Signal Processing Approaches to Minimize or Suppress Calibration Time in Oscillatory Activity-Based Brain–Computer Interfaces

Abstract

One of the major limitations of Brain-Computer Interfaces (BCI) is their long calibration time, which limits their use in practice, both by patients and healthy users alike. Such long calibration times are due to the large between-user variability and thus to the need to collect numerous training electroencephalography (EEG) trials for the machine learning algorithms used in BCI design. In this paper, we first survey existing approaches to reduce or suppress calibration time, these approaches being notably based on regularization, user-to-user transfer, semi-supervised learning and apriori physiological information. We then propose new tools to reduce BCI calibration time. In particular, we propose to generate artificial EEG trials from the few EEG trials initially available, in order to augment the training set size. These artificial EEG trials are obtained by relevant combinations and distortions of the original trials available. We propose 3 different methods to do so. We also propose a new, fast and simple approach to perform user-to-user transfer for BCI. Finally, we study and compare offline different approaches, both old and new ones, on the data of 50 users from 3 different BCI data sets. This enables us to identify guidelines about how to reduce or suppress calibration time for BCI.

Heading for new shores! Overcoming pitfalls in BCI design

“…It is therefore essential to estimate the generalization of selected models. Methods such as regularization, shrinkage or cross validation may prevent over fitting to the data [81-84]. Overfitting means that PR models memorize the training data.…”

Online and Offline Domain Adaptation for Reducing BCI Calibration Effort

“…Many signal processing and machine learning approaches have been proposed to reduce the BCI calibration effort [29], [30]. They may be grouped into five categories [29]: 1) Regularization, which is a very effective machine learning approach for constructing robust models [41], especially when the training data size is small.…”

“…They may be grouped into five categories [29]: 1) Regularization, which is a very effective machine learning approach for constructing robust models [41], especially when the training data size is small. A popular regularization approach in BCI calibration is shrinkage [27], which gives a regularized estimate of the covariance matrices.…”

Evolving Training Sets for Improved Transfer Learning in Brain Computer Interfaces

“…Both cases however, rely upon the natural distribution of the data as grouped by their original participant. Some attempts have been made to group datasets by known variants such as gender [8], and others using the information extracted from the trained models [9]; but little has been done in regards to instance selection for each model.…”

“…This, like most BCI ensembles [11], used naive partitioning in which the instances were divided by their associated labels, whether it be by source domain or by stimuli. This proves useful for weighting classifiers within the ensembles; allowing information regarding the appropriateness of each model and the test-domain to be extracted [9]. It was demonstrated in [11] that overlapping these naive divisions can actually increase accuracy, suggesting that having the same training data duplicated amongst the classifiers can benefit the overall performance.…”