IEEE Computational Intelligence Magazine - November 2022 - 24

TABLE VII Mean standard deviation (%) of offline cross-subject unsupervised classification accuracies on MI1.
APPROACH
S0
S1
S2
S3
S4
EEGNet
Baseline+SHOT 60.382.43 27.621.07 73.961.82 41.491.93 29.971.32 38.252.10 42.592.06 70.022.30 68.041.51 50.251.01
MDMAML+SHOT 62.222.01 29.681.56 75.212.21 45.812.06 29.521.42 40.301.57 43.921.45 71.671.95 69.802.03 52.011.98
Shallow ConvNet
Baseline+SHOT 58.373.16 24.341.75 61.601.62 37.151.91 25.951.52 33.981.51 32.932.69 63.922.23 64.102.82 44.701.60
MDMAML+SHOT 60.663.30 26.631.90 62.991.65 38.551.94 27.611.72 35.631.71 34.512.81 65.502.36 66.122.97 46.471.75
S5
S6
S7
S8
AVG
TABLE VIII Mean standard deviation (%) of offline cross-subject unsupervised classification accuracies on MI2.
APPROACH
S0
Baseline+SHOT
MDMAML+SHOT
S1
S2
S3
S4
S5
S6
EEGNet
63.30 79.40 86.10 69.10 57.80 56.80 83.50 49.60 94.30 68.50 78.10 61.00 56.30 50.0 68.13
3.80 2.37 3.25 3.25 5.65 6.47 4.07 6.59 2.34 1.41 3.28 2.62 4.70 2.56 2.01
67.00 81.00 87.80 69.14 60.90 60.00 83.40 52.10 94.00 70.80 80.40 63.90 58.00 49.90 69.88
4.41 2.56 3.87 3.11 4.57 4.59 4.67 5.01 3.01 2.41 4.09 2.00 3.98 3.52 2.34
Shallow ConvNet
Baseline+SHOT
MDMAML+SHOT
59.00 72.60 85.00 63.80 56.40 59.40 75.20 49.00 83.70 69.30 65.20 58.70 59.50 56.00 65.20
3.16 3.10 5.33 4.58 5.30 2.46 4.51 4.54 19.49 3.98 2.99 3.41 2.73 3.19 1.03
60.74 74.34 86.93 65.73 58.46 61.46 76.83 50.63 84.85 70.45 67.39 60.89 60.74 57.24 66.91
3.18 3.12 5.47 4.71 5.35 2.51 4.61 4.64 19.63 4.11 3.10 3.51 2.75 3.21 1.44
S7
S8
S9
S10
S11
S12
S13
AVG
2) MDMAML is simpler to implement as it does not include
random task sampling. MAML requires more complicated
settings of hyperparameters, including task numbers, task
size, etc. Additionally, MAML generally assumes the dataset
is class-balanced, and hence each sampled task contains
approximately equal number ofsamples from different classes.
In EEG-based BCIs, paradigms such as ERP are intrinsically
class-imbalanced. Thus, an under-sampling/upsampling
preprocessing step is needed before MAML.
3) MDMAML handles negative transfer by preventing
gradient updates if updating u in (1) decreases the model
performance on Sj. MAML does not particularly consider
negative transfer, and hence it would suffer from bad data
or irrelevant/dissimilar source domains. An illustration of
the impact on the model performance is shown in Fig. 5.
Finally, the Expected Calibration Error (ECE) and Maximum
Calibration Error (MCE) [65] were used to evaluate the
gap between a model's predicted probabilities and the true
results. Predictions are grouped into 10 bins with different levels
ofprediction confidence. The difference between accuracy and
confidence for a given bin represents the calibration gap. ECE
and MCE represent the average and maximum gap ofthe bins,
respectively. Lower ECE and MCE scores indicate better performance.
ECE and MCE scores on MI1 are shown in Fig. 6.
MDMAML achieved lower ECE and MCE scores, indicating
that its predicted probabilities were closer to the true labels.
I. Privacy-Protection Validation
MDMAML only saves the model initialization parameters u,
instead of the source domain raw EEG data, for adapting to a
new subject. Thus, it protects the privacy of the source data
24 IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE | NOVEMBER 2022
FIGURE 5 Test classification accuracy during MDMAML training without
negative transfer mitigation. Lines in red show adapting to Subject s1 in
MI1 during stochastic inner loop updates. Model performance usually
drops when transferring across dissimilar domains.
providers. To further demonstrate it, Model Inversion
Attack [66], which aims to reconstruct training data utilizing
only the model, was performed to examine ifthe identity ofthe
source data providers can be revealed fromfu.
EEGNet was first trained using MDMAML to obtain fu,
and then a state-of-the-art approach, DeepInversion [67], was
adopted to reconstruct the training data fromfu. More specifically,
data were randomly initialized in the input space and
their output probability ofa specific class was maximized using
modelfu. To force the reconstructed data to resemble the original
training data distribution, the running means and variances
ofthe data batch were restricted to match those ofthe model's
batch normalization layers.
IEEE Computational Intelligence Magazine - November 2022

Table of Contents for the Digital Edition of IEEE Computational Intelligence Magazine - November 2022
