Introduction
Gaze tracking is a vital tool for understanding human cognition, decision-making, and health. While traditional infrared eye trackers are highly accurate, they often require complex calibration and are sensitive to physical variations like pupil-iris contrast. Electroencephalography (EEG) presents a non-invasive alternative with exceptional temporal resolution. However, EEG-based gaze estimation faces two primary challenges: the "black box" nature of deep learning models, which hinders neuroscientific trust, and the high level of signal noise caused by muscle artifacts or poor sensor contact.
To address these issues, researchers have developed Attention-CNN, a framework designed to be both highly accurate and inherently interpretable. By integrating specific attention mechanisms, the model can explain its predictions and robustly handle noisy data.
The Attention-CNN Architecture
The core of the proposed method is the Attention-CNN, which enhances a standard convolutional neural network with two specialized modules: Squeeze-and-Excitation (SE) blocks and Self-Attention (SA) blocks.
- Squeeze-and-Excitation (SE) Blocks: These are embedded within each residual block to perform electrode-wise weighting. By compressing temporal data, the SE blocks identify which electrodes provide the most reliable signals at any given moment, effectively acting as an automated denoising filter.
- Self-Attention (SA) Blocks: Positioned before the final prediction layer, these modules capture global dependencies across the entire signal. This allows the network to prioritize features relevant to specific tasks, such as distinguishing between a steady fixation and a rapid saccade.
Experimental Framework
The model was evaluated using the EEGEyeNet dataset, a benchmark containing synchronized 128-channel EEG and high-precision infrared eye-tracking data. The study focused on two primary tasks:
- Position Task: Estimating absolute (x, y) gaze coordinates during fixation.
- Direction Task: Predicting the amplitude and angle of saccades.
To ensure the model's ability to generalize to new users, the researchers employed a strict 70/15/15 subject-wise split, preventing the network from simply memorizing individual participant patterns.
Interpretability and Biological Validation
A key innovation of this framework is treating individual electrodes as the primary entities for interpretation. Because electrodes correspond to specific cortical regions, the weights assigned by the attention blocks can be visualized as "heatmaps" of brain activity.
These heatmaps often highlight pre-frontal areas, which aligns with biological understanding of the cognitive control of eye movements and muscle activity. This transparency allows researchers to validate that the model is making decisions based on relevant physiological signals rather than spurious correlations.
Performance and Robustness
The Attention-CNN consistently outperformed baseline models by 5% to 10%. In the Position Task, it achieved a Euclidean distance error of 110.05 pixels (~2.28° visual angle). In the Direction Task, it reached an angle RMSE of 0.1707 radians.
Beyond raw accuracy, the model demonstrated remarkable robustness. In cases of extreme sensor noise (amplitudes >100 µV), the Attention-CNN reduced prediction error by over 200 pixels compared to standard CNNs. It achieved this by automatically assigning near-zero weights to malfunctioning electrodes, such as those near the ears that often pick up non-neural artifacts.
Conclusion
The Attention-CNN framework represents a significant step forward in making EEG-based gaze estimation practical and trustworthy. By combining high-performance deep learning with built-in interpretability, the model offers a tool that neuroscientists can use to both track gaze and understand the underlying neural drivers. Future work will focus on refining the mathematical metrics for explainability and applying this robust architecture to other clinical EEG tasks like seizure detection.
Summary updated