Meta-Optimization of Initial Weights for More Effective Few- and Zero-Shot Learning in BCI Classification

Brain-computer interfaces (BCIs) are heavily reliant on the underlying data classification. Neural network classifiers are often used for this purpose, but their performance is dependent, in turn, on the availability of large training sets, which are difficult to record. Hence, arises the necessity to employ methods capable of operating with limited sample sizes or leveraging experience acquired with different BCI users. Here, we explore the ability of meta-learning algorithms to enable neural network classifiers to leverage experience acquired with EEG data recorded in other BCI users to support few-shot or even zero-shot learning for BCI classifiers. We conducted experiments to assess the quality of EEG data classification using neural networks that were pre-trained on various users who were different from the test user. In these experiments we compared neural networks pre-trained with meta-learning algorithms and with traditional transfer learning, with further fine-tuning on small data amounts and even without fine-tuning. The experiments demonstrated the potential for classification quality improvement through meta-learning in few- and even zero-shot learning scenarios.