Convolutional neural networks (CNNs) are emerging as powerful tools for EEG decoding: these techniques, by automatically learning relevant features for class discrimination, improve EEG decoding performances without relying on handcrafted features. Nevertheless, the learned features are difficult to interpret and most of the existing CNNs introduce many trainable parameters. Here, we propose a lightweight and interpretable shallow CNN (Sinc-ShallowNet), by stacking a temporal sinc-convolutional layer (designed to learn band-pass filters, each having only the two cut-off frequencies as trainable parameters), a spatial depthwise convolutional layer (reducing channel connectivity and learning spatial filters tied to each band-pass filter), and a fully-connected layer finalizing the classification. This convolutional module limits the number of trainable parameters and allows direct interpretation of the learned spectral–spatial​ features via simple kernel visualizations. Furthermore, we designed a post-hoc gradient-based technique to enhance interpretation by identifying the more relevant and more class-specific features. Sinc-ShallowNet was evaluated on benchmark motor-execution and motor-imagery datasets and against different design choices and training strategies. Results show that (i) Sinc-ShallowNet outperformed a traditional machine learning algorithm and other CNNs for EEG decoding; (ii) The learned spectral–spatial features matched well-known EEG motor-related activity; (iii) The proposed architecture performed better with a larger number of temporal kernels still maintaining a good compromise between accuracy and parsimony, and with a trialwise rather than a cropped training strategy. In perspective, the proposed approach, with its interpretative capacity, can be exploited to investigate cognitive/motor aspects whose EEG correlates are yet scarcely known, potentially characterizing their relevant features.

卷积神经网络（CNN）逐渐成为功能强大的EEG解码工具：这些技术通过自动学习有关类别识别的相关功能，在不依赖手工功能的情况下提高了EEG解码性能。然而，学习到的特征很难解释，并且大多数现有的CNN引入了许多可训练的参数。在这里，我们通过在空间深度方向上堆叠一个时间正弦卷积层（旨在学习带通滤波器，每个都只具有两个截止频率作为可训练参数），提出了一种轻量且可解释的浅层CNN（Sinc-ShallowNet）卷积层（减少通道连通性并学习与每个带通滤波器相关的空间滤波器），以及完全连接的层，最终确定分类。这个卷积模块限制了可训练参数的数量，并允许通过简单的内核可视化直接解释所学的光谱空间特征。此外，我们设计了一种基于事后梯度的技术，以通过识别更相关且更特定于类的功能来增强解释。Sinc-ShallowNet在基准运动执行和运动图像数据集上进行了评估，并针对不同的设计选择和训练策略进行了评估。结果表明：（i）Sinc-ShallowNet在脑电图解码方面优于传统的机器学习算法和其他CNN；（ii）所学习的光谱空间特征与众所周知的脑电运动相关活动相匹配；（iii）所建议的架构在拥有更多时态核的情况下仍表现更好，但仍在准确性和简约性之间保持了良好的折衷，并采用试行而不是严格的培训策略。从角度上讲，可以利用提出的方法及其解释能力来研究其脑电图相关性尚不为人所知的认知/运动方面，从而可能表征其相关特征。