MI-EEGNET: A novel Convolutional Neural Network for motor imagery classification

[jinglescode]

Paper

Link: https://www.sciencedirect.com/science/article/pii/S016502702030460X
Year: 2020

Summary

• outperformed existing methods on BCI Competition IV dataset IIa

Contributions and Distinctions from Previous Works

• RNN -> hard to train
• transform data into the frequency domain or tomography images -> increase dimensionality and require high sampling signals
• do not use any pooling layer, reason is to learn a frequency-specific spatial filter that can help to understand the behavior of the ConvNet

Methods

• size of the kernel corresponds to a time duration of 0.25s where the EEG signal is considered as quasi-stationary signals
• Exponential Linear Unit (ELU) that leads to better results than ReLU, avoids from its inconvenient, and protects against the vanishing gradient
• parallel pipeline where each one will extract features from different filters of different scales

Results

Comments

• this paper has a few papers for justify why params are used