Recruitment vs. Backpropagation Learning: An empirical study on re-learning in connectionist networks

Abstract

This paper describes a first comparison between two connectionist learning techniques: backpropagation and recruitment learning. The task is to re-learn a conceptual representation, i.e. to significantly change a representation in an additional training period by the use of new data. Backpropagation denotes to a widely known, supervised learning technique which requires the repeated presentation of a set of training instances. Recruitment learning denotes to a technique which converts network units from a pool of free units into units which carry meaningful information, and can be used for both, instruction-based and similarity-based learning. It will be shown that a learning technique which makes use of structured knowledge (i.e. recruitment learning), re-learns and modifies a connectionist representation faster than backpropagation.