Radio telescopes must throw away the parts of their data corrupted by human-made radio frequency interference (RFI): phone networks, digital TV, satellites, lightning. Next-generation arrays produce too much data to ship it all to a GPU cluster, so the cheap option is to flag interference at the antenna. Spiking neural networks (SNNs) are networks whose neurons accumulate input in a slowly leaking internal state and emit binary pulses. They can run on milliwatt neuromorphic chips, and a spiking RFI detector has already run on physical hardware at about 100 mW. These detectors work well on simulated data and fail on real telescopes. On the real LOFAR dataset, the best supervised SNN reaches F1 0.162 against 0.942 on simulated HERA data, and the classical AOFlagger heuristic beats every spiking model reported there.
This note proposes a specific way to close that gap. A detector trained in simulation should be moved to a new telescope by retuning only its neurons' leak rates, meaning how quickly each neuron forgets past input. Its connection weights stay fixed. This retuning would use a handful of labeled spectrograms from the target telescope. The reasoning is physical. Different instruments sample time differently, and different sites suffer interference of different durations, so much of the shift between simulation and reality is a shift in timescale. Leak rates are a few hundred numbers. They map onto per-neuron registers on neuromorphic chips, and they can be re-flashed per site without retraining.
Core Claim
Prior spiking RFI work fixes neuron decay by hand or by black-box search. The learnable-decay literature trains decay by gradient descent but has only been evaluated on vision, event-camera and spoken-digit benchmarks. We claim two things. First, an SNN detector trained on simulated HERA data with gradient-learned, per-neuron decay can be adapted to real LOFAR data, and to the moving-satellite Tabascal simulation, by updating only its decay parameters on 1 to 20 labeled target spectrograms. Second, this decay-only adaptation recovers most of the gain of full fine-tuning and beats both AOFlagger and adapting the input gains. The comparison with input gains is the falsifiable core. If simply rescaling inputs per frequency channel recovers the same gain, then the domain shift is about amplitude, not timescale, and the claim fails.
Relation to Prior Work
The spiking RFI line of work supplies both the problem and the motivating evidence.
- Spiking neural networks for radio frequency interference detection in radio astronomy recasts flagging as per-timestep segmentation of a multichannel time series. It grid-searches fixed first- and second-order neuron parameters and reports the HERA-to-LOFAR collapse. Its authors say transfer strategies were not tried.
- Supervised Radio Frequency Interference Detection with SNNs tunes the decay factor $\beta$ externally with Optuna and evaluates only on HERA.
- Advancing RFI-Detection in Radio Astronomy with Liquid State Machines uses a frozen spiking reservoir. Its fixed temporal memory smears single-timestep wide-band bursts, and the authors call for multiple timescales. This is a constraint on our design: a fixed slow memory demonstrably fails on one interference class, so our model must be able to learn near-zero memory in some neurons.
- RFI detection with spiking neural networks converts an anomaly-detection autoencoder into an SNN after training. It trails AOFlagger on LOFAR and Tabascal, and its authors attribute this to the conversion ignoring neuron dynamics.
- Neuromorphic Astronomy: An End-to-End SNN Pipeline for RFI Detection Hardware deploys a hardware-aware SNN on the SynSense Xylo 2 chip. It finds that compact single-chip models beat models split across chips, and it leaves real-data validation and learnable dynamics open.
None of these five papers trains decay by gradient, and none adapts any part of the model to a new telescope.
The method comes from the learnable-decay literature.
- Incorporating Learnable Membrane Time Constant to Enhance Learning of Spiking Neural Networks (PLIF) learns a per-layer time constant through a sigmoid reparameterization. It shows that the network self-corrects from a poor initial time constant.
- DA-LIF separates the input weighting $\alpha$ from the decay $\beta$. It finds that layers specialize, with shallow layers retaining more memory.
- HetSyn learns a decay per synapse. Its networks keep accuracy with 80% of time constants frozen, which suggests timescales are a compact, robust parameter set.
We reuse PLIF's parameterization at per-neuron rather than per-layer granularity, and DA-LIF's separation of $\alpha$ from $\beta$. We avoid HetSyn's per-synapse decay because it doubles the parameter count and does not fit chip memory. What none of these papers tests is whether decay parameters are the right place to absorb a domain shift. That question is ours.
Research Program
The detector follows the time-series formulation of prior work. At each time sample, the input is the vector of visibility amplitudes across frequency channels. These are fed as direct driving currents, which the reservoir study found best, and latency encoding, which the supervised SNN studies found best, is tested as a control. The network is a compact two- to three-layer spiking network, chosen so that it fits within a single neuromorphic chip. Each hidden neuron $i$ follows
$$u_{i,t} = \beta_i\, u_{i,t-1} + \alpha_i \textstyle\sum_j W_{ij} x_{j,t} - V_{th}\, s_{i,t-1}, \qquad \beta_i = \sigma(\theta_i),$$
where $u_{i,t}$ is the neuron's internal state at time $t$, and $\beta_i$ is its learned leak rate. A value of $\beta_i$ near 0 means the neuron forgets immediately; a value near 1 means it remembers long. The weight $\alpha_i$ scales the neuron's input, $W$ is the fixed-structure weight matrix, $x_{j,t}$ is input channel $j$, $V_{th}$ is the firing threshold, $s_{i,t-1}\in\{0,1\}$ is the previous spike, and $\sigma$ is the sigmoid. Training uses surrogate-gradient backpropagation through time on simulated HERA, with a class-weighted loss for the roughly 3% interference rate.
Adaptation to a target telescope updates only $\theta$, a few hundred scalars, on $k$ labeled target spectrograms, with $k \in \{1, 5, 20\}$. Two further variants are tested. One is label-free: it adapts on target data using AOFlagger masks as pseudo-labels, restricted to regions where AOFlagger is confident. The other is hardware-constrained: $\beta$ is restricted during adaptation to a small discrete set of values, with straight-through gradients, matching what the Xylo 2 configuration allows. We will check the exact constraint on the chip rather than assume it.
The analysis has two parts. First, we inspect how adaptation moves the learned decays, for example whether LOFAR pushes some neurons toward shorter memory. Second, we stratify detection gains by interference morphology: narrow-band persistent, wide-band instantaneous, and moving satellite tracks.
Baselines
The do-nothing reference is the simulation-trained SNN applied unchanged to the target data. It is the same network, so the comparison directly measures what adaptation buys. The strongest existing method on real data is AOFlagger. RFI detection with spiking neural networks reports AOFlagger at F1 0.57 on LOFAR and 0.87 on Tabascal, against 0.40 and 0.74 for its converted SNN. The best published supervised SNN on LOFAR scores F1 0.162 (AUPRC 0.241) in Spiking neural networks for radio frequency interference detection in radio astronomy. Mask conventions differ across papers, machine-generated versus expert-labeled. We will therefore rerun AOFlagger with its standard strategy on our exact test split and score it against expert masks, and we will also tune it per telescope on the same $k$ labeled spectrograms. That gives it the same supervision budget as our method. Beating it means higher F1 and AUPRC on held-out expert-labeled LOFAR spectrograms with $k \le 20$, averaged over at least five random draws of the adaptation set.
On HERA, the in-domain reference is F1 0.973 from Neuromorphic Astronomy. Our model should stay near this figure. That checks that per-neuron decay does not cost in-domain accuracy.
Three parameter-efficient competitors form the key ablations. Each uses the same $k$ spectrograms, the same optimizer, and an equal tuning budget of the same number of gradient steps and learning-rate trials:
- Readout-only: updates only the output layer.
- Gain-only: updates a per-channel input scale and offset, which is the analogue of recalibrating normalization. It is matched to the decay-only method in parameter count.
- Full fine-tuning: updates all weights and serves as an upper reference.
A fourth ablation trains with fixed, Optuna-searched $\beta$ as in Supervised Radio Frequency Interference Detection with SNNs, then re-searches $\beta$ on the target with the same number of trials. This isolates whether gradients matter or merely having a timescale knob.
Why not just…?
Why not just use AOFlagger on real data, since it already wins? On LOFAR it does win today, at 0.57 against at most 0.48 for any learned spiking or autoencoder model in the file. Two weaknesses leave room for a learned detector. AOFlagger needs manual retuning per site, and it has documented trouble with sub-second broadband bursts from atmospheric ducting at the Murchison Widefield Array (The Low-Frequency Environment of the Murchison Widefield Array). Our per-morphology breakdown will show whether we beat it overall or only on burst classes. If the adapted SNN matches AOFlagger but does not exceed it, the honest result is "AOFlagger-level accuracy at milliwatt power," and we will report it that way.
Why not just recalibrate input gains or normalization? This is the real threat, because real telescopes have time-varying antenna gains, and the divisive normalization step in the 2025 SNN paper already helped. The gain-only baseline is matched in parameter count, and the combined gain-plus-decay condition shows whether the two are complementary. If gain-only matches decay-only, we drop the timescale claim. A secondary test strengthens the case if decay still wins: we resample HERA in time to mimic a different instrument cadence, a pure timescale shift. Decay adaptation should recover this shift and gain adaptation should not.
Why not just fine-tune the whole network on the few target examples? With 1 to 20 spectrograms and roughly 3% positive pixels, full fine-tuning risks overfitting. It also requires rewriting every weight on the deployed chip, whereas decay-only adaptation changes per-neuron registers. Full fine-tuning is run as a baseline. Our claim is that decay-only adaptation reaches at least 80% of its improvement over no adaptation at $k \le 5$, and overtakes it at $k=1$.
Why not just search fixed decays per telescope with Optuna? Black-box search scales poorly to hundreds of per-neuron decays, and PLIF's results show gradient-learned decay escapes poor initializations that fixed choices get stuck in. The re-searched-$\beta$ ablation tests this directly under an equal compute budget.
Proposed Experiments and Decisive Checks
We train on simulated HERA with its standard split of 420 training and 140 test spectrograms. We adapt to real LOFAR, holding out its expert-labeled masks for testing, and to Tabascal for moving-satellite interference. We report F1, AUROC and AUPRC, with AUPRC primary because of class imbalance, plus per-morphology recall and flagged-data fraction.
Three checks decide the work:
- The core claim. Decay-only adaptation at $k=5$ must beat the unadapted model, gain-only and readout-only on LOFAR AUPRC across five adaptation-set draws, and must reach or exceed AOFlagger's F1 under our protocol. If it does not, the claim is refuted.
- The wide-band failure mode. Recall on single-timestep wide-band bursts must exceed the frozen-reservoir behavior described in the Liquid State Machine paper. We expect adapted decays to show some neurons with near-zero $\beta$.
- Hardware. The discretized-decay variant must lose less than 0.02 AUPRC relative to continuous decay. It must also deploy on Xylo 2 through the published export pipeline at power comparable to the roughly 100 mW measured in Neuromorphic Astronomy.
References
- Spiking neural networks for radio frequency interference detection in radio astronomy
- Supervised Radio Frequency Interference Detection with SNNs
- Advancing RFI-Detection in Radio Astronomy with Liquid State Machines
- RFI detection with spiking neural networks
- Neuromorphic Astronomy: An End-to-End SNN Pipeline for RFI Detection Hardware
- Incorporating Learnable Membrane Time Constant to Enhance Learning of Spiking Neural Networks
- DA-LIF: Dual Adaptive Leaky Integrate-and-Fire Model for Deep Spiking Neural Networks
- HetSyn: Versatile Timescale Integration in Spiking Neural Networks via Heterogeneous Synapses
- The Low-Frequency Environment of the Murchison Widefield Array: Radio-Frequency Interference Analysis and Mitigation