Abstract
Perceptual processes can be probed by fitting stimulus-response models that relate measured brain signals such as electroencephalography (EEG) to the stimuli that evoke them. These models have also found application for the control of devices such as hearing aids. The quality of the fit, as measured by correlation, classification, or information rate metrics, indicates the value of the model and the usefulness of the device. Models based on Canonical Correlation Analysis (CCA) achieve a quality of fit that surpasses that of commonly-used linear forward and backward models. Here, we show that their performance can be further improved using several techniques, including adaptive beamforming, CCA weight optimization, and recurrent neural networks that capture the time-varying and context-dependent relationships within the data. We demonstrate these results using a match-vs-mismatch classification paradigm, in which the classifier must decide which of two stimulus samples produced a given EEG response and which is a randomly chosen stimulus sample. This task captures the essential features of the more complex auditory attention decoding (AAD) task explored in many other studies. The new techniques yield a significant decrease in classification errors and an increase in information transfer rate, suggesting that these models better fit the perceptual processes reflected by the data. This is useful for improving brain-computer interface (BCI) applications.
