Figure 1.Virtual.
evolution of perception and saccade with different visibility maps, eye movement models and configurations.
a. Ventral (perception) and dorsal (action) streams projecting from the primary visual cortex (V1). b. Flow chart for two models of human eye-movement search: Ideal Bayesian Searcher (IS) and the Saccadic targeting model (maximum a posteriori probability model, MAP). c. 8 alternative forced choice target search for steep visibility map. d. 8 alternative forced choice target search for broad visibility map. e. 4 alternative forced choice target search for broad visibility map. Light blue circles outline possible target locations. Location of fixations for 1st (blue) and 2nd saccades (red) for three models: IS, MAP and Entropy Limit Minimization (ELM) in white noise The MAP model simulations include small random saccade endpoint errors to facilitate visualization of the different fixations. Central cross indicates initial fixation point for all models.
Figure 2.
Virtual evolution of two separate streams with the genetic algorithms for three different targets.
a. Virtual evolution of the perception (ventral stream) and saccade (dorsal stream) templates constructed from different linear combinations of twenty four different V1 simple cells which spanned the target (Gabor functions with center frequencies, 0.5, 1, 2, 4 cycles/degree for 6 different orientations, 30 degrees apart, and octave bandwidths). Probability of survival of an individual depends on search accuracy of the ideal searcher approximation (ELM model) with the two templates. b. Top row three different targets (from right to left: isotropic Gaussian, vertical elongated Gaussian and the difference of a vertical and horizontal elongated Gaussians) used in different evolution simulations for search in 1/f noise and a steep visibility map (See Figure 1c, left). All targets are luminance grey patterns but are shown in pseudo-color and scaled for each image to maximize the use of the color scale.
Figure 3.
Evolution plots for detecting the isotropic Gaussian target embedded in three different backgrounds.
1st row: Sample images for the 8 alternative forced choice (AFC) search task for an isotropic Gaussian shaped luminance target with a steep visibility map (Figure 1c left) added to white noise, 1/f noise, and natural images. Center of circles indicate the possible target locations and the central cross is the initial fixation position for the models. 2nd row: Distribution of search accuracies for simulated individuals as a function of generation. 3rd row: Distribution of correlations between perception and saccade templates of individuals in each generation. Bottom row: Perception (red) and saccade (blue) templates radial profiles (averaged across all angles) of best performing simulated individual for each background type. Results are averages across ten different virtual evolution runs each with 500 generations. Plots only show data up to the 200th generation for which convergence has occurred. Radial profile of the Gaussian signal is shown in a dashed line for comparison.
Figure 4.
Evolution plots for different models and scenarios detecting the elongated Gaussian target (Figure 2b; middle).
a. 8 AFC search with a broad visibility map using 1/f noise for the Entropy Limit Minimization model (ELM) and the Saccade Targeting model (MAP); b. 4 AFC with broad visibility map using 1/f noise for the ELM and MAP model. All results based on averages across 10 virtual evolution runs.
Figure 5.
Evolution plots for a model with changing V1 receptive field size/spatial frequency with retinal eccentricity.
a. 8 AFC search task in 1/f noise (left) and graph (right) showing the change in central spatial frequency and width of channel in the frequency domain of oriented Gabor functions with retinal eccentricity. b. Radial profiles in the frequency domain of Gaussian target (left) and DoG target (right) with a center frequency of 8 cycles/degree. c. Distribution of correlations between perception and saccade templates of individuals in each generation for Gaussian target (left) and DoG target (right). d. Perception and saccade templates radial profiles (averaged across all angles) of best performing simulated individual for low-frequency Gaussian target (left) and higher frequency DoG target (right).
Figure 6.
Evolution plots for scenarios which resulted in partial or no convergence of two templates.
All proportion correct and correlation plots shows the distribution for individuals in each generation. All results based on averages across 10 virtual evolution runs. a. 8 AFC search of the elongated Gaussian signal for a flat visibility map (ELM model); b. 8 AFC search of the elongated Gaussian signal for a broad visibility map, natural images, but with 8 eye movements which allows the model to fixate on all possible target locations (ELM model); c. 8 AFC search of an isotropic Gaussian signal for a steep visibility map using 1/f noise for the ELM model and considering two visual processing streams with different spatial pre-filtering based on LGN parvocellular and magnocelluar properties; d. Normalized frequency amplitude for Gaussian target, parvocellular LGN cell and magnocellular LGN cell; e. Perception and saccade templates radial profiles (averaged across all angles) of best performing simulated individual for the model with pathway LGN pre-filtering.