Resonant Spiking Neurons Enhance Energy Efficiency for Wireless Time-Series Analysis.

The demand for energy-efficient computation at the network edge continues to drive innovation in neural network architectures, particularly for applications involving continuous data streams like wireless sensing and audio analysis. Conventional deep learning relies on substantial computational resources, prompting research into spiking neural networks (SNNs) as a potentially more power-conscious alternative. These networks mimic biological neurons, communicating via discrete ‘spikes’ of information, and offer the possibility of reduced energy consumption. A new investigation, detailed in the article ‘Neuromorphic Wireless Split Computing with Resonate-and-Fire Neurons’, explores a system utilising a specific type of spiking neuron, known as resonate-and-fire (RF) neurons, to directly process time-varying signals without the need for complex pre-processing.

Neuromorphic systems offer a potentially efficient route towards wireless communication with substantially reduced energy demands, and recent research demonstrates a functional computing system achieving competitive performance on signal processing tasks while minimising power consumption. The core innovation resides in an implementation of a spiking neural network (SNN) that directly processes raw radio frequency (RF) signals, circumventing the need for energy-intensive analog-to-digital conversion and establishing RF-native processing. Coupled with techniques maximising sparsity in neuronal spiking activity, the system delivers considerable power savings and presents a viable alternative to conventional deep learning accelerators for real-time time-series processing.

This research details a complete design utilising an Orthogonal Frequency-Division Multiplexing (OFDM) based analog wireless interface for spike transmission, minimising both computational load and the amount of data requiring transmission. OFDM is a method for encoding digital data on multiple carrier frequencies, improving transmission efficiency. By employing techniques such as non-negative matrix factorisation, the network generates sparse spikes, meaning neurons fire less frequently, and this temporal sparsity directly translates into energy savings, particularly crucial for applications like the Internet of Things (IoT) and edge computing where power constraints are paramount. Evaluation on both audio and modulation classification tasks confirms the efficacy of the RF-SNN architecture, indicating comparable accuracy to conventional leaky integrate-and-fire (LIF) SNNs and artificial neural networks (ANNs). LIF neurons, a common model in SNNs, accumulate input until a threshold is reached, triggering a spike.

The system utilises resonate-and-fire (RF) neurons, a specific type of spiking neuron, to effectively capture time-varying spectral features present in wireless signals, unlike conventional LIF neurons. RF neurons oscillate at tunable frequencies, enabling them to extract relevant spectral information directly from the time domain, and this temporal sparsity translates into reduced computational demands and lower energy expenditure for both processing and data transmission. The ability to process signals in the time domain, rather than converting them to a digital format, is a key advantage of this approach.

Evaluations conducted using realistic wireless channel models, specifically 3GPP TR 38.901, and datasets demonstrate the system’s efficacy on tasks such as 5G/NR waveform classification and modulation recognition. 3GPP TR 38.901 defines channel models for evaluating wireless communication systems, ensuring realistic testing conditions. Implementation and testing occur on Intel’s Loihi neuromorphic processor, demonstrating the feasibility of the proposed system, and the research also explores split computing, distributing processing between the edge, near the antenna, and the cloud, to further optimise energy usage and reduce latency.

Researchers validate the system’s robustness by testing it with realistic wireless channel models, specifically those defined in 3GPP TR 38.901, ensuring performance in practical scenarios. This work represents a significant step towards more sustainable wireless communication systems, offering a viable alternative to conventional deep learning accelerators for real-time time-series processing.