SimNeXt-EEG

Status

Item	Status
Research	Published · IEEE Signal Processing Letters 2026
Implementation	Model, training, and evaluation code

Compact Motor Imagery EEG Decoding via Parameter-Free Temporal Attention and Separable Refinement

Dae Hyeon Kim and Young-Seok Choi
IEEE Signal Processing Letters, 2026. Paper

Architecture

Multi-scale temporal fusion, spatial encoding, separable refinement, and classification. Input: four-second EEG at 250 Hz.

Stage	Operation
Multi-scale temporal feature fusion	Parallel kernels of 32 and 80 samples; four filters per branch; concatenate to eight maps
Spatio-temporal encoding	Depthwise spatial convolution across C electrodes; 16 components; 1D-SimAM, ELU, average pooling, dropout
Spatio-temporal refinement	Depthwise temporal kernel of 16 samples and pointwise convolution; 1D-SimAM, ELU, average pooling, dropout
Classification	Flatten and linear projection to class logits

1D-SimAM computes temporal gains from trace-wise mean and variance without learnable attention parameters. The implementation uses residual recalibration in both stages.

Results

Accuracy (%), mean ± standard deviation as reported in paper Table I. SD uses within-session five-fold evaluation; SI tests cross-session transfer within a subject; LOSO holds out a subject. The paper uses five seeds and training-partition normalization.

Dataset	SD	SI	LOSO
BCIC-IV-2a	70.80 ± 0.54	70.81 ± 0.49	52.49 ± 0.51
BCIC-IV-2b	76.04 ± 0.34	80.96 ± 0.31	79.02 ± 0.65
OpenBMI	67.97 ± 18.02	66.69 ± 17.30	76.34 ± 13.03

Accuracy and computational cost

BCIC-IV-2a SI, paper Table II. Latency is a batch-size-one desktop CPU measurement.

Model	Parameters (K)	FLOPs (M)	CPU latency (ms)	Accuracy (%)
EEGNet	3.44	23.52	0.79	58.60
FBCNet	8.07	37.62	0.87	67.91
LightConvNet	3.19	12.24	0.79	57.33
MSVTNet	75.49	98.35	3.78	65.93
SimNeXt-EEG	1.84	20.69	1.46	70.81

SimNeXt-EEG uses 1.84 K parameters and 20.69 M FLOPs, with 1.46 ms CPU latency in this configuration.

Ablation

SI accuracy (%), paper Table III.

Variant	BCIC-IV-2a	BCIC-IV-2b
Full convolution in STR	71.49	79.91
Full SimNeXt-EEG	70.81	80.96
Without SimAM-1	70.84	80.75
Without SimAM-2	69.30	79.93
Without both SimAM modules	68.26	77.31

Data

The datasets are not redistributed here. Download them and place the files as below.

BCI Competition IV 2a and 2b: https://www.bbci.de/competition/iv/. The competition ships GDF; the loader reads the MATLAB versions.

OpenBMI: Lee et al., GigaScience 8(5):giz002, 2019, doi:10.1093/gigascience/giz002. Fetch the MI recordings from the public repository named in that paper.

data/bci_iv_2a/   A01T.mat A01E.mat ... A09T.mat A09E.mat
data/bci_iv_2b/   B01T.mat B01E.mat ... B09T.mat B09E.mat
data/openbmi/     sess01_subj01_EEG_MI.mat ... sess02_subj54_EEG_MI.mat

Running

pip install -r requirements.txt

python train.py --dataset {2a|2b|openbmi} --protocol {SD|SI|LOSO} --seed N \
                [--gpu 0] [--subjects LO:HI] [--band 4,40] [--force]

--gpu selects the CUDA device and is applied before torch initialises the driver. --subjects 1:5 runs a range of subjects and writes its own result file, so a long run can be split across GPUs and the parts read back separately. --band affects OpenBMI only. --force overwrites an existing result file instead of skipping the run.

python evaluate.py [--tag TAG] [--format table|json] [--detail]

Citation

@article{kim2026simnext,
  author  = {Kim, Dae Hyeon and Choi, Young-Seok},
  journal = {IEEE Signal Processing Letters},
  title   = {SimNeXt-EEG: Compact Motor Imagery EEG Decoding via Parameter-Free
             Temporal Attention and Separable Refinement},
  year    = {2026},
  pages   = {1-5},
  doi     = {10.1109/LSP.2026.3727949}
}

License

The source code is released under the MIT License, see LICENSE. Documentation and non-code research materials are covered by LICENSE-DOCS.md (CC BY 4.0). The datasets are not covered by either and keep the terms set by their own distributors.