Memory Augmented Neural Networks in Brain-Computer Interfaces
Motor-imagery Brain-Computer Interfaces (MI-BCIs) aim to establish a communication channel between the human brain and an external device, solely based on human’s thought. Due to high inter-session and inter-user variance, however, users may need several training sessions while wearing EEG electrodes to reach an acceptable accuracy. This makes the usage of pure convolutional neural network (CNN) architectures very limited in this area, and less practical for multi-users deployment. A viable option are memory augmented neural networks (MANNs)  which augment CNNs with an external binary memory. This opens up a large variety of options that can significantly improve the applicability of MI-BCI, e.g.,
- Add a new MI class on the fly without retraining the whole model.
- Calibrate the model at the beginning of a new session to mitigate high inter-session variance in EEG.
- Calibrate the model on a new, unseen user due to high inter-user variance.
In this project, you start with a given CNN controller , augment it with an external memory, and explore all aforementioned options.
Status: In Progress
- Machine Learning
- 20% Theory
- 80% Implementation
-  Laguna et al., Design of Hardware-Friendly Memory Enhanced Neural Networks
-  Lawhern et. al., EEGNet: A Compact Convolutional Network for EEG-based Brain-Computer Interfaces