Generalization Ability and Matrix Learning in Prototype-Based Classification

Project Members:

  • Prof. Dr. Barbara Hammer


Duration: since 01/2006

Project Description: Prototype-based classification offers intuitive and powerful machine learning tools, which is particularly interesting for interdisciplinary applications due to easy interpretability of the results. Research has been conducted to exactly investigate the learning behavior of popular heuristic learning rules in relevant model situations by means of statistical physics.

Further, extended learning rules have been developed which are based on a clear mathematical objective and which allow a general matrix adaptation, taking relevance weighting as well as correlations into account. Interestingly, learning theoretical generalization bounds can be derived which show that the method can be interpreted as large margin optimization.