Radio telescopes must throw away the parts of their data corrupted by human transmitters (phones, satellites, TV), a step called radio frequency interference (RFI) flagging. Future arrays will need this done in real time at thousands of antennas, so power per flagger matters. Spiking neural networks, which communicate with sparse binary pulses and replace multiplications with additions, have been shown to flag RFI on milliwatt neuromorphic chips. There is a catch. Every strong spiking result is on simulated data. On real LOFAR observations with hand-drawn labels, the best directly trained spiking flagger scores F1 0.162, while the decades-old heuristic AOFlagger scores 0.57.
This proposal argues that the collapse comes mostly from the input rather than the spiking model. Real data has a sky background and receiver gain that drift across channels and over time. Simulated training data does not. AOFlagger survives this because it repeatedly fits and subtracts a smooth background while ignoring pixels it has already flagged. We propose to build that same loop into the spiking network. Each frequency channel gets a slow "background" neuron that tracks the channel's baseline and inhibits the fast detector neurons. The flag outputs feed back to freeze that tracking whenever interference is present. The network is then trained on real data, using AOFlagger masks as imperfect labels. The goal is one causal, streaming network that runs entirely on the chip and matches or beats AOFlagger on real hand-labelled data at milliwatt power.
Core Claim
A spiking flagger whose per-channel background estimate adapts over time and pauses while interference is flagged will close most of the sim-to-real gap on LOFAR. The test is the hand-labelled F1 score: it should move from the 0.162 reported for the best spiking model toward or above AOFlagger's 0.57. Three ingredients are separable, and the experiments are built to measure each one's contribution: adaptive normalization, gating that normalization on the model's own flags, and weak labels from real data. We also claim the whole loop fits within the fan-in and neuron budgets of a single SynSense Xylo 2 chip. That matters because sharded multi-chip models have been shown to perform worse than small monolithic ones.
Where This Sits
The evidence for the gap comes from a single research group's line of work. RFI detection with spiking neural networks converts a trained autoencoder into a spiking one and uses spike noise in place of a nearest-neighbour search. It reaches F1 0.953 on simulated HERA data but about 0.40 on LOFAR, below AOFlagger (0.57), and 0.74 on the Tabascal satellite set, below AOFlagger (0.87).
Spiking neural networks for radio frequency interference detection in radio astronomy recasts flagging as segmenting a multichannel time series and adds divisive normalization, a contrast-enhancing preprocessing step. It reaches F1 0.942 on HERA but only 0.162 on LOFAR (AUPRC 0.241), and it reused the HERA training recipe unchanged. We read that result as the clearest sign that the input statistics, not the neuron model, are what break. That paper's normalization is a fixed preprocessing step. Ours is causal, changes over time, runs on the chip, and responds to the network's own flags.
The hardware constraints on our design also come from this line of work. Polarisation-Inclusive Spiking Neural Networks measured 11.8–53.6 mW on Xylo, but only by cutting inputs into 16×16 patches. Neuromorphic Astronomy: An End-to-End SNN Pipeline supplies the method we reuse: snnTorch training, a penalty on neurons whose input count exceeds what a chip core allows, export through the Neural Intermediate Representation (NIR), and deployment on Xylo 2 at about 100 mW with F1 0.973 on HERA. Its finding that sharded large models underperform small single-chip models is a negative result we take seriously. We therefore do not propose a bigger network such as a spiking U-Net (Deep multi-threshold spiking-UNet). We propose a better-conditioned small one.
Advancing RFI-Detection with Liquid State Machines uses a fixed random recurrent spiking network with a trained readout. It reports trouble with instantaneous wide-band bursts. That warns us that slow adaptation must not smear sudden broadband events into the background, which is exactly what the flag gate is designed to prevent.
On the classical side, AOFlagger, as described in The Low-Frequency Environment of the Murchison Widefield Array, fits a smooth background surface and applies an iterative threshold. It is both our teacher and our target. Our mechanism is a spiking, streaming analogue of its fit-then-mask loop.
Mechanism
The model takes as input the visibility amplitude $x_c(t)$ of channel $c$ at time sample $t$ from a 64-channel stream. Each channel has a slow integrator whose target behaviour is a gated running baseline:
$$b_c(t) = b_c(t-1) + \frac{1 - f_c(t-1)}{\tau_s}\,\big(x_c(t) - b_c(t-1)\big),$$
Here $\tau_s$ is a long learned time constant, measured in time samples, and $f_c(t-1)\in\{0,1\}$ is the network's own flag output for that channel at the previous step. The detector neurons receive the current $x_c(t) - b_c(t)$, injected directly as a driving current rather than encoded into spikes. The liquid state machine study found that this direct encoding beat latency and rate encoding.
In the spiking implementation, $b_c$ is carried by a slow leaky neuron that makes subtractive inhibitory connections onto that channel's detectors. The gate is implemented as inhibitory feedback from the channel's flag neuron onto its background neuron. Both use only ordinary synapses and recurrent connections, with no custom neuron features. The detector layer uses second-order leaky integrate-and-fire neurons, which art_0349 found better suited to noisy LOFAR data. The detector layer also includes one pooling neuron that sums activity across all channels. That neuron lets a burst lighting up many channels at once trigger flags in a single step, which addresses the liquid state machine's reported blind spot. Training uses surrogate-gradient backpropagation through time, meaning the step function is given a smooth stand-in derivative so gradients can flow, together with the fan-in penalty from art_bb006.
Research Program
Training data comes from real observations. On the LOFAR training portion, AOFlagger masks serve as weak labels; we generate them ourselves if the release does not include them. HERA simulations are used for pretraining. Evaluation uses only the held-out hand-labelled LOFAR spectrograms. Tabascal tests the moving-satellite case, and HERA checks that nothing regresses on the setting where earlier work was strong.
The analysis goes beyond headline scores. We will split errors by interference type: narrow-band persistent, wide-band instantaneous, and slowly drifting. We will also check where the student disagrees with its AOFlagger teacher, to see whether a student can exceed a noisy teacher on hand labels. That is plausible only if the teacher's errors are unstructured, so we measure it directly rather than assume it. Finally, the chosen model goes through the NIR-to-Xylo 2 flow, and we measure power on the chip.
Baselines
The naive reference is the same small spiking network trained on HERA and applied to LOFAR, with the fixed divisive normalization of art_0349. Its reported LOFAR result is F1 0.162, AUROC 0.633, and AUPRC 0.241. We will re-run it in our pipeline so the scores are comparable.
The strongest existing method on real data is AOFlagger. On LOFAR it reports F1 0.57 against hand labels, the figure given in RFI detection with spiking neural networks, and 0.87 on Tabascal. We will run it with its default LOFAR strategy on the identical test spectrograms. It needs no training budget, but we will give it the same limited hand-labelled validation split to tune its threshold, so neither side gets an advantage from tuning. Beating it means an F1 above 0.57 on hand labels. Matching it within noise at milliwatt power, with no host-side preprocessing, would still be a meaningful result, and we would report it as such rather than as a win.
We also include the unsupervised spiking method (SNLN, about 0.40) and its non-spiking parent (NLN, 0.48) from the same paper. We include a continuous U-Net in the style of Radio frequency interference mitigation using deep convolutional neural networks, trained on the same weak labels. No LOFAR result is reported on Lacuna for the U-Net; its reported figure is F1 about 0.85 on HIDE simulations.
Three ablations isolate our key ingredient. The first removes the flag gate, so the background always adapts. The second replaces the spiking background neurons with an exact running median computed on the host. The third trains the full architecture on HERA only, without the LOFAR weak labels. All spiking variants share neuron count, fan-in limit, number of training epochs, and hyperparameter search budget. The U-Net gets the same search budget.
Why not just…?
Why not just run AOFlagger, which needs no training? On real data this is the bar to clear, and today it wins. The proposal succeeds only by matching its accuracy while running as a streaming, milliwatt process at every antenna station. AOFlagger's host-CPU cost at scale is not quantified in the file, so we will measure its throughput per CPU core on the test set. That lets the energy comparison be stated honestly rather than asserted.
Why not just train the existing spiking flagger on AOFlagger labels from real data? This is the most serious threat. It may account for most of the improvement. The "HERA only" ablation and the "no adaptive background" configuration trained on the same weak labels measure it directly. If weak labels alone reach AOFlagger's level, we will report that the architectural contribution is minor and reframe the result as a data-recipe finding.
Why not just normalize on the host with a running median before the chip? For accuracy this may be equivalent, and the host-median ablation tests it. Our argument for doing it on the chip is about the whole system, not accuracy. It removes a second compute device per stream, and the flag gating is awkward to implement on the host because it needs the model's flags from the previous step. If host normalization is both as accurate and cheap in measured power, we will say so.
Why not just use a small quantized CNN on an edge processor? The file's argument against this rests on power: about 5 W for a Jetson-class device versus tens of milliwatts on Xylo. We will train a CNN with matched parameter count on the same labels and report its accuracy alongside an energy estimate from operation counts. If it is much more accurate, the case for spiking hardware narrows to deployments where power is the binding constraint.
Decisive Checks
The central result is hand-labelled LOFAR F1, AUROC, and AUPRC, with confidence intervals from repeated splits and random seeds. The proposal is supported if the full model beats both the 0.162 naive result and the "no gate" ablation by a clear margin and reaches AOFlagger's 0.57. It also needs to hold its HERA F1 near the reported 0.94–0.97, and to fit and run on a single Xylo 2 chip at power comparable to the roughly 100 mW measured in art_bb006.
On Tabascal, we check whether the flag gate prevents satellite passes from being absorbed into the background. The result to beat is AOFlagger's 0.87.
The result that would falsify the core claim is the model with the adaptive background but no gate performing as well as the gated one. That would mean slow drift, not interference leaking into the background estimate, was the whole story.
References
- RFI detection with spiking neural networks
- Spiking neural networks for radio frequency interference detection in radio astronomy
- Polarisation-Inclusive Spiking Neural Networks for Real-Time RFI Detection in Modern Radio Telescopes
- Neuromorphic Astronomy: An End-to-End SNN Pipeline for RFI Detection Hardware
- Deep multi-threshold spiking-UNet for image processing
- Advancing RFI-Detection in Radio Astronomy with Liquid State Machines
- The Low-Frequency Environment of the Murchison Widefield Array: Radio-Frequency Interference Analysis and Mitigation
- Radio frequency interference mitigation using deep convolutional neural networks