Neuromorphic Breakthrough: Spiking AI Detects Radio Interference at Sub-Watt Power

In the race to curb AI's voracious energy appetite, a new study from the University of Western Australia demonstrates how spiking neural networks (SNNs) deployed on neuromorphic hardware could revolutionize data-intensive fields like radio astronomy. Published in November 2025, the research tackles the existential of Radio Frequency Interference (RFI), which corrupts faint cosmic signals from next-generation telescopes. By partitioning large SNNs onto SynSense Xylo chips using a novel 'maximal splitting' algorithm, the team achieved instrument-scaled inference at just 100 milliwatts—a fraction of conventional AI power demands. This work not only provides a blueprint for low-energy AI but also uncovers a critical insight: smaller, un-split models often outperform their larger, partitioned counterparts, highlighting that hardware co-design is paramount for optimal performance.

Ology centered on an end-to-end pipeline that trained SNNs using Backpropagation-Through-Time (BPTT) with surrogate gradients on a synthetic RFI dataset from the Hydrogen Epoch of Reionisation Array (HERA). The models were designed as fully-connected feedforward networks, with inputs encoded into spike trains using latency encoding to handle the spectro-temporal nature of radio astronomy data. A key innovation was 'maximal splitting,' a greedy algorithm that shards pre-trained SNNs into hardware-compatible sub-modules by selecting strongly connected neurons and iteratively culling the weakest ones to meet constraints like a maximum fan-in of 63 on the Xylo processor. This approach was benchmarked against naive and random splitting s, with the pipeline seamlessly converting models from snnTorch to the Neural Intermediate Representation (NIR) format and then deploying them on physical neuromorphic hardware via the Rockpool library.

Revealed that the unsplit 64-channel SNN achieved state-of-the-art accuracy among SNN baselines, with an F1-score of 0.983 and AUROC of 0.988, competitive with some artificial neural network (ANN) approaches. However, splitting these models led to significant performance drops; for instance, maximal splitting on a 128-channel model retained an accuracy of 0.958 but saw AUPRC plummet to 0.317. Power measurements on the Xylo hardware were strikingly low, with scaled consumption as minimal as 1.42 milliwatts per spectrogram at real-time frequencies, underscoring the efficiency gains. The study also found that hardware-aware regularisation—penalising excess fan-in during training—yielded modest improvements, particularly for mid-sized models, though it couldn't fully mitigate the trade-offs introduced by partitioning.

Of this research extend beyond astronomy, offering a template for deploying SNNs in other spectro-temporal domains like seismic monitoring or autonomous systems where real-time, low-power processing is critical. By validating RFI detection as a benchmark for neuromorphic computing, the study reinforces the potential of brain-inspired hardware to address the energy crises in data centers. However, the performance degradation from model splitting signals that current neuromorphic platforms may not yet support the complexity of large-scale SNNs without sacrificing accuracy, pointing to a need for more advanced hardware-aware training and retraining techniques to bridge this gap.

Despite its promise, the work has limitations, including a focus on synthetic data that may not capture the variability of real-world RFI, and the use of feedforward architectures that limit model expressiveness. The authors caution that unsupervised learning s and validation on observational telescope data are essential next steps. Nevertheless, this pipeline marks a significant leap toward making neuromorphic computing a practical tool for scientific , proving that with careful co-design, AI can indeed operate at the edge of efficiency without compromising on intelligence.