MI-EEGNet: Compact CNN for MI EEG Decoding

• MI-EEGNet is a compact convolutional neural network optimized for motor imagery EEG decoding, featuring temporal, depthwise spatial, and separable convolutions.
• It employs standardized preprocessing and domain adaptation techniques to enhance robustness across sessions and subjects in brain-computer interface applications.
• Its efficient design enables real-time, low-power deployment on embedded devices while delivering competitive classification accuracy on public benchmarks.

MI-EEGNet refers to a family of compact, convolutional neural network (CNN) architectures specifically optimized for motor imagery (MI) EEG decoding in brain-computer interface (BCI) applications. Defined most transparently as an adaptation or extension of EEGNet (Lawhern et al., 2018), MI-EEGNet architectures are characterized by a streamlined sequence of temporal and spatial convolutions, minimal parameterization, and tailored pre-/post-processing to robustly classify low-SNR, nonstationary, trial-based MI signals. The model has become a widely used baseline for both algorithmic research and low-power embedded deployment in MI-BCI systems, and is often the foundation for more sophisticated domain-adaptive, interpretable, or hardware-efficient variants (Miao et al., 2022, Aktar et al., 1 Nov 2025, Andrikopoulos et al., 7 Apr 2026, Zancanaro et al., 2021, Schneider et al., 2020, Wang et al., 2020, Köllőd et al., 2023, Bian et al., 2024, Parashiva et al., 3 Jan 2025).

1. MI-EEGNet Core Architecture

The canonical MI-EEGNet architecture inherits its structure from EEGNet, comprising three main projection blocks and a softmax classifier:

1. Temporal Convolution: A 2D convolution with kernel size (1 × k), where $k$ typically ranges between $T/8$ and $T/2$ (e.g., 32–128 samples), acts as band-pass filter across time for each channel and outputs $F_1$ feature maps (Miao et al., 2022, Aktar et al., 1 Nov 2025).
2. Depthwise Spatial Convolution: Each temporal feature map is spatially convolved using a depthwise grouped convolution with kernel size ($E$ × 1) (where $E$ is the number of channels), independently learning spatial filters per temporal feature. Batch normalization and ELU activation follow; average pooling and dropout regularize (Zancanaro et al., 2021, Wang et al., 2020).
3. Separable Convolution: A combination of depthwise temporal convolution (small kernel) and 1×1 pointwise convolution (mixing channels) aggregates higher-level features; batch normalization, ELU, pooling, and dropout are again applied (Aktar et al., 1 Nov 2025, Wang et al., 2020).
4. Dense Classifier: Flattened outputs pass to a fully connected layer and a softmax for final class prediction. Output size matches number of MI classes (commonly 2–4) (Köllőd et al., 2023, Wang et al., 2020).

Variations adjust filter sizes, depths, and other hyperparameters based on dataset, computational constraints, or specific MI paradigms.

2. Preprocessing and Domain Adaptation

MI-EEGNet performance is strongly influenced by signal preprocessing and domain-alignment strategies:

• Standard Preprocessing: Includes notch filtering (e.g., 50 Hz for line noise), band-pass (e.g., 0.5–100 Hz), and trialwise Z-score normalization. ICA-based artifact removal and FASTER pipeline may be incorporated for improved data quality (Aktar et al., 1 Nov 2025, Köllőd et al., 2023).
• Domain-Invariant Alignment: For cross-session or cross-subject generalization, preprocessing may include min-max scaling to [–1,1] and Euclidean Alignment: whitening each trial’s covariance to a session mean, thus harmonizing between-session covariance structures (Miao et al., 2022).
• Advanced Domain Adaptation: Siamese Deep Domain Adaptation (SDDA) integrates MI-EEGNet with dual network streams over source and target domains, enforcing feature distribution alignment in Reproducing Kernel Hilbert Space (via MMD loss) and intra-class compactness (center loss), yielding significantly improved cross-session accuracy (Miao et al., 2022).

3. Embedded and Hardware-Aware Variants

MI-EEGNet’s compactness enables real-time deployment on edge devices:

• Quantized Implementations: Q-EEGNet maps all weights and activations to 8-bit fixed-point, with less than 0.4% accuracy loss on 4-class MI tasks, resulting in sub-6 ms inference time and <1 mJ energy per decision on parallel RISC-V PULP SoCs (Schneider et al., 2020).
• Resource Scaling: Temporal downsampling, channel selection, and trial-duration truncation offer explicit control over memory, parameter count, and latency with minor loss in accuracy (e.g., 62.51% at 70 KB RAM for a model with 38 channels, 1 s window) (Wang et al., 2020).
• On-Device Learning: Real-time adaptation via online SGD of the dense classification layer on wearable hardware recovers up to 7.3% accuracy lost to subject-induced feature drift, all within 16–25 KB SRAM, ≈15 ms inference, and 0.8 mJ per trial on GAP9 RISC-V clusters (Bian et al., 2024).

4. Performance and Comparative Evaluation

MI-EEGNet consistently achieves competitive or superior MI classification on public benchmarks:

Dataset	Classes	Protocol	Accuracy (%)	Comp. Baseline	Ref
BCI-IV-2a	4	LOSO	68.2 ±12.1	ANFIS: 65.7	(Aktar et al., 1 Nov 2025)
BCI-IV-2a	4	Mixed	72.5 ±1.8	FBCSP: 45.9	(Zancanaro et al., 2021)
Physionet	4	5-fold CV	65.07	CNN: 58.58	(Wang et al., 2020)
BCI-IV-IIA	4	session DA	79.43	EEGNet vanilla	(Miao et al., 2022)

• Cross-Session/Subject Robustness: MI-EEGNet outperforms FBCSP and classic CNN baselines by 5–25% in cross-subject and cross-session scenarios (Miao et al., 2022, Zancanaro et al., 2021).
• Interpretability: While “black-box” compared to CSP-based or fuzzy systems, architectural extensions such as Squeeze-and-Excitation (SE) blocks provide filter and electrode rankings, enhancing physiological insight (Parashiva et al., 3 Jan 2025).
• Transfer Learning: Within-subject MI-EEGNet achieves highest improvement from chance, while in transfer learning, Shallow and Deep ConvNets sometimes show larger inter-subject gains (Köllőd et al., 2023).

5. Extensions: Multi-Scale, Attention, and Domain Adaptation

Recent advances extend the MI-EEGNet paradigm:

• Multi-Scale Convolution: Implementations such as EEG-MFTNet incorporate parallel temporal convolutions at different scales (e.g., kernel sizes 5–125), capturing both fast and slow MI EEG oscillations, improving accuracy and robustness to nonstationarity (Andrikopoulos et al., 7 Apr 2026).
• Transformer-Based Fusion: Adding lightweight Transformer encoders to MI-EEGNet artifacts captures long-range temporal dependencies; fusion of multi-scale and attention streams achieves 58.9% cross-session mean accuracy vs. 53.7% for EEGNet baseline, with <50 ms latency (Andrikopoulos et al., 7 Apr 2026).
• Siamese, Center Loss, and RKHS Alignment: The SDDA framework applies on MI-EEGNet to minimize feature distribution divergence and intra-class scatter, with joint optimization over source (labeled) and target (unlabeled) sessions (Miao et al., 2022).

6. Interpretability, Trade-Offs, and Future Directions

• Interpretability vs. Generalization: ANFIS–FBCSP–PSO offers superior per-subject interpretability; MI-EEGNet’s deep filters scale and generalize better but are less physically interpretable unless augmented with mechanisms like SE layers or attention heatmaps (Aktar et al., 1 Nov 2025, Parashiva et al., 3 Jan 2025).
• Efficiency Profiles: Edge optimizations (quantization, pruning, scaling) enable deployment under 100 ms and sub-mJ energy budgets without critical degradation (Schneider et al., 2020, Wang et al., 2020, Bian et al., 2024).
• Hybrid Networks: Anticipated directions include transformer-based attention, graph embeddings, and neuro-symbolic hybrids blending CNN front-ends with rule-based or fuzzy back-ends for explainable BCI decoding (Aktar et al., 1 Nov 2025, Andrikopoulos et al., 7 Apr 2026).
• Domain Robustness: Integration of adaptive normalization, domain-adaptive training, and session-invariant embeddings remain active research threads for deployment in persistent BCI applications (Miao et al., 2022, Andrikopoulos et al., 7 Apr 2026).

7. Summary and Outlook

MI-EEGNet establishes a performant, compact, and extensible architecture for motor imagery EEG decoding, achieving state-of-the-art performance on diverse public datasets across within- and cross-subject/session scenarios. Its core structure—stacked temporal, spatial, and separable convolutions with minimal parameterization—underpins numerous adaptations: domain-adaptive models for robust cross-session operation; hardware-aware scaling for wearable BCI; and interpretable variants with electrode and filter ranking. Ongoing research focuses on bridging the gap between model transparency and generalization, leveraging hybrid attention modules, domain alignment, and on-device learning to expand the frontier of deployable MI-BCI technology (Miao et al., 2022, Aktar et al., 1 Nov 2025, Andrikopoulos et al., 7 Apr 2026, Parashiva et al., 3 Jan 2025, Bian et al., 2024).