Systems Architecture and Pattern Recognition for an Intelligent, Pediatric, Non-Contact Communication Aid

Abstract

The system architecture for an intelligent, pediatric, non-contact communication aid is proposed for temporarily mute patients.  It consists of user profile, knowledge base and pattern prediction modules interfacing through a central information manager.  The information manager communicates with the user via a dialogue manager and visual interface. 
  To ensure that the system is easy to use, several learning algorithms are proposed for the pattern prediction module to be able to predict future inputs and display them for easy location and selection by the user.  A backpropagation algorithm is explored in detail, where data triplets are used to train the network.  The data is generated through a simulator that creates sets of data based on biases for time of day, previous selection(s), and time gap between selections. 
  Although the prediction rate is heavily tied to the simulator biases (representing underlying behavioral patterns), it is significantly increased compared to random guessing.  The prediction rate is also closely linked to the number of training cases that the network gets to see.
  Implementation of such learning algorithms within the structure proposed may lead to improved speed and accuracy of communication, and less support required to teach and maintain the system.