An Approach to Unseen Classs Classification with In-Service Predictors

Abstract

The goal of zero-shot learning is to recognize a novel class that did not appear in training. In this talk, we introduce a novel approach to zero-shot learning. Our approach reuses in-service predictors which are often available in practice. Unlike most of the existing methods, our method does not require to replace in-service predictors with new predictors specifically designed for zero-shot learning.