Neuromorphic Astronomy: An End-to-End SNN Pipeline for RFI Detection Hardware

TL;DR

The paper tackles real-time, energy-efficient RFI detection in radio astronomy by deploying deep Spiking Neural Networks on constrained neuromorphic hardware. It introduces an end-to-end pipeline—training large SNNs in snnTorch, converting to NIR, splitting with a novel maximal-splitting algorithm, and deploying on SynSense Xylo via Rockpool—to evaluate hardware-aware performance and energy use. Key contributions include a fan-in regularisation scheme, an end-to-end deployment blueprint, and critical insights into how hardware constraints shape model scaling, revealing that smaller, hardware-tuned models can outperform larger split networks. The work demonstrates substantial energy efficiency (sub-100 mW per spectrogram) and positions neuromorphic computing as a viable pathway for spectro-temporal data processing in data-intensive radio astronomy, while outlining limitations and directions for further validation and hardware-aware training. Overall, it provides a practical framework for translating high-performing SNNs into real-world neuromorphic deployments and highlights the need for hardware-conscious model design in scientific applications.

Abstract

Imminent radio telescope observatories provide massive data rates making deep learning based processing appealing while simultaneously demanding real-time performance at low-energy; prohibiting the use of many artificial neural network based approaches. We begin tackling the scientifically existential challenge of Radio Frequency Interference (RFI) detection by deploying deep Spiking Neural Networks (SNNs) on resource-constrained neuromorphic hardware. Our approach partitions large, pre-trained networks onto SynSense Xylo hardware using maximal splitting, a novel greedy algorithm. We validate this pipeline with on-chip power measurements, achieving instrument-scaled inference at 100mW. While our full-scale SNN achieves state-of-the-art accuracy among SNN baselines, our experiments reveal a more important insight that a smaller un-partitioned model significantly outperforms larger, split models. This finding highlights that hardware co-design is paramount for optimal performance. Our work thus provides a practical deployment blueprint, a key insight into the challenges of model scaling, and reinforces radio astronomy as a demanding yet ideal domain for advancing applied neuromorphic computing.

Figures (4)

• Figure 1: Overall methodology for RFI detection with spiking neural networks trained as a large single model and then split for inference on several neuromorphic chipsets. Spectrograms are split and latency encoded before feeding through the SNN, models are split in NIR format and deployed in snnTorch or to SynSense Xylo hardware for power measurement.
• Figure 2: Model splitting approaches. Our model splitting algorithm greedily selects a neuron bundle such that each neuron satisfies the hardware platform restrictions, then iteratively removes neurons until chipset restrictions are met. This is in contrast to a random or naive (geometric) splitting approach also depicted here.
• Figure 3: An original full-size spectrogram and its corresponding ground-truth label.
• Figure 4: Inference examples across different model sizes. Examples of the original spectrogram in Figure \ref{['fig:xylo:example:orig']} with varying trained model widths and their corresponding hardware reconstructions. We see that in all widths, hardware faithfully reconstructs the output of the software-trained models but that smaller models outperform larger input widths.