Supervised Radio Frequency Interference Detection with SNNs

TL;DR

The paper reframes RFI detection in radio astronomy as a supervised time-series segmentation problem suitable for Spiking Neural Networks (SNNs) and systematically evaluates six encoding schemes to translate visibility data into spike trains. Through a small two-layer SNN trained with surrogate-gradient backpropagation on simulated HERA data, latency-based and step-forward encodings emerge as the most effective, achieving high per-pixel accuracy and competitive F1-scores while preserving a compact network. Although the approach trails state-of-the-art non-SNN methods on the same dataset, its competitive performance using a simple architecture highlights the potential of SNNs for time-series RFI detection and as a benchmark for neuromorphic computing. The study also emphasizes the benefits of reframing two-dimensional spectrogram flagging into a one-dimensional temporal problem, enabling efficient SNN inference and future hardware-oriented optimization.

Abstract

Radio Frequency Interference (RFI) poses a significant challenge in radio astronomy, arising from terrestrial and celestial sources, disrupting observations conducted by radio telescopes. Addressing RFI involves intricate heuristic algorithms, manual examination, and, increasingly, machine learning methods. Given the dynamic and temporal nature of radio astronomy observations, Spiking Neural Networks (SNNs) emerge as a promising approach. In this study, we cast RFI detection as a supervised multi-variate time-series segmentation problem. Notably, our investigation explores the encoding of radio astronomy visibility data for SNN inference, considering six encoding schemes: rate, latency, delta-modulation, and three variations of the step-forward algorithm. We train a small twolayer fully connected SNN on simulated data derived from the Hydrogen Epoch of Reionization Array (HERA) telescope and perform extensive hyper-parameter optimization. Results reveal that latency encoding exhibits superior performance, achieving a per-pixel accuracy of 98.8% and an f1-score of 0.761. Remarkably, these metrics approach those of contemporary RFI detection algorithms, notwithstanding the simplicity and compactness of our proposed network architecture. This study underscores the potential of RFI detection as a benchmark problem for SNN researchers, emphasizing the efficacy of SNNs in addressing complex time-series segmentation tasks in radio astronomy.

Figures (2)

• Figure 1: Example spike raster plots for all explored encodings, the original spectrogram, and associated RFI mask. Encodings are exposed for four time-steps where applicable.
• Figure 2: Example HERA Spectrogram, the output of a latency-encoded trained SNN, and the original RFI mask.