ATM data classification with an artificial neural network

The ability to classify different data types in real time has many potential applications. Implemented in a switching element, data recognized to be time sensitive, such as voice, could be given higher priority or other special handling. Using competitive learning, it is possible to train a network to distinguish between audio, video and FTP data without an external supervisor. Unprocessed audio, video and FTP data exhibit strong linear trends which make it impossible for the neural network to learn the subtle variations in the data. Noticing that the three data types could be distinguished by their degree of correlation, a simple method of detrending was devised which allowed the data to be prepared for presentation to the neural network in real time. For this problem, the architecture best suited to classify the data types was determined to be an unsupervised, feed forward network. Using a competitive learning algorithm, the network was able to learn to classify the different input patterns without an external supervisor.