BiViMA - A Biologically Motivated Visual Memory Architecture for Life-Long Learning of Objects and Categories


 

The overall goal of this project is the development of a biologically inspired architecture for representing complex shaped objects and visually derivable categories like the basic color of an object. The main focus of this architecture is the ability of interactively training objects and their categories, while representing the objects in a resource-efficient way. These requirements are contradictory in the sense that interactively training requires fast learning methods, whereas resource efficiency typically requires methods with high computational costs.

The approach that is taken in this project is a separation of visual representation into a short-term and long-term memory. Additionally there is a focus on incremental learning methods also dealing with the stability-plasticity-dilemma, which are necessary requirements for a life-long learning system.