🔭 astrophysics

Spiking Neural Networks for Radio Frequency Interference Detection in Radio Astronomy

This paper demonstrates that Spiking Neural Networks, enhanced by divisive normalization preprocessing, offer an energy-efficient and effective solution for real-time Radio Frequency Interference detection in radio astronomy by reformulating the problem as a time-series segmentation task.

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

Published 2026-01-23

📖 4 min read☕ Coffee break read

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, ancient whisper from deep space. This is what radio astronomers do: they use giant telescopes to catch radio waves from the universe. But there's a problem. Just like trying to hear a whisper at a rock concert, these cosmic signals are constantly drowned out by "noise" from human-made sources like satellites, cell phones, and radar. This noise is called Radio Frequency Interference (RFI).

To save the data, astronomers have to find the noise and cut it out. Usually, they treat the radio data like a picture (a spectrogram) and use standard computer brains (Artificial Neural Networks) to find the bad spots. But this takes a lot of energy and time, and often requires waiting until the whole "picture" is taken before they can start cleaning it.

This paper proposes a different kind of computer brain: a Spiking Neural Network (SNN).

The Biological Brain vs. The Digital Brain

Think of a standard computer brain (ANN) like a person who never stops talking. They are constantly sending a stream of information, whether there is something important to say or not. This is accurate but very tiring (energy-intensive).

A Spiking Neural Network is more like a biological brain. It stays silent until something specific happens, and then it sends a quick "spike" or a single tap to say, "Hey, I noticed something!" This makes it incredibly energy-efficient and fast at reacting to changes over time.

The New Approach: Listening to the Rhythm, Not Just the Picture

The authors realized that radio data isn't just a static picture; it's a movie that changes every second. Instead of treating the data like a photo to be analyzed all at once, they treated it like a time-series segmentation task.

The Old Way: Wait for the whole movie to finish, then look at the whole frame to find the noise.
The New Way: Watch the movie frame-by-frame as it happens. If a "spike" of noise appears, flag it immediately.

To do this, they had to translate the radio data into a language the spiking brain could understand. They tried several "encoding" methods, which are like different ways of translating a song into Morse code:

Latency Encoding: If the signal is loud, the "tap" happens immediately. If it's quiet, the tap is delayed.
Delta-Exposure: A mix of methods that lets the brain "look" at the signal for a few moments before deciding to tap.

The Secret Sauce: "Divisive Normalization"

One of the biggest hurdles was that the background noise in the radio data was like a foggy window, making it hard to see the actual interference. The authors introduced a pre-processing step inspired by how our eyes adapt to bright light.

Imagine you are in a dark room, and a bright light turns on. Your eyes adjust so you can see the light without being blinded. The authors created a digital version of this adjustment. They subtracted the "foggy" background from the data, making the "bad" interference stand out clearly against a clean background. This simple step made the spiking brain much better at spotting the noise.

The Results: A Tale of Two Datasets

The team tested their system on two types of data:

The Synthetic Test (HERA): This was a clean, computer-generated simulation of radio data.
- Result: The spiking brain performed almost as well as the best existing methods, but with a much simpler structure. It successfully learned to spot the interference in the "movie" of data.
The Real-World Test (LOFAR): This was actual data from a real telescope in the Netherlands, filled with messy, unpredictable real-world noise.
- Result: This was much harder. The spiking brain didn't quite reach the performance of the most advanced current tools, but it did better than previous attempts to convert standard brains into spiking ones. It showed that while the method works, real-world data is a tough opponent that needs more training and refinement.

Why This Matters

The paper concludes that Spiking Neural Networks are a viable path forward for radio astronomy. They offer a way to process massive amounts of data in real-time with minimal energy.

Think of it this way: If standard computers are like a heavy, fuel-guzzling truck that carries the whole load at once, these Spiking Neural Networks are like a fleet of agile, electric scooters that pick up and drop off small pieces of information exactly when needed.

While the technology isn't perfect yet (especially with the messy real-world data), this study proves that using "biological-style" computing to listen to the universe is possible and holds great promise for the future of how we explore space.

Technical Summary: Spiking Neural Networks for Radio Frequency Interference Detection in Radio Astronomy

Problem Statement
Radio astronomy faces a critical challenge in Radio Frequency Interference (RFI) detection, where unwanted terrestrial or human-made signals contaminate observational data. Traditional detection methods often rely on cumulative-sum algorithms or machine learning models based on Convolutional Neural Networks (CNNs) and Vision Transformers. While effective, these approaches typically treat spectrograms as static 2D images, requiring the completion of an entire observation before processing can begin. This limitation necessitates storing and processing vast amounts of data in high-performance computing facilities, many of which are destined to be flagged as RFI and discarded. The field requires adaptive, data-driven detection methods capable of near real-time operation with minimal energy consumption and reduced training data requirements.

Methodology
The authors reformulate RFI detection as a time-series segmentation task rather than a 2D semantic segmentation problem. This approach leverages the temporal dynamics inherent to Spiking Neural Networks (SNNs). The core methodology involves:

Data Representation: Input "visibilities" (complex-valued data varying in frequency, time, and baseline) are converted into spike trains. The study investigates seven encoding methods, including latency, rate, delta-modulation, delta-exposure, and three variations of step-forward encoding.
Network Architecture: The authors employ feed-forward SNNs utilizing first-order and second-order Leaky Integrate and Fire (LiF) neurons. The networks are trained using Backpropagation Through Time (BPTT). To manage memory constraints associated with long spike trains, spectrograms are sliced into $32 \times 32$ pixel patches.
Divisive Normalisation Pre-processing: A novel pre-processing step inspired by divisive normalisation in neuroscience is introduced. This technique subtracts a weighted sum of the preceding time step's frequency channels from the current step. This acts as a lightweight kernel that reduces background gradients, increases spike sparsity, and enhances the contrast between RFI features and background noise.
Training and Optimization: Extensive hyper-parameter optimization was conducted using Optuna, varying network depth, width, neuron decay rates ( $\beta$ ), and exposure times. The study compares SNN performance against identically sized Artificial Neural Networks (ANNs) and existing algorithmic baselines.

Key Contributions

Task Reformulation: This work presents the first application of SNNs to RFI detection formulated specifically as a time-series segmentation problem, requiring the network to output predictions spanning the same temporal and spatial resolution as the input.
Divisive Normalisation: The introduction of a divisive normalisation-inspired pre-processing step that universally improves detection performance across multiple encoding strategies by increasing spike sparsity and feature contrast.
Direct Training on Real Data: To the authors' knowledge, this is the first study to successfully train SNNs from scratch on real radio astronomy data (from the Low-Frequency Array, LOFAR), moving beyond previous ANN-to-SNN conversion approaches.
Comprehensive Encoding Analysis: A systematic evaluation of various spike encoding methods (latency, delta-exposure, step-forward) and their interaction with neuron types (first vs. second-order LiF) and pre-processing techniques.

Results
The study evaluates performance on two datasets: a synthetic dataset simulating the Hydrogen Epoch of Reionisation Array (HERA) and a real dataset from the LOFAR instrument.

Synthetic (HERA) Dataset: The SNN approach achieved competitive, near state-of-the-art performance. Specifically, latency encoding with divisive normalisation yielded an AUROC of 0.968, AUPRC of 0.914, and F1-Score of 0.907. These results are comparable to traditional algorithms and ANN-based methods, despite the SNNs having significantly smaller model sizes (e.g., ~74k parameters vs. ~800k for UNet-style models).
Real (LOFAR) Dataset: Performance on real data was lower than on synthetic data, reflecting the increased noise and variability of real-world observations. However, the SNNs still outperformed equivalently sized ANNs. The best-performing SNN configuration (Step-Forward direct encoding with divisive normalisation) achieved an AUPRC of 0.693, surpassing previous ANN-to-SNN conversion results.
Neuron Types: Interestingly, first-order LiF neurons performed better on the synthetic dataset, while second-order LiF neurons showed better relative performance on the noisy real-world dataset, suggesting the more complex model handles real-world noise more adeptly.
Efficiency: Preliminary estimates suggest that deploying this time-series segmentation approach on neuromorphic hardware (e.g., SynSense Xylo) could achieve power consumption in the range of <50 mW per baseline, significantly lower than conventional high-performance computing requirements.

Significance and Claims
The paper claims that SNNs offer a viable path toward efficient, real-time RFI detection in radio astronomy. By reformulating the problem as time-series segmentation, the authors demonstrate that SNNs can capture complex temporal patterns without the computational overhead of 2D image-based processing. The study highlights that while a performance gap remains between synthetic and real-world results, the approach is promising for future deployment. The authors position this work as a foundational step, suggesting that integrating neuromorphic computing into radio astronomy could provide unique advantages in energy efficiency and real-time processing capabilities, potentially transforming how data-intensive scientific fields handle streaming observational data. The paper modestly acknowledges that further developments in architecture (e.g., spiking LSTMs, de-noising capabilities) and training strategies (e.g., self-supervised methods) are necessary to close the gap with state-of-the-art performance on real-world data.

The Biological Brain vs. The Digital Brain

The New Approach: Listening to the Rhythm, Not Just the Picture

The Secret Sauce: "Divisive Normalization"

The Results: A Tale of Two Datasets

Why This Matters

More like this