A Novel Parameter-Free Attention-based Multiscale Convolution Prototype Networks for P300 Brain-computer Interface

The brain-computer interface (BCI), facilitating sophisticated direct human-machine interaction, has garnered significant research attention over the past decade due to its vast potential in diverse applications, notably in rehabilitation and communication. As one of the most common brain-computer interface systems, the P300 speller is renowned for its ability to accurately identify characters. However, the applicability of the P300 speller is hampered for the low recognition rate partially attributed to the complex spatial-temporal characteristics of the electroencephalogram (EEG) signals, and its requirement for numerous data to train the system while data collection makes the task tedious. To cope with these issues, a novel parameter-free attention-based multiscale convolutional prototype network with prototype calibration model PampNet is proposed in this work for improvement of the P300 BCI systems with limited amount of training data. Specifically, the multiple scale convolutional neural network was developed to obtain refined discriminative feature. Furthermore, the parameter-free attention (PFA) module can find the importance of each neuron by optimizing an energy function without adding extra model parameters and generates 3D attentions weights for the feature map. In addition, a prototypical network with a prototype calibration to mitigate the effect of data noise on prototype generation was introduced to classify P300 signals with small samples. The performance of the proposed PampNet was tested on two publicly available datasets (i.e., Dataset IIb of BCI Competition 2003 and Dataset II of BCI Competition III). The results of multiple evaluation indicators such as accuracy of character recognition and ITR demonstrate that the proposed algorithm outperforms other state-of-the-art methods with limited training data.