Performance Comparison of Neural Network Backpropagation Algorithms in Detecting P300 Signals from Mind-Speller Data

Abstract

Visual P300 mind-speller refers to a category of braincomputer interfaces that facilitate its users to spell words or characters using brain signals, specifically the P300 waves. These devices prefer the artificial neural network classifier for the P300 signal detection, as it produces consistently high accuracy in this scenario. The ability of a neural network classifier to detect patterns depends on the number of hidden layers as well as the number of neurons in them, and the training function. This work analyses the performances of multi-layer neural networks corresponding to some training functions, which include gradient descent, conjugate gradient, one-step secant, and resilient algorithms, in detecting the P300 signals from the mind-speller data. All the algorithms were evaluated using 10-fold cross-validation with the classification accuracy and time consumption as the metrics.