# MI-004: EEGNet: A compact convolutional neural network for EEG-based brain-computer interfaces > 本页由本地论文卡片生成。请不要直接编辑本页；修改源卡片后运行 `vp run docs:generate`。 > 中文支持说明：本页是中文站点镜像，保留论文标题、证据状态和来源字段的原始表述，避免未经核实的翻译改变证据边界。 > 内部 wiki 边界：本页提供论文卡片证据和用户提供的 PDF 下载入口，仅用于私有阅读；全文抽取文本不发布。 > 来源: `library/paper_cards/MI-004.md` ## 论文访问 * 内部 PDF: 下载论文 PDF * DOI / 官方页面: [10.1088/1741-2552/aace8c](https://doi.org/10.1088/1741-2552/aace8c) * 部署边界: 这些 PDF 链接只适合私有/受保护的内部 wiki。 ## MI-004: EEGNet: A compact convolutional neural network for EEG-based brain-computer interfaces ## Metadata * ID: MI-004 * Title: EEGNet: A compact convolutional neural network for EEG-based brain-computer interfaces * Year: 2018 * DOI / URL: 10.1088/1741-2552/aace8c * Local PDF: 见上方论文访问区块 * Text artifact: local-only path withheld from docs site * Review status: `extracted` ## Study Type * Track: EEG decoding / motor imagery / compact neural models * Task: validate a compact CNN architecture across multiple EEG-BCI paradigms * Participants or dataset: four datasets covering P300, ERN, MRCP, and SMR; the SMR dataset is BCI Competition IV Dataset 2A with nine subjects and four imagined movement classes * Hardware: varies by dataset; SMR data used 22 Ag/AgCl electrodes sampled at 250 Hz before resampling * Channels or sensors: full-channel EEG inputs per dataset; traditional baselines also include channel-selection and FBCSP pipelines ## Methods * Paradigm: EEGNet is compared across ERP and oscillatory BCI paradigms, including sensory motor rhythm classification * Signal processing or model: compact CNN with depthwise and separable convolutions; compared with DeepConvNet, ShallowConvNet, xDAWN plus Riemannian geometry, and FBCSP * Training/calibration: within-subject and cross-subject evaluations with controlled folds; limited training data is a central condition * Online/offline: offline model validation; no live BRI or robot execution experiment ## Results * Metrics: AUC for P300, ERN, and MRCP; accuracy for SMR; trainable parameter counts * Main findings: EEGNet generalized across all tested paradigms and achieved comparable or better performance than reference approaches, especially with limited data; EEGNet parameter counts were up to two orders of magnitude smaller than DeepConvNet and ShallowConvNet * Reported limitations: cross-subject ERN favored xDAWN plus Riemannian geometry; SMR remains weak and variable across users; offline dataset results do not establish online robot-control performance ## Relevance To This Project * Supports: compact EEG decoder rationale, cross-paradigm EEG feature learning, and low-resource model comparison for MI/SMR-related control * Conflicts with: does not evaluate SSVEP-MI hybrid control, false activation, no-control states, or robotic grasping * Design implication: EEGNet can be treated as an offline baseline or ablation candidate, but SAH-BRI-Grasp still needs project-specific MI/no-control validation ## Extracted Evidence | Claim | Status | Evidence Note | Page/Section | | --- | --- | --- | --- | | EEGNet is a compact CNN intended for multiple EEG-BCI paradigms rather than one task-specific signal. | verified | The abstract and introduction define EEGNet as a compact architecture tested on P300, ERN, MRCP, and SMR datasets. | Abstract; Section 1 | | The SMR benchmark used nine subjects and four imagined movement classes from BCI Competition IV Dataset 2A. | verified | The SMR dataset section lists left hand, right hand, feet, and tongue imagery from nine subjects using 22 electrodes. | Section 2.1.4 | | EEGNet uses depthwise and separable convolutions to reduce parameters while preserving EEG-specific feature extraction. | verified | The methods section describes temporal filters, depthwise spatial filters, separable convolutions, and a softmax classification block. | Section 2.2 | | EEGNet models are up to two orders of magnitude smaller than DeepConvNet and ShallowConvNet. | verified | Table 3 reports trainable parameter counts across datasets and the text states the two-orders-of-magnitude comparison. | Section 2.2; Table 3 | | EEGNet is a plausible compact decoder baseline for SAH-BRI-Grasp MI/SMR experiments. | inferred | The paper validates compact EEG decoding across datasets, including SMR, but does not test this project's task or online control state machine. | Abstract; Sections 3-4 | | Online false activation and no-control behavior remain unverified by this paper. | needs confirmation | The validation is offline and does not include asynchronous online BRI operation. | Sections 2-4 | ## Open Questions * Should SAH-BRI-Grasp compare CSP/FBCSP, EEGNet, and a simple linear baseline under the same no-control protocol? * Which low-channel montage preserves enough SMR evidence for the project task? * Does the project need online adaptation or confidence thresholds beyond the offline EEGNet benchmark?