi-decadal lifespan of radio observatories, the authors posit that such efficiency could lead to substantial reductions in operational costs and power consumption. Consequently, they conclude that neuromorphic computing warrants serious consideration for high-performance scientific signal processing in power-constrained environments. Future work is identified as focusing on more sophisticated architectures and validation on real neuromorphic hardware.