Investigation of Attention Based Convolutional Neural Network for Motor Imagery Classification

Attention-based blocks have conquered the realm of deep learning since they advanced the state-of-the-art of pattern recognition. They can extract features more efficiently than the standard blocks by assigning attention scores to features. Brain-Computer Interface (BCI) detection systems adopted many of those blocks in many architectures to enhance the performance. But, many focus on one rather than testing many, despite their diversity. This paper proposes to study different blocks like Squeeze and Excitation (SE), efficient Squeeze and Excitation (eSE), Convolutional Block Attention Module (CBAM), and a modified CBAM. Also, we propose two architectures, like a multi-branch architecture that fed from multi-scale versions of the input signal, and a one pipeline architecture. We tested on the BCI competition IV dataset 2b, a Motor Imagery Dataset. The results demonstrate that our propositions work better than EEGNet and that the attention modules can improve the performance in many configurations. Also, we established that the attention module can raise the performance of a class at the expense of the other.