Transfer Learning in Brain-Computer Interfaces: Language-Pretrained Transformers for Classifying Electroencephalography

Created on April 29|Last edited on June 30

Comment

﻿
A summary of the results of the empirical evaluations of my master's thesis at the Vrije Universiteit Brussel.
﻿
﻿
﻿
Subject-wise average training loss
Subject-wise average training loss
051015202530Epochs0.20.40.60.811.21.4Loss
Data: features  Decay: 1  
Dropout: 0.1 
Pretrained: true  Frozen: 1  to 12
Data: time-series  Decay: 1  
Dropout: 0.1 
Pretrained: true  Frozen: 1  to 12
Subject-wise average training classification accuracy
Subject-wise average training classification accuracy
051015202530Epochs0.40.60.8Score
Data: features  Decay: 1  
Dropout: 0.1 
Pretrained: true  Frozen: 1  to 12
Data: time-series  Decay: 1  
Dropout: 0.1 
Pretrained: true  Frozen: 1  to 12
Hyperparameter optimisation bests18
 
Addressing overfitting9
 
Randomly initialised18
 
Unfreezing deep-to-shallow features99
 
Unfreezing shallow-to-deep features99
 
Unfreezing deep-to-shallow time-series99
 
Unfreezing shallow-to-deep time-series99
 
Completely unfrozen features9
 
Completely unfrozen time-series9
﻿
﻿

Add a comment