Training restricted Boltzmann machines to generate human-like eye movements

Ch. 2. P. 18–18.

Krasovskaya S., Zhulikov G., MacInnes W.

Approximately twenty years ago, Laurent Itti and Christof Koch created a saliency map of visual attention in an attempt to recreate the work of biological pyramidal neurons by mimicking neurons with centre-surround receptive fields. The Saliency Model launched many studies that contributed to the understanding of layers of vision and the sphere of visual attention. The aim of the current study is to create an artificial network that is able to learn to generate saccades similar to a human being, but with more accurate prediction and in a more biologically plausible way as compared to the Saliency Model. The methods of the current study will use a similar Leaky Integrate and Fire layer, but will replace salience map creation with a Restricted Boltzmann Machine in order to create a generative model that is biologically precise for both spatial and temporal output. The initial results of the study involve a Restricted Boltzmann Machine able to generate eye movements based on general temporal and spatial parameters of saccadic eye movements from a twodimensional array dataset as input. The results imply that salience modelling can be improved by matching of spatial and temporal distributions of the model to spatial and temporal distributions of human participants.

Keywords: саккады, нейросети, deep learning, salience model, Restricted Boltzman machines, saccade generation, Ограниченная машина Больцмана, модели заметности

Publication based on the results of: A computational model to simulate visual stability from eye movements and spatial attention (2017)

In book

European Conference on Visual Perception 2017 Abstract Book

Busch N., Hesselmann G., Maertens M., Ostendorf F., Rolfs M., Sterzer P. [б.и.], 2017.