Preserving Semantic Relations for Zero-Shot Learning

  • 2018-03-08 11:39:30
  • Yashas Annadani, Soma Biswas
  • 21


Zero-shot learning has gained popularity due to its potential to scalerecognition models without requiring additional training data. This is usuallyachieved by associating categories with their semantic information likeattributes. However, we believe that the potential offered by this paradigm isnot yet fully exploited. In this work, we propose to utilize the structure ofthe space spanned by the attributes using a set of relations. We deviseobjective functions to preserve these relations in the embedding space, therebyinducing semanticity to the embedding space. Through extensive experimentalevaluation on five benchmark datasets, we demonstrate that inducing semanticityto the embedding space is beneficial for zero-shot learning. The proposedapproach outperforms the state-of-the-art on the standard zero-shot setting aswell as the more realistic generalized zero-shot setting. We also demonstratehow the proposed approach can be useful for making approximate semanticinferences about an image belonging to a category for which attributeinformation is not available.


Introduction (beta)



Conclusion (beta)