Matthias Platho, "Learning of sensory representations for task-dependent control of search processes", Learning of sensory representations for task-dependent control of search processes, 2010.

Fulfilling demanding tasks in complex, dynamic environments poses a tough challenge for nowadays systems. A system relying on visual information has to handle a lot of data concerning the environment, the task and relevant objects. In order to reduce the amount of data a compression method for object datasets is proposed, that aims for maintaining task performance. An object dataset comprises numerical representations of object features in different feature dimensions. Current approaches for compression remove the least important feature dimensions as a whole and thus fail to preserve distinctive features of individual objects. In this work the compression is conducted by storing similar features of different objects in a coarse, shared representation, while distinctive features remain unchanged. It is shown that this method allows for a high compression rate while impairing the distinguishability of objects only minor.