End-to-End Radar and Communication Modulation Recognition with Neuromorphic Computing

arXiv:2606.24075 · cs.CV, cs.AI · Submitted 2026-06-23 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "End-to-End Radar and Communication Modulation Recognition with Neuromorphic Computing".

Jane: Here's the gist of the paper. The gist The proposed EMRFormer is a novel end-to-end spiking neural network architecture that applies spike-driven transformer to constraints of neuromorphic hardware for AMR, achieving state-of-the-art accuracy while reducing theoretical energy consumption by over 90% How it works The EMRFormer architecture is designed to overcome the high computational cost of deep learning methods in Automatic Modulation Recognition (AMR) by leveraging neuromorphic computing principles The model incorporates several key components to enhance information representation and temporal modeling 1) ILIF Neurons are used to mitigate the degradation of effective information and enhance SNN representational capacity These Integer Leaky Integrate-and-Fire (ILIF) neurons can implement multi-spiking emission, allowing the neuron to fire spikes up to D times rather than only once The formulation of an I-LIF neuron is given by equations (1), (2), and (3) In the inference phase, the time step T is extended to T × D steps, rewriting S l

t: as S l

t, d: ∈

0, 1: for binary spikes 2) Adaptive Spike Encoder addresses information loss caused by static encoding The encoder network encodes the input signals in a data-dependent fashion, mitigating information loss caused by the prior biases inherent in static encoding It formalizes the raw data of IQ waveform into XIQ =

XI, XQ: ∈ R 2×N, where N is the signal length This approach eliminates the additional data-preprocessing step, significantly reducing computational and storage demands on hardware resources 3) SSCNN Modules are integrated for efficient feature extraction The Spike Separable Convolutional Neural Network (SSCNN) module works as eq.

Tom: 8 and eq.

Jane: 9 to capture local signal features It employs a stacked structure of pointwise and depthwise convolution to perform this extraction 4) SpikeFormer Module captures long-term dependencies Inspired by the Transformer model, the SpikeFormer module is adopted to capture long-range dependency through a spike version of transformer

13: Its workflow involves calculating Q, K, and V via SpikeSepConv using ILIF neurons and computing attention scores using the spike attention operator This self-attention mechanism complements SSCNN by modeling dependency features across long-distance signals Experimental Validation The EMRFormer model was validated across various mainstream datasets, demonstrating superior performance compared to all baselines The experimental results show that EMRFormer achieves state-of-the-art in terms of accuracy, outperforming all the baselines It exhibits stronger robustness under low signal-to-noise Ratio (SNR) environments and cross-domain (radar/communication) conditions For instance, on DeepRadar2022, EMRFormer achieves the accuracy of 87.4%, a 2–10% improvement over other baseline models Furthermore, the model maintains an accuracy edge in low-SNR conditions (<0 dB), showcasing robust feature extraction under heavy noise Energy Efficiency and Hardware Deployment The theoretical energy consumption is reduced by over 90% compared to conventional computing platforms The model converts the integer activations to binary activations via virtual time steps, tailored to work with the neuromorphic chips in the inference phase When deployed on the KA200 neuromorphic chip, it boasts roughly 5 times lower power than GPU and Orin NX platforms Table III demonstrates that EMRFormer's Algorithm Power on KA200 outperforms the GPU and Orin NX by 5× This confirms the efficient hardware mapping of SNN models on neuromorphic chips Contributions The contributions of this work are summarized as follows 1) To the best of our knowledge, this is the first end-toend SNN-based AMR framework that directly processes raw IQ signals to conduct a complete signal-to-category recognition 2) We develop the novel EMRFormer to effectively handle information loss by static spike encoder and capture long-term temporal dependencies in radar and communication signals 3) We conduct the end-toend training and testing on general-purpose platforms with model deployment to a KA200 chip for inference validation Ablation Study Findings The ablation study on EMRFormer using the Deepradar dataset reveals that altering any component leads to a performance drop, as summarized in II Switching Conv1d to Conv2d boosts parameters and energy, while it cuts the performance of accuracy Axing PosEnc further hits accuracy Swapping the Adaptive spike encoder for static encoder (poisson encoder) slashes accuracy and hikes power draw Figure 4 indicates that variants match the full model with the high SNR, but they fall behind at low SNR The full model, packing only 0.

Paper summary: Tom: We've covered the high-level concept of EMRFormer and how it sets itself against deep learning methods on resource-constrained platforms. Now we’re diving a little deeper into what they actually propose in this paper, "End-to-End Radar and Communication Modulation Recognition with Neuromorphic Computing."

Jane: So, essentially the thesis is that by replacing conventional deep learning with an end-to-end spiking neural network architecture that uses spike-driven transformers on neuromorphic hardware, we can achieve high accuracy for automatic modulation recognition while keeping energy consumption low enough for practical deployment.

Lu: They argue that this approach addresses the high computational cost of deep learning by leveraging neuromorphic computing principles, specifically using spike-driven inference with modest energy budgets for vision and timeseries tasks.

Meng: The summary says they propose EMRFormer, which incorporates an adaptive spike encoder and Integer Leaky Integrate-and-Fire neurons to mitigate the degradation of effective information and enhance the spiking neural network's representational capacity.

Tom: That means they are specifically targeting how the spikes behave inside the network to ensure that even with limited energy, we don't lose too much meaningful data from those raw IQ signals.

Jane: They also integrate spike-separable Convolutional Neural Networks into Spike-Driven Transformers, which they call SpikeFormer, to effectively extract multi-scale temporal features from the raw IQ waveforms.

Lu: The SpikeFormer module is designed to capture long-term dependencies through a spike version of the transformer, and its workflow involves calculating Q, K, and V via SpikeSepConv using ILIF neurons and computing attention scores using the spike attention operator.

Meng: So they’re combining local feature extraction from SSCNNs with this long-range temporal modeling from SpikeFormer to process those signals efficiently.

Tom: And they claim these experiments demonstrate that EMRFormer achieves state-of-the-art performance among the end-to-end baselines across various datasets.

Jane: But the real meat for us is how it performs when conditions aren't perfect, specifically under low signal to noise ratio and in cross domain conditions between radar and communication signals.

Lu: They also validate this model on the KA200 neuromorphic chip, showing a power reduction of up to five relative to conventional computing platforms like the NVIDIA GeForce RTX three thousand ninety GPU or the NVIDIA Jetson Orin NX <ref:2606.24075#pg1>.

Meng: So, if you’re looking at deployment, this is really telling because they show it works on actual hardware and scales down the energy requirement significantly compared to what we usually ship.

Tom: This paper essentially presents a promising pathway for automatic modulation recognition on resource-constrained devices by showing a path from raw IQ signals directly into an efficient spiking neural network.

Conclusion: Jane: So, to wrap up this discussion on "End-to-End Radar and Communication Modulation Recognition with Neuromorphic Computing," the authors are proposing EMRFormer as the solution for running accurate modulation recognition on devices that have limited power and memory.

Tom: The main contribution they point out is that this is the first end-to-end SNN based AMR framework that directly processes raw IQ signals to conduct a complete signal to category recognition.

Lu: They also developed EMRFormer specifically to handle information loss by static spike encoders and capture long-term temporal dependencies in radar and communication signals.

Meng: From a deployment view, the validation on the KA200 chip confirms that this model has strong performance when deployed, showing roughly five times lower power than platforms like the GPU or Orin NX.

Jane: So what does this mean for the general listener? It means there’s a viable direction for applying these complex signal recognition tasks—like understanding modulation patterns in radar or communication signals—to everyday devices without needing massive computational resources.

Tom: It moves the focus from just achieving high accuracy on powerful hardware to finding efficient ways to get that accuracy where it actually needs to be used, which is the core of this work.

Lu: The future work suggested involves continuing the end-to-end training and testing on general-purpose platforms with model deployment validation on neuromorphic chips.

Meng: If this approach holds up when we move beyond these tested datasets and into real traffic scenarios, it opens up a lot of potential for embedding highly efficient AI into edge devices.

Xiaohu Lia, Chongxiao Qub, Caiyong Lina, Chenxiao Doua, *Wei Huaa

China Nanhu Academy of Electronics and Information Technology · University of Electronic Science and Technology of China

cs.CV, cs.AI

Submitted: 2026-06-23

Updated: 2026-10-19

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 86/100

The gist: The gist The proposed EMRFormer is a novel end-to-end spiking neural network architecture that applies spike-driven transformer to constraints of neuromorphic hardware for AMR, achieving

Key concepts

ILIF Neurons
Integer Leaky Integrate-and-Fire neurons are used in the SNN to improve information representation. They can emit spikes multiple times, up to D times, which helps the network capture richer temporal information from the input signals by allowing for more complex firing patterns.
Adaptive Spike Encoder
This component addresses information loss from static encoding by encoding raw IQ waveforms into a data-dependent format (XIQ). This eliminates extra preprocessing steps and reduces hardware demands while ensuring that the most relevant data features are preserved for the subsequent network layers.
SpikeFormer Module
Inspired by the Transformer model, this module uses a spike version of attention to capture long-range dependencies in radar and communication signals. It calculates attention scores using ILIF neurons and spike-separable convolutions, effectively modeling relationships across long distances in the signal.
SSCNN Modules
Spike Separable Convolutional Neural Networks are integrated to efficiently extract local signal features from the raw IQ waveforms. They use a stacked structure of pointwise and depthwise convolution to capture relevant local patterns while maintaining computational efficiency for SNN hardware.

Terminology

Summary

The gist The proposed EMRFormer is a novel end-to-end spiking neural network architecture that applies spike-driven transformer to constraints of neuromorphic hardware for AMR, achieving state-of-the-art accuracy while reducing theoretical energy consumption by over 90%

How it works

The EMRFormer architecture is designed to overcome the high computational cost of deep learning methods in Automatic Modulation Recognition (AMR) by leveraging neuromorphic computing principles The model incorporates several key components to enhance information representation and temporal modeling

  1. ILIF Neurons are used to mitigate the degradation of effective information and enhance SNN representational capacity These Integer Leaky Integrate-and-Fire (ILIF) neurons can implement multi-spiking emission, allowing the neuron to fire spikes up to D times rather than only once The formulation of an I-LIF neuron is given by equations (1), (2), and (3) In the inference phase, the time step T is extended to T × D steps, rewriting S l[t] as S l[t, d] ∈ [0, 1] for binary spikes

  2. Adaptive Spike Encoder addresses information loss caused by static encoding The encoder network encodes the input signals in a data-dependent fashion, mitigating information loss caused by the prior biases inherent in static encoding It formalizes the raw data of IQ waveform into XIQ = [XI, XQ] ∈ R 2×N, where N is the signal length This approach eliminates the additional data-preprocessing step, significantly reducing computational and storage demands on hardware resources

  3. SSCNN Modules are integrated for efficient feature extraction The Spike Separable Convolutional Neural Network (SSCNN) module works as eq. 8 and eq. 9 to capture local signal features It employs a stacked structure of pointwise and depthwise convolution to perform this extraction

  4. SpikeFormer Module captures long-term dependencies Inspired by the Transformer model, the SpikeFormer module is adopted to capture long-range dependency through a spike version of transformer [13] Its workflow involves calculating Q, K, and V via SpikeSepConv using ILIF neurons and computing attention scores using the spike attention operator This self-attention mechanism complements SSCNN by modeling dependency features across long-distance signals

Experimental Validation

The EMRFormer model was validated across various mainstream datasets, demonstrating superior performance compared to all baselines The experimental results show that EMRFormer achieves state-of-the-art in terms of accuracy, outperforming all the baselines It exhibits stronger robustness under low signal-to-noise Ratio (SNR) environments and cross-domain (radar/communication) conditions For instance, on DeepRadar2022, EMRFormer achieves the accuracy of 87.4%, a 2–10% improvement over other baseline models Furthermore, the model maintains an accuracy edge in low-SNR conditions (<0 dB), showcasing robust feature extraction under heavy noise

Energy Efficiency and Hardware Deployment

The theoretical energy consumption is reduced by over 90% compared to conventional computing platforms The model converts the integer activations to binary activations via virtual time steps, tailored to work with the neuromorphic chips in the inference phase When deployed on the KA200 neuromorphic chip, it boasts roughly 5 times lower power than GPU and Orin NX platforms Table III demonstrates that EMRFormer's Algorithm Power on KA200 outperforms the GPU and Orin NX by 5× This confirms the efficient hardware mapping of SNN models on neuromorphic chips

Contributions

The contributions of this work are summarized as follows

  1. To the best of our knowledge, this is the first end-toend SNN-based AMR framework that directly processes raw IQ signals to conduct a complete signal-to-category recognition

  2. We develop the novel EMRFormer to effectively handle information loss by static spike encoder and capture long-term temporal dependencies in radar and communication signals

  3. We conduct the end-toend training and testing on general-purpose platforms with model deployment to a KA200 chip for inference validation

Ablation Study Findings

The ablation study on EMRFormer using the Deepradar dataset reveals that altering any component leads to a performance drop, as summarized in II Switching Conv1d to Conv2d boosts parameters and energy, while it cuts the performance of accuracy Axing PosEnc further hits accuracy Swapping the Adaptive spike encoder for static encoder (poisson encoder) slashes accuracy and hikes power draw Figure 4 indicates that variants match the full model with the high SNR, but they fall behind at low SNR The full model, packing only 0.

Improvements for AI systems

  1. Adaptive spike encoder can learn from IQ data to alleviate information loss caused by static encoding, mitigating degradation by being a learned model that encodes input signals in a data-dependent fashion. This allows the AI system to maintain high accuracy even in complex signal environments where prior biases are inherent in static encoding.

  2. Integer Leaky Integrate-and-Fire (ILIF) neurons enhance the representational capacity of SNNs by implementing multi-spiking emission when SNN learning, allowing a neuron to fire spikes up to D times rather than only once. This enables the system to transmit more information through a multi-fire mechanism while ensuring compatibility with neuromorphic chips.

  3. Integration of Spike-Separable Convolutional Neural Networks (SSCNN) into SpikeFormer allows for the extraction of multiscale temporal features from raw IQ waveforms, effectively capturing both short-range local features and long-range dependencies. This enables EMRFormer to learn both short- and long-range dependencies from raw IQ sequences in an eventdriven and energy-efficient manner.

  4. The model achieves a significant reduction in theoretical energy consumption by over 90% compared to deep learning methods, with a 5× reduction in power compared to running on a 3090 GPU or an Orin NX when deployed on neuromorphic chips like the KA200. This enables deployment of AMR applications on resource-constrained devices with low power and high energy efficiency.

Sources

Related papers