[2027 AAAI] DAMC-Net: Dual-Branch Attention-Guided Multi-Scale Convolutional Network for EEG-Based Imagined Speech Decoding

Seung Won Kim, Dae Hyeon Kim, Young-Seok Choi^*

Department of Electronics and Communications Engineering, Kwangwoon University, Seoul, South Korea

---

📢 News

• [Aug. 2026] 📄 Our paper "DAMC-Net: Dual-Branch Attention-Guided Multi-Scale Convolutional Network for EEG-Based Imagined Speech Decoding" has been submitted to AAAI 2027. The code will be officially released upon acceptance.

---

📝 Abstract

Imagined speech decoding from electroencephalography (EEG) offers a promising foundation for intuitive, silent speech interfaces. However, the discriminative neural patterns of imagined speech unfold across multiple temporal scales and are distributed over spatially distinct electrodes, making them difficult to capture with conventional single-scale architectures.

We propose DAMC-Net, a dual-branch attention-guided multi-scale convolutional network for EEG-based imagined speech decoding. The two branches are designed to be complementary: a depthwise temporal branch models channel-independent temporal dynamics after cross-channel mixing, whereas a cross-channel branch jointly captures spatial dependency across electrodes. Within each branch, parallel dilated convolutions extract temporal features at multiple resolutions while preserving the original temporal resolution. A parameter-free attention mechanism then recalibrates salient activations from local first- and second-order statistics, with the comparison axis matched to each branch, and the recalibrated multi-scale responses are summarized as log-power representations. These representations are adaptively integrated through feature-specific scale weights, and the complementary features from the two branches are fused for final classification.

On two public imagined-speech EEG datasets, DAMC-Net attains the highest mean accuracy among the competing baselines, and ablation studies confirm the contribution of each component—multi-scale convolution, parameter-free recalibration, and log-power summarization. These results demonstrate that attention-guided multi-scale representation learning is an effective approach to EEG-based imagined speech decoding.

---

📊 Experimental Results

Subject-dependent evaluation on Track 3 of the 2020 International BCI Competition (5 classes, 15 subjects × 10 seeds). Standard deviation is the across-subject sample standard deviation (ddof=1) of per-subject seed means.

Dataset	Params	Accuracy (Mean $\pm$ std)
2020 International BCI Competition (Track 3)	89,029	0.8413 $\pm$ 0.0597

Complexity

Metric	Value
Parameters	89,029
MACs	69.401 M

MACs are counted for batch 1 with input (1, 64, 795) over Conv1d and Linear only; one multiply-accumulate is one MAC. The parameter-free recalibration, reductions, log, softmax, and element-wise operations are excluded.

Ablation studies

All configurations share the same data splits and training protocol (15 subjects × 10 seeds × 200 epochs).

Component ablation (MS-conv × PF-Attn. × LP-pooling)

MS-conv	PF-Attn.	LP-pooling	Accuracy (Mean $\pm$ std)
$\times$	$\times$	$\times$	0.4881 $\pm$ 0.0538
$\checkmark$	$\times$	$\times$	0.5001 $\pm$ 0.0538
$\times$	$\checkmark$	$\times$	0.5023 $\pm$ 0.0513
$\checkmark$	$\checkmark$	$\times$	0.5181 $\pm$ 0.0535
$\times$	$\times$	$\checkmark$	0.7965 $\pm$ 0.0697
$\checkmark$	$\times$	$\checkmark$	0.8326 $\pm$ 0.0673
$\times$	$\checkmark$	$\checkmark$	0.8016 $\pm$ 0.0643
$\checkmark$	$\checkmark$	$\checkmark$	0.8413 $\pm$ 0.0597

Key Findings:

1. The log-power readout carries most of the accuracy. Its main effect is +0.316 (0.8180 with LP vs. 0.5022 without, averaged over the other two factors), and every configuration without it collapses to ~0.50 regardless of what else is enabled. Replacing the log-power readout with time-axis mean pooling is the single most damaging change to the model.
2. Multi-scale convolution adds a further 0.036. With the log-power readout in place, enabling the parallel dilated scales lifts accuracy from 0.7965 to 0.8326 (+0.0361), and to 0.8413 (+0.0397) when the recalibration is also on.
3. The parameter-free recalibration is a small but cost-free gain. It contributes +0.0087 on top of MS-conv + LP-pooling (0.8326 → 0.8413) while adding zero parameters and zero MACs. Its benefit depends on the multi-scale scales being present: without MS-conv it is worth only +0.0051 (0.7965 → 0.8016), consistent with its role of recalibrating activations before the per-scale readout.
4. Lowest variability at the best accuracy. The full model also has the smallest across-subject standard deviation of all eight configurations (0.0597), i.e. the components do not trade stability for mean accuracy.

Branch contribution and recalibration axis

Configuration	Params	Accuracy (Mean $\pm$ std)
Single branch
T-AMC only	6,277	0.6924 $\pm$ 0.0870
S-AMC only	82,757	0.8251 $\pm$ 0.0668
Both branches; axis in T-AMC / S-AMC
Temporal / temporal	89,029	0.8360 $\pm$ 0.0620
Channel / channel	89,029	0.8343 $\pm$ 0.0647
Temporal / channel (Ours)	89,029	0.8413 $\pm$ 0.0597

The two branches are complementary: neither alone suffices (0.6924 and 0.8251 vs. 0.8413), and adding T-AMC on top of S-AMC costs only 6,272 further parameters. The recalibration axis must be matched to each branch — recalibrating both branches along the same axis costs 0.005–0.007 accuracy at identical parameter count and MACs, so the gain comes from where the recalibration is applied, not from added capacity.

---

📦 Repository Structure

This repository contains the minimal, self-contained code required to reproduce training and evaluation on Track 3 of the 2020 International BCI Competition.

main.py             Full protocol runner (15 subjects x 10 seeds) and result aggregation
train_eval.py       One run (single subject, single seed): preprocessing, training, evaluation
preprocessing.py    .mat loading, signal preprocessing, seed-wise data splitting
model.py            Model definition

File	Role
model.py	Model definition (DualBranchSimAMNet). Contains the parameter-free recalibration module, the multi-scale dilated branch with log-power-domain weighted-sum fusion, and the two branches: T-AMC (1x1 cross-channel mixing → depthwise multi-scale convolution, recalibrated along time) and S-AMC (full multi-scale convolution over all electrodes, recalibrated across channels). Input (B, 64, 795) → class logits (B, 5). Running it standalone prints the output shape and the trainable-parameter count.
preprocessing.py	Loads the per-subject .mat files (epo_train / epo_validation / epo_test), concatenates them, and re-splits them per seed into 60/10/10 trials per class. Applies common average reference, a 60 Hz Butterworth low-pass filter (order 4), and channel-wise z-scoring whose statistics are fitted on the training split only. Also provides global seed fixing.
train_eval.py	One complete run: seeding → preprocessing → training (AdamW, polynomial LR decay, mixed precision, cross-entropy with label smoothing) → best-validation checkpoint restoration → test evaluation, with learning curves, confusion matrix, and classification report written to disk.
main.py	Runs the full protocol over 15 subjects × 10 seeds, aggregates per-subject and overall accuracy.

Data

Place Data_Sample{1..15}.mat (containing epo_train / epo_validation / epo_test) in the following directories under DATA_ROOT, which defaults to the parent directory of this folder:

<DATA_ROOT>/
├── Training set/     Data_Sample1.mat ... Data_Sample15.mat
├── Validation set/   Data_Sample1.mat ... Data_Sample15.mat
└── Test set/         Data_Sample1.mat ... Data_Sample15.mat

Each trial uses the full epoch (−0.5 s to 2.6 s at 256 Hz, T = 795) without cropping.

Quick Start

main.py is the only entry point.

python main.py

The GPU and the data location are set through environment variables:

CUDA_VISIBLE_DEVICES=1 DATA_ROOT=/path/to/data python main.py

Results are written to results/subject_XX/seed_YYYY/, with SUBJECT_SUMMARY.txt per subject and ALL_SUBJECTS_SUMMARY.txt, seed_ranking.csv, subject_summary_top{k}seeds.csv, and RESULT_top{k}_{mean}_{std}.txt at the top level.

Key Hyperparameters

Run settings are defined as constants at the top of main.py.

Argument	Value	Description
NUM_EPOCHS	200	Training epochs
BATCH_SIZE	64	Mini-batch size
FS	256	Sampling rate (Hz)
SUBJECT_IDS	1–15	Subject-dependent evaluation
SEEDS	10 seeds	Each seed defines a different 60/10/10 per-class split
N_TRAIN/VAL/TEST_PER_CLASS	60 / 10 / 10	Stratified per-class split sizes
kernel_size, dilations	5, (1, 2, 3, 4)	Multi-scale dilated convolutions in both branches
recalibration axis	time (T-AMC) / channel (S-AMC)	Axis of the parameter-free recalibration in each branch
simam_lambda	1e-3	Stabilizing term of the parameter-free recalibration

Optimizer: AdamW, learning rate 1e-3 → 1e-6 (polynomial decay, power 2), weight decay 1e-2, cross-entropy with label smoothing 0.01, mixed precision on CUDA. The model with the best validation accuracy is restored before test evaluation.

Note on the released code. The model.py in this folder is an earlier configuration (kernel_size=3, both branches recalibrated along the time axis, 55,749 parameters). Reproducing the reported 0.8413 requires kernel_size=5 and channel-axis recalibration in the S-AMC branch; the final version will be released with the paper.

---

📄 License

This project is released under the MIT License. See LICENSE for details.

---

Citation

The paper is currently under review at AAAI 2027. A citation entry will be added upon acceptance.