emrformer-neuromorphic-amr

name	emrformer-neuromorphic-amr
description	EMRFormer: End-to-End Radar and Communication Modulation Recognition with Neuromorphic Computing using spike-driven transformers on KA200 neuromorphic chip. Achieves SOTA accuracy with 90%+ energy reduction. (arXiv: 2606.24075)
tags	["neuromorphic","SNN","spike-driven-transformer","modulation-recognition","edge-AI","KA200","LIF-neurons"]

EMRFormer: Neuromorphic Modulation Recognition

Paper Reference

• Title: End-to-End Radar and Communication Modulation Recognition with Neuromorphic Computing
• arXiv: 2606.24075
• Authors: Xiaohu Li, Chongxiao Qu, Caiyong Lin, Chenxiao Dou, Wei Hua
• Submitted: June 23, 2026
• Categories: cs.CV, cs.AI

Core Methodology

Architecture: EMRFormer

End-to-end spiking neural network applying spike-driven transformer to automatic modulation recognition (AMR) under neuromorphic hardware constraints.

Key Components

1. Adaptive Spike Encoder: Converts raw IQ waveforms into spike representations while preserving temporal information
2. Integer Leaky Integrate-and-Fire (I-LIF) Neurons: Mitigates effective information degradation in spike encoding
3. Spike-Separable CNN (SSCNN): Multi-scale temporal feature extraction from raw IQ data
4. SpikeFormer Integration: Spike-driven transformer backbone combining SSCNN with temporal attention

Innovation Points

• First application of spike-driven transformers to AMR tasks
• Integer LIF neurons enable hardware-friendly computation
• End-to-end pipeline from raw IQ waveforms to modulation classification
• Deployed and validated on KA200 neuromorphic chip

Results

• Accuracy: SOTA across multiple modulation datasets, outperforming all baselines
• Low SNR Robustness: Maintains strong performance in challenging signal conditions
• Energy Efficiency: 90%+ theoretical energy reduction vs ANN counterparts
• Hardware Deployment: 5x power reduction on KA200 vs 3090 GPU or Orin NX

Practical Applications

• Resource-constrained edge devices for radar/communication
• Military/defense signal intelligence
• IoT modulation classification
• Real-time spectrum monitoring

Activation Keywords

• EMRFormer
• spike-driven transformer
• neuromorphic modulation recognition
• KA200 chip
• SNN automatic modulation recognition
• edge AI radar
• IQ waveform classification

Related Skills

• [[spiketimer-snn-copyright-protection]] - SNN temporal coding
• [[neuromorphic-lidar-bev-snn]] - SNN for sensor processing
• [[edgespike-edge-iot-snn]] - Edge SNN deployment
• [[frequency-matching-snn-mmwave]] - SNN for mmWave signals