Saccadic search in large data spaces

We introduce a novel search mechanism for large data spaces based on principals of support vector machines and information geometry. The search paradigm called saccadic search is a controlled, goal-directed, multi-scale statistical search with the objective of maximizing information about a given target set given finite time and resources. We define the notion of statistical search and controlled information maximization and discuss the applications and contexts in which the paradigm is particularly suitable. Our research is motivated by saccades in the human visual system. Saccades are rapid high-velocity motions of the eye that refocus the high-resolution area of the eye known as fovea to a new target. Saccades represent an intelligent search and synthesis strategy by the visual system to discover features of the visual field using a combination of nonlinear sampling, feature discovery, and categorization. Independent of a specific system, any efficient search strategy needs to address at least two problems. The representation of information in the search domain has a significant impact on the efficiency of search. An effective search strategy often represents only the goal dependent "features" that need to be searched a minimal description length representation. Additionally, a consistent global description of the search target needs to be registered and maintained as a means of retaining information about searched locations. We call this task feature fusion. Formally, we consider a nonlinear sampling strategy where a current target area of search is sampled with a higher frequency while sample rates away from the current target diminish at an exponential rate. A set of statistical descriptors for the highly sampled area is paired with a feature descriptor derived using a support vector machine. A task specific loss function is then used to determine the next search step which maximizes the expectation of information gain. We show simulations that demonstrate the effectiveness of the strategy in visual search of images or image sequences.