DL Models
Model 1: EEGNet
Application: General EEG classification and Brain-Computer Interface (BCI) tasks.
Paper/GitHub:
EEGNet: A Compact Convolutional Network for EEG-based Brain-Computer Interfaces, Model Description:
Initial Convolutional Layer: 1 layer Depthwise Convolutional Layer: 1 layer Separable Convolutional Layer: 1 layer Classification Layer: 1 layer
Justification: Designed to be compact and efficient for EEG data processing, capturing both spatial and temporal features. Types: Conv2D, DepthwiseConv2D, SeparableConv2D, Linear Loss Function: Cross-Entropy Loss Metrics: Accuracy, Precision, Recall, F1-Score Performance: Achieves high accuracy on multiple BCI datasets, often above 90% for binary classification tasks, 60-80% for multiclass classification Customized code on my data:
Model 2: DeepConvNet
Application: General EEG tasks, more complex and deeper representations.
Paper/GitHub:
Deep learning with convolutional neural networks for EEG decoding and visualization, Model Description:
Several convolutional layers followed by fully connected layers. Typically includes 4 Conv2D layers and 2 fully connected layers. Justification: Deeper layers allow for capturing more complex patterns in EEG data. Types: Conv2D, MaxPooling, Dense (Fully Connected) Loss Function: Cross-Entropy Loss Metrics: Accuracy, Precision, Recall, F1-Score Performance: Known for high accuracy, robustness across different EEG tasks. Customized code on my data:
Model 3: CNN-LSTM Hybrid
Model 4: Spatio-Temporal Convolutional Networks (STCNN)
Model 5: Convolutional Bi-Directional LSTM (Conv-BiLSTM)
Model 6: 3D Convolutional Neural Networks (3D-CNN)
Model 7: Deep Residual Networks (ResNet) for SSVEP
Model 8: Temporal Convolutional Networks (TCN)
Model 9: EEGTransformer
Dropout Rate (dropoutRate)
Explanation: Dropout is a regularization technique used to prevent overfitting by randomly setting a fraction of input units to 0 during training. Typical Values: Common values are between 0.2 and 0.5. High Dropout Rate (e.g., 0.5): Use if you have a small dataset or observe overfitting. Low Dropout Rate (e.g., 0.2): Use if you have a large dataset or observe underfitting. Importance: Dropout helps in generalizing the model by preventing it from relying too much on any single neuron. Temporal Kernel Length (kernLength)
Explanation: This defines the size of the convolutional filter applied over the temporal dimension of the input data. Rule of Thumb: It can be set to approximately half the sampling rate. For instance, if your sampling rate is 128 Hz, a kernel length of 64 is a reasonable starting point. Importance: This affects how the model captures temporal features. Longer kernels can capture more extended temporal dependencies.
Number of Temporal Filters (F1)
Explanation: The number of filters in the first convolutional layer, capturing different temporal features. Typical Values: Common values are 8 or 16. Importance: More filters can capture more diverse features but also increase the computational load.
Depth Multiplier (D)
Explanation: This defines the number of spatial filters learned within each temporal convolution. Typical Values: Common values are 1 or 2. Importance: Higher values allow capturing more spatial features but increase model complexity.
Number of Pointwise Filters (F2)
Explanation: This is usually set to F1 * D, ensuring the separable convolution has the same number of filters as the combined depthwise convolution. Importance: Determines the complexity and capacity of the separable convolution layer. 
