Identification of fixations in reading eye movements by a multi-layer neural network

An approach for processing eye-movement data based on neural network is proposed. A cascade network divides the task into two stages: first, a three-layer perceptron to detect the fixation positions from the two-dimensional eye movement trajectory image, second, a single layer adaline to compute time duration for each fixation. The weights are updated according to the Least Mean Square algorithm in the training phase. The network was simulated by a 486DX/66 computer, performance of which was exhibited as results of processed eye-movement data. The identification error for missing and false fixations is less than 0.1%, the mean error for fixation duration less than 0.5%. The method can thus be conveniently and reliably used in an eye-movement laboratory