SimNeXt-EEG

SimNeXt-EEG is a compact convolutional network for motor-imagery EEG decoding. It combines a multi-scale temporal front-end, depthwise spatial encoding, two parameter-free attention stages, and a separable temporal refinement block. The configurations reported in the paper hold between 1,314 and 3,380 parameters. The aim is accuracy comparable to larger motor-imagery decoders at a small fraction of their parameter count.

This repository is the reference implementation for:

D. H. Kim and Y.-S. Choi, "SimNeXt-EEG: Compact Motor Imagery EEG Decoding via Parameter-Free Temporal Attention and Separable Refinement," submitted to IEEE Signal Processing Letters, 2026.

The manuscript is under review.

Model

model/simnext_eeg.py holds the whole network. Input is one trial of shape (batch, 1, channels, 1000), that is 4 s at 250 Hz.

temporal      Conv(1->4, k=32) || Conv(1->4, k=80) -> concat -> BN
spatial       depthwise Conv(8->16, (C,1)) -> BN -> 1D-SimAM + residual -> ELU -> avgpool(4) -> dropout
refinement    depthwise Conv(16, k=16) -> pointwise Conv(16) -> BN -> 1D-SimAM + residual -> ELU -> avgpool(pool2_k) -> dropout
classifier    flatten -> linear

1D-SimAM derives its gain analytically from the mean and variance of each temporal trace, so it holds no learnable weights; the residual around it lets the network bypass it if the gain is unhelpful. The refinement block is separable, and the two halves do different jobs: the depthwise stage refines temporal dynamics within each component, the pointwise stage mixes information across components. A full temporal convolution in their place couples every pair densely at F2^2 * K_sep weights, where the factorisation costs F2 * K_sep + F2^2 -- 512 against 4,096 at F2=16, K_sep=16. That gap is most of what keeps the model in the low thousands.

Measured parameter counts:

Dataset	Channels	Classes	SD / SI	LOSO
BCIC-IV-2a	22	4	1,844	3,380
BCIC-IV-2b	3	2	1,314	1,570
OpenBMI	20	2	1,586	1,842

from model.simnext_eeg import SimNeXtEEG

model = SimNeXtEEG(in_channels=22, num_classes=4, pool2_k=32)  # 2a, SD/SI
model = SimNeXtEEG(in_channels=22, num_classes=4, pool2_k=8)   # 2a, LOSO

1D-SimAM adapts SimAM (Yang et al., ICML 2021) to single-channel temporal traces.

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]

Protocols

Protocol	Test set	Train / validation
SD	one fold of the session	the other folds, 5-fold
SI	session 2	session 1, 5-fold
LOSO	the held-out subject	the other subjects

Citation

The paper is under review. Complete this entry once it is accepted.

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

Results

Table 1 (SD)

Complete comparisons under the session-dependent (SD) setting. Accuracy, Cohen's κ × 100, and weighted F1 are reported as mean ± standard deviation. Paired tests, BH-adjusted q-values, and Cohen's d_z are computed from subject-level accuracy differences after averaging repeated folds and the five random seeds as applicable.

Dataset	Method	Acc.	κ × 100	wF1	Sig.	q	dz
BCIC-IV-2a	FBCSP-SVM	35.77 ± 4.71	14.35 ± 4.96	35.77 ± 3.11	**	0.0002	+2.63
BCIC-IV-2a	ShallowConvNet	56.45 ± 0.61	41.94 ± 0.82	55.95 ± 0.66	**	0.0002	+2.40
BCIC-IV-2a	EEGNet	54.46 ± 0.82	39.27 ± 1.09	53.58 ± 0.91	**	0.0011	+1.76
BCIC-IV-2a	FBCNet	70.30 ± 0.32	60.41 ± 0.43	69.83 ± 0.34	--	0.8942	+0.05
BCIC-IV-2a	EEGConformer	60.24 ± 1.22	46.99 ± 1.60	57.42 ± 1.42	**	0.0011	+1.72
BCIC-IV-2a	LightConvNet	53.39 ± 0.94	37.86 ± 1.25	52.25 ± 1.05	**	0.0006	+2.02
BCIC-IV-2a	IFNet	49.99 ± 0.34	33.30 ± 0.45	48.77 ± 0.37	**	0.0001	+3.06
BCIC-IV-2a	MSVTNet	65.55 ± 0.63	54.07 ± 0.84	64.50 ± 0.85	*	0.0105	+1.14
BCIC-IV-2a	SimNeXt-EEG	70.80 ± 0.54	61.08 ± 0.72	70.42 ± 0.56	Ref.	Ref.	Ref.
BCIC-IV-2b	FBCSP-SVM	61.42 ± 1.91	23.68 ± 2.09	61.42 ± 1.69	**	<0.0001	+4.10
BCIC-IV-2b	ShallowConvNet	69.94 ± 0.31	39.87 ± 0.62	69.47 ± 0.27	**	0.0003	+2.31
BCIC-IV-2b	EEGNet	73.62 ± 0.53	47.23 ± 1.06	73.03 ± 0.72	**	0.0049	+1.28
BCIC-IV-2b	FBCNet	68.26 ± 0.29	36.52 ± 0.58	67.88 ± 0.42	**	0.0001	+2.77
BCIC-IV-2b	EEGConformer	69.22 ± 0.48	38.45 ± 0.95	67.23 ± 0.37	**	0.0004	+2.17
BCIC-IV-2b	LightConvNet	72.09 ± 0.65	44.19 ± 1.31	71.61 ± 0.69	**	0.0026	+1.50
BCIC-IV-2b	IFNet	71.40 ± 0.33	42.79 ± 0.65	71.05 ± 0.35	**	0.0015	+1.69
BCIC-IV-2b	MSVTNet	71.69 ± 0.21	43.39 ± 0.43	70.93 ± 0.13	**	0.0031	+1.44
BCIC-IV-2b	SimNeXt-EEG	76.04 ± 0.34	52.07 ± 0.69	75.87 ± 0.33	Ref.	Ref.	Ref.
OpenBMI	FBCSP-SVM	60.60 ± 11.73	22.43 ± 21.91	60.60 ± 11.73	**	0.0001	+0.63
OpenBMI	ShallowConvNet	66.19 ± 15.80	32.39 ± 31.61	64.95 ± 16.78	--	0.0977	+0.26
OpenBMI	EEGNet	61.52 ± 17.30	23.04 ± 34.60	59.23 ± 18.56	**	<0.0001	+0.73
OpenBMI	FBCNet	66.91 ± 17.96	33.82 ± 35.92	66.53 ± 17.78	--	0.3215	+0.14
OpenBMI	EEGConformer	67.01 ± 17.25	34.02 ± 34.50	61.95 ± 21.11	--	0.3358	+0.13
OpenBMI	LightConvNet	62.45 ± 18.31	24.91 ± 36.61	60.59 ± 19.51	**	<0.0001	+0.75
OpenBMI	IFNet	60.62 ± 15.69	21.24 ± 31.38	58.74 ± 16.67	**	<0.0001	+0.72
OpenBMI	MSVTNet	68.70 ± 19.08	37.41 ± 38.15	64.50 ± 22.62	--	0.2993	−0.12
OpenBMI	SimNeXt-EEG	67.97 ± 18.02	35.94 ± 36.03	66.22 ± 19.31	Ref.	Ref.	Ref.

Note. The q-values are BH-adjusted p-values within each dataset–protocol family of eight comparisons. * denotes q < 0.05, ** denotes q < 0.01, and -- denotes q ≥ 0.05. ^† indicates a significant difference favoring the baseline; positive d_z favors SimNeXt-EEG.

Table 2 (SI)

Complete comparisons under the session-independent (SI) setting. Accuracy, Cohen's κ × 100, and weighted F1 are reported as mean ± standard deviation. Paired tests, BH-adjusted q-values, and Cohen's d_z are computed from subject-level accuracy differences after averaging repeated folds and the five random seeds as applicable.

Dataset	Method	Acc.	κ × 100	wF1	Sig.	q	dz
BCIC-IV-2a	FBCSP-SVM	41.55 ± 2.84	22.07 ± 3.88	41.55 ± 3.29	**	0.0001	+2.92
BCIC-IV-2a	ShallowConvNet	57.63 ± 1.10	43.51 ± 1.47	56.46 ± 1.39	**	0.0001	+2.65
BCIC-IV-2a	EEGNet	58.60 ± 1.16	44.79 ± 1.55	57.70 ± 1.45	**	0.0071	+1.31
BCIC-IV-2a	FBCNet	67.91 ± 0.53	57.21 ± 0.71	66.67 ± 0.61	--	0.1229	+0.57
BCIC-IV-2a	EEGConformer	60.05 ± 1.99	46.73 ± 2.64	57.48 ± 2.89	**	0.0081	+1.23
BCIC-IV-2a	LightConvNet	57.33 ± 1.85	43.11 ± 2.47	56.63 ± 2.13	**	0.0003	+2.23
BCIC-IV-2a	IFNet	55.35 ± 1.62	40.46 ± 2.16	54.71 ± 1.83	**	0.0001	+3.20
BCIC-IV-2a	MSVTNet	65.93 ± 1.38	54.58 ± 1.85	64.26 ± 1.48	--	0.0523	+0.79
BCIC-IV-2a	SimNeXt-EEG	70.81 ± 0.49	60.80 ± 1.22	69.88 ± 1.14	Ref.	Ref.	Ref.
BCIC-IV-2b	FBCSP-SVM	68.12 ± 1.94	36.24 ± 3.01	68.12 ± 1.66	**	0.0042	+1.49
BCIC-IV-2b	ShallowConvNet	75.05 ± 0.60	50.11 ± 1.19	73.17 ± 1.22	**	0.0012	+2.23
BCIC-IV-2b	EEGNet	79.22 ± 0.86	58.43 ± 1.72	78.77 ± 1.12	--	0.2048	+0.46
BCIC-IV-2b	FBCNet	72.08 ± 0.48	44.16 ± 0.97	71.76 ± 0.56	**	0.0050	+1.39
BCIC-IV-2b	EEGConformer	73.31 ± 0.97	46.61 ± 1.93	71.97 ± 0.95	**	0.0029	+1.66
BCIC-IV-2b	LightConvNet	76.64 ± 1.05	53.28 ± 2.10	76.14 ± 1.28	**	0.0042	+1.67
BCIC-IV-2b	IFNet	77.93 ± 0.69	55.86 ± 1.39	77.43 ± 1.01	--	0.0792	+0.70
BCIC-IV-2b	MSVTNet	77.59 ± 1.29	55.17 ± 2.57	76.78 ± 1.37	*	0.0203	+1.03
BCIC-IV-2b	SimNeXt-EEG	80.96 ± 0.31	60.38 ± 1.37	79.92 ± 0.67	Ref.	Ref.	Ref.
OpenBMI	FBCSP-SVM	59.79 ± 11.26	19.98 ± 22.11	59.79 ± 11.26	**	0.0013	+0.53
OpenBMI	ShallowConvNet	65.06 ± 14.97	30.12 ± 29.93	63.34 ± 16.21	--	0.4003	+0.09
OpenBMI	EEGNet	61.68 ± 17.13	23.36 ± 34.27	59.53 ± 18.36	**	0.0013	+0.51
OpenBMI	FBCNet	64.95 ± 17.43	29.90 ± 34.86	64.63 ± 17.15	--	0.3285	+0.08
OpenBMI	EEGConformer	64.84 ± 16.92	29.69 ± 33.83	58.96 ± 22.17	--	0.3091	+0.10
OpenBMI	LightConvNet	62.69 ± 17.96	25.37 ± 35.92	61.40 ± 18.93	*	0.0265	+0.35
OpenBMI	IFNet	61.17 ± 15.74	22.34 ± 31.47	59.46 ± 16.68	**	0.0013	+0.52
OpenBMI	MSVTNet	68.06 ± 18.48	36.11 ± 36.95	64.29 ± 22.10	*†	0.0388	−0.32
OpenBMI	SimNeXt-EEG	66.69 ± 17.30	31.37 ± 35.89	63.63 ± 19.62	Ref.	Ref.	Ref.

Note. The q-values are BH-adjusted p-values within each dataset–protocol family of eight comparisons. * denotes q < 0.05, ** denotes q < 0.01, and -- denotes q ≥ 0.05. ^† indicates a significant difference favoring the baseline; positive d_z favors SimNeXt-EEG.

Table 3 (LOSO)

Complete comparisons under the leave-one-subject-out (LOSO) setting. Accuracy, Cohen's κ × 100, and weighted F1 are reported as mean ± standard deviation. Paired tests, BH-adjusted q-values, and Cohen's d_z are computed from subject-level accuracy differences after averaging the five random seeds.

Dataset	Method	Acc.	κ × 100	wF1	Sig.	q	dz
BCIC-IV-2a	FBCSP-SVM	28.40 ± 4.13	4.94 ± 2.10	28.40 ± 3.08	**	0.0018	+1.90
BCIC-IV-2a	ShallowConvNet	47.67 ± 1.16	30.23 ± 1.55	44.16 ± 1.51	**	0.0046	+1.54
BCIC-IV-2a	EEGNet	51.64 ± 1.66	35.52 ± 2.22	47.43 ± 2.03	--	0.4832	+0.28
BCIC-IV-2a	FBCNet	36.17 ± 0.40	14.90 ± 0.53	28.52 ± 0.58	**	0.0018	+1.97
BCIC-IV-2a	EEGConformer	42.22 ± 1.19	22.96 ± 1.58	37.85 ± 1.34	**	0.0054	+1.43
BCIC-IV-2a	LightConvNet	52.49 ± 1.01	36.66 ± 1.34	49.06 ± 0.79	--	1.0000	+0.00
BCIC-IV-2a	IFNet	50.52 ± 1.45	34.02 ± 1.93	48.32 ± 1.60	--	0.2327	+0.50
BCIC-IV-2a	MSVTNet	54.65 ± 0.44	39.54 ± 0.58	50.76 ± 0.75	--	0.2936	−0.39
BCIC-IV-2a	SimNeXt-EEG	52.49 ± 0.51	36.65 ± 0.67	48.39 ± 0.88	Ref.	Ref.	Ref.
BCIC-IV-2b	FBCSP-SVM	64.95 ± 1.82	29.89 ± 4.13	64.95 ± 1.56	**	0.0007	+2.20
BCIC-IV-2b	ShallowConvNet	78.32 ± 0.38	56.64 ± 0.77	78.11 ± 0.43	--	0.5763	+0.28
BCIC-IV-2b	EEGNet	79.07 ± 0.12	58.15 ± 0.23	78.86 ± 0.10	--	0.9269	−0.03
BCIC-IV-2b	FBCNet	67.14 ± 0.63	34.29 ± 1.25	65.94 ± 0.96	**	0.0003	+2.67
BCIC-IV-2b	EEGConformer	76.05 ± 1.32	52.11 ± 2.64	75.52 ± 1.45	--	0.0752	+0.89
BCIC-IV-2b	LightConvNet	78.40 ± 0.39	56.80 ± 0.77	78.08 ± 0.45	--	0.5211	+0.35
BCIC-IV-2b	IFNet	77.02 ± 0.31	54.04 ± 0.61	76.71 ± 0.25	--	0.1729	+0.65
BCIC-IV-2b	MSVTNet	78.44 ± 0.50	56.88 ± 0.99	78.24 ± 0.56	--	0.5872	+0.23
BCIC-IV-2b	SimNeXt-EEG	79.02 ± 0.65	58.04 ± 1.29	78.66 ± 0.64	Ref.	Ref.	Ref.
OpenBMI	FBCSP-SVM	60.50 ± 8.69	21.19 ± 17.12	60.50 ± 8.69	**	<0.0001	+2.17
OpenBMI	ShallowConvNet	74.08 ± 12.49	48.16 ± 24.99	73.21 ± 13.39	**	0.0003	+0.57
OpenBMI	EEGNet	76.86 ± 12.95	53.72 ± 25.89	76.30 ± 13.58	--	0.1962	−0.19
OpenBMI	FBCNet	70.86 ± 15.42	41.72 ± 30.84	69.66 ± 15.10	**	<0.0001	+1.32
OpenBMI	EEGConformer	74.35 ± 12.26	48.70 ± 24.51	73.62 ± 13.05	**	0.0011	+0.49
OpenBMI	LightConvNet	78.00 ± 11.66	56.00 ± 23.32	77.61 ± 12.02	**†	0.0011	−0.50
OpenBMI	IFNet	75.99 ± 11.59	51.97 ± 23.18	75.58 ± 11.99	--	0.5739	+0.08
OpenBMI	MSVTNet	78.66 ± 11.14	57.32 ± 22.28	78.25 ± 11.57	**†	0.0012	−0.48
OpenBMI	SimNeXt-EEG	76.34 ± 13.03	52.69 ± 26.05	75.66 ± 13.74	Ref.	Ref.	Ref.

Note. The q-values are BH-adjusted p-values within each dataset–protocol family of eight comparisons. * denotes q < 0.05, ** denotes q < 0.01, and -- denotes q ≥ 0.05. ^† indicates a significant difference favoring the baseline; positive d_z favors SimNeXt-EEG.

License

MIT, see LICENSE. The datasets are not covered by it and keep the terms set by their own distributors.