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

TL;DR

This work tackles cognitive load classification in air traffic control using multimodal EEG and eye-tracking features. It compares conventional ML baselines with spiking neural networks trained via a delta-rule and demonstrates a hardware-enabled, quantized SNN deployed on the DYNAP-SE neuromorphic chip. The minimal single-layer SNN achieves 80.6% accuracy in software, and 73.5% on hardware after quantization, illustrating the feasibility of ultra-low-power, event-driven cognitive monitoring in real-world settings. The findings underscore the potential of neuromorphic approaches for real-time, energy-efficient cognitive-state monitoring in dynamic environments like ATC, with implications for wearable neurotechnology and safety-critical applications.

Abstract

This paper presents a neuromorphic system for cognitive load classification in a real-world setting, an Air Traffic Control (ATC) task, using a hardware implementation of Spiking Neural Networks (SNNs). Electroencephalogram (EEG) and eye-tracking features, extracted from an open-source dataset, were used to train and evaluate both conventional machine learning models and SNNs. Among the SNN architectures explored, a minimalistic, single-layer model trained with a biologically inspired delta-rule learning algorithm achieved competitive performance (80.6%). To enable deployment on neuromorphic hardware, the model was quantized and implemented on the mixed-signal DYNAP-SE chip. Despite hardware constraints and analog variability, the chip-deployed SNN maintained a classification accuracy of up to 73.5% using spike-based input. These results demonstrate the feasibility of event-driven neuromorphic systems for ultra-low-power, embedded cognitive state monitoring in dynamic real-world scenarios.

Figures (3)

• Figure 1: System overview of the multimodal cognitive load classification setup. (a) The EEG electrode layout follows the standard 10–20 system, with Emotiv EEG sensor positions highlighted in orange. (b) EEG signals were recorded using Emotiv EPOC+ EEG headset, while gaze data were captured via a screen-based eye tracker. The panel also shows representative EEG and gaze-derived physiological features, including gaze position (X, Y), and oscillatory brain activity across standard frequency bands. (c) These features were used as input to a neuromorphic model. The extracted features were processed using the mixed-signal spiking neural network processor DYNAP-SE, which contains four neuromorphic cores, each supporting up to 256 analog neurons.
• Figure 2: Architectures of the SNN used for cognitive load classification. (a) SNN with hidden layer: Input features are spike-encoded using LIF neurons and passed through a sparse (P_conn) fixed-weight layer (fc1), followed by a trainable output layer (fc2). (b) SNN without hidden layer: Spike-encoded inputs are directly connected to the output layer (fc1), with no intermediate hidden neurons. In both models, the output layer consists of two LIF neurons with weak mutual inhibition and a softmax readout, enabling WTA behavior. All output weights are trained using a delta learning rule adapted to spiking activity.
• Figure 3: Example Dynapse trial result illustrating the spike-based classification process. The raster plot (top) shows spike events from output neurons of Class 0 (blue) and Class 1 (red). The bottom panel shows sliding-window firing rates computed over 2.5 ms intervals. In this trial, stronger activity from the Class 1 population led to a correct Class 1 classification.