A meta-learning BCI for estimating decision confidence.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Brain-computer interfaces (BCIs) typically require extensive user-specific training data.
• Traditional machine learning methods face challenges with noisy and variable electroencephalography (EEG) and electro-oculogram (EOG) data.
• Improving decision-confidence prediction in BCIs is crucial for reliable human-machine interaction.

Purpose of the Study:

• To evaluate a novel meta-learning transfer technique for enhancing BCI performance in decision-confidence prediction.
• To compare the meta-learning approach against traditional single-subject training and other advanced transfer learning methods.
• To assess the efficacy of meta-learning in scenarios with limited new participant data.

Main Methods:

• Adapted the meta-learning by biased regularization algorithm for predicting decision confidence from EEG and EOG data.
• Utilized a difficult target discrimination task based on video feeds.
• Compared the meta-learning approach with single-subject training, domain adversarial neural networks, and a zero-training method adaptation.

Main Results:

• The meta-learning approach demonstrated significantly superior performance compared to other methods across most conditions.
• Meta-learning particularly excelled when limited data from new participants was available for training/tuning.
• The method effectively integrated prior participant data with new user data, yielding high-performance predictors.

Conclusions:

• Meta-learning by biased regularization offers a robust and efficient solution for BCI training, especially with limited data.
• This technique significantly enhances decision-confidence prediction accuracy in BCIs, outperforming traditional methods.
• The findings suggest a promising direction for developing more adaptive and user-friendly BCIs.