Cognitive load estimation from eye-tracking data via cross-domain & test-time adaptation

Abstract

While the estimation of cognitive load through eye-tracker data has proven to hold immense potential for Human-Computer Interaction, existing implementations are often hindered by significant inter-subject variability and the need for calibration. Domain adaptation techniques attempt to mitigate these shifts by aligning source & target distributions globally. However, this creates a static model that is unable to account for instance-specific variations, namely physiological differences, lighting changes, sensor dropouts, and changes in hardware setups. Integrating test-time adaptation into the pipeline allows for real-time normalization of the model’s statistics using unlabeled streaming data from the test subject, which has been a challenge in prior studies on creating impartial cognitive load classifiers. This makes the model dynamic, which is validated by benchmarking using the COLET & ADABase datasets, representing varied scenarios, subjects, and setups. Our results show that adding the dimension of test-time adaptation improves generalization accuracy without requiring explicit recalibration, which is crucial as the use of extended reality (XR) headsets and the applications of situationally-aware interfaces are rapidly becoming widespread.