🏆 Best Paper Nominee April - August 2025

Edge-AI Infrastructure for Real-time Cognitive State Monitoring

Neuromorphic deployment of spiking neural networks for cognitive load classification in Air Traffic Control applications

View Code Read Paper

Overview

This project addresses the critical need for real-time cognitive state monitoring in safety-critical applications. We developed a complete pipeline from physiological signal acquisition to on-chip inference using neuromorphic hardware.

Ultra-Low Latency

Real-time inference suitable for time-critical decisions

μW Power Consumption

Neuromorphic design enables always-on monitoring

ATC Domain

Validated on Air Traffic Control cognitive load dataset

System Architecture

End-to-End Neuromorphic Cognitive Load Classification Pipeline ATC Operator Air Traffic Control EEG 32ch @ 256Hz Eye Tracker Pupil + Gaze Preprocessing • Bandpass filter • Band powers • α/β/θ/δ • Eye features Spike Encoder Rate coding Poisson spike DYNAP-SE Neuromorphic Chip Core 1 Core 2 Core 3 Core 4 1.2 mW Output Low High < 10ms μW power 85%+ acc

Figure 1: End-to-end pipeline from multimodal sensor acquisition (EEG, eye-tracking) through feature extraction and spike encoding to neuromorphic inference on the DYNAP-SE chip.

My Role

First Author & Lead Researcher
  • Algorithm Design: Developed the spiking neural network architecture optimized for neuromorphic deployment
  • Hardware Integration: Configured and deployed models on the DYNAP-SE mixed-signal neuromorphic processor
  • Evaluation: Conducted comprehensive benchmarking of latency, power, and accuracy trade-offs
  • Paper Writing: Led the manuscript preparation for ACM/IEEE ICONS 2025

Methodology

Feature Extraction Pipeline Raw EEG 32 channels Eye Data Pupil + Fixation Bandpass 0.5-45 Hz Notch 50/60 Hz Band Powers δ θ α β + Eye features blink, pupil, saccade Feature Vector 5 features / window 4sec sliding window Single-Layer SNN with Winner-Takes-All 5 Features f1 f2 f3 f4 f5 Spike Injection 5 LIF Neurons DYNAP-SE Core 2 Readout Low High Mutual Inhibition Winner-Takes-All

Figure 2: (Left) Feature extraction pipeline with EEG band powers and eye-tracking metrics. (Right) Single-layer SNN: 5 features → 5 LIF neurons → 2 readout with mutual inhibition (Winner-Takes-All).

What's Innovative

This work represents the first demonstration of cognitive workload classification on mixed-signal neuromorphic hardware (DYNAP-SE). By leveraging the event-driven nature of spiking neural networks, we achieve:

μW Power Range
<10ms Inference Latency
73.5% On-Chip Accuracy

This enables always-on monitoring for safety-critical applications where continuous operator state assessment is essential but power/form-factor constraints prohibit traditional GPU-based approaches.

Results

Power vs Accuracy Trade-off 90% 85% 80% 75% Accuracy Power Consumption μW mW W GPU 89% CPU 87% Quant 86% SE 73.5% Power Usage 1.2mW (ours) ~5W CPU ~100W GPU On-Chip Classification Results Predicted Actual Low High Low High 87 87% 13 13% 17 17% 83 83% On-Chip: 73.5%

Figure 3: (Left) Power vs accuracy comparison showing DYNAP-SE achieves competitive accuracy at 4000× lower power than GPU. (Right) Binary classification confusion matrix on neuromorphic hardware.

Publication

🏆 Best Paper Award Nominee

Neuromorphic Deployment of Spiking Neural Networks for Cognitive Load Classification in Air Traffic Control

An, J., Cai, C., Gallou, O., Fabrikant, S. I., Indiveri, G., & Donati, E.

2025 International Conference on Neuromorphic Systems (ICONS), 795–134

https://doi.org/10.1109/ICONS69015.2025.00028

PDF Code Cite

Related Projects

Multimodal SNN for Workload Monitoring

Fusion architectures for heterogeneous biosignals

Neuroadaptive Learning System

Closed-loop cognitive load adaptation