Online eye-movement classification with temporal convolutional networks.

Area of Science:

• Computer Science
• Biomedical Engineering
• Human-Computer Interaction

Background:

• The ternary eye-movement classification problem (3EMCP) is crucial for real-time interactive applications.
• Existing deep learning models (1D-CNN-BiLSTM, TCN) excel in offline 3EMCP but are not optimized for online use.
• There is a need for fast, low-latency eye-movement classification methods for dynamic applications.

Purpose of the Study:

• To propose a Temporal Convolutional Network (TCN) classifier for the 3EMCP adapted for online applications.
• To develop a lightweight preprocessing technique enabling real-time TCN predictions.
• To evaluate the TCN's performance against other deep learning and Bayesian classifiers in an online setting.

Main Methods:

• A Temporal Convolutional Network (TCN) classifier was adapted for online 3EMCP without look-ahead buffers.
• A novel lightweight preprocessing technique was introduced to facilitate real-time predictions.
• The TCN was evaluated against CNN-LSTM, CNN-BiLSTM, and an I-BDT Bayesian classifier using two public datasets.

Main Results:

• The proposed TCN model demonstrated consistent performance superiority across all eye-movement classes compared to other evaluated methods.
• Real-time predictions were achieved at approximately 500 Hz with low latency on commodity hardware.
• All methods showed improved performance with the inclusion of look-ahead information, although reasonable accuracy was achieved with zero look-ahead.

Conclusions:

• The TCN classifier offers an accurate, robust, and fast solution for online ternary eye-movement classification.
• The lightweight preprocessing and TCN architecture enable efficient real-time eye-tracking applications.
• Further improvements in 3EMCP accuracy are possible by incorporating look-ahead data, but the TCN provides strong performance even without it.