A DNN based Brain Computational Model of Flicker Fusion

Abstract

Models of psychophysics tasks can aid in understanding the information processing mechanisms in the human brain as well in the development of models for more complex tasks. Flicker fusion is a well studied domain in visual psychophysics, useful for its apparent simplicity. Experiments on flicker fusion phenomena involve presenting a human subject with an intermittent stimulus and the subject classifying the stimulus as either flickering or steady. Mathematical models that seek to account for flicker perception have been proposed, though many of these models were put forward before the advanced computational techniques, available today, were developed. The present work attempts to develop a DNN based Brain Computational Model of flicker fusion that could be trained and tested on pscyhophysics data of a single subject. The mapping of flicker stimuli to the classifications made by a human subject, allows a network to model the phenomena via supervised learning. Since the flicker stimulus is a time series data, we set out to find an appropriate recurrent neural network to model the phenomena. We propose a Convolutional Recurrent Neural Network (CRNN) along with logarithmic transformations of the input representation to model the data. We show that the model that has been pretrained on a particular subject could be used to train data of another subject much faster. The proposed model is trained using stimuli that could be defined with a few parameters. We tested the model on different categories of flicker stimuli that it was not trained with, and made subjective comparisons of the results with previously published psychophysics data. The model has not been evaluated with standard techniques, but the subjective comparisons show that the proposed model will be useful for developing more psychophysically faithful models in the future.