Weighted Ensemble of Deep Convolution Neural Networks for Single-Trial Character Detection in Devanagari-Script-Based P300 Speller

Abstract

The existing Devanagari-script-input-based P300 speller (DS-P3S) performs better mostly with 3–15 trials. This leads to poor information transfer rate (ITR) and a major concern in its real-time adaptation. In DS-P3S, the display paradigm is a matrix of $8\times 8$ size which has 28 more characters than the $6\times 6$ English paradigm. The increased number of characters leads to user-related issues such as a crowding effect, double flashing, adjacency distraction, task difficulty, and fatigue which increases the false detection rate. To tackle this, we propose an efficient single-trial character detection approach for DS-P3S using weighted ensemble of deep convolution neural networks (WE-DCNNs). The weighted strategy is constructed based on measured ensemble diversity to counter the instability by the individual classifier. Additionally, to reduce the false detection rate arising from a single trial, a new channel dropout-based character detection approach is introduced first time in this article. The ITR of 55.45 b/min and an average P300 classification accuracy of 92.64% achieved are comparatively higher than existing methods of DS-P3S. The significant reduction in tradeoff between bias and variance for the different subjects affirms the ease of applicability of the proposed model with just a single trial.