What Helps Where - And Why? Semantic Relatedness for Knowledge Transfer

Abstract

Remarkable performance has been reported to recognize single object classes. \nScalability to large numbers of classes however remains an important challenge \nfor today's recognition methods. Several authors have promoted knowledge \ntransfer between classes as a key ingredient to address this challenge. \nHowever, in previous work the decision which knowledge to transfer has required \neither manual supervision or at least a few training examples limiting the \nscalability of these approaches. In this work we explicitly address the \nquestion of how to automatically decide which information to transfer between \nclasses without the need of any human intervention. For this we tap into \nlinguistic knowledge bases to provide the semantic link between sources (what) \nand targets (where) of knowledge transfer. We provide a rigorous experimental \nevaluation of different knowledge bases and state-of-the-art techniques from \nNatural Language Processing which goes far beyond the limited use of language \nin related work. We also give insights into the applicability (why) of \ndifferent knowledge sources and similarity measures for knowledge transfer.