Efficient Spike Encoding for Sigma-Delta RF Receivers: An End-to-End Neuromorphic Radio Classification System

Overview

Automatic Modulation Classification (AMC) is a fundamental task in Radio Frequency (RF) spectrum monitoring and cognitive wireless systems, where real-time inference under strict latency, throughput, and energy constraints is required at the network edge. Conventional AMC pipelines rely on time-driven signal processing and dense neural network inference, which incur substantial computational and energy overhead when deployed on edge hardware. These limitations motivate the exploration of neuromorphic approaches that leverage event-driven computation and temporal processing.

First, an end-to-end neuromorphic AMC system is developed by directly interfacing RF front-end outputs with an SNN classifier using a spike-based signal representation. By eliminating explicit spike encoding and intermediate preprocessing stages, the proposed system simplifies the signal processing pipeline while maintaining competitive classification accuracy in both software and FPGA implementations.

Second, to address the gap between algorithm-level sparsity in SNNs and hardware-level efficiency, a sparsity-aware output-channel dataflow streaming execution model is proposed. By explicitly integrating temporal and spatial sparsity into a deterministic streaming framework, the proposed accelerator reduces redundant synaptic accumulations and memory accesses while preserving stable throughput and predictable latency. FPGA evaluations demonstrate significant improvements in accumulation efficiency, energy efficiency, and overall hardware performance compared with existing streaming SNN accelerators.

Finally, this work addresses workload imbalance arising from fixed parallelism configurations in streaming architectures. A configurable streaming SNN accelerator is introduced that decouples pixel-level and output-channel parallelism, enabling execution to be adapted to heterogeneous layer characteristics and sparsity distributions. By exposing parallelism as a configurable architectural dimension, the proposed design improves pipeline utilization and maintains high throughput across diverse operating conditions without compromising deterministic execution.

Together, these contributions demonstrate that efficient neuromorphic RF inference requires systematic co-design across signal representation, execution models, and hardware architectures. The proposed sparsity-aware and configurable streaming SNN accelerators provide a scalable and energy-efficient solution for real-time RF AMC at the edge.