🤖 machine learning

Neuromorphic Wireless Split Computing with Resonate-and-Fire Neurons

This paper proposes a neuromorphic wireless split computing architecture utilizing resonate-and-fire neurons to directly process streaming time-domain signals with tunable spectral resonance, achieving comparable accuracy to conventional models while significantly reducing spike rates and energy consumption in both computation and transmission.

Original authors: Dengyu Wu, Jiechen Chen, H. Vincent Poor, Bipin Rajendran, Osvaldo Simeone

Published 2026-06-16

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Dengyu Wu, Jiechen Chen, H. Vincent Poor, Bipin Rajendran, Osvaldo Simeone

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to send a long, complex story to a friend over a walkie-talkie. The story is full of specific sounds, like a bird chirping at a high pitch or a drum beating at a low rhythm.

The Old Way (LIF Neurons):
Currently, most "smart" devices use a method called Leaky Integrate-and-Fire (LIF). Think of this like a person who listens to your story but only understands simple "on/off" switches. They don't naturally hear the pitch of the bird or the rhythm of the drum. To understand the story, they have to stop, write down a detailed musical score (a spectral analysis) for every single sound, and then send that massive score to their friend. This takes a lot of energy and time, and the walkie-talkie gets clogged with too much data.

The New Way (Resonate-and-Fire Neurons):
This paper introduces a smarter approach using Resonate-and-Fire (RF) neurons. Imagine these neurons are like tuning forks.

Instead of listening to everything and then writing it down, each neuron is tuned to a specific frequency (like a specific musical note).
If you speak a word that matches that note, the tuning fork vibrates strongly and sends a signal.
If the sound doesn't match, the tuning fork stays quiet.

Because these neurons only "fire" (send a signal) when they hear the exact sound they are tuned for, they stay silent most of the time. This is called sparsity.

The Wireless Split System:
The researchers built a system where this "tuning fork" brain is split in two:

The Transmitter (The Ear): It listens to the raw sound (like audio or radio waves) and uses these tuning-fork neurons to pick out the important frequencies. Because they are so efficient, they only send a few "beeps" (spikes) over the air.
The Receiver (The Brain): It receives those few "beeps" and figures out what the sound was (e.g., "That was a 5G signal" or "That was the word 'Hello'").

Why is this a big deal?

Energy Savings: In the old system, the transmitter had to send a huge amount of data (the musical score). In this new system, it only sends a few "beeps." Since sending data over the air uses the most battery power, sending fewer beeps saves a massive amount of energy.
No Extra Steps: The old way required a complicated pre-processing step to turn sound into a score before the computer could understand it. The new way does this automatically inside the neurons themselves.
Better Accuracy: The paper tested this on recognizing spoken digits and identifying different types of radio signals. The new "tuning fork" system was just as accurate as the old, heavy systems but used significantly less energy.

The Bottom Line:
The paper shows that by using neurons that act like musical tuning forks, we can build wireless devices that listen to the world, pick out the important sounds, and send them to a server using a tiny fraction of the battery power required by current technology. It's like switching from mailing a 500-page book to sending a single postcard that says, "It's a bird!"

Technical Summary: Neuromorphic Wireless Split Computing with Resonate-and-Fire Neurons

Problem Statement
Modern edge intelligence faces a critical challenge in processing streaming time-series data, such as radio-frequency (RF) and audio signals, which possess rich spectral features. Conventional deep neural network (DNN) accelerators are often ill-suited for these always-on, continuous tasks due to high power consumption. While neuromorphic computing using Spiking Neural Networks (SNNs) offers an energy-efficient, event-driven alternative, standard implementations relying on Leaky Integrate-and-Fire (LIF) neurons lack inherent frequency selectivity. Consequently, LIF-based SNNs typically require costly pre-processing steps, such as Fast Fourier Transforms (FFTs) or spectrograms, to extract frequency-domain features before spike-based processing. This pre-processing generates high-dimensional inputs that necessitate large models, potentially negating the energy and latency benefits of neuromorphic inference, particularly in split computing architectures where spike transmission over wireless channels constitutes a significant energy bottleneck.

Methodology
The paper proposes a neuromorphic wireless split computing architecture that replaces conventional LIF neurons with Balanced Resonate-and-Fire (BRF) neurons. This approach aims to process raw time-domain signals directly, eliminating the need for explicit spectral pre-processing.

Neuron Model: The system utilizes BRF neurons, which exhibit sub-threshold membrane potential oscillations and resonate at tunable frequencies. Functionally, they act as band-pass filters, accumulating oscillatory input and firing spikes only when the input frequency matches the neuron's natural resonance. The model incorporates adaptive refractory mechanisms to control spiking activity and preserve memory of past inputs, ensuring sparse firing.
System Architecture: The SNN is partitioned between a transmitter (Tx) and a receiver (Rx). The Tx encodes raw signals (audio or baseband) into spike trains using an encoding SNN with BRF neurons. These spikes are transmitted over a wireless channel using an OFDM-based analog interface, where spike events modulate OFDM subcarriers. The Rx decodes the received signal using a decoding SNN to make a final classification decision.
Optimization: The system is trained using a loss function that combines cross-entropy loss with a sparsity-promoting regularizer. This regularizer, based on the ratio of the $\ell_1$ -norm to the $\ell_2$ -norm of normalized membrane potentials, encourages fewer spikes to reduce both computational and transmission energy.
Energy Modeling: The authors develop a comprehensive energy model that accounts for both synaptic operations (weight accumulation) and somatic operations (neuron dynamics). Unlike prior works that often overlook somatic costs, this model explicitly calculates the energy of additions and multiplications required for the complex dynamics of BRF neurons versus the real-valued dynamics of Adaptive LIF (ALIF) and LIF neurons.

Key Contributions

RF Neurons for Wireless Split Computing: The paper introduces the use of Resonate-and-Fire (RF) and BRF neurons in a split SNN architecture. These neurons inherently extract time-localized spectral features from raw signals, yielding lower-dimensional representations and significantly reduced spiking activity compared to LIF-based SNNs, with or without FFT pre-processing.
System Design and Energy Analysis: A complete system design is presented, featuring an OFDM-based analog interface for spike transmission. The authors provide an analytical characterization of total energy consumption, encompassing both computation (synaptic and somatic) and wireless transmission, demonstrating how spike sparsity directly translates to transmission energy savings.
Experimental Evaluation: The proposed RF-SNN split computing system is evaluated on two distinct tasks:
- Audio Classification: Using the Spiking Heidelberg Digits (SHD) dataset.
- RF Modulation Classification: Using the ITS radio signal dataset (complex baseband waveforms).
  The study compares BRF-SNNs against LIF-SNNs, ALIF-SNNs, and conventional Artificial Neural Networks (ANNs).

Results
Experimental results on both datasets demonstrate that the BRF-SNN architecture achieves competitive accuracy while substantially reducing energy consumption:

Audio (SHD): The BRF model achieved up to 93.3% accuracy with a compute energy of 8.77 µJ (RF) and 4.33 µJ (BRF), significantly outperforming LIF (82.0% accuracy, 30.44 µJ) and ALIF (89.7% accuracy, 22.01 µJ). In split computing scenarios at 100m distance, BRF achieved 92.9% accuracy with a total energy (compute + transmission) of 6.73 µJ, compared to 41.39 µJ for LIF.
Radio Frequency (ITS): BRF neurons achieved 87.7% accuracy with 1.61 µJ compute energy, outperforming ALIF and LIF models which required higher energy for lower accuracy.
Energy Breakdown: The analysis confirms that while BRF neurons have slightly higher somatic complexity, the massive reduction in spike rates leads to dominant savings in synaptic and transmission energy. The system effectively leverages the "resonance" property to filter noise and focus on task-relevant spectral features.
Quantization: The BRF models demonstrated stronger resilience to weight quantization (4-bit to 8-bit) compared to ALIF models, maintaining high accuracy even with reduced precision.

Significance and Claims
The paper claims that by embedding learned frequency analysis directly into the dynamics of individual neurons, the proposed architecture enables efficient processing of spectrally rich time-series signals without the latency and energy overhead of external spectral pre-processing. The authors assert that the combination of RF neurons' spectral sensitivity with a low-overhead analog wireless interface creates a uniquely efficient solution for edge applications like wireless sensing and audio recognition.

The work underscores that the intrinsic frequency selectivity of RF and BRF neurons allows for higher spectral discrimination, translating to improved accuracy-energy efficiency trade-offs compared to conventional LIF and ALIF models. The authors conclude that RF-based neuromorphic models hold significant potential for enabling low-power, low-latency intelligence at the wireless edge, particularly in scenarios where minimizing transmission energy is critical. Future work is suggested to extend this architecture to more complex sensory modalities and to investigate analog circuit implementations of BRF dynamics.

More like this