IEEE Computational Intelligence Magazine - November 2022 - 22

ProtoNets only used the feature extractor of the networks,
and was trained for 100 epochs. The three metalearning
algorithms (MAML, ANIL and MDMAML) were
all trained for 200 epochs with meta-learning rate 0.001
and inner-loop learning rate 0.001. The task number of
MDMAML was 1, i.e., all trials in each domain was used.
For MAML and ANIL, the task number was 128 for n-way
k-shot randomly formed small batches, where n was the
number of classes, and k wascalculatedsuchthatevery sample
in the source domains was sampled once. All inner loop
adaptation steps were 1, and the calibration adaptation step
was 5. First-order approximates were used to replace second-order
gradient calculations to reduce the training time,
as suggested in [62]. For the two ERP datasets, the inverse
classratio of thetrainingset wasusedasclass weightsinthe
loss function to mitigate class-imbalance. The Area Under
the Receiver Operating Characteristic Curve (AUC) was
used as the evaluation metric for the ERP datasets, instead
of accuracy, due to the class-imbalance.
All computations were performed on a single GeForce
RTX 3090 GPU. PyTorch and learn2learn4
were used to
implement all algorithms. Meta training took approximately
1 second per epoch. All source code is available on GitHub.5
FIGURE 4 Cosine similarity of layer parameters before and after
adaptation. (a) EEGNet; (b) Shallow ConvNet.
domain. The amount ofparameter changes should reflect each
layer's sensitivity when facing domain shift.
Fig. 4 shows the cosine similarity of the layer parameters
before and after adaptation on MI1. Clearly, the Conv1 layer
ofEEGNet and all four Conv layers ofShallow ConvNet had
little change after adaptation, so they were frozen during inner
loop training. Again, this does not prevent these layers from
learning, because they are still updated in the outer loop.
E.Algorithms
The following algorithms, which can be used in online EEGbased
BCI applications, were compared with MDMAML:
1) Baseline, which was trained for 100 epochs with learning
rate 0.001 on all source domain data combined. The model
is further fine-tuned for 5 more epochs using few-shot
labeled calibration data from the test subject.
2) MAML, which forms the batch pairs randomly.
3) Almost No Inner Loop (ANIL) [55], which only adapts the
head ofthe network in the inner loop.
4) Prototypical Networks (ProtoNets) [46], a classical metalearning
approach. Note that ProtoNets is not applicable to
the 0-shot scenario.
22 IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE | NOVEMBER 2022
F. Online Experimental Results
Leave-one-subject-out cross-validation was performed, i.e.,
each subject in the corresponding dataset was treated as the test
subject once, with all remaining subjects as the source
domains in training. At the calibration stage, k (a very
small number) labeled samples from each of the n classes
were randomly provided for adaptation, usually known as
n-way k-shot learning. The model was then tested on the
remaining trials of the test subject. All experiments were
repeated 10 times.
Note that the batch normalization layers were deactivated
during testing, ensuring the outputs can be evaluated in realtime
in online applications. In the original MAML setting,
transductive batch normalization is used at test time [63].
However, exploiting unlabeled test data distribution to boost
the performance is impossible in online applications. The test
data distribution information was prevented from being
accessed through any means, ensuring the causality of online
experiments.
The average classification accuracies for MI and AUC
scores for ERP on the four datasets are shown in Tables II-V,
respectively. The best result for each k is marked in bold. Our
proposed MDMAML outperformed all other approaches.
To check if the performance improvement ofMDMAML
over every other algorithm was statistically significant,
Wilcoxon signed-rank tests [64] were performed, and the
p-values are shown in Table VI. The performance improvements
of MDMAML over Baseline, MAML and ProtoNets
4[Online]. Available: http://learn2learn.net/
5[Online]. Available: https://github.com/sylyoung/MetaEEG
http://learn2learn.net/ https://www.github.com/sylyoung/MetaEEG
IEEE Computational Intelligence Magazine - November 2022

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