MODELING OF TOP-DOWN OBJECT-BASED ATTENTION USING PROBABILISTIC NEURAL NETWORK

Object-based attention theory posits that attention is directed towards one object at a time. This paper attempts to simulate top-down influences. Five components of top-down influences are modeled: structure of object representation for long-term memory (LTM), learning of object representations, deduction of task-relevant features, estimation of top-down biases, mediation between bottom-up and top-down fashions, and perceptual completion. This model builds a dual-coding object representation for LTM. It consists of local and global codings, characterizing internal properties and global attributes of an object. Probabilistic neural networks (PNNs) are used for object representation in that they can model probabilistic distribution of an object through combination of confident instances. A dynamically constructive learning algorithm is developed to train PNNs when an object is attended. Given a task-specific object, this proposed model recalls the corresponding object representation from PNNs, deduces the task-relevant feature dimensions and evaluates top-down biases. Bottom-up and top-down biases are mediated to yield a primitive grouping based saliency map. The most salient primitive grouping is finally put into the perceptual completion processing module to yield an accurate and complete object representation for attention. This model has been applied into the robotic task: detection of task-specific multi-part objects.