A few filters are enough: Convolutional Neural Network for P300 Detection

• DOI: 10.48550/arxiv.1909.06970
• Publication Date: 2019-01-01
• Authors: Alicia Montserrat...; Gibrán Fuentes-Pi...; Jorge Cervantes‐O...
• Keywords: Sigmoid function

Over the past decade, convolutional neural networks (CNNs) have become driving force of an ever-increasing set applications, achieving state-of-the-art performance. Most modern CNN architectures are composed many and fully connected layers typically require thousands or millions parameters to learn. CNNs also been effective in detection Event-Related Potentials from electroencephalogram (EEG) signals, notably P300 component which is frequently employed Brain-Computer Interfaces (BCIs). However, for this task, increase rates compared approaches based on human-engineered features has not as impressive other areas might justify such a large number parameters. In paper, we study performances existing with diverse complexities single-trial within-subject cross-subject four different datasets. We proposed SepConv1D, very simple architecture consisting single depthwise separable 1D layer followed by Sigmoid classification neuron. found that few filters its small overall parameters, SepConv1D obtained competitive believe may represent important step towards building simpler, cheaper, faster, more portable BCIs.