Abstract
Decoding cognitive states from Electroencephalography (EEG) signals is complicated by inherent biological processing delays and the high energy demands of traditional neural networks. This paper introduces the Delayed Knowledge Transfer (DKT) framework, which utilizes biologically inspired Spiking Neural Networks (SNNs) to improve attention detection. By employing a Teacher-Student learning paradigm with a novel, learnable Delay Layer, the framework aligns external audiovisual stimuli with delayed neural responses. Experimental results on the "WithMeAttention" dataset demonstrate a 3% accuracy improvement over baseline models and a 36.39% reduction in energy consumption, highlighting the potential for efficient, real-time Brain-Computer Interfaces (BCIs).
Introduction
Brain-Computer Interfaces (BCIs) often rely on EEG to monitor attention, but mapping these signals to external events is hindered by two primary factors. First, the biological lag, known as Event-Related Potentials (ERPs), creates a temporal misalignment between a stimulus and the brain's electrical response. Second, traditional Artificial Neural Networks (ANNs) are computationally expensive, limiting their utility in wearable, battery-powered BCI devices.
Methods
The DKT framework addresses these challenges through a Teacher-Student (T-S) paradigm built on Spiking Neural Networks (SNNs). SNNs utilize Leaky Integrate-and-Fire (LIF) neurons that only compute when a voltage threshold is reached, mimicking the energy efficiency of the human brain.
The DKT Framework
The system consists of two primary components:
- Teacher Model: Trained on visual and auditory sequences to extract features of "attention-grabbing" stimuli.
- Student Model: An SNN that processes raw 64-channel EEG signals.
The core innovation is the learnable Delay Layer. This layer mathematically determines the optimal time offset (τ) to align stimulus features from the Teacher with the corresponding brain responses in the Student. To handle the non-differentiable nature of discrete time delays, the authors employed a Straight-Through Estimator (STE) during backpropagation.
Dataset: WithMeAttention
The researchers developed the "WithMeAttention" benchmark, involving 20 participants performing a digit-memorization task under four conditions: purely visual (C1), rhythmic visual (C2), visual with non-rhythmic audio (C3), and rhythmic visual with rhythmic audio (C4).
Results and Discussion
The DKT framework achieved a peak accuracy of 59.73% using a 0.17-second decision window, outperforming established models like EEGNet and CNNNet.
Key Findings
- Rhythmic Synchronization: Rhythmic stimuli (C2 and C4) significantly enhanced decoding accuracy. Rhythm appears to act as a "neural hook," synchronizing attentional oscillations and producing more predictable EEG patterns.
- Bimodal Processing Delays: The learned delays typically clustered around 200ms and 450ms, reflecting the biological reality that auditory information is generally processed faster than complex visual data.
- Energy Efficiency: The spiking implementation reduced computational costs by 36.39% (measured in picojoules) compared to standard ANNs, making it highly suitable for edge BCI hardware.
Conclusion
The Delayed Knowledge Transfer framework successfully bridges the gap between external stimuli and internal neural states by accounting for biological latency. By combining cross-modal knowledge transfer with energy-efficient spiking neurons, DKT provides a robust foundation for the next generation of real-time, wearable attention-monitoring systems. Future work will focus on refining individual variability in audiovisual binding and integrating more complex attention mechanisms within the SNN architecture.
Reference: Sun, P., et al. (2024). Delayed Knowledge Transfer: Cross-Modal Knowledge Transfer from Delayed Stimulus to EEG for Continuous Attention Detection Based on Spike-Represented EEG Signals. Preprint submitted to Elsevier.
Summary updated