Dae Hyeon Kim RESEARCH 한국어

EEG graph learning and compact decoding

Graph contrastive learning and compact temporal models were developed for EEG emotion recognition and motor-imagery decoding.

GitHub: ASGCRL Paper DOI Paper DOI
Role
EEG research lead
Contribution
Designed SGCL, ASGCRL and SimNeXt-EEG. Implemented models, controlled comparisons and ablation experiments, and wrote first-author papers.
Period & affiliation
2022.07 ~ Present ITRC

01 / RESEARCH DETAILASGCRL: adaptive graphs and contrastive learning

Challenge
Incorrect connections in fixed similarity graphs can propagate errors when labels are scarce.
Approach
Updated connections with learned representations and used adaptive augmentation to preserve class relationships during contrastive learning.
Contribution
Designed the method, implemented the model, conducted ablations and wrote the first-author KBS 2026 paper as a follow-up to SGCL.
ASGCRL: adaptive graphs, label-guided augmentation and contrastive loss
ASGCRL: adaptive graphs, label-guided augmentation and contrastive loss
View full-size figure
SEED-IV participant 15: Static GCN embedding, 84.29% accuracy
SEED-IV participant 15: Static GCN embedding, 84.29% accuracy
View full-size figure
SEED-IV participant 15: ASGCRL embedding, 94.23% accuracy
SEED-IV participant 15: ASGCRL embedding, 94.23% accuracy
View full-size figure
SEED-IV Case 1 accuracy comparison
SEED-IV Case 1 Accuracy
Static GCN 75.75 ± 8.95%
ASGCRL 87.31 ± 3.84%
Matched-protocol difference +11.56%p
Evaluation conditions and scopeSEED-IV Case 1, 2.4% labels, subject-wise transductive evaluation. Embedding example: subject 15.

02 / RESEARCH DETAILSimNeXt-EEG: compact temporal attention

Challenge
A small motor-imagery model must capture multiple temporal scales and suppress noise.
Approach
Extracted temporal features with multiple kernels and recalibrated time points using parameter-free SimAM attention.
Contribution
Designed the model, measured classification accuracy and CPU latency, and conducted ablations for the first-author IEEE SPL 2026 paper.
Classification accuracy (%) Accuracy: EEGNet 58.6, MSVTNet 65.93, SimNeXt 70.81 0 22 43 65 86 58.6 EEGNet 65.93 MSVTNet 70.81 SimNeXt Accuracy
Classification accuracy (%)
Temporal attention ablation (%) Accuracy: Without both modules 68.26, Full model 70.81 0 22 43 65 86 68.26 Without both modules 70.81 Full model Accuracy
Temporal attention ablation (%)
SimNeXt-EEG architecture
Stage Operation / output
MSTF Temporal kernels 32, 80 / 8 channels
STE Spatial separable convolution + SimAM
STR Temporal separable convolution + SimAM
Classification Flatten, Linear / 4 classes
Parameter count and CPU latency
Model Parameters CPU latency
EEGNet 3.44K 0.79 ms
MSVTNet 75.49K 3.78 ms
SimNeXt 1.84K 1.46 ms
Evaluation conditions and scopeBCI Competition IV 2a, session-independent accuracy: 70.81 ± 0.49%. Complexity: 1.84K parameters and 20.69 MFLOPs. Latency: 1.46 ms on a desktop CPU, batch size 1; not a wearable-device measurement.