Region selection: segmentation, classification and task relevance in a single grouping mechanism

Abstract

We introduce an approach for the fusion of segmentation, classification and examination of task relevance into a grouping mechanism performed by a competitive neural relaxation network. This means, information extracted from the environment is selected according to the relevance of the systems intended action. Due to the fact of a task-specific focus of attention, we avoid the separation of perception and generation of behavior. Our network for the selection of action relevant visual regions consists of interacting columns with local excitatory, and global inhibitory coupled feedback. The lateral cooperation is used as a way to integrate task pertinent subgoals. Possible subgoals are the size of regions, the security of a classification hypothesis and the valuation of the hypothesis for the task. The input activity, received from different hypothesis-layers, evokes several activation areas, which compete in a few iteration cycles. When equilibrium is attained cooperating neurons in one layer remain active, others are suppressed with regard to the relevant subgoals. The performance is demonstrated on a real-world selection of textured objects for a robot grasping task.