A comparison of advanced learning models for the P300 brain computer interfaces

The goal to converting brain activity into instructions for control and communication, predicting human mental intent has arisen as a prominent focus of research in Brain-Computer Interface research. These research entail gathering electroencephalographic data from participants in order to build classifiers that can decode users’ mental states. However, training classifiers in BCI systems is difficult due to many source among the inter-subject or intra-subject variability in brain signals. This model training typically follows a common technique in terms of machine learning: i) feature extraction, needs effort and may also need domain expertise, ii) train a classifier using the retrieved features. Not only are very accurate classifiers now achievable thanks to deep learning algorithms, additionally incorporate the feature extraction stage into the process of creating classifiers. However deep learning models have the benefit of incorporating domain-dependent feature extraction during classifier construction, architectural selection procedure for BCIs is often dependent on domain knowledge. In this paper, we look at whether it’s possible to build reliable classifier for deciphering P300 event-related potentials utilising a systematic deep learning model selection. The findings of Gated Recurrent Unit (GRU), long short-term memory cells (LSTMs) and convolutional neural networks (CNN) and of varied complexity are presented. Our empirical findings demonstrate which is the best model for categorization of P300 waveforms.