🔭 astrophysics

Advancing RFI-Detection in Radio Astronomy with Liquid State Machines

This paper introduces a novel Liquid State Machine architecture utilizing second-order Leaky Integrate-and-Fire neurons and hybrid SNN-ANN readout layers to achieve competitive 98% per-pixel accuracy in detecting Radio Frequency Interference on HERA spectrograms, effectively advancing spiking neural networks for fine-grained spatio-temporal segmentation in radio astronomy.

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, distant whisper from a friend (a signal from deep space) while standing in a crowded, noisy stadium. The stadium is full of people shouting, music blaring, and phones ringing. In the world of radio astronomy, these "shouts" are called Radio Frequency Interference (RFI). They come from our own technology—cell towers, satellites, and Wi-Fi—and they drown out the precious signals from the universe that scientists want to study.

For a long time, scientists have used complex rulebooks and manual checking to find and ignore these "shouts." But as telescopes become more sensitive and collect more data, these old methods are struggling to keep up. They are too slow and require too much human effort.

This paper introduces a new way to solve this problem using a type of computer brain inspired by how real biological brains work. Here is a breakdown of their approach:

1. The New "Brain": Liquid State Machines

Instead of using standard computer chips that process data like a calculator (step-by-step), the authors used Spiking Neural Networks (SNNs). Think of these like a real brain: neurons don't just fire numbers; they fire tiny electrical "spikes" when they get enough stimulation.

Specifically, they used a system called a Liquid State Machine (LSM).

The Analogy: Imagine dropping a pebble into a pond. The water ripples, swirls, and changes shape based on the pebble's size and where it landed. The water is the "liquid" (the network), and the ripples are the "spikes."
How it works: The radio data is dropped into this "liquid pond." The way the water moves (the ripples) changes over time. The system doesn't need to learn how the water moves; it just needs to learn how to read the ripples to figure out if a "shout" (interference) happened. This makes it very fast and efficient.

2. Speaking the Language of Spikes

To get the radio data into this "liquid brain," the authors had to translate it into a language the brain understands. They tested three different ways to do this translation (encoding):

Latency Encoding: Like saying, "The louder the sound, the sooner you shout."
Rate Encoding: Like saying, "The louder the sound, the more often you shout."
Direct Encoding: Like turning the volume knob directly into an electrical current.

The Surprise: Previous studies thought "Latency" (timing) was the best way to talk to these brains. However, the authors found that Direct Encoding (turning the signal directly into current) worked the best in their specific setup. It was like finding that speaking directly to the brain was more effective than trying to time your words perfectly.

3. The "Reader": A Hybrid Brain

Once the "liquid" creates its ripples, a "reader" (called a readout layer) has to interpret them to say, "Yes, that's interference," or "No, that's a real signal."

The authors tested three types of readers: a simple one (Linear), a slightly smarter one (ReLU), and a very advanced one based on Transformers (the same technology that powers modern AI chatbots).
The Innovation: They used the "spikes" from the liquid brain as the memory for the advanced Transformer reader. It's like having a librarian who not only reads the books but also remembers the entire history of the library's ripples to make a better decision.
The Result: The more complex the reader, the better the system performed. The combination of the "Liquid Brain" and the "Transformer Reader" was the most successful.

4. The Results

The team tested their system on a simulated dataset that mimics a real radio telescope (HERA).

Accuracy: The system correctly identified about 98% of the pixels in the data (meaning it knew what was signal and what was noise).
The Catch: While it was very accurate overall, it still struggled a bit with very sudden, wide-band "shouts" (interference that happens instantly and covers many frequencies).
Comparison: While other AI methods (like standard deep learning) are currently slightly better at this specific task, those methods are much heavier and require more computing power. The Liquid State Machine approach is much lighter, faster to train, and uses less energy because it only needs to train the "reader," not the whole "liquid brain."

Summary

In short, this paper shows that we can use a "liquid" computer brain to filter out the noise of our modern world to hear the whispers of the universe. By using a specific type of neuron (second-order Leaky Integrate-and-Fire) and pairing it with a smart "Transformer" reader, they created a system that is highly accurate and much more efficient than current methods. It's a promising step toward building radio telescopes that can automatically clean their own data in real-time, without needing a human to sit there and watch the screens.

Technical Summary: Advancing RFI-Detection in Radio Astronomy with Liquid State Machines

Problem Statement
Radio Frequency Interference (RFI) from anthropogenic sources poses a critical challenge to current and future radio telescopes, such as the Square Kilometer Array (SKA). Traditional RFI detection ("flagging") relies on complex heuristic algorithms and manual inspection, which struggle to scale with the increasing data volume and sensitivity of modern instruments. While machine learning approaches, particularly Convolutional Neural Networks (CNNs), have shown success in treating RFI detection as a 2D semantic segmentation problem, they often incur high computational costs that limit real-time operational applicability. Furthermore, existing Spiking Neural Network (SNN) attempts at RFI detection have largely been limited to first-order Leaky Integrate-and-Fire (LiF) neurons and have not fully leveraged the temporal dynamics inherent in radio astronomy spectrograms.

Methodology
This work formulates RFI detection as a fine-grained spatio-temporal time-series segmentation problem rather than a static 2D image segmentation task. The authors employ Liquid State Machines (LSMs), a class of reservoir computing using SNNs, to process the time-varying nature of radio spectrograms.

Architecture: The core model consists of a recurrently connected reservoir of spiking neurons and a trainable readout layer. The reservoir is initialized with an 80/20 ratio of excitatory to inhibitory connections.
Neuron Model: A key innovation is the use of second-order Leaky Integrate-and-Fire (LiF) neurons, characterized by separate synaptic and membrane time constants, noise injection, and bias currents. This marks the first application of this specific neuron architecture to RFI detection.
Encoding Strategies: Three spike-encoding methods were tested to convert input visibility data into spike trains:
1. Latency Encoding: Spike delay proportional to input value.
2. Rate Encoding: Input intensity converted to spike frequency.
3. Direct Encoding: Input values treated as driving currents (identity function) for the exposure time.
Readout Layers: Three increasingly complex readout architectures were evaluated to map the reservoir's spiking activity to the output mask:
1. Linear Layer.
2. Linear Layer with ReLU activation.
3. Transformer Decoder (utilizing the spiking history as memory).
Dataset: The models were trained and tested on synthetic data derived from the Hydrogen Epoch of Reionization Array (HERA) simulator. The data consists of 512×512 spectrograms split into 32×32 chunks, with approximately 3% RFI contamination.
Training: Only the readout layer is trained, significantly reducing computational complexity compared to training all network parameters. Hyper-parameters (input sparsity, exposure time, reservoir size) were optimized using Optuna to maximize the F1-Score.

Key Results
The study evaluated performance using Accuracy, AUROC, AUPRC, and F1-Score across 25 trials per configuration.

Encoding Performance: Direct encoding consistently outperformed latency and rate encoding across all readout layers. Latency encoding performed significantly worse, a deviation from prior SNN studies, likely due to the fixed nature of the reservoir parameters in this formulation.
Readout Complexity: Increasing the complexity of the readout layer (Linear $\to$ ReLU $\to$ Transformer) yielded significant performance improvements. The Transformer decoder achieved the highest scores, particularly when paired with direct encoding.
Best Performance: The optimal configuration (Direct Encoding + Transformer Readout) achieved a per-pixel accuracy of 98.24% and an F1-Score of 0.743.
Comparison to State-of-the-Art: While the LSM approach did not surpass the F1-Scores of fully trained ANN-based methods (e.g., ANN Auto-Encoder with F1=0.910) or ANN2SNN conversions, it offers a distinct advantage in training simplicity. The LSM approach requires training only the readout layer, whereas competing methods often involve training millions of parameters in convolutional or auto-encoder architectures.
Limitations: The authors note that the LSM struggles with instantaneous, wide-band, and short-duration RFI events, as visualized in the inference examples.

Significance and Claims
The paper claims to advance the sophistication of SNN techniques applied to radio astronomy by:

Expanding the Application Domain: Successfully applying LSMs to fine-grained spatio-temporal segmentation, a task previously dominated by classification problems.
Architectural Innovation: Introducing second-order LiF neurons and a hybrid SNN-ANN architecture (LSM reservoir with Transformer readout) to RFI detection.
Efficiency: Demonstrating that competitive performance can be achieved with significantly reduced training complexity by leveraging the reservoir's dynamics and training only the readout layer.
Biological Plausibility: Moving beyond simple first-order neuron models to more biologically plausible second-order dynamics.

The authors conclude that while current LSM performance lags behind fully trained deep learning models, the approach offers a promising, computationally efficient direction for real-time RFI detection. They suggest future work could involve multi-column reservoirs to handle varying RFI timescales and unsupervised learning paradigms.