🔭 astrophysics

RFI Detection with Spiking Neural Networks

This study presents the first application of Spiking Neural Networks (SNNs) to radio frequency interference (RFI) detection in radio astronomy, demonstrating that converting a nearest-latent-neighbours auto-encoder to an SNN via ANN2SNN achieves competitive performance on HERA data while eliminating computationally expensive latent sampling steps, though challenges remain on LOFAR and satellite-based datasets.

Original authors: Nicholas J. Pritchard, Andreas Wicenec, Mohammed Bennamoun, Richard Dodson

Published 2026-01-23

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Nicholas J. Pritchard, Andreas Wicenec, Mohammed Bennamoun, Richard Dodson

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to listen to a very faint whisper from a distant star using a giant radio telescope. The problem is that the universe is incredibly noisy. Not just with cosmic noise, but with "Radio Frequency Interference" (RFI)—think of it as a chaotic mix of static, cell phone signals, satellite beeps, and airplane radar that drowns out the cosmic whispers.

To hear the stars, astronomers need a way to automatically spot and silence this interference without accidentally deleting the star signals. This is where Spiking Neural Networks (SNNs) come in, and this paper is the first time anyone has tried using them for this specific job in astronomy.

Here is a breakdown of what the researchers did, using simple analogies:

1. The Problem: Finding the Needle in the Haystack

Traditionally, astronomers use computer programs to flag (mark) the interference. Newer methods use Artificial Neural Networks (ANNs), which are like smart students that learn by looking at thousands of examples of "clean" radio data and "dirty" radio data.

However, there's a catch. To teach these students, you need a massive library of examples. A recent method called NLN (Nearest-Latent-Neighbours) tried to solve this by teaching the computer to recognize what "clean" data looks like. If the computer sees something that doesn't match its memory of "clean," it flags it as interference.

The Flaw in the Old Method:
Imagine the NLN method is like a librarian trying to find a book that doesn't belong. To do this, the librarian has to pull every single book off the shelves, compare them one by one to the new book, and find the closest matches. This is slow and requires a huge amount of memory (shelf space) to hold all those books while they are being compared.

2. The New Idea: The "Spiking" Solution

The researchers wanted to try something different: Spiking Neural Networks (SNNs).

The Analogy: Think of a traditional computer (ANN) like a steady stream of water flowing through a pipe. It's constant. An SNN, however, is like a drumbeat. It only "fires" or "spikes" when a specific threshold is reached, just like a biological neuron in your brain.
Why it matters: Because they fire only when needed, SNNs are much more energy-efficient. They are also naturally good at handling data that changes over time (like a radio signal moving across a frequency chart).

3. The Innovation: Turning the Library into a Drum

The team took the existing "Librarian" (the NLN algorithm) and converted it into a "Drummer" (an SNN). This is called ANN2SNN conversion.

How the new method (SNLN) works:
Instead of pulling every book off the shelf to compare (which is slow), the new method uses the "drumbeat" nature of the SNN.

They feed the radio data into the SNN.
Because the SNN is "spiking," it creates a natural, shifting internal pattern over time.
Instead of searching for similar data in a huge library, the SNN generates its own "neighbors" just by letting the data run through it for a few seconds. The internal "noise" of the drumming creates enough variation to act as a comparison group.
The Result: You don't need to store a massive library of data in memory anymore. The computer generates the comparison on the fly, saving a huge amount of computing power and memory.

4. The Test Drive

The researchers tested this new "Drummer" method against the old "Librarian" method and a standard tool called AOFlagger using three different datasets:

HERA (Simulated): A clean, simulated environment.
LOFAR (Real): Real data from a telescope in Europe, full of messy, real-world interference.
Tabascal (New Simulation): A new dataset created specifically for this paper, focusing on interference from satellites (a growing problem as more satellites launch).

5. The Results: A Mixed Bag

In the Clean Room (HERA): The new SNN method performed just as well as the old method. It successfully identified the interference. Crucially, it did this without needing the heavy memory load of the old method. It was a "win" for efficiency.
In the Real World (LOFAR & Tabascal): The SNN method struggled a bit more. It wasn't as sharp at finding the subtle, messy interference in real data or the satellite signals as the old method was. The "masks" it produced were a bit blurrier.

The Bottom Line

This paper is a "proof of concept." It's the first time anyone has tried using Spiking Neural Networks to clean up radio telescope data.

What they achieved: They proved that you can convert a standard AI into a "spiking" AI and use it to detect radio interference. In ideal conditions, it works just as well as the current best methods but uses a clever trick to save memory and computing power.
What they didn't achieve: It's not perfect yet. In messy, real-world scenarios, the current "spiking" version is slightly less accurate than the traditional version.

The Takeaway: The researchers aren't saying "Throw away your old tools and use this new one tomorrow." Instead, they are saying, "We've built a prototype engine that runs on a different, more efficient fuel. It works well in the lab, and while it needs tuning for the rough roads of real life, it opens up a promising new path for the future of radio astronomy."

Technical Summary: RFI Detection with Spiking Neural Networks

Problem Statement
Radio Frequency Interference (RFI) poses a critical threat to the scientific output of radio telescopes, necessitating robust detection and mitigation strategies. While machine learning (ML) methods, particularly deep learning, have shown promise in outperforming traditional analytical approaches, they face significant hurdles: the requirement for vast amounts of high-quality labelled training data and substantial computational costs. Furthermore, existing ML-based RFI detection schemes often struggle with the "reality gap" when moving from simulated to real-world data. The Nearest-Latent-Neighbours (NLN) algorithm (Mesarcik et al., 2022a) offered a novel unsupervised approach by framing RFI detection as an anomaly detection task, reducing reliance on perfect labels. However, the NLN method requires storing and sampling a large pre-loaded test set in memory during inference to find latent neighbours, creating a computational bottleneck that limits scalability.

Methodology
This study introduces the Spiking Nearest Latent Neighbours (SNLN) algorithm, the first application of Spiking Neural Networks (SNNs) to astronomical data processing. The methodology follows a three-step pipeline:

Training an ANN Auto-Encoder: The authors first replicate the NLN approach by training a standard Artificial Neural Network (ANN) auto-encoder. This model is trained on uncontaminated radio data patches (selected via a traditional flagging algorithm like AOFlagger) to learn a representation of "noise-only" data.
ANN-to-SNN Conversion: The trained ANN is converted into a Spiking Neural Network using the SpikingJelly library. This conversion utilizes the equivalence between sigmoidal neurons in ANNs and Leaky-Integrate-and-Fire (LiF) spiking neurons. The process involves replacing neuron layers with spiking neurons and scaling weights to appropriate voltage values, requiring ReLU activation functions and the absence of MaxPool layers.
Modified Inference (SNLN): Unlike the original NLN, which searches for similar latent representations across a stored dataset, the SNLN leverages the intrinsic time-varying nature of SNNs. During inference, a single test patch is presented to the SNN auto-encoder for multiple time steps ( $T$ ). The stochastic spiking behavior of the neurons causes the internal latent state to vary over time. The algorithm averages the final $k$ output frames from the auto-encoder to generate "latent neighbours" implicitly. The difference between the original input and this averaged reconstruction is binarized to produce the RFI mask.

This approach eliminates the need to store a large reference dataset in memory during inference, as the SNN generates its own latent variations through temporal dynamics.

Key Contributions
The paper outlines four primary contributions to radio astronomy and RFI detection:

First Application of SNNs in Astronomy: This work marks the inaugural use of Spiking Neural Networks for astronomical data processing, opening new avenues for biologically inspired computation in the field.
The SNLN Algorithm: A novel RFI detection method that adapts the unsupervised NLN framework to SNNs, utilizing time-varying inference to simplify the downstream detection task and remove memory-intensive latent sampling steps.
New Tabascal Dataset: The authors introduce and publish a new, challenging dataset based on the MeerKAT telescope, simulated using the Tabascal package. This dataset focuses on satellite-based RFI (including moving low-orbit sources), representing an increasingly critical class of interference.
Performance Baseline: The study establishes a minimal performance baseline for SNNs in RFI detection, comparing them against the original ANN-based NLN and the traditional AOFlagger algorithm across diverse datasets.

Results
The authors evaluated SNLN against NLN and AOFlagger using three datasets: simulated HERA data, hand-labelled LOFAR observations, and the new Tabascal (MeerKAT-inspired) dataset. Performance was measured using AUROC, AUPRC, and F1-Score.

HERA Dataset: The SNLN approach remained competitive with the original NLN algorithm. In many configurations, SNLN achieved F1-scores comparable to or slightly exceeding NLN (e.g., 0.953 vs 0.939), despite the typical accuracy loss associated with ANN-to-SNN conversion. The results suggest that the binarization inherent in the SNN output may enhance resilience.
LOFAR Dataset: On real-world LOFAR data, SNLN exhibited a performance drop compared to the original NLN. While NLN outperformed AOFlagger in AUROC, SNLN struggled to match the original ANN's performance, producing less sharply defined masks.
Tabascal Dataset: On the satellite-focused Tabascal dataset, SNLN again showed a performance degradation compared to NLN, though the drop was less dramatic than in the LOFAR case. The dataset's specific challenge—RFI localised in time but spanning all frequencies—proved difficult for both methods, particularly where training data lacked heavily contaminated patches.

Across all datasets, the SNLN method successfully maintained accuracy while completely removing the compute and memory-intensive latent sampling step required by NLN.

Significance and Claims
The paper modestly claims that this work demonstrates the viability of SNNs as a promising avenue for machine-learning-based RFI detection. The authors emphasize that while the current ANN-to-SNN conversion does not exceed the performance of the original ANN (a known limitation of conversion methods), it successfully establishes a baseline for future research.

The primary significance lies in proving that SNNs can be applied to astronomical tasks and that their unique time-varying properties can be exploited to simplify inference pipelines, potentially offering orders-of-magnitude improvements in energy efficiency when deployed on neuromorphic hardware. The authors position this work not as a final solution, but as a foundational step toward training fully spiking networks from scratch, which could better exploit SNN dynamics for more sophisticated input encodings and improved performance on real-world, energy-constrained radio telescope systems.