Investigating Ensemble Learning and Classifier Generalization in a Hybrid, Passive Brain-Computer Interface for...

Area of Science:

• Neuroscience and Cognitive Science
• Machine Learning and Artificial Intelligence
• Human-Computer Interaction

Background:

• Hybrid, passive brain-computer interfaces (h/pBCI) are crucial for measuring mental states, particularly operator workload.
• Accurate classification of workload is essential for enhancing operator performance.
• Physiological measures used in machine learning for workload classification are prone to nonstationarity, limiting generalizability.

Purpose of the Study:

• To investigate ensemble learning methods for creating more generalizable brain-computer interface classifiers.
• To improve classification accuracy and reduce variance in workload state detection.
• To leverage ensemble theoretical performance for inferring generalizability in h/pBCI systems.

Main Methods:

• Implemented ensemble learning, specifically adaptive boosting (AdaBoost), to train multiple base learning algorithms.
• Utilized three base learning algorithms: artificial neural network (ANN), support vector machine (SVM), and linear discriminant analysis (LDA).
• Compared classifier performance using single-day versus multi-day training paradigms to assess generalizability.

Main Results:

• Ensemble learning approaches demonstrated potential for improving accuracies and reducing variance in pBCI applications.
• The ensemble model converged on theoretical performance regarding error and variance exclusively when training sets utilized a multi-day paradigm.
• Observed that ensemble convergence on theoretical performance can provide insights into generalizability, especially when simple accuracy is misleading.

Conclusions:

• Ensemble learning methods are effective for enhancing brain-computer interface (BCI) classification, but their generalizability is influenced by physiological nonstationarity.
• The multi-day learning paradigm is critical for achieving ensemble convergence on theoretical performance, indicating improved generalizability.
• Evaluating ensemble convergence offers a robust method for inferring classifier generalizability in BCI systems susceptible to nonstationary physiological data.