Computational modelling of visual search

Visual search traditionally has two main competing theories of parallel and serial search and this architectural issue has not been solved to this day. The latest developments in the field have suggested a possibility that response time distributions may aid in differentiating the two competing theories. For this purpose we have used the best available serial model Competitive Guided Search and two biologically-plausible parallel models inspired by the theory of biased competition. The parallel models adopted a winner-take-all mechanism from Selective Attention for Identification Model as base model that was extended to form a novel model for explaining response time distributions. These models are analytically intractable, therefore we adopted a more accurate kernel density estimator for representing unknown probability density function. Introduced robustness properties to the fitness method and developed a more efficient algorithm for finding the parameter solutions. Then these methods were applied for comparison of the respective models and concluded that winner-takes-all model poorly generalises to response time distributions. The results were followed by introducing a novel Asymmetrical Dynamic Neural Network model that managed to explain distributional changes better than Competitive Guided Search model